blk-cgroup: Fix RCU correctness warning in cfq_init_queue()
[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/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk-cgroup.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
35
36 /*
37  * offset from end of service tree
38  */
39 #define CFQ_IDLE_DELAY          (HZ / 5)
40
41 /*
42  * below this threshold, we consider thinktime immediate
43  */
44 #define CFQ_MIN_TT              (2)
45
46 #define CFQ_SLICE_SCALE         (5)
47 #define CFQ_HW_QUEUE_MIN        (5)
48 #define CFQ_SERVICE_SHIFT       12
49
50 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
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         unsigned total_weight;
84         u64 min_vdisktime;
85         struct rb_node *active;
86 };
87 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
88                         .count = 0, .min_vdisktime = 0, }
89
90 /*
91  * Per process-grouping structure
92  */
93 struct cfq_queue {
94         /* reference count */
95         atomic_t ref;
96         /* various state flags, see below */
97         unsigned int flags;
98         /* parent cfq_data */
99         struct cfq_data *cfqd;
100         /* service_tree member */
101         struct rb_node rb_node;
102         /* service_tree key */
103         unsigned long rb_key;
104         /* prio tree member */
105         struct rb_node p_node;
106         /* prio tree root we belong to, if any */
107         struct rb_root *p_root;
108         /* sorted list of pending requests */
109         struct rb_root sort_list;
110         /* if fifo isn't expired, next request to serve */
111         struct request *next_rq;
112         /* requests queued in sort_list */
113         int queued[2];
114         /* currently allocated requests */
115         int allocated[2];
116         /* fifo list of requests in sort_list */
117         struct list_head fifo;
118
119         /* time when queue got scheduled in to dispatch first request. */
120         unsigned long dispatch_start;
121         unsigned int allocated_slice;
122         unsigned int slice_dispatch;
123         /* time when first request from queue completed and slice started. */
124         unsigned long slice_start;
125         unsigned long slice_end;
126         long slice_resid;
127
128         /* pending metadata requests */
129         int meta_pending;
130         /* number of requests that are on the dispatch list or inside driver */
131         int dispatched;
132
133         /* io prio of this group */
134         unsigned short ioprio, org_ioprio;
135         unsigned short ioprio_class, org_ioprio_class;
136
137         pid_t pid;
138
139         u32 seek_history;
140         sector_t last_request_pos;
141
142         struct cfq_rb_root *service_tree;
143         struct cfq_queue *new_cfqq;
144         struct cfq_group *cfqg;
145         struct cfq_group *orig_cfqg;
146         /* Sectors dispatched in current dispatch round */
147         unsigned long nr_sectors;
148 };
149
150 /*
151  * First index in the service_trees.
152  * IDLE is handled separately, so it has negative index
153  */
154 enum wl_prio_t {
155         BE_WORKLOAD = 0,
156         RT_WORKLOAD = 1,
157         IDLE_WORKLOAD = 2,
158 };
159
160 /*
161  * Second index in the service_trees.
162  */
163 enum wl_type_t {
164         ASYNC_WORKLOAD = 0,
165         SYNC_NOIDLE_WORKLOAD = 1,
166         SYNC_WORKLOAD = 2
167 };
168
169 /* This is per cgroup per device grouping structure */
170 struct cfq_group {
171         /* group service_tree member */
172         struct rb_node rb_node;
173
174         /* group service_tree key */
175         u64 vdisktime;
176         unsigned int weight;
177         bool on_st;
178
179         /* number of cfqq currently on this group */
180         int nr_cfqq;
181
182         /* Per group busy queus average. Useful for workload slice calc. */
183         unsigned int busy_queues_avg[2];
184         /*
185          * rr lists of queues with requests, onle rr for each priority class.
186          * Counts are embedded in the cfq_rb_root
187          */
188         struct cfq_rb_root service_trees[2][3];
189         struct cfq_rb_root service_tree_idle;
190
191         unsigned long saved_workload_slice;
192         enum wl_type_t saved_workload;
193         enum wl_prio_t saved_serving_prio;
194         struct blkio_group blkg;
195 #ifdef CONFIG_CFQ_GROUP_IOSCHED
196         struct hlist_node cfqd_node;
197         atomic_t ref;
198 #endif
199 };
200
201 /*
202  * Per block device queue structure
203  */
204 struct cfq_data {
205         struct request_queue *queue;
206         /* Root service tree for cfq_groups */
207         struct cfq_rb_root grp_service_tree;
208         struct cfq_group root_group;
209
210         /*
211          * The priority currently being served
212          */
213         enum wl_prio_t serving_prio;
214         enum wl_type_t serving_type;
215         unsigned long workload_expires;
216         struct cfq_group *serving_group;
217         bool noidle_tree_requires_idle;
218
219         /*
220          * Each priority tree is sorted by next_request position.  These
221          * trees are used when determining if two or more queues are
222          * interleaving requests (see cfq_close_cooperator).
223          */
224         struct rb_root prio_trees[CFQ_PRIO_LISTS];
225
226         unsigned int busy_queues;
227
228         int rq_in_driver;
229         int rq_in_flight[2];
230
231         /*
232          * queue-depth detection
233          */
234         int rq_queued;
235         int hw_tag;
236         /*
237          * hw_tag can be
238          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
239          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
240          *  0 => no NCQ
241          */
242         int hw_tag_est_depth;
243         unsigned int hw_tag_samples;
244
245         /*
246          * idle window management
247          */
248         struct timer_list idle_slice_timer;
249         struct work_struct unplug_work;
250
251         struct cfq_queue *active_queue;
252         struct cfq_io_context *active_cic;
253
254         /*
255          * async queue for each priority case
256          */
257         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
258         struct cfq_queue *async_idle_cfqq;
259
260         sector_t last_position;
261
262         /*
263          * tunables, see top of file
264          */
265         unsigned int cfq_quantum;
266         unsigned int cfq_fifo_expire[2];
267         unsigned int cfq_back_penalty;
268         unsigned int cfq_back_max;
269         unsigned int cfq_slice[2];
270         unsigned int cfq_slice_async_rq;
271         unsigned int cfq_slice_idle;
272         unsigned int cfq_latency;
273         unsigned int cfq_group_isolation;
274
275         struct list_head cic_list;
276
277         /*
278          * Fallback dummy cfqq for extreme OOM conditions
279          */
280         struct cfq_queue oom_cfqq;
281
282         unsigned long last_delayed_sync;
283
284         /* List of cfq groups being managed on this device*/
285         struct hlist_head cfqg_list;
286         struct rcu_head rcu;
287 };
288
289 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
290
291 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
292                                             enum wl_prio_t prio,
293                                             enum wl_type_t type)
294 {
295         if (!cfqg)
296                 return NULL;
297
298         if (prio == IDLE_WORKLOAD)
299                 return &cfqg->service_tree_idle;
300
301         return &cfqg->service_trees[prio][type];
302 }
303
304 enum cfqq_state_flags {
305         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
306         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
307         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
308         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
309         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
310         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
311         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
312         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
313         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
314         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
315         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
316         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
317         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
318 };
319
320 #define CFQ_CFQQ_FNS(name)                                              \
321 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
322 {                                                                       \
323         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
324 }                                                                       \
325 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
326 {                                                                       \
327         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
328 }                                                                       \
329 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
330 {                                                                       \
331         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
332 }
333
334 CFQ_CFQQ_FNS(on_rr);
335 CFQ_CFQQ_FNS(wait_request);
336 CFQ_CFQQ_FNS(must_dispatch);
337 CFQ_CFQQ_FNS(must_alloc_slice);
338 CFQ_CFQQ_FNS(fifo_expire);
339 CFQ_CFQQ_FNS(idle_window);
340 CFQ_CFQQ_FNS(prio_changed);
341 CFQ_CFQQ_FNS(slice_new);
342 CFQ_CFQQ_FNS(sync);
343 CFQ_CFQQ_FNS(coop);
344 CFQ_CFQQ_FNS(split_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 struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
423                                             bool is_sync)
424 {
425         return cic->cfqq[is_sync];
426 }
427
428 static inline void cic_set_cfqq(struct cfq_io_context *cic,
429                                 struct cfq_queue *cfqq, bool is_sync)
430 {
431         cic->cfqq[is_sync] = cfqq;
432 }
433
434 /*
435  * We regard a request as SYNC, if it's either a read or has the SYNC bit
436  * set (in which case it could also be direct WRITE).
437  */
438 static inline bool cfq_bio_sync(struct bio *bio)
439 {
440         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
441 }
442
443 /*
444  * scheduler run of queue, if there are requests pending and no one in the
445  * driver that will restart queueing
446  */
447 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
448 {
449         if (cfqd->busy_queues) {
450                 cfq_log(cfqd, "schedule dispatch");
451                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
452         }
453 }
454
455 static int cfq_queue_empty(struct request_queue *q)
456 {
457         struct cfq_data *cfqd = q->elevator->elevator_data;
458
459         return !cfqd->rq_queued;
460 }
461
462 /*
463  * Scale schedule slice based on io priority. Use the sync time slice only
464  * if a queue is marked sync and has sync io queued. A sync queue with async
465  * io only, should not get full sync slice length.
466  */
467 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
468                                  unsigned short prio)
469 {
470         const int base_slice = cfqd->cfq_slice[sync];
471
472         WARN_ON(prio >= IOPRIO_BE_NR);
473
474         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
475 }
476
477 static inline int
478 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
479 {
480         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
481 }
482
483 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
484 {
485         u64 d = delta << CFQ_SERVICE_SHIFT;
486
487         d = d * BLKIO_WEIGHT_DEFAULT;
488         do_div(d, cfqg->weight);
489         return d;
490 }
491
492 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
493 {
494         s64 delta = (s64)(vdisktime - min_vdisktime);
495         if (delta > 0)
496                 min_vdisktime = vdisktime;
497
498         return min_vdisktime;
499 }
500
501 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
502 {
503         s64 delta = (s64)(vdisktime - min_vdisktime);
504         if (delta < 0)
505                 min_vdisktime = vdisktime;
506
507         return min_vdisktime;
508 }
509
510 static void update_min_vdisktime(struct cfq_rb_root *st)
511 {
512         u64 vdisktime = st->min_vdisktime;
513         struct cfq_group *cfqg;
514
515         if (st->active) {
516                 cfqg = rb_entry_cfqg(st->active);
517                 vdisktime = cfqg->vdisktime;
518         }
519
520         if (st->left) {
521                 cfqg = rb_entry_cfqg(st->left);
522                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
523         }
524
525         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
526 }
527
528 /*
529  * get averaged number of queues of RT/BE priority.
