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