5617ae030b15864f2d01b3669eabbcd9856922ad
[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/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
17
18 /*
19  * tunables
20  */
21 /* max queue in one round of service */
22 static const int cfq_quantum = 8;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
34
35 /*
36  * offset from end of service tree
37  */
38 #define CFQ_IDLE_DELAY          (HZ / 5)
39
40 /*
41  * below this threshold, we consider thinktime immediate
42  */
43 #define CFQ_MIN_TT              (2)
44
45 #define CFQ_SLICE_SCALE         (5)
46 #define CFQ_HW_QUEUE_MIN        (5)
47 #define CFQ_SERVICE_SHIFT       12
48
49 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
50 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
51 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
52 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
53
54 #define RQ_CIC(rq)              \
55         ((struct cfq_io_context *) (rq)->elevator_private)
56 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
57
58 static struct kmem_cache *cfq_pool;
59 static struct kmem_cache *cfq_ioc_pool;
60
61 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
62 static struct completion *ioc_gone;
63 static DEFINE_SPINLOCK(ioc_gone_lock);
64
65 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68
69 #define sample_valid(samples)   ((samples) > 80)
70 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
71
72 /*
73  * Most of our rbtree usage is for sorting with min extraction, so
74  * if we cache the leftmost node we don't have to walk down the tree
75  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76  * move this into the elevator for the rq sorting as well.
77  */
78 struct cfq_rb_root {
79         struct rb_root rb;
80         struct rb_node *left;
81         unsigned count;
82         unsigned total_weight;
83         u64 min_vdisktime;
84         struct rb_node *active;
85 };
86 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87                         .count = 0, .min_vdisktime = 0, }
88
89 /*
90  * Per process-grouping structure
91  */
92 struct cfq_queue {
93         /* reference count */
94         atomic_t ref;
95         /* various state flags, see below */
96         unsigned int flags;
97         /* parent cfq_data */
98         struct cfq_data *cfqd;
99         /* service_tree member */
100         struct rb_node rb_node;
101         /* service_tree key */
102         unsigned long rb_key;
103         /* prio tree member */
104         struct rb_node p_node;
105         /* prio tree root we belong to, if any */
106         struct rb_root *p_root;
107         /* sorted list of pending requests */
108         struct rb_root sort_list;
109         /* if fifo isn't expired, next request to serve */
110         struct request *next_rq;
111         /* requests queued in sort_list */
112         int queued[2];
113         /* currently allocated requests */
114         int allocated[2];
115         /* fifo list of requests in sort_list */
116         struct list_head fifo;
117
118         /* time when queue got scheduled in to dispatch first request. */
119         unsigned long dispatch_start;
120         unsigned int allocated_slice;
121         unsigned int slice_dispatch;
122         /* time when first request from queue completed and slice started. */
123         unsigned long slice_start;
124         unsigned long slice_end;
125         long slice_resid;
126
127         /* pending metadata requests */
128         int meta_pending;
129         /* number of requests that are on the dispatch list or inside driver */
130         int dispatched;
131
132         /* io prio of this group */
133         unsigned short ioprio, org_ioprio;
134         unsigned short ioprio_class, org_ioprio_class;
135
136         pid_t pid;
137
138         u32 seek_history;
139         sector_t last_request_pos;
140
141         struct cfq_rb_root *service_tree;
142         struct cfq_queue *new_cfqq;
143         struct cfq_group *cfqg;
144         struct cfq_group *orig_cfqg;
145 };
146
147 /*
148  * First index in the service_trees.
149  * IDLE is handled separately, so it has negative index
150  */
151 enum wl_prio_t {
152         BE_WORKLOAD = 0,
153         RT_WORKLOAD = 1,
154         IDLE_WORKLOAD = 2,
155 };
156
157 /*
158  * Second index in the service_trees.
159  */
160 enum wl_type_t {
161         ASYNC_WORKLOAD = 0,
162         SYNC_NOIDLE_WORKLOAD = 1,
163         SYNC_WORKLOAD = 2
164 };
165
166 /* This is per cgroup per device grouping structure */
167 struct cfq_group {
168         /* group service_tree member */
169         struct rb_node rb_node;
170
171         /* group service_tree key */
172         u64 vdisktime;
173         unsigned int weight;
174         bool on_st;
175
176         /* number of cfqq currently on this group */
177         int nr_cfqq;
178
179         /* Per group busy queus average. Useful for workload slice calc. */
180         unsigned int busy_queues_avg[2];
181         /*
182          * rr lists of queues with requests, onle rr for each priority class.
183          * Counts are embedded in the cfq_rb_root
184          */
185         struct cfq_rb_root service_trees[2][3];
186         struct cfq_rb_root service_tree_idle;
187
188         unsigned long saved_workload_slice;
189         enum wl_type_t saved_workload;
190         enum wl_prio_t saved_serving_prio;
191         struct blkio_group blkg;
192 #ifdef CONFIG_CFQ_GROUP_IOSCHED
193         struct hlist_node cfqd_node;
194         atomic_t ref;
195 #endif
196 };
197
198 /*
199  * Per block device queue structure
200  */
201 struct cfq_data {
202         struct request_queue *queue;
203         /* Root service tree for cfq_groups */
204         struct cfq_rb_root grp_service_tree;
205         struct cfq_group root_group;
206
207         /*
208          * The priority currently being served
209          */
210         enum wl_prio_t serving_prio;
211         enum wl_type_t serving_type;
212         unsigned long workload_expires;
213         struct cfq_group *serving_group;
214         bool noidle_tree_requires_idle;
215
216         /*
217          * Each priority tree is sorted by next_request position.  These
218          * trees are used when determining if two or more queues are
219          * interleaving requests (see cfq_close_cooperator).
220          */
221         struct rb_root prio_trees[CFQ_PRIO_LISTS];
222
223         unsigned int busy_queues;
224
225         int rq_in_driver;
226         int rq_in_flight[2];
227
228         /*
229          * queue-depth detection
230          */
231         int rq_queued;
232         int hw_tag;
233         /*
234          * hw_tag can be
235          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
236          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
237          *  0 => no NCQ
238          */
239         int hw_tag_est_depth;
240         unsigned int hw_tag_samples;
241
242         /*
243          * idle window management
244          */
245         struct timer_list idle_slice_timer;
246         struct work_struct unplug_work;
247
248         struct cfq_queue *active_queue;
249         struct cfq_io_context *active_cic;
250
251         /*
252          * async queue for each priority case
253          */
254         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
255         struct cfq_queue *async_idle_cfqq;
256
257         sector_t last_position;
258
259         /*
260          * tunables, see top of file
261          */
262         unsigned int cfq_quantum;
263         unsigned int cfq_fifo_expire[2];
264         unsigned int cfq_back_penalty;
265         unsigned int cfq_back_max;
266         unsigned int cfq_slice[2];
267         unsigned int cfq_slice_async_rq;
268         unsigned int cfq_slice_idle;
269         unsigned int cfq_latency;
270         unsigned int cfq_group_isolation;
271
272         struct list_head cic_list;
273
274         /*
275          * Fallback dummy cfqq for extreme OOM conditions
276          */
277         struct cfq_queue oom_cfqq;
278
279         unsigned long last_delayed_sync;
280
281         /* List of cfq groups being managed on this device*/
282         struct hlist_head cfqg_list;
283         struct rcu_head rcu;
284 };
285
286 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
287
288 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
289                                             enum wl_prio_t prio,
290                                             enum wl_type_t type)
291 {
292         if (!cfqg)
293                 return NULL;
294
295         if (prio == IDLE_WORKLOAD)
296                 return &cfqg->service_tree_idle;
297
298         return &cfqg->service_trees[prio][type];
299 }
300
301 enum cfqq_state_flags {
302         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
303         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
304         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
305         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
306         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
307         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
308         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
309         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
310         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
311         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
312         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
313         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
314         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
315 };
316
317 #define CFQ_CFQQ_FNS(name)                                              \
318 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
319 {                                                                       \
320         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
321 }                                                                       \
322 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
323 {                                                                       \
324         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
325 }                                                                       \
326 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
327 {                                                                       \
328         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
329 }
330
331 CFQ_CFQQ_FNS(on_rr);
332 CFQ_CFQQ_FNS(wait_request);
333 CFQ_CFQQ_FNS(must_dispatch);
334 CFQ_CFQQ_FNS(must_alloc_slice);
335 CFQ_CFQQ_FNS(fifo_expire);
336 CFQ_CFQQ_FNS(idle_window);
337 CFQ_CFQQ_FNS(prio_changed);
338 CFQ_CFQQ_FNS(slice_new);
339 CFQ_CFQQ_FNS(sync);
340 CFQ_CFQQ_FNS(coop);
341 CFQ_CFQQ_FNS(split_coop);
342 CFQ_CFQQ_FNS(deep);
343 CFQ_CFQQ_FNS(wait_busy);
344 #undef CFQ_CFQQ_FNS
345
346 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
347 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
348         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
349                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
350                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
351
352 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
353         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
354                                 blkg_path(&(cfqg)->blkg), ##args);      \
355
356 #else
357 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
358         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
359 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
360 #endif
361 #define cfq_log(cfqd, fmt, args...)     \
362         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
363
364 /* Traverses through cfq group service trees */
365 #define for_each_cfqg_st(cfqg, i, j, st) \
366         for (i = 0; i <= IDLE_WORKLOAD; i++) \
367                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
368                         : &cfqg->service_tree_idle; \
369                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
370                         (i == IDLE_WORKLOAD && j == 0); \
371                         j++, st = i < IDLE_WORKLOAD ? \
372                         &cfqg->service_trees[i][j]: NULL) \
373
374
375 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
376 {
377         if (cfq_class_idle(cfqq))
378                 return IDLE_WORKLOAD;
379         if (cfq_class_rt(cfqq))
380                 return RT_WORKLOAD;
381         return BE_WORKLOAD;
382 }
383
384
385 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
386 {
387         if (!cfq_cfqq_sync(cfqq))
388                 return ASYNC_WORKLOAD;
389         if (!cfq_cfqq_idle_window(cfqq))
390                 return SYNC_NOIDLE_WORKLOAD;
391         return SYNC_WORKLOAD;
392 }
393
394 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
395                                         struct cfq_data *cfqd,
396                                         struct cfq_group *cfqg)
397 {
398         if (wl == IDLE_WORKLOAD)
399                 return cfqg->service_tree_idle.count;
400
401         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
402                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
403                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
404 }
405
406 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
407                                         struct cfq_group *cfqg)
408 {
409         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
410                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
411 }
412
413 static void cfq_dispatch_insert(struct request_queue *, struct request *);
414 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
415                                        struct io_context *, gfp_t);
416 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
417                                                 struct io_context *);
418
419 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
420                                             bool is_sync)
421 {
422         return cic->cfqq[is_sync];
423 }
424
425 static inline void cic_set_cfqq(struct cfq_io_context *cic,
426                                 struct cfq_queue *cfqq, bool is_sync)
427 {
428         cic->cfqq[is_sync] = cfqq;
429 }
430
431 /*
432  * We regard a request as SYNC, if it's either a read or has the SYNC bit
433  * set (in which case it could also be direct WRITE).
434  */
435 static inline bool cfq_bio_sync(struct bio *bio)
436 {
437         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
438 }
439
440 /*
441  * scheduler run of queue, if there are requests pending and no one in the
442  * driver that will restart queueing
443  */
444 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
445 {
446         if (cfqd->busy_queues) {
447                 cfq_log(cfqd, "schedule dispatch");
448                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
449         }
450 }
451
452 static int cfq_queue_empty(struct request_queue *q)
453 {
454         struct cfq_data *cfqd = q->elevator->elevator_data;
455
456         return !cfqd->rq_queued;
457 }
458
459 /*
460  * Scale schedule slice based on io priority. Use the sync time slice only
461  * if a queue is marked sync and has sync io queued. A sync queue with async
462  * io only, should not get full sync slice length.
463  */
464 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
465                                  unsigned short prio)
466 {
467         const int base_slice = cfqd->cfq_slice[sync];
468
469         WARN_ON(prio >= IOPRIO_BE_NR);
470
471         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
472 }
473
474 static inline int
475 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
476 {
477         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
478 }
479
480 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
481 {
482         u64 d = delta << CFQ_SERVICE_SHIFT;
483
484         d = d * BLKIO_WEIGHT_DEFAULT;
485         do_div(d, cfqg->weight);
486         return d;
487 }
488
489 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
490 {
491         s64 delta = (s64)(vdisktime - min_vdisktime);
492         if (delta > 0)
493                 min_vdisktime = vdisktime;
494
495         return min_vdisktime;
496 }
497
498 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
499 {
500         s64 delta = (s64)(vdisktime - min_vdisktime);
501         if (delta < 0)
502                 min_vdisktime = vdisktime;
503
504         return min_vdisktime;
505 }
506
507 static void update_min_vdisktime(struct cfq_rb_root *st)
508 {
509         u64 vdisktime = st->min_vdisktime;
510         struct cfq_group *cfqg;
511
512         if (st->active) {
513                 cfqg = rb_entry_cfqg(st->active);
514                 vdisktime = cfqg->vdisktime;
515         }
516
517         if (st->left) {
518                 cfqg = rb_entry_cfqg(st->left);
519                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
520         }
521
522         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
523 }
524
525 /*
526  * get averaged number of queues of RT/BE priority.