530  * average is updated, with a formula that gives more weight to higher numbers,
531  * to quickly follows sudden increases and decrease slowly
532  */
533
534 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
535                                         struct cfq_group *cfqg, bool rt)
536 {
537         unsigned min_q, max_q;
538         unsigned mult  = cfq_hist_divisor - 1;
539         unsigned round = cfq_hist_divisor / 2;
540         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
541
542         min_q = min(cfqg->busy_queues_avg[rt], busy);
543         max_q = max(cfqg->busy_queues_avg[rt], busy);
544         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
545                 cfq_hist_divisor;
546         return cfqg->busy_queues_avg[rt];
547 }
548
549 static inline unsigned
550 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
551 {
552         struct cfq_rb_root *st = &cfqd->grp_service_tree;
553
554         return cfq_target_latency * cfqg->weight / st->total_weight;
555 }
556
557 static inline void
558 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
559 {
560         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
561         if (cfqd->cfq_latency) {
562                 /*
563                  * interested queues (we consider only the ones with the same
564                  * priority class in the cfq group)
565                  */
566                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
567                                                 cfq_class_rt(cfqq));
568                 unsigned sync_slice = cfqd->cfq_slice[1];
569                 unsigned expect_latency = sync_slice * iq;
570                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
571
572                 if (expect_latency > group_slice) {
573                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
574                         /* scale low_slice according to IO priority
575                          * and sync vs async */
576                         unsigned low_slice =
577                                 min(slice, base_low_slice * slice / sync_slice);
578                         /* the adapted slice value is scaled to fit all iqs
579                          * into the target latency */
580                         slice = max(slice * group_slice / expect_latency,
581                                     low_slice);
582                 }
583         }
584         cfqq->slice_start = jiffies;
585         cfqq->slice_end = jiffies + slice;
586         cfqq->allocated_slice = slice;
587         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
588 }
589
590 /*
591  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
592  * isn't valid until the first request from the dispatch is activated
593  * and the slice time set.
594  */
595 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
596 {
597         if (cfq_cfqq_slice_new(cfqq))
598                 return 0;
599         if (time_before(jiffies, cfqq->slice_end))
600                 return 0;
601
602         return 1;
603 }
604
605 /*
606  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
607  * We choose the request that is closest to the head right now. Distance
608  * behind the head is penalized and only allowed to a certain extent.
609  */
610 static struct request *
611 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
612 {
613         sector_t s1, s2, d1 = 0, d2 = 0;
614         unsigned long back_max;
615 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
616 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
617         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
618
619         if (rq1 == NULL || rq1 == rq2)
620                 return rq2;
621         if (rq2 == NULL)
622                 return rq1;
623
624         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
625                 return rq1;
626         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
627                 return rq2;
628         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
629                 return rq1;
630         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
631                 return rq2;
632
633         s1 = blk_rq_pos(rq1);
634         s2 = blk_rq_pos(rq2);
635
636         /*
637          * by definition, 1KiB is 2 sectors
638          */
639         back_max = cfqd->cfq_back_max * 2;
640
641         /*
642          * Strict one way elevator _except_ in the case where we allow
643          * short backward seeks which are biased as twice the cost of a
644          * similar forward seek.
645          */
646         if (s1 >= last)
647                 d1 = s1 - last;
648         else if (s1 + back_max >= last)
649                 d1 = (last - s1) * cfqd->cfq_back_penalty;
650         else
651                 wrap |= CFQ_RQ1_WRAP;
652
653         if (s2 >= last)
654                 d2 = s2 - last;
655         else if (s2 + back_max >= last)
656                 d2 = (last - s2) * cfqd->cfq_back_penalty;
657         else
658                 wrap |= CFQ_RQ2_WRAP;
659
660         /* Found required data */
661
662         /*
663          * By doing switch() on the bit mask "wrap" we avoid having to
664          * check two variables for all permutations: --> faster!
665          */
666         switch (wrap) {
667         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
668                 if (d1 < d2)
669                         return rq1;
670                 else if (d2 < d1)
671                         return rq2;
672                 else {
673                         if (s1 >= s2)
674                                 return rq1;
675                         else
676                                 return rq2;
677                 }
678
679         case CFQ_RQ2_WRAP:
680                 return rq1;
681         case CFQ_RQ1_WRAP:
682                 return rq2;
683         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
684         default:
685                 /*
686                  * Since both rqs are wrapped,
687                  * start with the one that's further behind head
688                  * (--> only *one* back seek required),
689                  * since back seek takes more time than forward.
690                  */
691                 if (s1 <= s2)
692                         return rq1;
693                 else
694                         return rq2;
695         }
696 }
697
698 /*
699  * The below is leftmost cache rbtree addon
700  */
701 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
702 {
703         /* Service tree is empty */
704         if (!root->count)
705                 return NULL;
706
707         if (!root->left)
708                 root->left = rb_first(&root->rb);
709
710         if (root->left)
711                 return rb_entry(root->left, struct cfq_queue, rb_node);
712
713         return NULL;
714 }
715
716 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
717 {
718         if (!root->left)
719                 root->left = rb_first(&root->rb);
720
721         if (root->left)
722                 return rb_entry_cfqg(root->left);
723
724         return NULL;
725 }
726
727 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
728 {
729         rb_erase(n, root);
730         RB_CLEAR_NODE(n);
731 }
732
733 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
734 {
735         if (root->left == n)
736                 root->left = NULL;
737         rb_erase_init(n, &root->rb);
738         --root->count;
739 }
740
741 /*
742  * would be nice to take fifo expire time into account as well
743  */
744 static struct request *
745 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
746                   struct request *last)
747 {
748         struct rb_node *rbnext = rb_next(&last->rb_node);
749         struct rb_node *rbprev = rb_prev(&last->rb_node);
750         struct request *next = NULL, *prev = NULL;
751
752         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
753
754         if (rbprev)
755                 prev = rb_entry_rq(rbprev);
756
757         if (rbnext)
758                 next = rb_entry_rq(rbnext);
759         else {
760                 rbnext = rb_first(&cfqq->sort_list);
761                 if (rbnext && rbnext != &last->rb_node)
762                         next = rb_entry_rq(rbnext);
763         }
764
765         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
766 }
767
768 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
769                                       struct cfq_queue *cfqq)
770 {
771         /*
772          * just an approximation, should be ok.
773          */
774         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
775                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
776 }
777
778 static inline s64
779 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
780 {
781         return cfqg->vdisktime - st->min_vdisktime;
782 }
783
784 static void
785 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
786 {
787         struct rb_node **node = &st->rb.rb_node;
788         struct rb_node *parent = NULL;
789         struct cfq_group *__cfqg;
790         s64 key = cfqg_key(st, cfqg);
791         int left = 1;
792
793         while (*node != NULL) {
794                 parent = *node;
795                 __cfqg = rb_entry_cfqg(parent);
796
797                 if (key < cfqg_key(st, __cfqg))
798                         node = &parent->rb_left;
799                 else {
800                         node = &parent->rb_right;
801                         left = 0;
802                 }
803         }
804
805         if (left)
806                 st->left = &cfqg->rb_node;
807
808         rb_link_node(&cfqg->rb_node, parent, node);
809         rb_insert_color(&cfqg->rb_node, &st->rb);
810 }
811
812 static void
813 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
814 {
815         struct cfq_rb_root *st = &cfqd->grp_service_tree;
816         struct cfq_group *__cfqg;
817         struct rb_node *n;
818
819         cfqg->nr_cfqq++;
820         if (cfqg->on_st)
821                 return;
822
823         /*
824          * Currently put the group at the end. Later implement something
825          * so that groups get lesser vtime based on their weights, so that
826          * if group does not loose all if it was not continously backlogged.
827          */
828         n = rb_last(&st->rb);
829         if (n) {
830                 __cfqg = rb_entry_cfqg(n);
831                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
832         } else
833                 cfqg->vdisktime = st->min_vdisktime;
834
835         __cfq_group_service_tree_add(st, cfqg);
836         cfqg->on_st = true;
837         st->total_weight += cfqg->weight;
838 }
839
840 static void
841 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
842 {
843         struct cfq_rb_root *st = &cfqd->grp_service_tree;
844
845         if (st->active == &cfqg->rb_node)
846                 st->active = NULL;
847
848         BUG_ON(cfqg->nr_cfqq < 1);
849         cfqg->nr_cfqq--;
850
851         /* If there are other cfq queues under this group, don't delete it */
852         if (cfqg->nr_cfqq)
853                 return;
854
855         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
856         cfqg->on_st = false;
857         st->total_weight -= cfqg->weight;
858         if (!RB_EMPTY_NODE(&cfqg->rb_node))
859                 cfq_rb_erase(&cfqg->rb_node, st);
860         cfqg->saved_workload_slice = 0;
861         blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
862 }
863
864 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
865 {
866         unsigned int slice_used;
867
868         /*
869          * Queue got expired before even a single request completed or
870          * got expired immediately after first request completion.
871          */
872         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
873                 /*
874                  * Also charge the seek time incurred to the group, otherwise
875                  * if there are mutiple queues in the group, each can dispatch
876                  * a single request on seeky media and cause lots of seek time
877                  * and group will never know it.
878                  */
879                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
880                                         1);
881         } else {
882                 slice_used = jiffies - cfqq->slice_start;
883                 if (slice_used > cfqq->allocated_slice)
884                         slice_used = cfqq->allocated_slice;
885         }
886
887         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
888                                 cfqq->nr_sectors);
889         return slice_used;
890 }
891
892 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
893                                 struct cfq_queue *cfqq)
894 {
895         struct cfq_rb_root *st = &cfqd->grp_service_tree;
896         unsigned int used_sl, charge_sl;
897         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
898                         - cfqg->service_tree_idle.count;
899
900         BUG_ON(nr_sync < 0);
901         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
902
903         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
904                 charge_sl = cfqq->allocated_slice;
905
906         /* Can't update vdisktime while group is on service tree */
907         cfq_rb_erase(&cfqg->rb_node, st);
908         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
909         __cfq_group_service_tree_add(st, cfqg);
910
911         /* This group is being expired. Save the context */
912         if (time_after(cfqd->workload_expires, jiffies)) {
913                 cfqg->saved_workload_slice = cfqd->workload_expires
914                                                 - jiffies;
915                 cfqg->saved_workload = cfqd->serving_type;
916                 cfqg->saved_serving_prio = cfqd->serving_prio;
917         } else
918                 cfqg->saved_workload_slice = 0;
919
920         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
921                                         st->min_vdisktime);
922         blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
923                                                 cfqq->nr_sectors);
924 }
925
926 #ifdef CONFIG_CFQ_GROUP_IOSCHED
927 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
928 {
929         if (blkg)
930                 return container_of(blkg, struct cfq_group, blkg);
931         return NULL;
932 }
933
934 void
935 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
936 {
937         cfqg_of_blkg(blkg)->weight = weight;
938 }
939
940 static struct cfq_group *
941 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
942 {
943         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
944         struct cfq_group *cfqg = NULL;
945         void *key = cfqd;
946         int i, j;
947         struct cfq_rb_root *st;
948         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
949         unsigned int major, minor;
950
951         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
952         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
953                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
954                 cfqg->blkg.dev = MKDEV(major, minor);
955                 goto done;
956         }
957         if (cfqg || !create)
958                 goto done;
959
960         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
961         if (!cfqg)
962                 goto done;
963
964         cfqg->weight = blkcg->weight;
965         for_each_cfqg_st(cfqg, i, j, st)
966                 *st = CFQ_RB_ROOT;
967         RB_CLEAR_NODE(&cfqg->rb_node);
968
969         /*
970          * Take the initial reference that will be released on destroy
971          * This can be thought of a joint reference by cgroup and
972          * elevator which will be dropped by either elevator exit
973          * or cgroup deletion path depending on who is exiting first.
