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