527  * average is updated, with a formula that gives more weight to higher numbers,
528  * to quickly follows sudden increases and decrease slowly
529  */
530
531 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
532                                         struct cfq_group *cfqg, bool rt)
533 {
534         unsigned min_q, max_q;
535         unsigned mult  = cfq_hist_divisor - 1;
536         unsigned round = cfq_hist_divisor / 2;
537         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
538
539         min_q = min(cfqg->busy_queues_avg[rt], busy);
540         max_q = max(cfqg->busy_queues_avg[rt], busy);
541         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
542                 cfq_hist_divisor;
543         return cfqg->busy_queues_avg[rt];
544 }
545
546 static inline unsigned
547 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
548 {
549         struct cfq_rb_root *st = &cfqd->grp_service_tree;
550
551         return cfq_target_latency * cfqg->weight / st->total_weight;
552 }
553
554 static inline void
555 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
556 {
557         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
558         if (cfqd->cfq_latency) {
559                 /*
560                  * interested queues (we consider only the ones with the same
561                  * priority class in the cfq group)
562                  */
563                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
564                                                 cfq_class_rt(cfqq));
565                 unsigned sync_slice = cfqd->cfq_slice[1];
566                 unsigned expect_latency = sync_slice * iq;
567                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
568
569                 if (expect_latency > group_slice) {
570                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
571                         /* scale low_slice according to IO priority
572                          * and sync vs async */
573                         unsigned low_slice =
574                                 min(slice, base_low_slice * slice / sync_slice);
575                         /* the adapted slice value is scaled to fit all iqs
576                          * into the target latency */
577                         slice = max(slice * group_slice / expect_latency,
578                                     low_slice);
579                 }
580         }
581         cfqq->slice_start = jiffies;
582         cfqq->slice_end = jiffies + slice;
583         cfqq->allocated_slice = slice;
584         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
585 }
586
587 /*
588  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
589  * isn't valid until the first request from the dispatch is activated
590  * and the slice time set.
591  */
592 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
593 {
594         if (cfq_cfqq_slice_new(cfqq))
595                 return 0;
596         if (time_before(jiffies, cfqq->slice_end))
597                 return 0;
598
599         return 1;
600 }
601
602 /*
603  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
604  * We choose the request that is closest to the head right now. Distance
605  * behind the head is penalized and only allowed to a certain extent.
606  */
607 static struct request *
608 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
609 {
610         sector_t s1, s2, d1 = 0, d2 = 0;
611         unsigned long back_max;
612 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
613 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
614         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
615
616         if (rq1 == NULL || rq1 == rq2)
617                 return rq2;
618         if (rq2 == NULL)
619                 return rq1;
620
621         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
622                 return rq1;
623         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
624                 return rq2;
625         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
626                 return rq1;
627         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
628                 return rq2;
629
630         s1 = blk_rq_pos(rq1);
631         s2 = blk_rq_pos(rq2);
632
633         /*
634          * by definition, 1KiB is 2 sectors
635          */
636         back_max = cfqd->cfq_back_max * 2;
637
638         /*
639          * Strict one way elevator _except_ in the case where we allow
640          * short backward seeks which are biased as twice the cost of a
641          * similar forward seek.
642          */
643         if (s1 >= last)
644                 d1 = s1 - last;
645         else if (s1 + back_max >= last)
646                 d1 = (last - s1) * cfqd->cfq_back_penalty;
647         else
648                 wrap |= CFQ_RQ1_WRAP;
649
650         if (s2 >= last)
651                 d2 = s2 - last;
652         else if (s2 + back_max >= last)
653                 d2 = (last - s2) * cfqd->cfq_back_penalty;
654         else
655                 wrap |= CFQ_RQ2_WRAP;
656
657         /* Found required data */
658
659         /*
660          * By doing switch() on the bit mask "wrap" we avoid having to
661          * check two variables for all permutations: --> faster!
662          */
663         switch (wrap) {
664         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
665                 if (d1 < d2)
666                         return rq1;
667                 else if (d2 < d1)
668                         return rq2;
669                 else {
670                         if (s1 >= s2)
671                                 return rq1;
672                         else
673                                 return rq2;
674                 }
675
676         case CFQ_RQ2_WRAP:
677                 return rq1;
678         case CFQ_RQ1_WRAP:
679                 return rq2;
680         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
681         default:
682                 /*
683                  * Since both rqs are wrapped,
684                  * start with the one that's further behind head
685                  * (--> only *one* back seek required),
686                  * since back seek takes more time than forward.
687                  */
688                 if (s1 <= s2)
689                         return rq1;
690                 else
691                         return rq2;
692         }
693 }
694
695 /*
696  * The below is leftmost cache rbtree addon
697  */
698 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
699 {
700         /* Service tree is empty */
701         if (!root->count)
702                 return NULL;
703
704         if (!root->left)
705                 root->left = rb_first(&root->rb);
706
707         if (root->left)
708                 return rb_entry(root->left, struct cfq_queue, rb_node);
709
710         return NULL;
711 }
712
713 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
714 {
715         if (!root->left)
716                 root->left = rb_first(&root->rb);
717
718         if (root->left)
719                 return rb_entry_cfqg(root->left);
720
721         return NULL;
722 }
723
724 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
725 {
726         rb_erase(n, root);
727         RB_CLEAR_NODE(n);
728 }
729
730 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
731 {
732         if (root->left == n)
733                 root->left = NULL;
734         rb_erase_init(n, &root->rb);
735         --root->count;
736 }
737
738 /*
739  * would be nice to take fifo expire time into account as well
740  */
741 static struct request *
742 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
743                   struct request *last)
744 {
745         struct rb_node *rbnext = rb_next(&last->rb_node);
746         struct rb_node *rbprev = rb_prev(&last->rb_node);
747         struct request *next = NULL, *prev = NULL;
748
749         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
750
751         if (rbprev)
752                 prev = rb_entry_rq(rbprev);
753
754         if (rbnext)
755                 next = rb_entry_rq(rbnext);
756         else {
757                 rbnext = rb_first(&cfqq->sort_list);
758                 if (rbnext && rbnext != &last->rb_node)
759                         next = rb_entry_rq(rbnext);
760         }
761
762         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
763 }
764
765 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
766                                       struct cfq_queue *cfqq)
767 {
768         /*
769          * just an approximation, should be ok.
770          */
771         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
772                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
773 }
774
775 static inline s64
776 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
777 {
778         return cfqg->vdisktime - st->min_vdisktime;
779 }
780
781 static void
782 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
783 {
784         struct rb_node **node = &st->rb.rb_node;
785         struct rb_node *parent = NULL;
786         struct cfq_group *__cfqg;
787         s64 key = cfqg_key(st, cfqg);
788         int left = 1;
789
790         while (*node != NULL) {
791                 parent = *node;
792                 __cfqg = rb_entry_cfqg(parent);
793
794                 if (key < cfqg_key(st, __cfqg))
795                         node = &parent->rb_left;
796                 else {
797                         node = &parent->rb_right;
798                         left = 0;
799                 }
800         }
801
802         if (left)
803                 st->left = &cfqg->rb_node;
804
805         rb_link_node(&cfqg->rb_node, parent, node);
806         rb_insert_color(&cfqg->rb_node, &st->rb);
807 }
808
809 static void
810 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
811 {
812         struct cfq_rb_root *st = &cfqd->grp_service_tree;
813         struct cfq_group *__cfqg;
814         struct rb_node *n;
815
816         cfqg->nr_cfqq++;
817         if (cfqg->on_st)
818                 return;
819
820         /*
821          * Currently put the group at the end. Later implement something
822          * so that groups get lesser vtime based on their weights, so that
823          * if group does not loose all if it was not continously backlogged.
824          */
825         n = rb_last(&st->rb);
826         if (n) {
827                 __cfqg = rb_entry_cfqg(n);
828                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
829         } else
830                 cfqg->vdisktime = st->min_vdisktime;
831
832         __cfq_group_service_tree_add(st, cfqg);
833         cfqg->on_st = true;
834         st->total_weight += cfqg->weight;
835 }
836
837 static void
838 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
839 {
840         struct cfq_rb_root *st = &cfqd->grp_service_tree;
841
842         if (st->active == &cfqg->rb_node)
843                 st->active = NULL;
844
845         BUG_ON(cfqg->nr_cfqq < 1);
846         cfqg->nr_cfqq--;
847
848         /* If there are other cfq queues under this group, don't delete it */
849         if (cfqg->nr_cfqq)
850                 return;
851
852         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
853         cfqg->on_st = false;
854         st->total_weight -= cfqg->weight;
855         if (!RB_EMPTY_NODE(&cfqg->rb_node))
856                 cfq_rb_erase(&cfqg->rb_node, st);
857         cfqg->saved_workload_slice = 0;
858         blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
859 }
860
861 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
862 {
863         unsigned int slice_used;
864
865         /*
866          * Queue got expired before even a single request completed or
867          * got expired immediately after first request completion.
868          */
869         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
870                 /*
871                  * Also charge the seek time incurred to the group, otherwise
872                  * if there are mutiple queues in the group, each can dispatch
873                  * a single request on seeky media and cause lots of seek time
874                  * and group will never know it.
875                  */
876                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
877                                         1);
878         } else {
879                 slice_used = jiffies - cfqq->slice_start;
880                 if (slice_used > cfqq->allocated_slice)
881                         slice_used = cfqq->allocated_slice;
882         }
883
884         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
885         return slice_used;
886 }
887
888 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
889                                 struct cfq_queue *cfqq)
890 {
891         struct cfq_rb_root *st = &cfqd->grp_service_tree;
892         unsigned int used_sl, charge_sl;
893         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
894                         - cfqg->service_tree_idle.count;
895
896         BUG_ON(nr_sync < 0);
897         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
898
899         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
900                 charge_sl = cfqq->allocated_slice;
901
902         /* Can't update vdisktime while group is on service tree */
903         cfq_rb_erase(&cfqg->rb_node, st);
904         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
905         __cfq_group_service_tree_add(st, cfqg);
906
907         /* This group is being expired. Save the context */
908         if (time_after(cfqd->workload_expires, jiffies)) {
909                 cfqg->saved_workload_slice = cfqd->workload_expires
910                                                 - jiffies;
911                 cfqg->saved_workload = cfqd->serving_type;
912                 cfqg->saved_serving_prio = cfqd->serving_prio;
913         } else
914                 cfqg->saved_workload_slice = 0;
915
916         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
917                                         st->min_vdisktime);
918         blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
919 }
920
921 #ifdef CONFIG_CFQ_GROUP_IOSCHED
922 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
923 {
924         if (blkg)
925                 return container_of(blkg, struct cfq_group, blkg);
926         return NULL;
927 }
928
929 void
930 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
931 {
932         cfqg_of_blkg(blkg)->weight = weight;
933 }
934
935 static struct cfq_group *
936 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
937 {
938         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
939         struct cfq_group *cfqg = NULL;
940         void *key = cfqd;
941         int i, j;
942         struct cfq_rb_root *st;
943         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
944         unsigned int major, minor;
945
946         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
947         if (cfqg || !create)
948                 goto done;
949
950         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
951         if (!cfqg)
952                 goto done;
953
954         cfqg->weight = blkcg->weight;
955         for_each_cfqg_st(cfqg, i, j, st)
956                 *st = CFQ_RB_ROOT;
957         RB_CLEAR_NODE(&cfqg->rb_node);
958         blkio_group_init(&cfqg->blkg);
959
960         /*
961          * Take the initial reference that will be released on destroy
962          * This can be thought of a joint reference by cgroup and
963          * elevator which will be dropped by either elevator exit
964          * or cgroup deletion path depending on who is exiting first.
965          */
966         atomic_set(&cfqg->ref, 1);
967
968         /* Add group onto cgroup list */
969         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
970         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
971                                         MKDEV(major, minor));
972
973         /* Add group on cfqd list */
974         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
975
976 done:
977         return cfqg;
978 }
979
980 /*
981  * Search for the cfq group current task belongs to. If create = 1, then also
982  * create the cfq group if it does not exist. request_queue lock must be held.
983  */
984 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
985 {
986         struct cgroup *cgroup;
987         struct cfq_group *cfqg = NULL;
988
989         rcu_read_lock();
990         cgroup = task_cgroup(current, blkio_subsys_id);
991         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
992         if (!cfqg && create)
993                 cfqg = &cfqd->root_group;
994         rcu_read_unlock();
995         return cfqg;
996 }
997
998 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
999 {
1000         /* Currently, all async queues are mapped to root group */
1001         if (!cfq_cfqq_sync(cfqq))
1002                 cfqg = &cfqq->cfqd->root_group;
1003
1004         cfqq->cfqg = cfqg;
1005         /* cfqq reference on cfqg */
1006         atomic_inc(&cfqq->cfqg->ref);
1007 }
1008
1009 static void cfq_put_cfqg(struct cfq_group *cfqg)
1010 {
1011         struct cfq_rb_root *st;
1012         int i, j;
1013
1014         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1015         if (!atomic_dec_and_test(&cfqg->ref))
1016                 return;
1017         for_each_cfqg_st(cfqg, i, j, st)
1018                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1019         kfree(cfqg);
1020 }
1021
1022 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1023 {
1024         /* Something wrong if we are trying to remove same group twice */
1025         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1026
1027         hlist_del_init(&cfqg->cfqd_node);
1028
1029         /*
1030          * Put the reference taken at the time of creation so that when all
1031          * queues are gone, group can be destroyed.
1032          */
1033         cfq_put_cfqg(cfqg);
1034 }
1035
1036 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1037 {
1038         struct hlist_node *pos, *n;
1039         struct cfq_group *cfqg;
1040
1041         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1042                 /*
1043                  * If cgroup removal path got to blk_group first and removed
1044                  * it from cgroup list, then it will take care of destroying
1045                  * cfqg also.
1046                  */
1047                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1048                         cfq_destroy_cfqg(cfqd, cfqg);
1049         }
1050 }
1051
1052 /*
1053  * Blk cgroup controller notification saying that blkio_group object is being
1054  * delinked as associated cgroup object is going away. That also means that
1055  * no new IO will come in this group. So get rid of this group as soon as
1056  * any pending IO in the group is finished.
1057  *
1058  * This function is called under rcu_read_lock(). key is the rcu protected
1059  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1060  * read lock.
1061  *
1062  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1063  * it should not be NULL as even if elevator was exiting, cgroup deltion
1064  * path got to it first.