974          */
975         atomic_set(&cfqg->ref, 1);
976
977         /* Add group onto cgroup list */
978         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
979         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
980                                         MKDEV(major, minor));
981
982         /* Add group on cfqd list */
983         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
984
985 done:
986         return cfqg;
987 }
988
989 /*
990  * Search for the cfq group current task belongs to. If create = 1, then also
991  * create the cfq group if it does not exist. request_queue lock must be held.
992  */
993 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
994 {
995         struct cgroup *cgroup;
996         struct cfq_group *cfqg = NULL;
997
998         rcu_read_lock();
999         cgroup = task_cgroup(current, blkio_subsys_id);
1000         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1001         if (!cfqg && create)
1002                 cfqg = &cfqd->root_group;
1003         rcu_read_unlock();
1004         return cfqg;
1005 }
1006
1007 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1008 {
1009         /* Currently, all async queues are mapped to root group */
1010         if (!cfq_cfqq_sync(cfqq))
1011                 cfqg = &cfqq->cfqd->root_group;
1012
1013         cfqq->cfqg = cfqg;
1014         /* cfqq reference on cfqg */
1015         atomic_inc(&cfqq->cfqg->ref);
1016 }
1017
1018 static void cfq_put_cfqg(struct cfq_group *cfqg)
1019 {
1020         struct cfq_rb_root *st;
1021         int i, j;
1022
1023         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1024         if (!atomic_dec_and_test(&cfqg->ref))
1025                 return;
1026         for_each_cfqg_st(cfqg, i, j, st)
1027                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1028         kfree(cfqg);
1029 }
1030
1031 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1032 {
1033         /* Something wrong if we are trying to remove same group twice */
1034         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1035
1036         hlist_del_init(&cfqg->cfqd_node);
1037
1038         /*
1039          * Put the reference taken at the time of creation so that when all
1040          * queues are gone, group can be destroyed.
1041          */
1042         cfq_put_cfqg(cfqg);
1043 }
1044
1045 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1046 {
1047         struct hlist_node *pos, *n;
1048         struct cfq_group *cfqg;
1049
1050         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1051                 /*
1052                  * If cgroup removal path got to blk_group first and removed
1053                  * it from cgroup list, then it will take care of destroying
1054                  * cfqg also.
1055                  */
1056                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1057                         cfq_destroy_cfqg(cfqd, cfqg);
1058         }
1059 }
1060
1061 /*
1062  * Blk cgroup controller notification saying that blkio_group object is being
1063  * delinked as associated cgroup object is going away. That also means that
1064  * no new IO will come in this group. So get rid of this group as soon as
1065  * any pending IO in the group is finished.
1066  *
1067  * This function is called under rcu_read_lock(). key is the rcu protected
1068  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1069  * read lock.
1070  *
1071  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1072  * it should not be NULL as even if elevator was exiting, cgroup deltion
1073  * path got to it first.
1074  */
1075 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1076 {
1077         unsigned long  flags;
1078         struct cfq_data *cfqd = key;
1079
1080         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1081         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1082         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1083 }
1084
1085 #else /* GROUP_IOSCHED */
1086 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1087 {
1088         return &cfqd->root_group;
1089 }
1090 static inline void
1091 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1092         cfqq->cfqg = cfqg;
1093 }
1094
1095 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1096 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1097
1098 #endif /* GROUP_IOSCHED */
1099
1100 /*
1101  * The cfqd->service_trees holds all pending cfq_queue's that have
1102  * requests waiting to be processed. It is sorted in the order that
1103  * we will service the queues.
1104  */
1105 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1106                                  bool add_front)
1107 {
1108         struct rb_node **p, *parent;
1109         struct cfq_queue *__cfqq;
1110         unsigned long rb_key;
1111         struct cfq_rb_root *service_tree;
1112         int left;
1113         int new_cfqq = 1;
1114         int group_changed = 0;
1115
1116 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1117         if (!cfqd->cfq_group_isolation
1118             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1119             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1120                 /* Move this cfq to root group */
1121                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1122                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1123                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1124                 cfqq->orig_cfqg = cfqq->cfqg;
1125                 cfqq->cfqg = &cfqd->root_group;
1126                 atomic_inc(&cfqd->root_group.ref);
1127                 group_changed = 1;
1128         } else if (!cfqd->cfq_group_isolation
1129                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1130                 /* cfqq is sequential now needs to go to its original group */
1131                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1132                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134                 cfq_put_cfqg(cfqq->cfqg);
1135                 cfqq->cfqg = cfqq->orig_cfqg;
1136                 cfqq->orig_cfqg = NULL;
1137                 group_changed = 1;
1138                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1139         }
1140 #endif
1141
1142         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1143                                                 cfqq_type(cfqq));
1144         if (cfq_class_idle(cfqq)) {
1145                 rb_key = CFQ_IDLE_DELAY;
1146                 parent = rb_last(&service_tree->rb);
1147                 if (parent && parent != &cfqq->rb_node) {
1148                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1149                         rb_key += __cfqq->rb_key;
1150                 } else
1151                         rb_key += jiffies;
1152         } else if (!add_front) {
1153                 /*
1154                  * Get our rb key offset. Subtract any residual slice
1155                  * value carried from last service. A negative resid
1156                  * count indicates slice overrun, and this should position
1157                  * the next service time further away in the tree.
1158                  */
1159                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1160                 rb_key -= cfqq->slice_resid;
1161                 cfqq->slice_resid = 0;
1162         } else {
1163                 rb_key = -HZ;
1164                 __cfqq = cfq_rb_first(service_tree);
1165                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1166         }
1167
1168         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1169                 new_cfqq = 0;
1170                 /*
1171                  * same position, nothing more to do
1172                  */
1173                 if (rb_key == cfqq->rb_key &&
1174                     cfqq->service_tree == service_tree)
1175                         return;
1176
1177                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1178                 cfqq->service_tree = NULL;
1179         }
1180
1181         left = 1;
1182         parent = NULL;
1183         cfqq->service_tree = service_tree;
1184         p = &service_tree->rb.rb_node;
1185         while (*p) {
1186                 struct rb_node **n;
1187
1188                 parent = *p;
1189                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1190
1191                 /*
1192                  * sort by key, that represents service time.
1193                  */
1194                 if (time_before(rb_key, __cfqq->rb_key))
1195                         n = &(*p)->rb_left;
1196                 else {
1197                         n = &(*p)->rb_right;
1198                         left = 0;
1199                 }
1200
1201                 p = n;
1202         }
1203
1204         if (left)
1205                 service_tree->left = &cfqq->rb_node;
1206
1207         cfqq->rb_key = rb_key;
1208         rb_link_node(&cfqq->rb_node, parent, p);
1209         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1210         service_tree->count++;
1211         if ((add_front || !new_cfqq) && !group_changed)
1212                 return;
1213         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1214 }
1215
1216 static struct cfq_queue *
1217 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1218                      sector_t sector, struct rb_node **ret_parent,
1219                      struct rb_node ***rb_link)
1220 {
1221         struct rb_node **p, *parent;
1222         struct cfq_queue *cfqq = NULL;
1223
1224         parent = NULL;
1225         p = &root->rb_node;
1226         while (*p) {
1227                 struct rb_node **n;
1228
1229                 parent = *p;
1230                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1231
1232                 /*
1233                  * Sort strictly based on sector.  Smallest to the left,
1234                  * largest to the right.
1235                  */
1236                 if (sector > blk_rq_pos(cfqq->next_rq))
1237                         n = &(*p)->rb_right;
1238                 else if (sector < blk_rq_pos(cfqq->next_rq))
1239                         n = &(*p)->rb_left;
1240                 else
1241                         break;
1242                 p = n;
1243                 cfqq = NULL;
1244         }
1245
1246         *ret_parent = parent;
1247         if (rb_link)
1248                 *rb_link = p;
1249         return cfqq;
1250 }
1251
1252 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1253 {
1254         struct rb_node **p, *parent;
1255         struct cfq_queue *__cfqq;
1256
1257         if (cfqq->p_root) {
1258                 rb_erase(&cfqq->p_node, cfqq->p_root);
1259                 cfqq->p_root = NULL;
1260         }
1261
1262         if (cfq_class_idle(cfqq))
1263                 return;
1264         if (!cfqq->next_rq)
1265                 return;
1266
1267         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1268         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1269                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1270         if (!__cfqq) {
1271                 rb_link_node(&cfqq->p_node, parent, p);
1272                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1273         } else
1274                 cfqq->p_root = NULL;
1275 }
1276
1277 /*
1278  * Update cfqq's position in the service tree.
1279  */
1280 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1281 {
1282         /*
1283          * Resorting requires the cfqq to be on the RR list already.
1284          */
1285         if (cfq_cfqq_on_rr(cfqq)) {
1286                 cfq_service_tree_add(cfqd, cfqq, 0);
1287                 cfq_prio_tree_add(cfqd, cfqq);
1288         }
1289 }
1290
1291 /*
1292  * add to busy list of queues for service, trying to be fair in ordering
1293  * the pending list according to last request service
1294  */
1295 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1296 {
1297         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1298         BUG_ON(cfq_cfqq_on_rr(cfqq));
1299         cfq_mark_cfqq_on_rr(cfqq);
1300         cfqd->busy_queues++;
1301
1302         cfq_resort_rr_list(cfqd, cfqq);
1303 }
1304
1305 /*
1306  * Called when the cfqq no longer has requests pending, remove it from
1307  * the service tree.
1308  */
1309 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1310 {
1311         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1312         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1313         cfq_clear_cfqq_on_rr(cfqq);
1314
1315         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1316                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1317                 cfqq->service_tree = NULL;
1318         }
1319         if (cfqq->p_root) {
1320                 rb_erase(&cfqq->p_node, cfqq->p_root);
1321                 cfqq->p_root = NULL;
1322         }
1323
1324         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1325         BUG_ON(!cfqd->busy_queues);
1326         cfqd->busy_queues--;
1327 }
1328
1329 /*
1330  * rb tree support functions
1331  */
1332 static void cfq_del_rq_rb(struct request *rq)
1333 {
1334         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1335         const int sync = rq_is_sync(rq);
1336
1337         BUG_ON(!cfqq->queued[sync]);
1338         cfqq->queued[sync]--;
1339
1340         elv_rb_del(&cfqq->sort_list, rq);
1341
1342         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1343                 /*
1344                  * Queue will be deleted from service tree when we actually
1345                  * expire it later. Right now just remove it from prio tree
1346                  * as it is empty.
1347                  */
1348                 if (cfqq->p_root) {
1349                         rb_erase(&cfqq->p_node, cfqq->p_root);
1350                         cfqq->p_root = NULL;
1351                 }
1352         }
1353 }
1354
1355 static void cfq_add_rq_rb(struct request *rq)
1356 {
1357         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1358         struct cfq_data *cfqd = cfqq->cfqd;
1359         struct request *__alias, *prev;
1360
1361         cfqq->queued[rq_is_sync(rq)]++;
1362
1363         /*
1364          * looks a little odd, but the first insert might return an alias.
