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