1065  */
1066 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1067 {
1068         unsigned long  flags;
1069         struct cfq_data *cfqd = key;
1070
1071         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1072         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1073         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1074 }
1075
1076 #else /* GROUP_IOSCHED */
1077 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1078 {
1079         return &cfqd->root_group;
1080 }
1081 static inline void
1082 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1083         cfqq->cfqg = cfqg;
1084 }
1085
1086 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1087 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1088
1089 #endif /* GROUP_IOSCHED */
1090
1091 /*
1092  * The cfqd->service_trees holds all pending cfq_queue's that have
1093  * requests waiting to be processed. It is sorted in the order that
1094  * we will service the queues.
1095  */
1096 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1097                                  bool add_front)
1098 {
1099         struct rb_node **p, *parent;
1100         struct cfq_queue *__cfqq;
1101         unsigned long rb_key;
1102         struct cfq_rb_root *service_tree;
1103         int left;
1104         int new_cfqq = 1;
1105         int group_changed = 0;
1106
1107 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1108         if (!cfqd->cfq_group_isolation
1109             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1110             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1111                 /* Move this cfq to root group */
1112                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1113                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1114                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1115                 cfqq->orig_cfqg = cfqq->cfqg;
1116                 cfqq->cfqg = &cfqd->root_group;
1117                 atomic_inc(&cfqd->root_group.ref);
1118                 group_changed = 1;
1119         } else if (!cfqd->cfq_group_isolation
1120                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1121                 /* cfqq is sequential now needs to go to its original group */
1122                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1123                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1124                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1125                 cfq_put_cfqg(cfqq->cfqg);
1126                 cfqq->cfqg = cfqq->orig_cfqg;
1127                 cfqq->orig_cfqg = NULL;
1128                 group_changed = 1;
1129                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1130         }
1131 #endif
1132
1133         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1134                                                 cfqq_type(cfqq));
1135         if (cfq_class_idle(cfqq)) {
1136                 rb_key = CFQ_IDLE_DELAY;
1137                 parent = rb_last(&service_tree->rb);
1138                 if (parent && parent != &cfqq->rb_node) {
1139                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1140                         rb_key += __cfqq->rb_key;
1141                 } else
1142                         rb_key += jiffies;
1143         } else if (!add_front) {
1144                 /*
1145                  * Get our rb key offset. Subtract any residual slice
1146                  * value carried from last service. A negative resid
1147                  * count indicates slice overrun, and this should position
1148                  * the next service time further away in the tree.
1149                  */
1150                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1151                 rb_key -= cfqq->slice_resid;
1152                 cfqq->slice_resid = 0;
1153         } else {
1154                 rb_key = -HZ;
1155                 __cfqq = cfq_rb_first(service_tree);
1156                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1157         }
1158
1159         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1160                 new_cfqq = 0;
1161                 /*
1162                  * same position, nothing more to do
1163                  */
1164                 if (rb_key == cfqq->rb_key &&
1165                     cfqq->service_tree == service_tree)
1166                         return;
1167
1168                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1169                 cfqq->service_tree = NULL;
1170         }
1171
1172         left = 1;
1173         parent = NULL;
1174         cfqq->service_tree = service_tree;
1175         p = &service_tree->rb.rb_node;
1176         while (*p) {
1177                 struct rb_node **n;
1178
1179                 parent = *p;
1180                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181
1182                 /*
1183                  * sort by key, that represents service time.
1184                  */
1185                 if (time_before(rb_key, __cfqq->rb_key))
1186                         n = &(*p)->rb_left;
1187                 else {
1188                         n = &(*p)->rb_right;
1189                         left = 0;
1190                 }
1191
1192                 p = n;
1193         }
1194
1195         if (left)
1196                 service_tree->left = &cfqq->rb_node;
1197
1198         cfqq->rb_key = rb_key;
1199         rb_link_node(&cfqq->rb_node, parent, p);
1200         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1201         service_tree->count++;
1202         if ((add_front || !new_cfqq) && !group_changed)
1203                 return;
1204         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1205 }
1206
1207 static struct cfq_queue *
1208 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1209                      sector_t sector, struct rb_node **ret_parent,
1210                      struct rb_node ***rb_link)
1211 {
1212         struct rb_node **p, *parent;
1213         struct cfq_queue *cfqq = NULL;
1214
1215         parent = NULL;
1216         p = &root->rb_node;
1217         while (*p) {
1218                 struct rb_node **n;
1219
1220                 parent = *p;
1221                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1222
1223                 /*
1224                  * Sort strictly based on sector.  Smallest to the left,
1225                  * largest to the right.
1226                  */
1227                 if (sector > blk_rq_pos(cfqq->next_rq))
1228                         n = &(*p)->rb_right;
1229                 else if (sector < blk_rq_pos(cfqq->next_rq))
1230                         n = &(*p)->rb_left;
1231                 else
1232                         break;
1233                 p = n;
1234                 cfqq = NULL;
1235         }
1236
1237         *ret_parent = parent;
1238         if (rb_link)
1239                 *rb_link = p;
1240         return cfqq;
1241 }
1242
1243 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1244 {
1245         struct rb_node **p, *parent;
1246         struct cfq_queue *__cfqq;
1247
1248         if (cfqq->p_root) {
1249                 rb_erase(&cfqq->p_node, cfqq->p_root);
1250                 cfqq->p_root = NULL;
1251         }
1252
1253         if (cfq_class_idle(cfqq))
1254                 return;
1255         if (!cfqq->next_rq)
1256                 return;
1257
1258         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1259         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1260                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1261         if (!__cfqq) {
1262                 rb_link_node(&cfqq->p_node, parent, p);
1263                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1264         } else
1265                 cfqq->p_root = NULL;
1266 }
1267
1268 /*
1269  * Update cfqq's position in the service tree.
1270  */
1271 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1272 {
1273         /*
1274          * Resorting requires the cfqq to be on the RR list already.
1275          */
1276         if (cfq_cfqq_on_rr(cfqq)) {
1277                 cfq_service_tree_add(cfqd, cfqq, 0);
1278                 cfq_prio_tree_add(cfqd, cfqq);
1279         }
1280 }
1281
1282 /*
1283  * add to busy list of queues for service, trying to be fair in ordering
1284  * the pending list according to last request service
1285  */
1286 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1287 {
1288         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1289         BUG_ON(cfq_cfqq_on_rr(cfqq));
1290         cfq_mark_cfqq_on_rr(cfqq);
1291         cfqd->busy_queues++;
1292
1293         cfq_resort_rr_list(cfqd, cfqq);
1294 }
1295
1296 /*
1297  * Called when the cfqq no longer has requests pending, remove it from
1298  * the service tree.
1299  */
1300 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1301 {
1302         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1303         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1304         cfq_clear_cfqq_on_rr(cfqq);
1305
1306         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1307                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1308                 cfqq->service_tree = NULL;
1309         }
1310         if (cfqq->p_root) {
1311                 rb_erase(&cfqq->p_node, cfqq->p_root);
1312                 cfqq->p_root = NULL;
1313         }
1314
1315         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1316         BUG_ON(!cfqd->busy_queues);
1317         cfqd->busy_queues--;
1318 }
1319
1320 /*
1321  * rb tree support functions
1322  */
1323 static void cfq_del_rq_rb(struct request *rq)
1324 {
1325         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1326         const int sync = rq_is_sync(rq);
1327
1328         BUG_ON(!cfqq->queued[sync]);
1329         cfqq->queued[sync]--;
1330
1331         elv_rb_del(&cfqq->sort_list, rq);
1332
1333         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1334                 /*
1335                  * Queue will be deleted from service tree when we actually
1336                  * expire it later. Right now just remove it from prio tree
1337                  * as it is empty.
1338                  */
1339                 if (cfqq->p_root) {
1340                         rb_erase(&cfqq->p_node, cfqq->p_root);
1341                         cfqq->p_root = NULL;
1342                 }
1343         }
1344 }
1345
1346 static void cfq_add_rq_rb(struct request *rq)
1347 {
1348         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1349         struct cfq_data *cfqd = cfqq->cfqd;
1350         struct request *__alias, *prev;
1351
1352         cfqq->queued[rq_is_sync(rq)]++;
1353
1354         /*
1355          * looks a little odd, but the first insert might return an alias.
1356          * if that happens, put the alias on the dispatch list
1357          */
1358         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1359                 cfq_dispatch_insert(cfqd->queue, __alias);
1360
1361         if (!cfq_cfqq_on_rr(cfqq))
1362                 cfq_add_cfqq_rr(cfqd, cfqq);
1363
1364         /*
1365          * check if this request is a better next-serve candidate
1366          */
1367         prev = cfqq->next_rq;
1368         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1369
1370         /*
1371          * adjust priority tree position, if ->next_rq changes
1372          */
1373         if (prev != cfqq->next_rq)
1374                 cfq_prio_tree_add(cfqd, cfqq);
1375
1376         BUG_ON(!cfqq->next_rq);
1377 }
1378
1379 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1380 {
1381         elv_rb_del(&cfqq->sort_list, rq);
1382         cfqq->queued[rq_is_sync(rq)]--;
1383         cfq_add_rq_rb(rq);
1384 }
1385
1386 static struct request *
1387 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1388 {
1389         struct task_struct *tsk = current;
1390         struct cfq_io_context *cic;
1391         struct cfq_queue *cfqq;
1392
1393         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1394         if (!cic)
1395                 return NULL;
1396
1397         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1398         if (cfqq) {
1399                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1400
1401                 return elv_rb_find(&cfqq->sort_list, sector);
1402         }
1403
1404         return NULL;
1405 }
1406
1407 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1408 {
1409         struct cfq_data *cfqd = q->elevator->elevator_data;
1410
1411         cfqd->rq_in_driver++;
1412         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1413                                                 cfqd->rq_in_driver);
1414
1415         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1416 }
1417
1418 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1419 {
1420         struct cfq_data *cfqd = q->elevator->elevator_data;
1421
1422         WARN_ON(!cfqd->rq_in_driver);
1423         cfqd->rq_in_driver--;
1424         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1425                                                 cfqd->rq_in_driver);
1426 }
1427
1428 static void cfq_remove_request(struct request *rq)
1429 {
1430         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1431
1432         if (cfqq->next_rq == rq)
1433                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1434
1435         list_del_init(&rq->queuelist);
1436         cfq_del_rq_rb(rq);
1437
1438         cfqq->cfqd->rq_queued--;
1439         if (rq_is_meta(rq)) {
1440                 WARN_ON(!cfqq->meta_pending);
1441                 cfqq->meta_pending--;
1442         }
1443 }
1444
1445 static int cfq_merge(struct request_queue *q, struct request **req,
1446                      struct bio *bio)
1447 {
1448         struct cfq_data *cfqd = q->elevator->elevator_data;
1449         struct request *__rq;
1450
1451         __rq = cfq_find_rq_fmerge(cfqd, bio);
1452         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1453                 *req = __rq;
1454                 return ELEVATOR_FRONT_MERGE;
1455         }
1456
1457         return ELEVATOR_NO_MERGE;
1458 }
1459
1460 static void cfq_merged_request(struct request_queue *q, struct request *req,
1461                                int type)
1462 {
1463         if (type == ELEVATOR_FRONT_MERGE) {
1464                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1465
1466                 cfq_reposition_rq_rb(cfqq, req);
1467         }
1468 }
1469
1470 static void
1471 cfq_merged_requests(struct request_queue *q, struct request *rq,
1472                     struct request *next)
1473 {
1474         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1475         /*
1476          * reposition in fifo if next is older than rq
1477          */
1478         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1479             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1480                 list_move(&rq->queuelist, &next->queuelist);
1481                 rq_set_fifo_time(rq, rq_fifo_time(next));
1482         }
1483
1484         if (cfqq->next_rq == next)
1485                 cfqq->next_rq = rq;
1486         cfq_remove_request(next);
1487 }
1488
1489 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1490                            struct bio *bio)
1491 {
1492         struct cfq_data *cfqd = q->elevator->elevator_data;
1493         struct cfq_io_context *cic;
1494         struct cfq_queue *cfqq;
1495
1496         /*
1497          * Disallow merge of a sync bio into an async request.
1498          */
1499         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1500                 return false;
1501
1502         /*
1503          * Lookup the cfqq that this bio will be queued with. Allow
1504          * merge only if rq is queued there.
1505          */
1506         cic = cfq_cic_lookup(cfqd, current->io_context);
1507         if (!cic)
1508                 return false;
1509
1510         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1511         return cfqq == RQ_CFQQ(rq);
1512 }
1513
1514 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1515                                    struct cfq_queue *cfqq)
1516 {
1517         if (cfqq) {
1518                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1519                                 cfqd->serving_prio, cfqd->serving_type);
1520                 cfqq->slice_start = 0;
1521                 cfqq->dispatch_start = jiffies;
1522                 cfqq->allocated_slice = 0;
1523                 cfqq->slice_end = 0;
1524                 cfqq->slice_dispatch = 0;
1525
1526                 cfq_clear_cfqq_wait_request(cfqq);
1527                 cfq_clear_cfqq_must_dispatch(cfqq);
1528                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1529                 cfq_clear_cfqq_fifo_expire(cfqq);
1530                 cfq_mark_cfqq_slice_new(cfqq);
1531
1532                 del_timer(&cfqd->idle_slice_timer);
1533         }
1534
1535         cfqd->active_queue = cfqq;
1536 }
1537
1538 /*
1539  * current cfqq expired its slice (or was too idle), select new one
1540  */
1541 static void
1542 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1543                     bool timed_out)
1544 {
1545         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1546
1547         if (cfq_cfqq_wait_request(cfqq))
1548                 del_timer(&cfqd->idle_slice_timer);
1549
1550         cfq_clear_cfqq_wait_request(cfqq);
1551         cfq_clear_cfqq_wait_busy(cfqq);
1552
1553         /*
1554          * If this cfqq is shared between multiple processes, check to
1555          * make sure that those processes are still issuing I/Os within
1556          * the mean seek distance.  If not, it may be time to break the
1557          * queues apart again.