1365          * if that happens, put the alias on the dispatch list
1366          */
1367         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1368                 cfq_dispatch_insert(cfqd->queue, __alias);
1369
1370         if (!cfq_cfqq_on_rr(cfqq))
1371                 cfq_add_cfqq_rr(cfqd, cfqq);
1372
1373         /*
1374          * check if this request is a better next-serve candidate
1375          */
1376         prev = cfqq->next_rq;
1377         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1378
1379         /*
1380          * adjust priority tree position, if ->next_rq changes
1381          */
1382         if (prev != cfqq->next_rq)
1383                 cfq_prio_tree_add(cfqd, cfqq);
1384
1385         BUG_ON(!cfqq->next_rq);
1386 }
1387
1388 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1389 {
1390         elv_rb_del(&cfqq->sort_list, rq);
1391         cfqq->queued[rq_is_sync(rq)]--;
1392         cfq_add_rq_rb(rq);
1393 }
1394
1395 static struct request *
1396 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1397 {
1398         struct task_struct *tsk = current;
1399         struct cfq_io_context *cic;
1400         struct cfq_queue *cfqq;
1401
1402         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1403         if (!cic)
1404                 return NULL;
1405
1406         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1407         if (cfqq) {
1408                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1409
1410                 return elv_rb_find(&cfqq->sort_list, sector);
1411         }
1412
1413         return NULL;
1414 }
1415
1416 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1417 {
1418         struct cfq_data *cfqd = q->elevator->elevator_data;
1419
1420         cfqd->rq_in_driver++;
1421         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1422                                                 cfqd->rq_in_driver);
1423
1424         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1425 }
1426
1427 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1428 {
1429         struct cfq_data *cfqd = q->elevator->elevator_data;
1430
1431         WARN_ON(!cfqd->rq_in_driver);
1432         cfqd->rq_in_driver--;
1433         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1434                                                 cfqd->rq_in_driver);
1435 }
1436
1437 static void cfq_remove_request(struct request *rq)
1438 {
1439         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1440
1441         if (cfqq->next_rq == rq)
1442                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1443
1444         list_del_init(&rq->queuelist);
1445         cfq_del_rq_rb(rq);
1446
1447         cfqq->cfqd->rq_queued--;
1448         if (rq_is_meta(rq)) {
1449                 WARN_ON(!cfqq->meta_pending);
1450                 cfqq->meta_pending--;
1451         }
1452 }
1453
1454 static int cfq_merge(struct request_queue *q, struct request **req,
1455                      struct bio *bio)
1456 {
1457         struct cfq_data *cfqd = q->elevator->elevator_data;
1458         struct request *__rq;
1459
1460         __rq = cfq_find_rq_fmerge(cfqd, bio);
1461         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1462                 *req = __rq;
1463                 return ELEVATOR_FRONT_MERGE;
1464         }
1465
1466         return ELEVATOR_NO_MERGE;
1467 }
1468
1469 static void cfq_merged_request(struct request_queue *q, struct request *req,
1470                                int type)
1471 {
1472         if (type == ELEVATOR_FRONT_MERGE) {
1473                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1474
1475                 cfq_reposition_rq_rb(cfqq, req);
1476         }
1477 }
1478
1479 static void
1480 cfq_merged_requests(struct request_queue *q, struct request *rq,
1481                     struct request *next)
1482 {
1483         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1484         /*
1485          * reposition in fifo if next is older than rq
1486          */
1487         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1488             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1489                 list_move(&rq->queuelist, &next->queuelist);
1490                 rq_set_fifo_time(rq, rq_fifo_time(next));
1491         }
1492
1493         if (cfqq->next_rq == next)
1494                 cfqq->next_rq = rq;
1495         cfq_remove_request(next);
1496 }
1497
1498 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1499                            struct bio *bio)
1500 {
1501         struct cfq_data *cfqd = q->elevator->elevator_data;
1502         struct cfq_io_context *cic;
1503         struct cfq_queue *cfqq;
1504
1505         /*
1506          * Disallow merge of a sync bio into an async request.
1507          */
1508         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1509                 return false;
1510
1511         /*
1512          * Lookup the cfqq that this bio will be queued with. Allow
1513          * merge only if rq is queued there.
1514          */
1515         cic = cfq_cic_lookup(cfqd, current->io_context);
1516         if (!cic)
1517                 return false;
1518
1519         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1520         return cfqq == RQ_CFQQ(rq);
1521 }
1522
1523 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1524                                    struct cfq_queue *cfqq)
1525 {
1526         if (cfqq) {
1527                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1528                                 cfqd->serving_prio, cfqd->serving_type);
1529                 cfqq->slice_start = 0;
1530                 cfqq->dispatch_start = jiffies;
1531                 cfqq->allocated_slice = 0;
1532                 cfqq->slice_end = 0;
1533                 cfqq->slice_dispatch = 0;
1534                 cfqq->nr_sectors = 0;
1535
1536                 cfq_clear_cfqq_wait_request(cfqq);
1537                 cfq_clear_cfqq_must_dispatch(cfqq);
1538                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1539                 cfq_clear_cfqq_fifo_expire(cfqq);
1540                 cfq_mark_cfqq_slice_new(cfqq);
1541
1542                 del_timer(&cfqd->idle_slice_timer);
1543         }
1544
1545         cfqd->active_queue = cfqq;
1546 }
1547
1548 /*
1549  * current cfqq expired its slice (or was too idle), select new one
1550  */
1551 static void
1552 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1553                     bool timed_out)
1554 {
1555         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1556
1557         if (cfq_cfqq_wait_request(cfqq))
1558                 del_timer(&cfqd->idle_slice_timer);
1559
1560         cfq_clear_cfqq_wait_request(cfqq);
1561         cfq_clear_cfqq_wait_busy(cfqq);
1562
1563         /*
1564          * If this cfqq is shared between multiple processes, check to
1565          * make sure that those processes are still issuing I/Os within
1566          * the mean seek distance.  If not, it may be time to break the
1567          * queues apart again.
1568          */
1569         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1570                 cfq_mark_cfqq_split_coop(cfqq);
1571
1572         /*
1573          * store what was left of this slice, if the queue idled/timed out
1574          */
1575         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1576                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1577                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1578         }
1579
1580         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1581
1582         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1583                 cfq_del_cfqq_rr(cfqd, cfqq);
1584
1585         cfq_resort_rr_list(cfqd, cfqq);
1586
1587         if (cfqq == cfqd->active_queue)
1588                 cfqd->active_queue = NULL;
1589
1590         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1591                 cfqd->grp_service_tree.active = NULL;
1592
1593         if (cfqd->active_cic) {
1594                 put_io_context(cfqd->active_cic->ioc);
1595                 cfqd->active_cic = NULL;
1596         }
1597 }
1598
1599 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1600 {
1601         struct cfq_queue *cfqq = cfqd->active_queue;
1602
1603         if (cfqq)
1604                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1605 }
1606
1607 /*
1608  * Get next queue for service. Unless we have a queue preemption,
1609  * we'll simply select the first cfqq in the service tree.
1610  */
1611 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1612 {
1613         struct cfq_rb_root *service_tree =
1614                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1615                                         cfqd->serving_type);
1616
1617         if (!cfqd->rq_queued)
1618                 return NULL;
1619
1620         /* There is nothing to dispatch */
1621         if (!service_tree)
1622                 return NULL;
1623         if (RB_EMPTY_ROOT(&service_tree->rb))
1624                 return NULL;
1625         return cfq_rb_first(service_tree);
1626 }
1627
1628 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1629 {
1630         struct cfq_group *cfqg;
1631         struct cfq_queue *cfqq;
1632         int i, j;
1633         struct cfq_rb_root *st;
1634
1635         if (!cfqd->rq_queued)
1636                 return NULL;
1637
1638         cfqg = cfq_get_next_cfqg(cfqd);
1639         if (!cfqg)
1640                 return NULL;
1641
1642         for_each_cfqg_st(cfqg, i, j, st)
1643                 if ((cfqq = cfq_rb_first(st)) != NULL)
1644                         return cfqq;
1645         return NULL;
1646 }
1647
1648 /*
1649  * Get and set a new active queue for service.
1650  */
1651 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1652                                               struct cfq_queue *cfqq)
1653 {
1654         if (!cfqq)
1655                 cfqq = cfq_get_next_queue(cfqd);
1656
1657         __cfq_set_active_queue(cfqd, cfqq);
1658         return cfqq;
1659 }
1660
1661 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1662                                           struct request *rq)
1663 {
1664         if (blk_rq_pos(rq) >= cfqd->last_position)
1665                 return blk_rq_pos(rq) - cfqd->last_position;
1666         else
1667                 return cfqd->last_position - blk_rq_pos(rq);
1668 }
1669
1670 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1671                                struct request *rq)
1672 {
1673         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1674 }
1675
1676 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1677                                     struct cfq_queue *cur_cfqq)
1678 {
1679         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1680         struct rb_node *parent, *node;
1681         struct cfq_queue *__cfqq;
1682         sector_t sector = cfqd->last_position;
1683
1684         if (RB_EMPTY_ROOT(root))
1685                 return NULL;
1686
1687         /*
1688          * First, if we find a request starting at the end of the last
1689          * request, choose it.
1690          */
1691         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1692         if (__cfqq)
1693                 return __cfqq;
1694
1695         /*
1696          * If the exact sector wasn't found, the parent of the NULL leaf
1697          * will contain the closest sector.
1698          */
1699         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1700         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1701                 return __cfqq;
1702
1703         if (blk_rq_pos(__cfqq->next_rq) < sector)
1704                 node = rb_next(&__cfqq->p_node);
1705         else
1706                 node = rb_prev(&__cfqq->p_node);
1707         if (!node)
1708                 return NULL;
1709
1710         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1711         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1712                 return __cfqq;
1713
1714         return NULL;
1715 }
1716
1717 /*
1718  * cfqd - obvious
1719  * cur_cfqq - passed in so that we don't decide that the current queue is
1720  *            closely cooperating with itself.
1721  *
1722  * So, basically we're assuming that that cur_cfqq has dispatched at least
1723  * one request, and that cfqd->last_position reflects a position on the disk
1724  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1725  * assumption.
1726  */
1727 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1728                                               struct cfq_queue *cur_cfqq)
1729 {
1730         struct cfq_queue *cfqq;
1731
1732         if (cfq_class_idle(cur_cfqq))
1733                 return NULL;
1734         if (!cfq_cfqq_sync(cur_cfqq))
1735                 return NULL;
1736         if (CFQQ_SEEKY(cur_cfqq))
1737                 return NULL;
1738
1739         /*
1740          * Don't search priority tree if it's the only queue in the group.
1741          */
1742         if (cur_cfqq->cfqg->nr_cfqq == 1)
1743                 return NULL;
1744
1745         /*
1746          * We should notice if some of the queues are cooperating, eg
1747          * working closely on the same area of the disk. In that case,
1748          * we can group them together and don't waste time idling.
1749          */
1750         cfqq = cfqq_close(cfqd, cur_cfqq);
1751         if (!cfqq)
1752                 return NULL;
1753
1754         /* If new queue belongs to different cfq_group, don't choose it */
1755         if (cur_cfqq->cfqg != cfqq->cfqg)
1756                 return NULL;
1757
1758         /*
1759          * It only makes sense to merge sync queues.
