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