1558          */
1559         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1560                 cfq_mark_cfqq_split_coop(cfqq);
1561
1562         /*
1563          * store what was left of this slice, if the queue idled/timed out
1564          */
1565         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1566                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1567                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1568         }
1569
1570         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1571
1572         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1573                 cfq_del_cfqq_rr(cfqd, cfqq);
1574
1575         cfq_resort_rr_list(cfqd, cfqq);
1576
1577         if (cfqq == cfqd->active_queue)
1578                 cfqd->active_queue = NULL;
1579
1580         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1581                 cfqd->grp_service_tree.active = NULL;
1582
1583         if (cfqd->active_cic) {
1584                 put_io_context(cfqd->active_cic->ioc);
1585                 cfqd->active_cic = NULL;
1586         }
1587 }
1588
1589 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1590 {
1591         struct cfq_queue *cfqq = cfqd->active_queue;
1592
1593         if (cfqq)
1594                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1595 }
1596
1597 /*
1598  * Get next queue for service. Unless we have a queue preemption,
1599  * we'll simply select the first cfqq in the service tree.
1600  */
1601 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1602 {
1603         struct cfq_rb_root *service_tree =
1604                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1605                                         cfqd->serving_type);
1606
1607         if (!cfqd->rq_queued)
1608                 return NULL;
1609
1610         /* There is nothing to dispatch */
1611         if (!service_tree)
1612                 return NULL;
1613         if (RB_EMPTY_ROOT(&service_tree->rb))
1614                 return NULL;
1615         return cfq_rb_first(service_tree);
1616 }
1617
1618 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1619 {
1620         struct cfq_group *cfqg;
1621         struct cfq_queue *cfqq;
1622         int i, j;
1623         struct cfq_rb_root *st;
1624
1625         if (!cfqd->rq_queued)
1626                 return NULL;
1627
1628         cfqg = cfq_get_next_cfqg(cfqd);
1629         if (!cfqg)
1630                 return NULL;
1631
1632         for_each_cfqg_st(cfqg, i, j, st)
1633                 if ((cfqq = cfq_rb_first(st)) != NULL)
1634                         return cfqq;
1635         return NULL;
1636 }
1637
1638 /*
1639  * Get and set a new active queue for service.
1640  */
1641 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1642                                               struct cfq_queue *cfqq)
1643 {
1644         if (!cfqq)
1645                 cfqq = cfq_get_next_queue(cfqd);
1646
1647         __cfq_set_active_queue(cfqd, cfqq);
1648         return cfqq;
1649 }
1650
1651 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1652                                           struct request *rq)
1653 {
1654         if (blk_rq_pos(rq) >= cfqd->last_position)
1655                 return blk_rq_pos(rq) - cfqd->last_position;
1656         else
1657                 return cfqd->last_position - blk_rq_pos(rq);
1658 }
1659
1660 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1661                                struct request *rq)
1662 {
1663         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1664 }
1665
1666 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1667                                     struct cfq_queue *cur_cfqq)
1668 {
1669         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1670         struct rb_node *parent, *node;
1671         struct cfq_queue *__cfqq;
1672         sector_t sector = cfqd->last_position;
1673
1674         if (RB_EMPTY_ROOT(root))
1675                 return NULL;
1676
1677         /*
1678          * First, if we find a request starting at the end of the last
1679          * request, choose it.
1680          */
1681         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1682         if (__cfqq)
1683                 return __cfqq;
1684
1685         /*
1686          * If the exact sector wasn't found, the parent of the NULL leaf
1687          * will contain the closest sector.
1688          */
1689         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1690         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1691                 return __cfqq;
1692
1693         if (blk_rq_pos(__cfqq->next_rq) < sector)
1694                 node = rb_next(&__cfqq->p_node);
1695         else
1696                 node = rb_prev(&__cfqq->p_node);
1697         if (!node)
1698                 return NULL;
1699
1700         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1701         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1702                 return __cfqq;
1703
1704         return NULL;
1705 }
1706
1707 /*
1708  * cfqd - obvious
1709  * cur_cfqq - passed in so that we don't decide that the current queue is
1710  *            closely cooperating with itself.
1711  *
1712  * So, basically we're assuming that that cur_cfqq has dispatched at least
1713  * one request, and that cfqd->last_position reflects a position on the disk
1714  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1715  * assumption.
1716  */
1717 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1718                                               struct cfq_queue *cur_cfqq)
1719 {
1720         struct cfq_queue *cfqq;
1721
1722         if (cfq_class_idle(cur_cfqq))
1723                 return NULL;
1724         if (!cfq_cfqq_sync(cur_cfqq))
1725                 return NULL;
1726         if (CFQQ_SEEKY(cur_cfqq))
1727                 return NULL;
1728
1729         /*
1730          * Don't search priority tree if it's the only queue in the group.
1731          */
1732         if (cur_cfqq->cfqg->nr_cfqq == 1)
1733                 return NULL;
1734
1735         /*
1736          * We should notice if some of the queues are cooperating, eg
1737          * working closely on the same area of the disk. In that case,
1738          * we can group them together and don't waste time idling.
1739          */
1740         cfqq = cfqq_close(cfqd, cur_cfqq);
1741         if (!cfqq)
1742                 return NULL;
1743
1744         /* If new queue belongs to different cfq_group, don't choose it */
1745         if (cur_cfqq->cfqg != cfqq->cfqg)
1746                 return NULL;
1747
1748         /*
1749          * It only makes sense to merge sync queues.
1750          */
1751         if (!cfq_cfqq_sync(cfqq))
1752                 return NULL;
1753         if (CFQQ_SEEKY(cfqq))
1754                 return NULL;
1755
1756         /*
1757          * Do not merge queues of different priority classes
1758          */
1759         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1760                 return NULL;
1761
1762         return cfqq;
1763 }
1764
1765 /*
1766  * Determine whether we should enforce idle window for this queue.
1767  */
1768
1769 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1770 {
1771         enum wl_prio_t prio = cfqq_prio(cfqq);
1772         struct cfq_rb_root *service_tree = cfqq->service_tree;
1773
1774         BUG_ON(!service_tree);
1775         BUG_ON(!service_tree->count);
1776
1777         /* We never do for idle class queues. */
1778         if (prio == IDLE_WORKLOAD)
1779                 return false;
1780
1781         /* We do for queues that were marked with idle window flag. */
1782         if (cfq_cfqq_idle_window(cfqq) &&
1783            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1784                 return true;
1785
1786         /*
1787          * Otherwise, we do only if they are the last ones
1788          * in their service tree.
1789          */
1790         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1791                 return 1;
1792         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1793                         service_tree->count);
1794         return 0;
1795 }
1796
1797 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1798 {
1799         struct cfq_queue *cfqq = cfqd->active_queue;
1800         struct cfq_io_context *cic;
1801         unsigned long sl;
1802
1803         /*
1804          * SSD device without seek penalty, disable idling. But only do so
1805          * for devices that support queuing, otherwise we still have a problem
1806          * with sync vs async workloads.
1807          */
1808         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1809                 return;
1810
1811         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1812         WARN_ON(cfq_cfqq_slice_new(cfqq));
1813
1814         /*
1815          * idle is disabled, either manually or by past process history
1816          */
1817         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1818                 return;
1819
1820         /*
1821          * still active requests from this queue, don't idle
1822          */
1823         if (cfqq->dispatched)
1824                 return;
1825
1826         /*
1827          * task has exited, don't wait
1828          */
1829         cic = cfqd->active_cic;
1830         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1831                 return;
1832
1833         /*
1834          * If our average think time is larger than the remaining time
1835          * slice, then don't idle. This avoids overrunning the allotted
1836          * time slice.
1837          */
1838         if (sample_valid(cic->ttime_samples) &&
1839             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1840                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1841                                 cic->ttime_mean);
1842                 return;
1843         }
1844
1845         cfq_mark_cfqq_wait_request(cfqq);
1846
1847         sl = cfqd->cfq_slice_idle;
1848
1849         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1850         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1851 }
1852
1853 /*
1854  * Move request from internal lists to the request queue dispatch list.
1855  */
1856 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1857 {
1858         struct cfq_data *cfqd = q->elevator->elevator_data;
1859         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1860
1861         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1862
1863         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1864         cfq_remove_request(rq);
1865         cfqq->dispatched++;
1866         elv_dispatch_sort(q, rq);
1867
1868         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1869         blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1870                                         rq_data_dir(rq), rq_is_sync(rq));
1871 }
1872
1873 /*
1874  * return expired entry, or NULL to just start from scratch in rbtree
1875  */
1876 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1877 {
1878         struct request *rq = NULL;
1879
1880         if (cfq_cfqq_fifo_expire(cfqq))
1881                 return NULL;
1882
1883         cfq_mark_cfqq_fifo_expire(cfqq);
1884
1885         if (list_empty(&cfqq->fifo))
1886                 return NULL;
1887
1888         rq = rq_entry_fifo(cfqq->fifo.next);
1889         if (time_before(jiffies, rq_fifo_time(rq)))
1890                 rq = NULL;
1891
1892         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1893         return rq;
1894 }
1895
1896 static inline int
1897 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1898 {
1899         const int base_rq = cfqd->cfq_slice_async_rq;
1900
1901         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1902
1903         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1904 }
1905
1906 /*
1907  * Must be called with the queue_lock held.
1908  */
1909 static int cfqq_process_refs(struct cfq_queue *cfqq)
1910 {
1911         int process_refs, io_refs;
1912
1913         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1914         process_refs = atomic_read(&cfqq->ref) - io_refs;
1915         BUG_ON(process_refs < 0);
1916         return process_refs;
1917 }
1918
1919 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1920 {
1921         int process_refs, new_process_refs;
1922         struct cfq_queue *__cfqq;
1923
1924         /* Avoid a circular list and skip interim queue merges */
1925         while ((__cfqq = new_cfqq->new_cfqq)) {
1926                 if (__cfqq == cfqq)
1927                         return;
1928                 new_cfqq = __cfqq;
1929         }
1930
1931         process_refs = cfqq_process_refs(cfqq);
1932         /*
1933          * If the process for the cfqq has gone away, there is no
1934          * sense in merging the queues.
1935          */
1936         if (process_refs == 0)
1937                 return;
1938
1939         /*
1940          * Merge in the direction of the lesser amount of work.
1941          */
1942         new_process_refs = cfqq_process_refs(new_cfqq);
1943         if (new_process_refs >= process_refs) {
1944                 cfqq->new_cfqq = new_cfqq;
1945                 atomic_add(process_refs, &new_cfqq->ref);
1946         } else {
1947                 new_cfqq->new_cfqq = cfqq;
1948                 atomic_add(new_process_refs, &cfqq->ref);
1949         }
1950 }
1951
1952 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1953                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1954 {
1955         struct cfq_queue *queue;
1956         int i;
1957         bool key_valid = false;
1958         unsigned long lowest_key = 0;
1959         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1960
1961         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1962                 /* select the one with lowest rb_key */
1963                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1964                 if (queue &&
1965                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1966                         lowest_key = queue->rb_key;
1967                         cur_best = i;
1968                         key_valid = true;
1969                 }
1970         }
1971
1972         return cur_best;
1973 }
1974
1975 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1976 {
1977         unsigned slice;
1978         unsigned count;
1979         struct cfq_rb_root *st;
1980         unsigned group_slice;
1981
1982         if (!cfqg) {
1983                 cfqd->serving_prio = IDLE_WORKLOAD;
1984                 cfqd->workload_expires = jiffies + 1;
1985                 return;
1986         }
1987
1988         /* Choose next priority. RT > BE > IDLE */
1989         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1990                 cfqd->serving_prio = RT_WORKLOAD;
1991         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1992                 cfqd->serving_prio = BE_WORKLOAD;
1993         else {
1994                 cfqd->serving_prio = IDLE_WORKLOAD;
1995                 cfqd->workload_expires = jiffies + 1;
1996                 return;
1997         }
1998
1999         /*
2000          * For RT and BE, we have to choose also the type
2001          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2002          * expiration time
2003          */
2004         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2005         count = st->count;
2006
2007         /*
2008          * check workload expiration, and that we still have other queues ready
2009          */
2010         if (count && !time_after(jiffies, cfqd->workload_expires))
2011                 return;
2012
2013         /* otherwise select new workload type */
2014         cfqd->serving_type =
2015                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2016         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2017         count = st->count;
2018
2019         /*
2020          * the workload slice is computed as a fraction of target latency
2021          * proportional to the number of queues in that workload, over
2022          * all the queues in the same priority class
2023          */
2024         group_slice = cfq_group_slice(cfqd, cfqg);
2025
2026         slice = group_slice * count /
2027                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2028                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2029
2030         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2031                 unsigned int tmp;
2032
2033                 /*
2034                  * Async queues are currently system wide. Just taking
2035                  * proportion of queues with-in same group will lead to higher
2036                  * async ratio system wide as generally root group is going
2037                  * to have higher weight. A more accurate thing would be to
2038                  * calculate system wide asnc/sync ratio.