1760          */
1761         if (!cfq_cfqq_sync(cfqq))
1762                 return NULL;
1763         if (CFQQ_SEEKY(cfqq))
1764                 return NULL;
1765
1766         /*
1767          * Do not merge queues of different priority classes
1768          */
1769         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1770                 return NULL;
1771
1772         return cfqq;
1773 }
1774
1775 /*
1776  * Determine whether we should enforce idle window for this queue.
1777  */
1778
1779 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1780 {
1781         enum wl_prio_t prio = cfqq_prio(cfqq);
1782         struct cfq_rb_root *service_tree = cfqq->service_tree;
1783
1784         BUG_ON(!service_tree);
1785         BUG_ON(!service_tree->count);
1786
1787         /* We never do for idle class queues. */
1788         if (prio == IDLE_WORKLOAD)
1789                 return false;
1790
1791         /* We do for queues that were marked with idle window flag. */
1792         if (cfq_cfqq_idle_window(cfqq) &&
1793            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1794                 return true;
1795
1796         /*
1797          * Otherwise, we do only if they are the last ones
1798          * in their service tree.
1799          */
1800         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1801                 return 1;
1802         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1803                         service_tree->count);
1804         return 0;
1805 }
1806
1807 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1808 {
1809         struct cfq_queue *cfqq = cfqd->active_queue;
1810         struct cfq_io_context *cic;
1811         unsigned long sl;
1812
1813         /*
1814          * SSD device without seek penalty, disable idling. But only do so
1815          * for devices that support queuing, otherwise we still have a problem
1816          * with sync vs async workloads.
1817          */
1818         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1819                 return;
1820
1821         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1822         WARN_ON(cfq_cfqq_slice_new(cfqq));
1823
1824         /*
1825          * idle is disabled, either manually or by past process history
1826          */
1827         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1828                 return;
1829
1830         /*
1831          * still active requests from this queue, don't idle
1832          */
1833         if (cfqq->dispatched)
1834                 return;
1835
1836         /*
1837          * task has exited, don't wait
1838          */
1839         cic = cfqd->active_cic;
1840         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1841                 return;
1842
1843         /*
1844          * If our average think time is larger than the remaining time
1845          * slice, then don't idle. This avoids overrunning the allotted
1846          * time slice.
1847          */
1848         if (sample_valid(cic->ttime_samples) &&
1849             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1850                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1851                                 cic->ttime_mean);
1852                 return;
1853         }
1854
1855         cfq_mark_cfqq_wait_request(cfqq);
1856
1857         sl = cfqd->cfq_slice_idle;
1858
1859         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1860         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1861 }
1862
1863 /*
1864  * Move request from internal lists to the request queue dispatch list.
1865  */
1866 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1867 {
1868         struct cfq_data *cfqd = q->elevator->elevator_data;
1869         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1870
1871         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1872
1873         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1874         cfq_remove_request(rq);
1875         cfqq->dispatched++;
1876         elv_dispatch_sort(q, rq);
1877
1878         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
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         cfq_log(cfqd, "workload slice:%d", slice);
2062         cfqd->workload_expires = jiffies + slice;
2063         cfqd->noidle_tree_requires_idle = false;
2064 }
2065
2066 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2067 {
2068         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2069         struct cfq_group *cfqg;
2070
2071         if (RB_EMPTY_ROOT(&st->rb))
2072                 return NULL;
2073         cfqg = cfq_rb_first_group(st);
2074         st->active = &cfqg->rb_node;
2075         update_min_vdisktime(st);
2076         return cfqg;
2077 }
2078
2079 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2080 {
2081         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2082
2083         cfqd->serving_group = cfqg;
2084
2085         /* Restore the workload type data */
2086         if (cfqg->saved_workload_slice) {
2087                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2088                 cfqd->serving_type = cfqg->saved_workload;
2089                 cfqd->serving_prio = cfqg->saved_serving_prio;
2090         } else
2091                 cfqd->workload_expires = jiffies - 1;
2092
2093         choose_service_tree(cfqd, cfqg);
2094 }
2095
2096 /*
2097  * Select a queue for service. If we have a current active queue,
2098  * check whether to continue servicing it, or retrieve and set a new one.
2099  */
2100 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2101 {
2102         struct cfq_queue *cfqq, *new_cfqq = NULL;
2103
2104         cfqq = cfqd->active_queue;
2105         if (!cfqq)
2106                 goto new_queue;
2107
2108         if (!cfqd->rq_queued)
2109                 return NULL;
2110
2111         /*
2112          * We were waiting for group to get backlogged. Expire the queue
2113          */
2114         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2115                 goto expire;
2116
2117         /*
2118          * The active queue has run out of time, expire it and select new.
2119          */
2120         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2121                 /*
2122                  * If slice had not expired at the completion of last request
2123                  * we might not have turned on wait_busy flag. Don't expire
2124                  * the queue yet. Allow the group to get backlogged.
2125                  *
2126                  * The very fact that we have used the slice, that means we
2127                  * have been idling all along on this queue and it should be
2128                  * ok to wait for this request to complete.
2129                  */
2130                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2131                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2132                         cfqq = NULL;
2133                         goto keep_queue;
2134                 } else
2135                         goto expire;
2136         }
2137
2138         /*
2139          * The active queue has requests and isn't expired, allow it to
2140          * dispatch.
2141          */
2142         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2143                 goto keep_queue;
2144
2145         /*
2146          * If another queue has a request waiting within our mean seek
2147          * distance, let it run.  The expire code will check for close
2148          * cooperators and put the close queue at the front of the service
2149          * tree.  If possible, merge the expiring queue with the new cfqq.
2150          */
2151         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2152         if (new_cfqq) {
2153                 if (!cfqq->new_cfqq)
2154                         cfq_setup_merge(cfqq, new_cfqq);
2155                 goto expire;
2156         }
2157
2158         /*
2159          * No requests pending. If the active queue still has requests in
2160          * flight or is idling for a new request, allow either of these
2161          * conditions to happen (or time out) before selecting a new queue.
2162          */
2163         if (timer_pending(&cfqd->idle_slice_timer) ||
2164             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2165                 cfqq = NULL;
2166                 goto keep_queue;
2167         }
2168
2169 expire:
2170         cfq_slice_expired(cfqd, 0);
2171 new_queue:
2172         /*
2173          * Current queue expired. Check if we have to switch to a new
2174          * service tree
2175          */
2176         if (!new_cfqq)
2177                 cfq_choose_cfqg(cfqd);
2178
2179         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2180 keep_queue:
2181         return cfqq;
2182 }
2183
2184 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2185 {
2186         int dispatched = 0;
2187
2188         while (cfqq->next_rq) {
2189                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2190                 dispatched++;
2191         }
2192
2193         BUG_ON(!list_empty(&cfqq->fifo));
2194
2195         /* By default cfqq is not expired if it is empty. Do it explicitly */
2196         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2197         return dispatched;
2198 }
2199
2200 /*
2201  * Drain our current requests. Used for barriers and when switching
2202  * io schedulers on-the-fly.
2203  */
2204 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2205 {
2206         struct cfq_queue *cfqq;
2207         int dispatched = 0;
2208
2209         /* Expire the timeslice of the current active queue first */
2210         cfq_slice_expired(cfqd, 0);
2211         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2212                 __cfq_set_active_queue(cfqd, cfqq);
2213                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2214         }
2215
2216         BUG_ON(cfqd->busy_queues);
2217
2218         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2219         return dispatched;
2220 }
2221
2222 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2223         struct cfq_queue *cfqq)
2224 {
2225         /* the queue hasn't finished any request, can't estimate */
2226         if (cfq_cfqq_slice_new(cfqq))
2227                 return 1;
2228         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2229                 cfqq->slice_end))
2230                 return 1;
2231
2232         return 0;
2233 }
2234
2235 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2236 {
2237         unsigned int max_dispatch;
2238
2239         /*
2240          * Drain async requests before we start sync IO
2241          */
2242         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2243                 return false;
2244
2245         /*
2246          * If this is an async queue and we have sync IO in flight, let it wait
2247          */
2248         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2249                 return false;
2250
2251         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2252         if (cfq_class_idle(cfqq))
2253                 max_dispatch = 1;
2254
2255         /*
2256          * Does this cfqq already have too much IO in flight?
2257          */
2258         if (cfqq->dispatched >= max_dispatch) {
2259                 /*
2260                  * idle queue must always only have a single IO in flight
2261                  */
2262                 if (cfq_class_idle(cfqq))
2263                         return false;
2264
2265                 /*
2266                  * We have other queues, don't allow more IO from this one
2267                  */
2268                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2269                         return false;
2270
2271                 /*
2272                  * Sole queue user, no limit
2273                  */
2274                 if (cfqd->busy_queues == 1)
2275                         max_dispatch = -1;
2276                 else
2277                         /*
2278                          * Normally we start throttling cfqq when cfq_quantum/2
2279                          * requests have been dispatched. But we can drive
2280                          * deeper queue depths at the beginning of slice
2281                          * subjected to upper limit of cfq_quantum.
2282                          * */
2283                         max_dispatch = cfqd->cfq_quantum;
2284         }
2285
2286         /*
2287          * Async queues must wait a bit before being allowed dispatch.
2288          * We also ramp up the dispatch depth gradually for async IO,
2289          * based on the last sync IO we serviced
2290          */
2291         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2292                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2293                 unsigned int depth;
2294
2295                 depth = last_sync / cfqd->cfq_slice[1];
2296                 if (!depth && !cfqq->dispatched)
2297                         depth = 1;
2298                 if (depth < max_dispatch)
2299                         max_dispatch = depth;
2300         }
2301
2302         /*
2303          * If we're below the current max, allow a dispatch
2304          */
2305         return cfqq->dispatched < max_dispatch;
2306 }
2307
2308 /*
2309  * Dispatch a request from cfqq, moving them to the request queue
2310  * dispatch list.
2311  */
2312 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2313 {
2314         struct request *rq;
2315
2316         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2317
2318         if (!cfq_may_dispatch(cfqd, cfqq))
2319                 return false;
2320
2321         /*
2322          * follow expired path, else get first next available
2323          */
2324         rq = cfq_check_fifo(cfqq);
2325         if (!rq)
2326                 rq = cfqq->next_rq;
2327
2328         /*
2329          * insert request into driver dispatch list
2330          */
2331         cfq_dispatch_insert(cfqd->queue, rq);
2332
2333         if (!cfqd->active_cic) {
2334                 struct cfq_io_context *cic = RQ_CIC(rq);
2335
2336                 atomic_long_inc(&cic->ioc->refcount);
2337                 cfqd->active_cic = cic;
2338         }
2339
2340         return true;
2341 }
2342
2343 /*
2344  * Find the cfqq that we need to service and move a request from that to the
2345  * dispatch list
2346  */
2347 static int cfq_dispatch_requests(struct request_queue *q, int force)
2348 {
2349         struct cfq_data *cfqd = q->elevator->elevator_data;
2350         struct cfq_queue *cfqq;
2351
2352         if (!cfqd->busy_queues)
2353                 return 0;
2354
2355         if (unlikely(force))
2356                 return cfq_forced_dispatch(cfqd);
2357
2358         cfqq = cfq_select_queue(cfqd);
2359         if (!cfqq)
2360                 return 0;
2361
2362         /*
2363          * Dispatch a request from this cfqq, if it is allowed
2364          */
2365         if (!cfq_dispatch_request(cfqd, cfqq))
2366                 return 0;
2367
2368         cfqq->slice_dispatch++;
2369         cfq_clear_cfqq_must_dispatch(cfqq);
2370
2371         /*
2372          * expire an async queue immediately if it has used up its slice. idle
2373          * queue always expire after 1 dispatch round.