2039                  */
2040                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2041                 tmp = tmp/cfqd->busy_queues;
2042                 slice = min_t(unsigned, slice, tmp);
2043
2044                 /* async workload slice is scaled down according to
2045                  * the sync/async slice ratio. */
2046                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2047         } else
2048                 /* sync workload slice is at least 2 * cfq_slice_idle */
2049                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2050
2051         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2052         cfq_log(cfqd, "workload slice:%d", slice);
2053         cfqd->workload_expires = jiffies + slice;
2054         cfqd->noidle_tree_requires_idle = false;
2055 }
2056
2057 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2058 {
2059         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2060         struct cfq_group *cfqg;
2061
2062         if (RB_EMPTY_ROOT(&st->rb))
2063                 return NULL;
2064         cfqg = cfq_rb_first_group(st);
2065         st->active = &cfqg->rb_node;
2066         update_min_vdisktime(st);
2067         return cfqg;
2068 }
2069
2070 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2071 {
2072         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2073
2074         cfqd->serving_group = cfqg;
2075
2076         /* Restore the workload type data */
2077         if (cfqg->saved_workload_slice) {
2078                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2079                 cfqd->serving_type = cfqg->saved_workload;
2080                 cfqd->serving_prio = cfqg->saved_serving_prio;
2081         } else
2082                 cfqd->workload_expires = jiffies - 1;
2083
2084         choose_service_tree(cfqd, cfqg);
2085 }
2086
2087 /*
2088  * Select a queue for service. If we have a current active queue,
2089  * check whether to continue servicing it, or retrieve and set a new one.
2090  */
2091 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2092 {
2093         struct cfq_queue *cfqq, *new_cfqq = NULL;
2094
2095         cfqq = cfqd->active_queue;
2096         if (!cfqq)
2097                 goto new_queue;
2098
2099         if (!cfqd->rq_queued)
2100                 return NULL;
2101
2102         /*
2103          * We were waiting for group to get backlogged. Expire the queue
2104          */
2105         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2106                 goto expire;
2107
2108         /*
2109          * The active queue has run out of time, expire it and select new.
2110          */
2111         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2112                 /*
2113                  * If slice had not expired at the completion of last request
2114                  * we might not have turned on wait_busy flag. Don't expire
2115                  * the queue yet. Allow the group to get backlogged.
2116                  *
2117                  * The very fact that we have used the slice, that means we
2118                  * have been idling all along on this queue and it should be
2119                  * ok to wait for this request to complete.
2120                  */
2121                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2122                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2123                         cfqq = NULL;
2124                         goto keep_queue;
2125                 } else
2126                         goto expire;
2127         }
2128
2129         /*
2130          * The active queue has requests and isn't expired, allow it to
2131          * dispatch.
2132          */
2133         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2134                 goto keep_queue;
2135
2136         /*
2137          * If another queue has a request waiting within our mean seek
2138          * distance, let it run.  The expire code will check for close
2139          * cooperators and put the close queue at the front of the service
2140          * tree.  If possible, merge the expiring queue with the new cfqq.
2141          */
2142         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2143         if (new_cfqq) {
2144                 if (!cfqq->new_cfqq)
2145                         cfq_setup_merge(cfqq, new_cfqq);
2146                 goto expire;
2147         }
2148
2149         /*
2150          * No requests pending. If the active queue still has requests in
2151          * flight or is idling for a new request, allow either of these
2152          * conditions to happen (or time out) before selecting a new queue.
2153          */
2154         if (timer_pending(&cfqd->idle_slice_timer) ||
2155             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2156                 cfqq = NULL;
2157                 goto keep_queue;
2158         }
2159
2160 expire:
2161         cfq_slice_expired(cfqd, 0);
2162 new_queue:
2163         /*
2164          * Current queue expired. Check if we have to switch to a new
2165          * service tree
2166          */
2167         if (!new_cfqq)
2168                 cfq_choose_cfqg(cfqd);
2169
2170         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2171 keep_queue:
2172         return cfqq;
2173 }
2174
2175 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2176 {
2177         int dispatched = 0;
2178
2179         while (cfqq->next_rq) {
2180                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2181                 dispatched++;
2182         }
2183
2184         BUG_ON(!list_empty(&cfqq->fifo));
2185
2186         /* By default cfqq is not expired if it is empty. Do it explicitly */
2187         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2188         return dispatched;
2189 }
2190
2191 /*
2192  * Drain our current requests. Used for barriers and when switching
2193  * io schedulers on-the-fly.
2194  */
2195 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2196 {
2197         struct cfq_queue *cfqq;
2198         int dispatched = 0;
2199
2200         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2201                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2202
2203         cfq_slice_expired(cfqd, 0);
2204         BUG_ON(cfqd->busy_queues);
2205
2206         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2207         return dispatched;
2208 }
2209
2210 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2211         struct cfq_queue *cfqq)
2212 {
2213         /* the queue hasn't finished any request, can't estimate */
2214         if (cfq_cfqq_slice_new(cfqq))
2215                 return 1;
2216         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2217                 cfqq->slice_end))
2218                 return 1;
2219
2220         return 0;
2221 }
2222
2223 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2224 {
2225         unsigned int max_dispatch;
2226
2227         /*
2228          * Drain async requests before we start sync IO
2229          */
2230         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2231                 return false;
2232
2233         /*
2234          * If this is an async queue and we have sync IO in flight, let it wait
2235          */
2236         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2237                 return false;
2238
2239         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2240         if (cfq_class_idle(cfqq))
2241                 max_dispatch = 1;
2242
2243         /*
2244          * Does this cfqq already have too much IO in flight?
2245          */
2246         if (cfqq->dispatched >= max_dispatch) {
2247                 /*
2248                  * idle queue must always only have a single IO in flight
2249                  */
2250                 if (cfq_class_idle(cfqq))
2251                         return false;
2252
2253                 /*
2254                  * We have other queues, don't allow more IO from this one
2255                  */
2256                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2257                         return false;
2258
2259                 /*
2260                  * Sole queue user, no limit
2261                  */
2262                 if (cfqd->busy_queues == 1)
2263                         max_dispatch = -1;
2264                 else
2265                         /*
2266                          * Normally we start throttling cfqq when cfq_quantum/2
2267                          * requests have been dispatched. But we can drive
2268                          * deeper queue depths at the beginning of slice
2269                          * subjected to upper limit of cfq_quantum.
2270                          * */
2271                         max_dispatch = cfqd->cfq_quantum;
2272         }
2273
2274         /*
2275          * Async queues must wait a bit before being allowed dispatch.
2276          * We also ramp up the dispatch depth gradually for async IO,
2277          * based on the last sync IO we serviced
2278          */
2279         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2280                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2281                 unsigned int depth;
2282
2283                 depth = last_sync / cfqd->cfq_slice[1];
2284                 if (!depth && !cfqq->dispatched)
2285                         depth = 1;
2286                 if (depth < max_dispatch)
2287                         max_dispatch = depth;
2288         }
2289
2290         /*
2291          * If we're below the current max, allow a dispatch
2292          */
2293         return cfqq->dispatched < max_dispatch;
2294 }
2295
2296 /*
2297  * Dispatch a request from cfqq, moving them to the request queue
2298  * dispatch list.
2299  */
2300 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2301 {
2302         struct request *rq;
2303
2304         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2305
2306         if (!cfq_may_dispatch(cfqd, cfqq))
2307                 return false;
2308
2309         /*
2310          * follow expired path, else get first next available
2311          */
2312         rq = cfq_check_fifo(cfqq);
2313         if (!rq)
2314                 rq = cfqq->next_rq;
2315
2316         /*
2317          * insert request into driver dispatch list
2318          */
2319         cfq_dispatch_insert(cfqd->queue, rq);
2320
2321         if (!cfqd->active_cic) {
2322                 struct cfq_io_context *cic = RQ_CIC(rq);
2323
2324                 atomic_long_inc(&cic->ioc->refcount);
2325                 cfqd->active_cic = cic;
2326         }
2327
2328         return true;
2329 }
2330
2331 /*
2332  * Find the cfqq that we need to service and move a request from that to the
2333  * dispatch list
2334  */
2335 static int cfq_dispatch_requests(struct request_queue *q, int force)
2336 {
2337         struct cfq_data *cfqd = q->elevator->elevator_data;
2338         struct cfq_queue *cfqq;
2339
2340         if (!cfqd->busy_queues)
2341                 return 0;
2342
2343         if (unlikely(force))
2344                 return cfq_forced_dispatch(cfqd);
2345
2346         cfqq = cfq_select_queue(cfqd);
2347         if (!cfqq)
2348                 return 0;
2349
2350         /*
2351          * Dispatch a request from this cfqq, if it is allowed
2352          */
2353         if (!cfq_dispatch_request(cfqd, cfqq))
2354                 return 0;
2355
2356         cfqq->slice_dispatch++;
2357         cfq_clear_cfqq_must_dispatch(cfqq);
2358
2359         /*
2360          * expire an async queue immediately if it has used up its slice. idle
2361          * queue always expire after 1 dispatch round.
2362          */
2363         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2364             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2365             cfq_class_idle(cfqq))) {
2366                 cfqq->slice_end = jiffies + 1;
2367                 cfq_slice_expired(cfqd, 0);
2368         }
2369
2370         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2371         return 1;
2372 }
2373
2374 /*
2375  * task holds one reference to the queue, dropped when task exits. each rq
2376  * in-flight on this queue also holds a reference, dropped when rq is freed.
2377  *
2378  * Each cfq queue took a reference on the parent group. Drop it now.
2379  * queue lock must be held here.
2380  */
2381 static void cfq_put_queue(struct cfq_queue *cfqq)
2382 {
2383         struct cfq_data *cfqd = cfqq->cfqd;
2384         struct cfq_group *cfqg, *orig_cfqg;
2385
2386         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2387
2388         if (!atomic_dec_and_test(&cfqq->ref))
2389                 return;
2390
2391         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2392         BUG_ON(rb_first(&cfqq->sort_list));
2393         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2394         cfqg = cfqq->cfqg;
2395         orig_cfqg = cfqq->orig_cfqg;
2396
2397         if (unlikely(cfqd->active_queue == cfqq)) {
2398                 __cfq_slice_expired(cfqd, cfqq, 0);
2399                 cfq_schedule_dispatch(cfqd);
2400         }
2401
2402         BUG_ON(cfq_cfqq_on_rr(cfqq));
2403         kmem_cache_free(cfq_pool, cfqq);
2404         cfq_put_cfqg(cfqg);
2405         if (orig_cfqg)
2406                 cfq_put_cfqg(orig_cfqg);
2407 }
2408
2409 /*
2410  * Must always be called with the rcu_read_lock() held
2411  */
2412 static void
2413 __call_for_each_cic(struct io_context *ioc,
2414                     void (*func)(struct io_context *, struct cfq_io_context *))
2415 {
2416         struct cfq_io_context *cic;
2417         struct hlist_node *n;
2418
2419         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2420                 func(ioc, cic);
2421 }
2422
2423 /*
2424  * Call func for each cic attached to this ioc.
2425  */
2426 static void
2427 call_for_each_cic(struct io_context *ioc,
2428                   void (*func)(struct io_context *, struct cfq_io_context *))
2429 {
2430         rcu_read_lock();
2431         __call_for_each_cic(ioc, func);
2432         rcu_read_unlock();
2433 }
2434
2435 static void cfq_cic_free_rcu(struct rcu_head *head)
2436 {
2437         struct cfq_io_context *cic;
2438
2439         cic = container_of(head, struct cfq_io_context, rcu_head);
2440
2441         kmem_cache_free(cfq_ioc_pool, cic);
2442         elv_ioc_count_dec(cfq_ioc_count);
2443
2444         if (ioc_gone) {
2445                 /*
2446                  * CFQ scheduler is exiting, grab exit lock and check
2447                  * the pending io context count. If it hits zero,
2448                  * complete ioc_gone and set it back to NULL
2449                  */
2450                 spin_lock(&ioc_gone_lock);
2451                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2452                         complete(ioc_gone);
2453                         ioc_gone = NULL;
2454                 }
2455                 spin_unlock(&ioc_gone_lock);
2456         }
2457 }
2458
2459 static void cfq_cic_free(struct cfq_io_context *cic)
2460 {
2461         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2462 }
2463
2464 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2465 {
2466         unsigned long flags;
2467
2468         BUG_ON(!cic->dead_key);
2469
2470         spin_lock_irqsave(&ioc->lock, flags);
2471         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2472         hlist_del_rcu(&cic->cic_list);
2473         spin_unlock_irqrestore(&ioc->lock, flags);
2474
2475         cfq_cic_free(cic);
2476 }
2477
2478 /*
2479  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2480  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2481  * and ->trim() which is called with the task lock held
2482  */
2483 static void cfq_free_io_context(struct io_context *ioc)
2484 {
2485         /*
2486          * ioc->refcount is zero here, or we are called from elv_unregister(),
2487          * so no more cic's are allowed to be linked into this ioc.  So it
2488          * should be ok to iterate over the known list, we will see all cic's
2489          * since no new ones are added.
2490          */
2491         __call_for_each_cic(ioc, cic_free_func);
2492 }
2493
2494 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2495 {
2496         struct cfq_queue *__cfqq, *next;
2497
2498         if (unlikely(cfqq == cfqd->active_queue)) {
2499                 __cfq_slice_expired(cfqd, cfqq, 0);
2500                 cfq_schedule_dispatch(cfqd);
2501         }
2502
2503         /*
2504          * If this queue was scheduled to merge with another queue, be
2505          * sure to drop the reference taken on that queue (and others in
2506          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2507          */
2508         __cfqq = cfqq->new_cfqq;
2509         while (__cfqq) {
2510                 if (__cfqq == cfqq) {
2511                         WARN(1, "cfqq->new_cfqq loop detected\n");
2512                         break;
2513                 }
2514                 next = __cfqq->new_cfqq;
2515                 cfq_put_queue(__cfqq);
2516                 __cfqq = next;
2517         }
2518
2519         cfq_put_queue(cfqq);
2520 }
2521
2522 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2523                                          struct cfq_io_context *cic)
2524 {
2525         struct io_context *ioc = cic->ioc;
2526
2527         list_del_init(&cic->queue_list);
2528
2529         /*
2530          * Make sure key == NULL is seen for dead queues
2531          */
2532         smp_wmb();
2533         cic->dead_key = (unsigned long) cic->key;
2534         cic->key = NULL;
2535
2536         if (ioc->ioc_data == cic)
2537                 rcu_assign_pointer(ioc->ioc_data, NULL);
2538
2539         if (cic->cfqq[BLK_RW_ASYNC]) {
2540                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2541                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2542         }
2543
2544         if (cic->cfqq[BLK_RW_SYNC]) {
2545                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2546                 cic->cfqq[BLK_RW_SYNC] = NULL;
2547         }
2548 }
2549
2550 static void cfq_exit_single_io_context(struct io_context *ioc,
2551                                        struct cfq_io_context *cic)
2552 {
2553         struct cfq_data *cfqd = cic->key;
2554
2555         if (cfqd) {
2556                 struct request_queue *q = cfqd->queue;
2557                 unsigned long flags;
2558
2559                 spin_lock_irqsave(q->queue_lock, flags);
2560
2561                 /*
2562                  * Ensure we get a fresh copy of the ->key to prevent
2563                  * race between exiting task and queue
2564                  */
2565                 smp_read_barrier_depends();
2566                 if (cic->key)
2567                         __cfq_exit_single_io_context(cfqd, cic);
2568
2569                 spin_unlock_irqrestore(q->queue_lock, flags);
2570         }
2571 }
2572
2573 /*
2574  * The process that ioc belongs to has exited, we need to clean up
2575  * and put the internal structures we have that belongs to that process.
2576  */
2577 static void cfq_exit_io_context(struct io_context *ioc)
2578 {
2579         call_for_each_cic(ioc, cfq_exit_single_io_context);
2580 }
2581
2582 static struct cfq_io_context *
2583 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2584 {
2585         struct cfq_io_context *cic;
2586
2587         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2588                                                         cfqd->queue->node);
2589         if (cic) {
2590                 cic->last_end_request = jiffies;
2591                 INIT_LIST_HEAD(&cic->queue_list);
2592                 INIT_HLIST_NODE(&cic->cic_list);
2593                 cic->dtor = cfq_free_io_context;
2594                 cic->exit = cfq_exit_io_context;
2595                 elv_ioc_count_inc(cfq_ioc_count);
2596         }
2597
2598         return cic;
2599 }
2600
2601 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2602 {
2603         struct task_struct *tsk = current;
2604         int ioprio_class;
2605
2606         if (!cfq_cfqq_prio_changed(cfqq))
2607                 return;
2608
2609         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2610         switch (ioprio_class) {
2611         default:
2612                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2613         case IOPRIO_CLASS_NONE:
2614                 /*
2615                  * no prio set, inherit CPU scheduling settings
2616                  */
2617                 cfqq->ioprio = task_nice_ioprio(tsk);
2618                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2619                 break;
2620         case IOPRIO_CLASS_RT:
2621                 cfqq->ioprio = task_ioprio(ioc);
2622                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2623                 break;
2624         case IOPRIO_CLASS_BE:
2625                 cfqq->ioprio = task_ioprio(ioc);
2626                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2627                 break;
2628         case IOPRIO_CLASS_IDLE:
2629                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2630                 cfqq->ioprio = 7;
2631                 cfq_clear_cfqq_idle_window(cfqq);
2632                 break;
2633         }
2634
2635         /*
2636          * keep track of original prio settings in case we have to temporarily
2637          * elevate the priority of this queue
2638          */
2639         cfqq->org_ioprio = cfqq->ioprio;
2640         cfqq->org_ioprio_class = cfqq->ioprio_class;
2641         cfq_clear_cfqq_prio_changed(cfqq);
2642 }
2643
2644 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2645 {
2646         struct cfq_data *cfqd = cic->key;
2647         struct cfq_queue *cfqq;
2648         unsigned long flags;
2649
2650         if (unlikely(!cfqd))
2651                 return;
2652
2653         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2654
2655         cfqq = cic->cfqq[BLK_RW_ASYNC];
2656         if (cfqq) {
2657                 struct cfq_queue *new_cfqq;
2658                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2659                                                 GFP_ATOMIC);
2660                 if (new_cfqq) {
2661                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2662                         cfq_put_queue(cfqq);
2663                 }
2664         }
2665
2666         cfqq = cic->cfqq[BLK_RW_SYNC];
2667         if (cfqq)
2668                 cfq_mark_cfqq_prio_changed(cfqq);
2669
2670         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2671 }
2672
2673 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2674 {
2675         call_for_each_cic(ioc, changed_ioprio);
2676         ioc->ioprio_changed = 0;
2677 }
2678
2679 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2680                           pid_t pid, bool is_sync)
2681 {
2682         RB_CLEAR_NODE(&cfqq->rb_node);
2683         RB_CLEAR_NODE(&cfqq->p_node);
2684         INIT_LIST_HEAD(&cfqq->fifo);
2685
2686         atomic_set(&cfqq->ref, 0);
2687         cfqq->cfqd = cfqd;
2688
2689         cfq_mark_cfqq_prio_changed(cfqq);
2690
2691         if (is_sync) {
2692                 if (!cfq_class_idle(cfqq))
2693                         cfq_mark_cfqq_idle_window(cfqq);
2694                 cfq_mark_cfqq_sync(cfqq);
2695         }
2696         cfqq->pid = pid;
2697 }
2698
2699 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2700 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2701 {
2702         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2703         struct cfq_data *cfqd = cic->key;
2704         unsigned long flags;
2705         struct request_queue *q;
2706
2707         if (unlikely(!cfqd))
2708                 return;
2709
2710         q = cfqd->queue;
2711
2712         spin_lock_irqsave(q->queue_lock, flags);
2713
2714         if (sync_cfqq) {
2715                 /*
2716                  * Drop reference to sync queue. A new sync queue will be
2717                  * assigned in new group upon arrival of a fresh request.
2718                  */
2719                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2720                 cic_set_cfqq(cic, NULL, 1);
2721                 cfq_put_queue(sync_cfqq);
2722         }
2723
2724         spin_unlock_irqrestore(q->queue_lock, flags);
2725 }
2726
2727 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2728 {
2729         call_for_each_cic(ioc, changed_cgroup);
2730         ioc->cgroup_changed = 0;
2731 }
2732 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2733
2734 static struct cfq_queue *
2735 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2736                      struct io_context *ioc, gfp_t gfp_mask)
2737 {
2738         struct cfq_queue *cfqq, *new_cfqq = NULL;
2739         struct cfq_io_context *cic;
2740         struct cfq_group *cfqg;
2741
2742 retry:
2743         cfqg = cfq_get_cfqg(cfqd, 1);
2744         cic = cfq_cic_lookup(cfqd, ioc);
2745         /* cic always exists here */
2746         cfqq = cic_to_cfqq(cic, is_sync);
2747
2748         /*
2749          * Always try a new alloc if we fell back to the OOM cfqq
2750          * originally, since it should just be a temporary situation.
2751          */
2752         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2753                 cfqq = NULL;
2754                 if (new_cfqq) {
2755                         cfqq = new_cfqq;
2756                         new_cfqq = NULL;
2757                 } else if (gfp_mask & __GFP_WAIT) {
2758                         spin_unlock_irq(cfqd->queue->queue_lock);
2759                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2760                                         gfp_mask | __GFP_ZERO,
2761                                         cfqd->queue->node);
2762                         spin_lock_irq(cfqd->queue->queue_lock);
2763                         if (new_cfqq)
2764                                 goto retry;
2765                 } else {
2766                         cfqq = kmem_cache_alloc_node(cfq_pool,
2767                                         gfp_mask | __GFP_ZERO,
2768                                         cfqd->queue->node);
2769                 }
2770
2771                 if (cfqq) {
2772                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2773                         cfq_init_prio_data(cfqq, ioc);
2774                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2775                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2776                 } else
2777                         cfqq = &cfqd->oom_cfqq;
2778         }
2779
2780         if (new_cfqq)
2781                 kmem_cache_free(cfq_pool, new_cfqq);
2782
2783         return cfqq;
2784 }
2785
2786 static struct cfq_queue **
2787 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2788 {
2789         switch (ioprio_class) {
2790         case IOPRIO_CLASS_RT:
2791                 return &cfqd->async_cfqq[0][ioprio];
2792         case IOPRIO_CLASS_BE:
2793                 return &cfqd->async_cfqq[1][ioprio];
2794         case IOPRIO_CLASS_IDLE:
2795                 return &cfqd->async_idle_cfqq;
2796         default:
2797                 BUG();
2798         }
2799 }
2800
2801 static struct cfq_queue *
2802 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2803               gfp_t gfp_mask)
2804 {
2805         const int ioprio = task_ioprio(ioc);
2806         const int ioprio_class = task_ioprio_class(ioc);
2807         struct cfq_queue **async_cfqq = NULL;
2808         struct cfq_queue *cfqq = NULL;
2809
2810         if (!is_sync) {
2811                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2812                 cfqq = *async_cfqq;
2813         }
2814
2815         if (!cfqq)
2816                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2817
2818         /*
2819          * pin the queue now that it's allocated, scheduler exit will prune it
2820          */
2821         if (!is_sync && !(*async_cfqq)) {
2822                 atomic_inc(&cfqq->ref);
2823                 *async_cfqq = cfqq;
2824         }
2825
2826         atomic_inc(&cfqq->ref);
2827         return cfqq;
2828 }
2829
2830 /*
2831  * We drop cfq io contexts lazily, so we may find a dead one.
2832  */
2833 static void
2834 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2835                   struct cfq_io_context *cic)
2836 {
2837         unsigned long flags;
2838
2839         WARN_ON(!list_empty(&cic->queue_list));
2840
2841         spin_lock_irqsave(&ioc->lock, flags);
2842
2843         BUG_ON(ioc->ioc_data == cic);
2844
2845         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2846         hlist_del_rcu(&cic->cic_list);
2847         spin_unlock_irqrestore(&ioc->lock, flags);
2848
2849         cfq_cic_free(cic);
2850 }
2851
2852 static struct cfq_io_context *
2853 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2854 {
2855         struct cfq_io_context *cic;
2856         unsigned long flags;
2857         void *k;
2858
2859         if (unlikely(!ioc))
2860                 return NULL;
2861
2862         rcu_read_lock();
2863
2864         /*
2865          * we maintain a last-hit cache, to avoid browsing over the tree
2866          */
2867         cic = rcu_dereference(ioc->ioc_data);
2868         if (cic && cic->key == cfqd) {
2869                 rcu_read_unlock();
2870                 return cic;
2871         }
2872
2873         do {
2874                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2875                 rcu_read_unlock();
2876                 if (!cic)
2877                         break;
2878                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2879                 k = cic->key;
2880                 if (unlikely(!k)) {
2881                         cfq_drop_dead_cic(cfqd, ioc, cic);
2882                         rcu_read_lock();
2883                         continue;
2884                 }
2885
2886                 spin_lock_irqsave(&ioc->lock, flags);
2887                 rcu_assign_pointer(ioc->ioc_data, cic);
2888                 spin_unlock_irqrestore(&ioc->lock, flags);
2889                 break;
2890         } while (1);
2891
2892         return cic;
2893 }
2894
2895 /*
2896  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2897  * the process specific cfq io context when entered from the block layer.
2898  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2899  */
2900 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2901                         struct cfq_io_context *cic, gfp_t gfp_mask)
2902 {
2903         unsigned long flags;
2904         int ret;
2905
2906         ret = radix_tree_preload(gfp_mask);
2907         if (!ret) {
2908                 cic->ioc = ioc;
2909                 cic->key = cfqd;
2910
2911                 spin_lock_irqsave(&ioc->lock, flags);
2912                 ret = radix_tree_insert(&ioc->radix_root,
2913                                                 (unsigned long) cfqd, cic);
2914                 if (!ret)
2915                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2916                 spin_unlock_irqrestore(&ioc->lock, flags);
2917
2918                 radix_tree_preload_end();
2919
2920                 if (!ret) {
2921                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2922                         list_add(&cic->queue_list, &cfqd->cic_list);
2923                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2924                 }
2925         }
2926
2927         if (ret)
2928                 printk(KERN_ERR "cfq: cic link failed!\n");
2929
2930         return ret;
2931 }
2932
2933 /*
2934  * Setup general io context and cfq io context. There can be several cfq
2935  * io contexts per general io context, if this process is doing io to more
2936  * than one device managed by cfq.
2937  */
2938 static struct cfq_io_context *
2939 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2940 {
2941         struct io_context *ioc = NULL;
2942         struct cfq_io_context *cic;
2943
2944         might_sleep_if(gfp_mask & __GFP_WAIT);
2945
2946         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2947         if (!ioc)
2948                 return NULL;
2949
2950         cic = cfq_cic_lookup(cfqd, ioc);
2951         if (cic)
2952                 goto out;
2953
2954         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2955         if (cic == NULL)
2956                 goto err;
2957
2958         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2959                 goto err_free;
2960
2961 out:
2962         smp_read_barrier_depends();
2963         if (unlikely(ioc->ioprio_changed))
2964                 cfq_ioc_set_ioprio(ioc);
2965
2966 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2967         if (unlikely(ioc->cgroup_changed))
2968                 cfq_ioc_set_cgroup(ioc);
2969 #endif
2970         return cic;
2971 err_free:
2972         cfq_cic_free(cic);
2973 err:
2974         put_io_context(ioc);
2975         return NULL;
2976 }
2977
2978 static void
2979 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2980 {
2981         unsigned long elapsed = jiffies - cic->last_end_request;
2982         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2983
2984         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2985         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2986         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2987 }
2988
2989 static void
2990 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2991                        struct request *rq)
2992 {
2993         sector_t sdist = 0;
2994         sector_t n_sec = blk_rq_sectors(rq);
2995         if (cfqq->last_request_pos) {
2996                 if (cfqq->last_request_pos < blk_rq_pos(rq))
2997                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2998                 else
2999                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3000         }
3001
3002         cfqq->seek_history <<= 1;
3003         if (blk_queue_nonrot(cfqd->queue))
3004                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3005         else
3006                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3007 }
3008
3009 /*
3010  * Disable idle window if the process thinks too long or seeks so much that
3011  * it doesn't matter
3012  */
3013 static void
3014 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3015                        struct cfq_io_context *cic)
3016 {
3017         int old_idle, enable_idle;
3018
3019         /*
3020          * Don't idle for async or idle io prio class
3021          */
3022         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3023                 return;
3024
3025         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3026
3027         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3028                 cfq_mark_cfqq_deep(cfqq);
3029
3030         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3031             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3032                 enable_idle = 0;
3033         else if (sample_valid(cic->ttime_samples)) {
3034                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3035                         enable_idle = 0;
3036                 else
3037                         enable_idle = 1;
3038         }
3039
3040         if (old_idle != enable_idle) {
3041                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3042                 if (enable_idle)
3043                         cfq_mark_cfqq_idle_window(cfqq);
3044                 else
3045                         cfq_clear_cfqq_idle_window(cfqq);
3046         }
3047 }
3048
3049 /*
3050  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3051  * no or if we aren't sure, a 1 will cause a preempt.