2374          */
2375         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2376             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2377             cfq_class_idle(cfqq))) {
2378                 cfqq->slice_end = jiffies + 1;
2379                 cfq_slice_expired(cfqd, 0);
2380         }
2381
2382         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2383         return 1;
2384 }
2385
2386 /*
2387  * task holds one reference to the queue, dropped when task exits. each rq
2388  * in-flight on this queue also holds a reference, dropped when rq is freed.
2389  *
2390  * Each cfq queue took a reference on the parent group. Drop it now.
2391  * queue lock must be held here.
2392  */
2393 static void cfq_put_queue(struct cfq_queue *cfqq)
2394 {
2395         struct cfq_data *cfqd = cfqq->cfqd;
2396         struct cfq_group *cfqg, *orig_cfqg;
2397
2398         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2399
2400         if (!atomic_dec_and_test(&cfqq->ref))
2401                 return;
2402
2403         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2404         BUG_ON(rb_first(&cfqq->sort_list));
2405         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2406         cfqg = cfqq->cfqg;
2407         orig_cfqg = cfqq->orig_cfqg;
2408
2409         if (unlikely(cfqd->active_queue == cfqq)) {
2410                 __cfq_slice_expired(cfqd, cfqq, 0);
2411                 cfq_schedule_dispatch(cfqd);
2412         }
2413
2414         BUG_ON(cfq_cfqq_on_rr(cfqq));
2415         kmem_cache_free(cfq_pool, cfqq);
2416         cfq_put_cfqg(cfqg);
2417         if (orig_cfqg)
2418                 cfq_put_cfqg(orig_cfqg);
2419 }
2420
2421 /*
2422  * Must always be called with the rcu_read_lock() held
2423  */
2424 static void
2425 __call_for_each_cic(struct io_context *ioc,
2426                     void (*func)(struct io_context *, struct cfq_io_context *))
2427 {
2428         struct cfq_io_context *cic;
2429         struct hlist_node *n;
2430
2431         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2432                 func(ioc, cic);
2433 }
2434
2435 /*
2436  * Call func for each cic attached to this ioc.
2437  */
2438 static void
2439 call_for_each_cic(struct io_context *ioc,
2440                   void (*func)(struct io_context *, struct cfq_io_context *))
2441 {
2442         rcu_read_lock();
2443         __call_for_each_cic(ioc, func);
2444         rcu_read_unlock();
2445 }
2446
2447 static void cfq_cic_free_rcu(struct rcu_head *head)
2448 {
2449         struct cfq_io_context *cic;
2450
2451         cic = container_of(head, struct cfq_io_context, rcu_head);
2452
2453         kmem_cache_free(cfq_ioc_pool, cic);
2454         elv_ioc_count_dec(cfq_ioc_count);
2455
2456         if (ioc_gone) {
2457                 /*
2458                  * CFQ scheduler is exiting, grab exit lock and check
2459                  * the pending io context count. If it hits zero,
2460                  * complete ioc_gone and set it back to NULL
2461                  */
2462                 spin_lock(&ioc_gone_lock);
2463                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2464                         complete(ioc_gone);
2465                         ioc_gone = NULL;
2466                 }
2467                 spin_unlock(&ioc_gone_lock);
2468         }
2469 }
2470
2471 static void cfq_cic_free(struct cfq_io_context *cic)
2472 {
2473         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2474 }
2475
2476 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2477 {
2478         unsigned long flags;
2479
2480         BUG_ON(!cic->dead_key);
2481
2482         spin_lock_irqsave(&ioc->lock, flags);
2483         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2484         hlist_del_rcu(&cic->cic_list);
2485         spin_unlock_irqrestore(&ioc->lock, flags);
2486
2487         cfq_cic_free(cic);
2488 }
2489
2490 /*
2491  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2492  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2493  * and ->trim() which is called with the task lock held
2494  */
2495 static void cfq_free_io_context(struct io_context *ioc)
2496 {
2497         /*
2498          * ioc->refcount is zero here, or we are called from elv_unregister(),
2499          * so no more cic's are allowed to be linked into this ioc.  So it
2500          * should be ok to iterate over the known list, we will see all cic's
2501          * since no new ones are added.
2502          */
2503         __call_for_each_cic(ioc, cic_free_func);
2504 }
2505
2506 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2507 {
2508         struct cfq_queue *__cfqq, *next;
2509
2510         if (unlikely(cfqq == cfqd->active_queue)) {
2511                 __cfq_slice_expired(cfqd, cfqq, 0);
2512                 cfq_schedule_dispatch(cfqd);
2513         }
2514
2515         /*
2516          * If this queue was scheduled to merge with another queue, be
2517          * sure to drop the reference taken on that queue (and others in
2518          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2519          */
2520         __cfqq = cfqq->new_cfqq;
2521         while (__cfqq) {
2522                 if (__cfqq == cfqq) {
2523                         WARN(1, "cfqq->new_cfqq loop detected\n");
2524                         break;
2525                 }
2526                 next = __cfqq->new_cfqq;
2527                 cfq_put_queue(__cfqq);
2528                 __cfqq = next;
2529         }
2530
2531         cfq_put_queue(cfqq);
2532 }
2533
2534 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2535                                          struct cfq_io_context *cic)
2536 {
2537         struct io_context *ioc = cic->ioc;
2538
2539         list_del_init(&cic->queue_list);
2540
2541         /*
2542          * Make sure key == NULL is seen for dead queues
2543          */
2544         smp_wmb();
2545         cic->dead_key = (unsigned long) cic->key;
2546         cic->key = NULL;
2547
2548         if (ioc->ioc_data == cic)
2549                 rcu_assign_pointer(ioc->ioc_data, NULL);
2550
2551         if (cic->cfqq[BLK_RW_ASYNC]) {
2552                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2553                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2554         }
2555
2556         if (cic->cfqq[BLK_RW_SYNC]) {
2557                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2558                 cic->cfqq[BLK_RW_SYNC] = NULL;
2559         }
2560 }
2561
2562 static void cfq_exit_single_io_context(struct io_context *ioc,
2563                                        struct cfq_io_context *cic)
2564 {
2565         struct cfq_data *cfqd = cic->key;
2566
2567         if (cfqd) {
2568                 struct request_queue *q = cfqd->queue;
2569                 unsigned long flags;
2570
2571                 spin_lock_irqsave(q->queue_lock, flags);
2572
2573                 /*
2574                  * Ensure we get a fresh copy of the ->key to prevent
2575                  * race between exiting task and queue
2576                  */
2577                 smp_read_barrier_depends();
2578                 if (cic->key)
2579                         __cfq_exit_single_io_context(cfqd, cic);
2580
2581                 spin_unlock_irqrestore(q->queue_lock, flags);
2582         }
2583 }
2584
2585 /*
2586  * The process that ioc belongs to has exited, we need to clean up
2587  * and put the internal structures we have that belongs to that process.
2588  */
2589 static void cfq_exit_io_context(struct io_context *ioc)
2590 {
2591         call_for_each_cic(ioc, cfq_exit_single_io_context);
2592 }
2593
2594 static struct cfq_io_context *
2595 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2596 {
2597         struct cfq_io_context *cic;
2598
2599         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2600                                                         cfqd->queue->node);
2601         if (cic) {
2602                 cic->last_end_request = jiffies;
2603                 INIT_LIST_HEAD(&cic->queue_list);
2604                 INIT_HLIST_NODE(&cic->cic_list);
2605                 cic->dtor = cfq_free_io_context;
2606                 cic->exit = cfq_exit_io_context;
2607                 elv_ioc_count_inc(cfq_ioc_count);
2608         }
2609
2610         return cic;
2611 }
2612
2613 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2614 {
2615         struct task_struct *tsk = current;
2616         int ioprio_class;
2617
2618         if (!cfq_cfqq_prio_changed(cfqq))
2619                 return;
2620
2621         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2622         switch (ioprio_class) {
2623         default:
2624                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2625         case IOPRIO_CLASS_NONE:
2626                 /*
2627                  * no prio set, inherit CPU scheduling settings
2628                  */
2629                 cfqq->ioprio = task_nice_ioprio(tsk);
2630                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2631                 break;
2632         case IOPRIO_CLASS_RT:
2633                 cfqq->ioprio = task_ioprio(ioc);
2634                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2635                 break;
2636         case IOPRIO_CLASS_BE:
2637                 cfqq->ioprio = task_ioprio(ioc);
2638                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2639                 break;
2640         case IOPRIO_CLASS_IDLE:
2641                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2642                 cfqq->ioprio = 7;
2643                 cfq_clear_cfqq_idle_window(cfqq);
2644                 break;
2645         }
2646
2647         /*
2648          * keep track of original prio settings in case we have to temporarily
2649          * elevate the priority of this queue
2650          */
2651         cfqq->org_ioprio = cfqq->ioprio;
2652         cfqq->org_ioprio_class = cfqq->ioprio_class;
2653         cfq_clear_cfqq_prio_changed(cfqq);
2654 }
2655
2656 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2657 {
2658         struct cfq_data *cfqd = cic->key;
2659         struct cfq_queue *cfqq;
2660         unsigned long flags;
2661
2662         if (unlikely(!cfqd))
2663                 return;
2664
2665         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2666
2667         cfqq = cic->cfqq[BLK_RW_ASYNC];
2668         if (cfqq) {
2669                 struct cfq_queue *new_cfqq;
2670                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2671                                                 GFP_ATOMIC);
2672                 if (new_cfqq) {
2673                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2674                         cfq_put_queue(cfqq);
2675                 }
2676         }
2677
2678         cfqq = cic->cfqq[BLK_RW_SYNC];
2679         if (cfqq)
2680                 cfq_mark_cfqq_prio_changed(cfqq);
2681
2682         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2683 }
2684
2685 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2686 {
2687         call_for_each_cic(ioc, changed_ioprio);
2688         ioc->ioprio_changed = 0;
2689 }
2690
2691 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2692                           pid_t pid, bool is_sync)
2693 {
2694         RB_CLEAR_NODE(&cfqq->rb_node);
2695         RB_CLEAR_NODE(&cfqq->p_node);
2696         INIT_LIST_HEAD(&cfqq->fifo);
2697
2698         atomic_set(&cfqq->ref, 0);
2699         cfqq->cfqd = cfqd;
2700
2701         cfq_mark_cfqq_prio_changed(cfqq);
2702
2703         if (is_sync) {
2704                 if (!cfq_class_idle(cfqq))
2705                         cfq_mark_cfqq_idle_window(cfqq);
2706                 cfq_mark_cfqq_sync(cfqq);
2707         }
2708         cfqq->pid = pid;
2709 }
2710
2711 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2712 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2713 {
2714         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2715         struct cfq_data *cfqd = cic->key;
2716         unsigned long flags;
2717         struct request_queue *q;
2718
2719         if (unlikely(!cfqd))
2720                 return;
2721
2722         q = cfqd->queue;
2723
2724         spin_lock_irqsave(q->queue_lock, flags);
2725
2726         if (sync_cfqq) {
2727                 /*
2728                  * Drop reference to sync queue. A new sync queue will be
2729                  * assigned in new group upon arrival of a fresh request.