3052  */
3053 static bool
3054 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3055                    struct request *rq)
3056 {
3057         struct cfq_queue *cfqq;
3058
3059         cfqq = cfqd->active_queue;
3060         if (!cfqq)
3061                 return false;
3062
3063         if (cfq_class_idle(new_cfqq))
3064                 return false;
3065
3066         if (cfq_class_idle(cfqq))
3067                 return true;
3068
3069         /*
3070          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3071          */
3072         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3073                 return false;
3074
3075         /*
3076          * if the new request is sync, but the currently running queue is
3077          * not, let the sync request have priority.
3078          */
3079         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3080                 return true;
3081
3082         if (new_cfqq->cfqg != cfqq->cfqg)
3083                 return false;
3084
3085         if (cfq_slice_used(cfqq))
3086                 return true;
3087
3088         /* Allow preemption only if we are idling on sync-noidle tree */
3089         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3090             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3091             new_cfqq->service_tree->count == 2 &&
3092             RB_EMPTY_ROOT(&cfqq->sort_list))
3093                 return true;
3094
3095         /*
3096          * So both queues are sync. Let the new request get disk time if
3097          * it's a metadata request and the current queue is doing regular IO.
3098          */
3099         if (rq_is_meta(rq) && !cfqq->meta_pending)
3100                 return true;
3101
3102         /*
3103          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3104          */
3105         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3106                 return true;
3107
3108         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3109                 return false;
3110
3111         /*
3112          * if this request is as-good as one we would expect from the
3113          * current cfqq, let it preempt
3114          */
3115         if (cfq_rq_close(cfqd, cfqq, rq))
3116                 return true;
3117
3118         return false;
3119 }
3120
3121 /*
3122  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3123  * let it have half of its nominal slice.
3124  */
3125 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3126 {
3127         cfq_log_cfqq(cfqd, cfqq, "preempt");
3128         cfq_slice_expired(cfqd, 1);
3129
3130         /*
3131          * Put the new queue at the front of the of the current list,
3132          * so we know that it will be selected next.
3133          */
3134         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3135
3136         cfq_service_tree_add(cfqd, cfqq, 1);
3137
3138         cfqq->slice_end = 0;
3139         cfq_mark_cfqq_slice_new(cfqq);
3140 }
3141
3142 /*
3143  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3144  * something we should do about it
3145  */
3146 static void
3147 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3148                 struct request *rq)
3149 {
3150         struct cfq_io_context *cic = RQ_CIC(rq);
3151
3152         cfqd->rq_queued++;
3153         if (rq_is_meta(rq))
3154                 cfqq->meta_pending++;
3155
3156         cfq_update_io_thinktime(cfqd, cic);
3157         cfq_update_io_seektime(cfqd, cfqq, rq);
3158         cfq_update_idle_window(cfqd, cfqq, cic);
3159
3160         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3161
3162         if (cfqq == cfqd->active_queue) {
3163                 /*
3164                  * Remember that we saw a request from this process, but
3165                  * don't start queuing just yet. Otherwise we risk seeing lots
3166                  * of tiny requests, because we disrupt the normal plugging
3167                  * and merging. If the request is already larger than a single
3168                  * page, let it rip immediately. For that case we assume that
3169                  * merging is already done. Ditto for a busy system that
3170                  * has other work pending, don't risk delaying until the
3171                  * idle timer unplug to continue working.
3172                  */
3173                 if (cfq_cfqq_wait_request(cfqq)) {
3174                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3175                             cfqd->busy_queues > 1) {
3176                                 del_timer(&cfqd->idle_slice_timer);
3177                                 cfq_clear_cfqq_wait_request(cfqq);
3178                                 __blk_run_queue(cfqd->queue);
3179                         } else
3180                                 cfq_mark_cfqq_must_dispatch(cfqq);
3181                 }
3182         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3183                 /*
3184                  * not the active queue - expire current slice if it is
3185                  * idle and has expired it's mean thinktime or this new queue
3186                  * has some old slice time left and is of higher priority or
3187                  * this new queue is RT and the current one is BE
3188                  */
3189                 cfq_preempt_queue(cfqd, cfqq);
3190                 __blk_run_queue(cfqd->queue);
3191         }
3192 }
3193
3194 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3195 {
3196         struct cfq_data *cfqd = q->elevator->elevator_data;
3197         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3198
3199         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3200         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3201
3202         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3203         list_add_tail(&rq->queuelist, &cfqq->fifo);
3204         cfq_add_rq_rb(rq);
3205
3206         cfq_rq_enqueued(cfqd, cfqq, rq);
3207 }
3208
3209 /*
3210  * Update hw_tag based on peak queue depth over 50 samples under
3211  * sufficient load.
3212  */
3213 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3214 {
3215         struct cfq_queue *cfqq = cfqd->active_queue;
3216
3217         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3218                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3219
3220         if (cfqd->hw_tag == 1)
3221                 return;
3222
3223         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3224             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3225                 return;
3226
3227         /*
3228          * If active queue hasn't enough requests and can idle, cfq might not
3229          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3230          * case
3231          */
3232         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3233             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3234             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3235                 return;
3236
3237         if (cfqd->hw_tag_samples++ < 50)
3238                 return;
3239
3240         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3241                 cfqd->hw_tag = 1;
3242         else
3243                 cfqd->hw_tag = 0;
3244 }
3245
3246 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3247 {
3248         struct cfq_io_context *cic = cfqd->active_cic;
3249
3250         /* If there are other queues in the group, don't wait */
3251         if (cfqq->cfqg->nr_cfqq > 1)
3252                 return false;
3253
3254         if (cfq_slice_used(cfqq))
3255                 return true;
3256
3257         /* if slice left is less than think time, wait busy */
3258         if (cic && sample_valid(cic->ttime_samples)
3259             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3260                 return true;
3261
3262         /*
3263          * If think times is less than a jiffy than ttime_mean=0 and above
3264          * will not be true. It might happen that slice has not expired yet
3265          * but will expire soon (4-5 ns) during select_queue(). To cover the
3266          * case where think time is less than a jiffy, mark the queue wait
3267          * busy if only 1 jiffy is left in the slice.
3268          */
3269         if (cfqq->slice_end - jiffies == 1)
3270                 return true;
3271
3272         return false;
3273 }
3274
3275 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3276 {
3277         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3278         struct cfq_data *cfqd = cfqq->cfqd;
3279         const int sync = rq_is_sync(rq);
3280         unsigned long now;
3281
3282         now = jiffies;
3283         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3284
3285         cfq_update_hw_tag(cfqd);
3286
3287         WARN_ON(!cfqd->rq_in_driver);
3288         WARN_ON(!cfqq->dispatched);
3289         cfqd->rq_in_driver--;
3290         cfqq->dispatched--;
3291         blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3292                         rq_io_start_time_ns(rq), rq_data_dir(rq),
3293                         rq_is_sync(rq));
3294
3295         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3296
3297         if (sync) {
3298                 RQ_CIC(rq)->last_end_request = now;
3299                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3300                         cfqd->last_delayed_sync = now;
3301         }
3302
3303         /*
3304          * If this is the active queue, check if it needs to be expired,
3305          * or if we want to idle in case it has no pending requests.
3306          */
3307         if (cfqd->active_queue == cfqq) {
3308                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3309
3310                 if (cfq_cfqq_slice_new(cfqq)) {
3311                         cfq_set_prio_slice(cfqd, cfqq);
3312                         cfq_clear_cfqq_slice_new(cfqq);
3313                 }
3314
3315                 /*
3316                  * Should we wait for next request to come in before we expire
3317                  * the queue.
3318                  */
3319                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3320                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3321                         cfq_mark_cfqq_wait_busy(cfqq);
3322                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3323                 }
3324
3325                 /*
3326                  * Idling is not enabled on:
3327                  * - expired queues
3328                  * - idle-priority queues
3329                  * - async queues
3330                  * - queues with still some requests queued
3331                  * - when there is a close cooperator
3332                  */
3333                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3334                         cfq_slice_expired(cfqd, 1);
3335                 else if (sync && cfqq_empty &&
3336                          !cfq_close_cooperator(cfqd, cfqq)) {
3337                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3338                         /*
3339                          * Idling is enabled for SYNC_WORKLOAD.
3340                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3341                          * only if we processed at least one !rq_noidle request
3342                          */
3343                         if (cfqd->serving_type == SYNC_WORKLOAD
3344                             || cfqd->noidle_tree_requires_idle
3345                             || cfqq->cfqg->nr_cfqq == 1)
3346                                 cfq_arm_slice_timer(cfqd);
3347                 }
3348         }
3349
3350         if (!cfqd->rq_in_driver)
3351                 cfq_schedule_dispatch(cfqd);
3352 }
3353
3354 /*
3355  * we temporarily boost lower priority queues if they are holding fs exclusive
3356  * resources. they are boosted to normal prio (CLASS_BE/4)
3357  */
3358 static void cfq_prio_boost(struct cfq_queue *cfqq)
3359 {
3360         if (has_fs_excl()) {
3361                 /*
3362                  * boost idle prio on transactions that would lock out other
3363                  * users of the filesystem
3364                  */
3365                 if (cfq_class_idle(cfqq))
3366                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3367                 if (cfqq->ioprio > IOPRIO_NORM)
3368                         cfqq->ioprio = IOPRIO_NORM;
3369         } else {
3370                 /*
3371                  * unboost the queue (if needed)
3372                  */
3373                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3374                 cfqq->ioprio = cfqq->org_ioprio;
3375         }
3376 }
3377
3378 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3379 {
3380         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3381                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3382                 return ELV_MQUEUE_MUST;
3383         }
3384
3385         return ELV_MQUEUE_MAY;
3386 }
3387
3388 static int cfq_may_queue(struct request_queue *q, int rw)
3389 {
3390         struct cfq_data *cfqd = q->elevator->elevator_data;
3391         struct task_struct *tsk = current;
3392         struct cfq_io_context *cic;
3393         struct cfq_queue *cfqq;
3394
3395         /*
3396          * don't force setup of a queue from here, as a call to may_queue
3397          * does not necessarily imply that a request actually will be queued.
3398          * so just lookup a possibly existing queue, or return 'may queue'
3399          * if that fails
3400          */
3401         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3402         if (!cic)
3403                 return ELV_MQUEUE_MAY;
3404
3405         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3406         if (cfqq) {
3407                 cfq_init_prio_data(cfqq, cic->ioc);
3408                 cfq_prio_boost(cfqq);
3409
3410                 return __cfq_may_queue(cfqq);
3411         }
3412
3413         return ELV_MQUEUE_MAY;
3414 }
3415
3416 /*
3417  * queue lock held here
3418  */
3419 static void cfq_put_request(struct request *rq)
3420 {
3421         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3422
3423         if (cfqq) {
3424                 const int rw = rq_data_dir(rq);
3425
3426                 BUG_ON(!cfqq->allocated[rw]);
3427                 cfqq->allocated[rw]--;
3428
3429                 put_io_context(RQ_CIC(rq)->ioc);
3430
3431                 rq->elevator_private = NULL;
3432                 rq->elevator_private2 = NULL;
3433
3434                 cfq_put_queue(cfqq);
3435         }
3436 }
3437
3438 static struct cfq_queue *
3439 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3440                 struct cfq_queue *cfqq)
3441 {
3442         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3443         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3444         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3445         cfq_put_queue(cfqq);
3446         return cic_to_cfqq(cic, 1);
3447 }
3448
3449 /*
3450  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3451  * was the last process referring to said cfqq.
3452  */
3453 static struct cfq_queue *
3454 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3455 {
3456         if (cfqq_process_refs(cfqq) == 1) {
3457                 cfqq->pid = current->pid;
3458                 cfq_clear_cfqq_coop(cfqq);
3459                 cfq_clear_cfqq_split_coop(cfqq);
3460                 return cfqq;
3461         }
3462
3463         cic_set_cfqq(cic, NULL, 1);
3464         cfq_put_queue(cfqq);
3465         return NULL;
3466 }
3467 /*
3468  * Allocate cfq data structures associated with this request.
3469  */
3470 static int
3471 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3472 {
3473         struct cfq_data *cfqd = q->elevator->elevator_data;
3474         struct cfq_io_context *cic;
3475         const int rw = rq_data_dir(rq);
3476         const bool is_sync = rq_is_sync(rq);
3477         struct cfq_queue *cfqq;
3478         unsigned long flags;
3479
3480         might_sleep_if(gfp_mask & __GFP_WAIT);
3481
3482         cic = cfq_get_io_context(cfqd, gfp_mask);
3483
3484         spin_lock_irqsave(q->queue_lock, flags);
3485
3486         if (!cic)
3487                 goto queue_fail;
3488
3489 new_queue:
3490         cfqq = cic_to_cfqq(cic, is_sync);
3491         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3492                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3493                 cic_set_cfqq(cic, cfqq, is_sync);
3494         } else {
3495                 /*
3496                  * If the queue was seeky for too long, break it apart.
3497                  */
3498                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3499                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3500                         cfqq = split_cfqq(cic, cfqq);
3501                         if (!cfqq)
3502                                 goto new_queue;
3503                 }
3504
3505                 /*
3506                  * Check to see if this queue is scheduled to merge with
3507                  * another, closely cooperating queue.  The merging of
3508                  * queues happens here as it must be done in process context.
3509                  * The reference on new_cfqq was taken in merge_cfqqs.
3510                  */
3511                 if (cfqq->new_cfqq)
3512                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3513         }
3514
3515         cfqq->allocated[rw]++;
3516         atomic_inc(&cfqq->ref);
3517
3518         spin_unlock_irqrestore(q->queue_lock, flags);
3519
3520         rq->elevator_private = cic;
3521         rq->elevator_private2 = cfqq;
3522         return 0;
3523
3524 queue_fail:
3525         if (cic)
3526                 put_io_context(cic->ioc);
3527
3528         cfq_schedule_dispatch(cfqd);
3529         spin_unlock_irqrestore(q->queue_lock, flags);
3530         cfq_log(cfqd, "set_request fail");
3531         return 1;
3532 }
3533
3534 static void cfq_kick_queue(struct work_struct *work)
3535 {
3536         struct cfq_data *cfqd =
3537                 container_of(work, struct cfq_data, unplug_work);
3538         struct request_queue *q = cfqd->queue;
3539
3540         spin_lock_irq(q->queue_lock);
3541         __blk_run_queue(cfqd->queue);
3542         spin_unlock_irq(q->queue_lock);
3543 }
3544
3545 /*
3546  * Timer running if the active_queue is currently idling inside its time slice
3547  */
3548 static void cfq_idle_slice_timer(unsigned long data)
3549 {
3550         struct cfq_data *cfqd = (struct cfq_data *) data;
3551         struct cfq_queue *cfqq;
3552         unsigned long flags;
3553         int timed_out = 1;
3554
3555         cfq_log(cfqd, "idle timer fired");
3556
3557         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3558
3559         cfqq = cfqd->active_queue;
3560         if (cfqq) {
3561                 timed_out = 0;
3562
3563                 /*
3564                  * We saw a request before the queue expired, let it through
3565                  */
3566                 if (cfq_cfqq_must_dispatch(cfqq))
3567                         goto out_kick;
3568
3569                 /*
3570                  * expired
3571                  */
3572                 if (cfq_slice_used(cfqq))
3573                         goto expire;
3574
3575                 /*
3576                  * only expire and reinvoke request handler, if there are
3577                  * other queues with pending requests
3578                  */
3579                 if (!cfqd->busy_queues)
3580                         goto out_cont;
3581
3582                 /*
3583                  * not expired and it has a request pending, let it dispatch
3584                  */
3585                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3586                         goto out_kick;
3587
3588                 /*
3589                  * Queue depth flag is reset only when the idle didn't succeed
3590                  */
3591                 cfq_clear_cfqq_deep(cfqq);
3592         }
3593 expire:
3594         cfq_slice_expired(cfqd, timed_out);
3595 out_kick:
3596         cfq_schedule_dispatch(cfqd);
3597 out_cont:
3598         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3599 }
3600
3601 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3602 {
3603         del_timer_sync(&cfqd->idle_slice_timer);
3604         cancel_work_sync(&cfqd->unplug_work);
3605 }
3606
3607 static void cfq_put_async_queues(struct cfq_data *cfqd)
3608 {
3609         int i;
3610
3611         for (i = 0; i < IOPRIO_BE_NR; i++) {
3612                 if (cfqd->async_cfqq[0][i])
3613                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3614                 if (cfqd->async_cfqq[1][i])
3615                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3616         }
3617
3618         if (cfqd->async_idle_cfqq)
3619                 cfq_put_queue(cfqd->async_idle_cfqq);
3620 }
3621
3622 static void cfq_cfqd_free(struct rcu_head *head)
3623 {
3624         kfree(container_of(head, struct cfq_data, rcu));
3625 }
3626
3627 static void cfq_exit_queue(struct elevator_queue *e)
3628 {
3629         struct cfq_data *cfqd = e->elevator_data;
3630         struct request_queue *q = cfqd->queue;
3631
3632         cfq_shutdown_timer_wq(cfqd);
3633
3634         spin_lock_irq(q->queue_lock);
3635
3636         if (cfqd->active_queue)
3637                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3638
3639         while (!list_empty(&cfqd->cic_list)) {
3640                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3641                                                         struct cfq_io_context,
3642                                                         queue_list);
3643
3644                 __cfq_exit_single_io_context(cfqd, cic);
3645         }
3646
3647         cfq_put_async_queues(cfqd);
3648         cfq_release_cfq_groups(cfqd);
3649         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3650
3651         spin_unlock_irq(q->queue_lock);
3652
3653         cfq_shutdown_timer_wq(cfqd);
3654
3655         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3656         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3657 }
3658
3659 static void *cfq_init_queue(struct request_queue *q)
3660 {
3661         struct cfq_data *cfqd;
3662         int i, j;
3663         struct cfq_group *cfqg;
3664         struct cfq_rb_root *st;
3665
3666         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3667         if (!cfqd)
3668                 return NULL;
3669
3670         /* Init root service tree */
3671         cfqd->grp_service_tree = CFQ_RB_ROOT;
3672
3673         /* Init root group */
3674         cfqg = &cfqd->root_group;
3675         for_each_cfqg_st(cfqg, i, j, st)
3676                 *st = CFQ_RB_ROOT;
3677         RB_CLEAR_NODE(&cfqg->rb_node);
3678
3679         /* Give preference to root group over other groups */
3680         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3681
3682 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3683         /*
3684          * Take a reference to root group which we never drop. This is just
3685          * to make sure that cfq_put_cfqg() does not try to kfree root group
3686          */
3687         atomic_set(&cfqg->ref, 1);
3688         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3689                                         0);
3690 #endif
3691         /*
3692          * Not strictly needed (since RB_ROOT just clears the node and we
3693          * zeroed cfqd on alloc), but better be safe in case someone decides
3694          * to add magic to the rb code
3695          */
3696         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3697                 cfqd->prio_trees[i] = RB_ROOT;
3698
3699         /*
3700          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3701          * Grab a permanent reference to it, so that the normal code flow
3702          * will not attempt to free it.
3703          */
3704         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3705         atomic_inc(&cfqd->oom_cfqq.ref);
3706         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3707
3708         INIT_LIST_HEAD(&cfqd->cic_list);
3709
3710         cfqd->queue = q;
3711
3712         init_timer(&cfqd->idle_slice_timer);
3713         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3714         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3715
3716         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3717
3718         cfqd->cfq_quantum = cfq_quantum;
3719         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3720         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3721         cfqd->cfq_back_max = cfq_back_max;
3722         cfqd->cfq_back_penalty = cfq_back_penalty;
3723         cfqd->cfq_slice[0] = cfq_slice_async;
3724         cfqd->cfq_slice[1] = cfq_slice_sync;
3725         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3726         cfqd->cfq_slice_idle = cfq_slice_idle;
3727         cfqd->cfq_latency = 1;
3728         cfqd->cfq_group_isolation = 0;
3729         cfqd->hw_tag = -1;
3730         /*
3731          * we optimistically start assuming sync ops weren't delayed in last
3732          * second, in order to have larger depth for async operations.
3733          */
3734         cfqd->last_delayed_sync = jiffies - HZ;
3735         INIT_RCU_HEAD(&cfqd->rcu);
3736         return cfqd;
3737 }
3738
3739 static void cfq_slab_kill(void)
3740 {
3741         /*
3742          * Caller already ensured that pending RCU callbacks are completed,
3743          * so we should have no busy allocations at this point.
3744          */
3745         if (cfq_pool)
3746                 kmem_cache_destroy(cfq_pool);
3747         if (cfq_ioc_pool)
3748                 kmem_cache_destroy(cfq_ioc_pool);
3749 }
3750
3751 static int __init cfq_slab_setup(void)
3752 {
3753         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3754         if (!cfq_pool)
3755                 goto fail;
3756
3757         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3758         if (!cfq_ioc_pool)
3759                 goto fail;
3760
3761         return 0;
3762 fail:
3763         cfq_slab_kill();
3764         return -ENOMEM;
3765 }
3766
3767 /*
3768  * sysfs parts below -->
3769  */
3770 static ssize_t
3771 cfq_var_show(unsigned int var, char *page)
3772 {
3773         return sprintf(page, "%d\n", var);
3774 }
3775
3776 static ssize_t
3777 cfq_var_store(unsigned int *var, const char *page, size_t count)
3778 {
3779         char *p = (char *) page;
3780
3781         *var = simple_strtoul(p, &p, 10);
3782         return count;
3783 }
3784
3785 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3786 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3787 {                                                                       \
3788         struct cfq_data *cfqd = e->elevator_data;                       \
3789         unsigned int __data = __VAR;                                    \
3790         if (__CONV)                                                     \
3791                 __data = jiffies_to_msecs(__data);                      \
3792         return cfq_var_show(__data, (page));                            \
3793 }
3794 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3795 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3796 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3797 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3798 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3799 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3800 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3801 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3802 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3803 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3804 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3805 #undef SHOW_FUNCTION
3806
3807 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3808 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3809 {                                                                       \
3810         struct cfq_data *cfqd = e->elevator_data;                       \
3811         unsigned int __data;                                            \
3812         int ret = cfq_var_store(&__data, (page), count);                \
3813         if (__data < (MIN))                                             \
3814                 __data = (MIN);                                         \
3815         else if (__data > (MAX))                                        \
3816                 __data = (MAX);                                         \
3817         if (__CONV)                                                     \
3818                 *(__PTR) = msecs_to_jiffies(__data);                    \
3819         else                                                            \
3820                 *(__PTR) = __data;                                      \
3821         return ret;                                                     \
3822 }
3823 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3824 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3825                 UINT_MAX, 1);
3826 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3827                 UINT_MAX, 1);
3828 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3829 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3830                 UINT_MAX, 0);
3831 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3832 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3833 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3834 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3835                 UINT_MAX, 0);
3836 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3837 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3838 #undef STORE_FUNCTION
3839
3840 #define CFQ_ATTR(name) \
3841         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3842
3843 static struct elv_fs_entry cfq_attrs[] = {
3844         CFQ_ATTR(quantum),
3845         CFQ_ATTR(fifo_expire_sync),
3846         CFQ_ATTR(fifo_expire_async),
3847         CFQ_ATTR(back_seek_max),
3848         CFQ_ATTR(back_seek_penalty),
3849         CFQ_ATTR(slice_sync),
3850         CFQ_ATTR(slice_async),
3851         CFQ_ATTR(slice_async_rq),
3852         CFQ_ATTR(slice_idle),
3853         CFQ_ATTR(low_latency),
3854         CFQ_ATTR(group_isolation),
3855         __ATTR_NULL
3856 };
3857
3858 static struct elevator_type iosched_cfq = {
3859         .ops = {
3860                 .elevator_merge_fn =            cfq_merge,
3861                 .elevator_merged_fn =           cfq_merged_request,
3862                 .elevator_merge_req_fn =        cfq_merged_requests,
3863                 .elevator_allow_merge_fn =      cfq_allow_merge,
3864                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3865                 .elevator_add_req_fn =          cfq_insert_request,
3866                 .elevator_activate_req_fn =     cfq_activate_request,
3867                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3868                 .elevator_queue_empty_fn =      cfq_queue_empty,
3869                 .elevator_completed_req_fn =    cfq_completed_request,
3870                 .elevator_former_req_fn =       elv_rb_former_request,
3871                 .elevator_latter_req_fn =       elv_rb_latter_request,
3872                 .elevator_set_req_fn =          cfq_set_request,
3873                 .elevator_put_req_fn =          cfq_put_request,
3874                 .elevator_may_queue_fn =        cfq_may_queue,
3875                 .elevator_init_fn =             cfq_init_queue,
3876                 .elevator_exit_fn =             cfq_exit_queue,
3877                 .trim =                         cfq_free_io_context,
3878         },
3879         .elevator_attrs =       cfq_attrs,
3880         .elevator_name =        "cfq",
3881         .elevator_owner =       THIS_MODULE,
3882 };
3883
3884 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3885 static struct blkio_policy_type blkio_policy_cfq = {
3886         .ops = {
3887                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3888                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3889         },
3890 };
3891 #else
3892 static struct blkio_policy_type blkio_policy_cfq;
3893 #endif
3894
3895 static int __init cfq_init(void)
3896 {
3897         /*
3898          * could be 0 on HZ < 1000 setups
3899          */
3900         if (!cfq_slice_async)
3901                 cfq_slice_async = 1;
3902         if (!cfq_slice_idle)
3903                 cfq_slice_idle = 1;
3904
3905         if (cfq_slab_setup())
3906                 return -ENOMEM;
3907
3908         elv_register(&iosched_cfq);
3909         blkio_policy_register(&blkio_policy_cfq);
3910
3911         return 0;
3912 }
3913
3914 static void __exit cfq_exit(void)
3915 {
3916         DECLARE_COMPLETION_ONSTACK(all_gone);
3917         blkio_policy_unregister(&blkio_policy_cfq);
3918         elv_unregister(&iosched_cfq);
3919         ioc_gone = &all_gone;
3920         /* ioc_gone's update must be visible before reading ioc_count */
3921         smp_wmb();
3922
3923         /*
3924          * this also protects us from entering cfq_slab_kill() with
3925          * pending RCU callbacks
3926          */
3927         if (elv_ioc_count_read(cfq_ioc_count))
3928                 wait_for_completion(&all_gone);
3929         cfq_slab_kill();
3930 }
3931
3932 module_init(cfq_init);
3933 module_exit(cfq_exit);
3934
3935 MODULE_AUTHOR("Jens Axboe");
3936 MODULE_LICENSE("GPL");
3937 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");