2730                  */
2731                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2732                 cic_set_cfqq(cic, NULL, 1);
2733                 cfq_put_queue(sync_cfqq);
2734         }
2735
2736         spin_unlock_irqrestore(q->queue_lock, flags);
2737 }
2738
2739 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2740 {
2741         call_for_each_cic(ioc, changed_cgroup);
2742         ioc->cgroup_changed = 0;
2743 }
2744 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2745
2746 static struct cfq_queue *
2747 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2748                      struct io_context *ioc, gfp_t gfp_mask)
2749 {
2750         struct cfq_queue *cfqq, *new_cfqq = NULL;
2751         struct cfq_io_context *cic;
2752         struct cfq_group *cfqg;
2753
2754 retry:
2755         cfqg = cfq_get_cfqg(cfqd, 1);
2756         cic = cfq_cic_lookup(cfqd, ioc);
2757         /* cic always exists here */
2758         cfqq = cic_to_cfqq(cic, is_sync);
2759
2760         /*
2761          * Always try a new alloc if we fell back to the OOM cfqq
2762          * originally, since it should just be a temporary situation.
2763          */
2764         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2765                 cfqq = NULL;
2766                 if (new_cfqq) {
2767                         cfqq = new_cfqq;
2768                         new_cfqq = NULL;
2769                 } else if (gfp_mask & __GFP_WAIT) {
2770                         spin_unlock_irq(cfqd->queue->queue_lock);
2771                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2772                                         gfp_mask | __GFP_ZERO,
2773                                         cfqd->queue->node);
2774                         spin_lock_irq(cfqd->queue->queue_lock);
2775                         if (new_cfqq)
2776                                 goto retry;
2777                 } else {
2778                         cfqq = kmem_cache_alloc_node(cfq_pool,
2779                                         gfp_mask | __GFP_ZERO,
2780                                         cfqd->queue->node);
2781                 }
2782
2783                 if (cfqq) {
2784                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2785                         cfq_init_prio_data(cfqq, ioc);
2786                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2787                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2788                 } else
2789                         cfqq = &cfqd->oom_cfqq;
2790         }
2791
2792         if (new_cfqq)
2793                 kmem_cache_free(cfq_pool, new_cfqq);
2794
2795         return cfqq;
2796 }
2797
2798 static struct cfq_queue **
2799 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2800 {
2801         switch (ioprio_class) {
2802         case IOPRIO_CLASS_RT:
2803                 return &cfqd->async_cfqq[0][ioprio];
2804         case IOPRIO_CLASS_BE:
2805                 return &cfqd->async_cfqq[1][ioprio];
2806         case IOPRIO_CLASS_IDLE:
2807                 return &cfqd->async_idle_cfqq;
2808         default:
2809                 BUG();
2810         }
2811 }
2812
2813 static struct cfq_queue *
2814 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2815               gfp_t gfp_mask)
2816 {
2817         const int ioprio = task_ioprio(ioc);
2818         const int ioprio_class = task_ioprio_class(ioc);
2819         struct cfq_queue **async_cfqq = NULL;
2820         struct cfq_queue *cfqq = NULL;
2821
2822         if (!is_sync) {
2823                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2824                 cfqq = *async_cfqq;
2825         }
2826
2827         if (!cfqq)
2828                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2829
2830         /*
2831          * pin the queue now that it's allocated, scheduler exit will prune it
2832          */
2833         if (!is_sync && !(*async_cfqq)) {
2834                 atomic_inc(&cfqq->ref);
2835                 *async_cfqq = cfqq;
2836         }
2837
2838         atomic_inc(&cfqq->ref);
2839         return cfqq;
2840 }
2841
2842 /*
2843  * We drop cfq io contexts lazily, so we may find a dead one.
2844  */
2845 static void
2846 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2847                   struct cfq_io_context *cic)
2848 {
2849         unsigned long flags;
2850
2851         WARN_ON(!list_empty(&cic->queue_list));
2852
2853         spin_lock_irqsave(&ioc->lock, flags);
2854
2855         BUG_ON(ioc->ioc_data == cic);
2856
2857         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2858         hlist_del_rcu(&cic->cic_list);
2859         spin_unlock_irqrestore(&ioc->lock, flags);
2860
2861         cfq_cic_free(cic);
2862 }
2863
2864 static struct cfq_io_context *
2865 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2866 {
2867         struct cfq_io_context *cic;
2868         unsigned long flags;
2869         void *k;
2870
2871         if (unlikely(!ioc))
2872                 return NULL;
2873
2874         rcu_read_lock();
2875
2876         /*
2877          * we maintain a last-hit cache, to avoid browsing over the tree
2878          */
2879         cic = rcu_dereference(ioc->ioc_data);
2880         if (cic && cic->key == cfqd) {
2881                 rcu_read_unlock();
2882                 return cic;
2883         }
2884
2885         do {
2886                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2887                 rcu_read_unlock();
2888                 if (!cic)
2889                         break;
2890                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2891                 k = cic->key;
2892                 if (unlikely(!k)) {
2893                         cfq_drop_dead_cic(cfqd, ioc, cic);
2894                         rcu_read_lock();
2895                         continue;
2896                 }
2897
2898                 spin_lock_irqsave(&ioc->lock, flags);
2899                 rcu_assign_pointer(ioc->ioc_data, cic);
2900                 spin_unlock_irqrestore(&ioc->lock, flags);
2901                 break;
2902         } while (1);
2903
2904         return cic;
2905 }
2906
2907 /*
2908  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2909  * the process specific cfq io context when entered from the block layer.
2910  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2911  */
2912 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2913                         struct cfq_io_context *cic, gfp_t gfp_mask)
2914 {
2915         unsigned long flags;
2916         int ret;
2917
2918         ret = radix_tree_preload(gfp_mask);
2919         if (!ret) {
2920                 cic->ioc = ioc;
2921                 cic->key = cfqd;
2922
2923                 spin_lock_irqsave(&ioc->lock, flags);
2924                 ret = radix_tree_insert(&ioc->radix_root,
2925                                                 (unsigned long) cfqd, cic);
2926                 if (!ret)
2927                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2928                 spin_unlock_irqrestore(&ioc->lock, flags);
2929
2930                 radix_tree_preload_end();
2931
2932                 if (!ret) {
2933                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2934                         list_add(&cic->queue_list, &cfqd->cic_list);
2935                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2936                 }
2937         }
2938
2939         if (ret)
2940                 printk(KERN_ERR "cfq: cic link failed!\n");
2941
2942         return ret;
2943 }
2944
2945 /*
2946  * Setup general io context and cfq io context. There can be several cfq
2947  * io contexts per general io context, if this process is doing io to more
2948  * than one device managed by cfq.
2949  */
2950 static struct cfq_io_context *
2951 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2952 {
2953         struct io_context *ioc = NULL;
2954         struct cfq_io_context *cic;
2955
2956         might_sleep_if(gfp_mask & __GFP_WAIT);
2957
2958         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2959         if (!ioc)
2960                 return NULL;
2961
2962         cic = cfq_cic_lookup(cfqd, ioc);
2963         if (cic)
2964                 goto out;
2965
2966         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2967         if (cic == NULL)
2968                 goto err;
2969
2970         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2971                 goto err_free;
2972
2973 out:
2974         smp_read_barrier_depends();
2975         if (unlikely(ioc->ioprio_changed))
2976                 cfq_ioc_set_ioprio(ioc);
2977
2978 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2979         if (unlikely(ioc->cgroup_changed))
2980                 cfq_ioc_set_cgroup(ioc);
2981 #endif
2982         return cic;
2983 err_free:
2984         cfq_cic_free(cic);
2985 err:
2986         put_io_context(ioc);
2987         return NULL;
2988 }
2989
2990 static void
2991 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2992 {
2993         unsigned long elapsed = jiffies - cic->last_end_request;
2994         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2995
2996         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2997         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2998         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2999 }
3000
3001 static void
3002 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3003                        struct request *rq)
3004 {
3005         sector_t sdist = 0;
3006         sector_t n_sec = blk_rq_sectors(rq);
3007         if (cfqq->last_request_pos) {
3008                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3009                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3010                 else
3011                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3012         }
3013
3014         cfqq->seek_history <<= 1;
3015         if (blk_queue_nonrot(cfqd->queue))
3016                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3017         else
3018                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3019 }
3020
3021 /*
3022  * Disable idle window if the process thinks too long or seeks so much that
3023  * it doesn't matter
3024  */
3025 static void
3026 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3027                        struct cfq_io_context *cic)
3028 {
3029         int old_idle, enable_idle;
3030
3031         /*
3032          * Don't idle for async or idle io prio class
3033          */
3034         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3035                 return;
3036
3037         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3038
3039         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3040                 cfq_mark_cfqq_deep(cfqq);
3041
3042         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3043             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3044                 enable_idle = 0;
3045         else if (sample_valid(cic->ttime_samples)) {
3046                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3047                         enable_idle = 0;
3048                 else
3049                         enable_idle = 1;
3050         }
3051
3052         if (old_idle != enable_idle) {
3053                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3054                 if (enable_idle)
3055                         cfq_mark_cfqq_idle_window(cfqq);
3056                 else
3057                         cfq_clear_cfqq_idle_window(cfqq);
3058         }
3059 }
3060
3061 /*
3062  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3063  * no or if we aren't sure, a 1 will cause a preempt.
3064  */
3065 static bool
3066 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3067                    struct request *rq)
3068 {
3069         struct cfq_queue *cfqq;
3070
3071         cfqq = cfqd->active_queue;
3072         if (!cfqq)
3073                 return false;
3074
3075         if (cfq_class_idle(new_cfqq))
3076                 return false;
3077
3078         if (cfq_class_idle(cfqq))
3079                 return true;
3080
3081         /*
3082          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3083          */
3084         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3085                 return false;
3086
3087         /*
3088          * if the new request is sync, but the currently running queue is
3089          * not, let the sync request have priority.
3090          */
3091         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3092                 return true;
3093
3094         if (new_cfqq->cfqg != cfqq->cfqg)
3095                 return false;
3096
3097         if (cfq_slice_used(cfqq))
3098                 return true;
3099
3100         /* Allow preemption only if we are idling on sync-noidle tree */
3101         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3102             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3103             new_cfqq->service_tree->count == 2 &&
3104             RB_EMPTY_ROOT(&cfqq->sort_list))
3105                 return true;
3106
3107         /*
3108          * So both queues are sync. Let the new request get disk time if
3109          * it's a metadata request and the current queue is doing regular IO.
3110          */
3111         if (rq_is_meta(rq) && !cfqq->meta_pending)
3112                 return true;
3113
3114         /*
3115          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3116          */
3117         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3118                 return true;
3119
3120         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3121                 return false;
3122
3123         /*
3124          * if this request is as-good as one we would expect from the
3125          * current cfqq, let it preempt
3126          */
3127         if (cfq_rq_close(cfqd, cfqq, rq))
3128                 return true;
3129
3130         return false;
3131 }
3132
3133 /*
3134  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3135  * let it have half of its nominal slice.
3136  */
3137 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3138 {
3139         cfq_log_cfqq(cfqd, cfqq, "preempt");
3140         cfq_slice_expired(cfqd, 1);
3141
3142         /*
3143          * Put the new queue at the front of the of the current list,
3144          * so we know that it will be selected next.
3145          */
3146         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3147
3148         cfq_service_tree_add(cfqd, cfqq, 1);
3149
3150         cfqq->slice_end = 0;
3151         cfq_mark_cfqq_slice_new(cfqq);
3152 }
3153
3154 /*
3155  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3156  * something we should do about it
3157  */
3158 static void
3159 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3160                 struct request *rq)
3161 {
3162         struct cfq_io_context *cic = RQ_CIC(rq);
3163
3164         cfqd->rq_queued++;
3165         if (rq_is_meta(rq))
3166                 cfqq->meta_pending++;
3167
3168         cfq_update_io_thinktime(cfqd, cic);
3169         cfq_update_io_seektime(cfqd, cfqq, rq);
3170         cfq_update_idle_window(cfqd, cfqq, cic);
3171
3172         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3173
3174         if (cfqq == cfqd->active_queue) {
3175                 /*
3176                  * Remember that we saw a request from this process, but
3177                  * don't start queuing just yet. Otherwise we risk seeing lots
3178                  * of tiny requests, because we disrupt the normal plugging
3179                  * and merging. If the request is already larger than a single
3180                  * page, let it rip immediately. For that case we assume that
3181                  * merging is already done. Ditto for a busy system that
3182                  * has other work pending, don't risk delaying until the
3183                  * idle timer unplug to continue working.
3184                  */
3185                 if (cfq_cfqq_wait_request(cfqq)) {
3186                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3187                             cfqd->busy_queues > 1) {
3188                                 del_timer(&cfqd->idle_slice_timer);
3189                                 cfq_clear_cfqq_wait_request(cfqq);
3190                                 __blk_run_queue(cfqd->queue);
3191                         } else
3192                                 cfq_mark_cfqq_must_dispatch(cfqq);
3193                 }
3194         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3195                 /*
3196                  * not the active queue - expire current slice if it is
3197                  * idle and has expired it's mean thinktime or this new queue
3198                  * has some old slice time left and is of higher priority or
3199                  * this new queue is RT and the current one is BE
3200                  */
3201                 cfq_preempt_queue(cfqd, cfqq);
3202                 __blk_run_queue(cfqd->queue);
3203         }
3204 }
3205
3206 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3207 {
3208         struct cfq_data *cfqd = q->elevator->elevator_data;
3209         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3210
3211         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3212         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3213
3214         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3215         list_add_tail(&rq->queuelist, &cfqq->fifo);
3216         cfq_add_rq_rb(rq);
3217
3218         cfq_rq_enqueued(cfqd, cfqq, rq);
3219 }
3220
3221 /*
3222  * Update hw_tag based on peak queue depth over 50 samples under
3223  * sufficient load.
3224  */
3225 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3226 {
3227         struct cfq_queue *cfqq = cfqd->active_queue;
3228
3229         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3230                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3231
3232         if (cfqd->hw_tag == 1)
3233                 return;
3234
3235         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3236             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3237                 return;
3238
3239         /*
3240          * If active queue hasn't enough requests and can idle, cfq might not
3241          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3242          * case
3243          */
3244         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3245             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3246             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3247                 return;
3248
3249         if (cfqd->hw_tag_samples++ < 50)
3250                 return;
3251
3252         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3253                 cfqd->hw_tag = 1;
3254         else
3255                 cfqd->hw_tag = 0;
3256 }
3257
3258 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3259 {
3260         struct cfq_io_context *cic = cfqd->active_cic;
3261
3262         /* If there are other queues in the group, don't wait */
3263         if (cfqq->cfqg->nr_cfqq > 1)
3264                 return false;
3265
3266         if (cfq_slice_used(cfqq))
3267                 return true;
3268
3269         /* if slice left is less than think time, wait busy */
3270         if (cic && sample_valid(cic->ttime_samples)
3271             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3272                 return true;
3273
3274         /*
3275          * If think times is less than a jiffy than ttime_mean=0 and above
3276          * will not be true. It might happen that slice has not expired yet
3277          * but will expire soon (4-5 ns) during select_queue(). To cover the
3278          * case where think time is less than a jiffy, mark the queue wait
3279          * busy if only 1 jiffy is left in the slice.
3280          */
3281         if (cfqq->slice_end - jiffies == 1)
3282                 return true;
3283
3284         return false;
3285 }
3286
3287 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3288 {
3289         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3290         struct cfq_data *cfqd = cfqq->cfqd;
3291         const int sync = rq_is_sync(rq);
3292         unsigned long now;
3293
3294         now = jiffies;
3295         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3296
3297         cfq_update_hw_tag(cfqd);
3298
3299         WARN_ON(!cfqd->rq_in_driver);
3300         WARN_ON(!cfqq->dispatched);
3301         cfqd->rq_in_driver--;
3302         cfqq->dispatched--;
3303
3304         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3305
3306         if (sync) {
3307                 RQ_CIC(rq)->last_end_request = now;
3308                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3309                         cfqd->last_delayed_sync = now;
3310         }
3311
3312         /*
3313          * If this is the active queue, check if it needs to be expired,
3314          * or if we want to idle in case it has no pending requests.
3315          */
3316         if (cfqd->active_queue == cfqq) {
3317                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3318
3319                 if (cfq_cfqq_slice_new(cfqq)) {
3320                         cfq_set_prio_slice(cfqd, cfqq);
3321                         cfq_clear_cfqq_slice_new(cfqq);
3322                 }
3323
3324                 /*
3325                  * Should we wait for next request to come in before we expire
3326                  * the queue.
3327                  */
3328                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3329                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3330                         cfq_mark_cfqq_wait_busy(cfqq);
3331                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3332                 }
3333
3334                 /*
3335                  * Idling is not enabled on:
3336                  * - expired queues
3337                  * - idle-priority queues
3338                  * - async queues
3339                  * - queues with still some requests queued
3340                  * - when there is a close cooperator
3341                  */
3342                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3343                         cfq_slice_expired(cfqd, 1);
3344                 else if (sync && cfqq_empty &&
3345                          !cfq_close_cooperator(cfqd, cfqq)) {
3346                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3347                         /*
3348                          * Idling is enabled for SYNC_WORKLOAD.
3349                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3350                          * only if we processed at least one !rq_noidle request
3351                          */
3352                         if (cfqd->serving_type == SYNC_WORKLOAD
3353                             || cfqd->noidle_tree_requires_idle
3354                             || cfqq->cfqg->nr_cfqq == 1)
3355                                 cfq_arm_slice_timer(cfqd);
3356                 }
3357         }
3358
3359         if (!cfqd->rq_in_driver)
3360                 cfq_schedule_dispatch(cfqd);
3361 }
3362
3363 /*
3364  * we temporarily boost lower priority queues if they are holding fs exclusive
3365  * resources. they are boosted to normal prio (CLASS_BE/4)
3366  */
3367 static void cfq_prio_boost(struct cfq_queue *cfqq)
3368 {
3369         if (has_fs_excl()) {
3370                 /*
3371                  * boost idle prio on transactions that would lock out other
3372                  * users of the filesystem
3373                  */
3374                 if (cfq_class_idle(cfqq))
3375                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3376                 if (cfqq->ioprio > IOPRIO_NORM)
3377                         cfqq->ioprio = IOPRIO_NORM;
3378         } else {
3379                 /*
3380                  * unboost the queue (if needed)
3381                  */
3382                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3383                 cfqq->ioprio = cfqq->org_ioprio;
3384         }
3385 }
3386
3387 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3388 {
3389         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3390                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3391                 return ELV_MQUEUE_MUST;
3392         }
3393
3394         return ELV_MQUEUE_MAY;
3395 }
3396
3397 static int cfq_may_queue(struct request_queue *q, int rw)
3398 {
3399         struct cfq_data *cfqd = q->elevator->elevator_data;
3400         struct task_struct *tsk = current;
3401         struct cfq_io_context *cic;
3402         struct cfq_queue *cfqq;
3403
3404         /*
3405          * don't force setup of a queue from here, as a call to may_queue
3406          * does not necessarily imply that a request actually will be queued.
3407          * so just lookup a possibly existing queue, or return 'may queue'
3408          * if that fails
3409          */
3410         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3411         if (!cic)
3412                 return ELV_MQUEUE_MAY;
3413
3414         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3415         if (cfqq) {
3416                 cfq_init_prio_data(cfqq, cic->ioc);
3417                 cfq_prio_boost(cfqq);
3418
3419                 return __cfq_may_queue(cfqq);
3420         }
3421
3422         return ELV_MQUEUE_MAY;
3423 }
3424
3425 /*
3426  * queue lock held here
3427  */
3428 static void cfq_put_request(struct request *rq)
3429 {
3430         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3431
3432         if (cfqq) {
3433                 const int rw = rq_data_dir(rq);
3434
3435                 BUG_ON(!cfqq->allocated[rw]);
3436                 cfqq->allocated[rw]--;
3437
3438                 put_io_context(RQ_CIC(rq)->ioc);
3439
3440                 rq->elevator_private = NULL;
3441                 rq->elevator_private2 = NULL;
3442
3443                 cfq_put_queue(cfqq);
3444         }
3445 }
3446
3447 static struct cfq_queue *
3448 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3449                 struct cfq_queue *cfqq)
3450 {
3451         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3452         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3453         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3454         cfq_put_queue(cfqq);
3455         return cic_to_cfqq(cic, 1);
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->pid = current->pid;
3467                 cfq_clear_cfqq_coop(cfqq);
3468                 cfq_clear_cfqq_split_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) && cfq_cfqq_split_coop(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         rcu_read_lock();
3698         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3699                                         0);
3700         rcu_read_unlock();
3701 #endif
3702         /*
3703          * Not strictly needed (since RB_ROOT just clears the node and we
3704          * zeroed cfqd on alloc), but better be safe in case someone decides
3705          * to add magic to the rb code
3706          */
3707         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3708                 cfqd->prio_trees[i] = RB_ROOT;
3709
3710         /*
3711          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3712          * Grab a permanent reference to it, so that the normal code flow
3713          * will not attempt to free it.
3714          */
3715         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3716         atomic_inc(&cfqd->oom_cfqq.ref);
3717         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3718
3719         INIT_LIST_HEAD(&cfqd->cic_list);
3720
3721         cfqd->queue = q;
3722
3723         init_timer(&cfqd->idle_slice_timer);
3724         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3725         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3726
3727         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3728
3729         cfqd->cfq_quantum = cfq_quantum;
3730         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3731         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3732         cfqd->cfq_back_max = cfq_back_max;
3733         cfqd->cfq_back_penalty = cfq_back_penalty;
3734         cfqd->cfq_slice[0] = cfq_slice_async;
3735         cfqd->cfq_slice[1] = cfq_slice_sync;
3736         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3737         cfqd->cfq_slice_idle = cfq_slice_idle;
3738         cfqd->cfq_latency = 1;
3739         cfqd->cfq_group_isolation = 0;
3740         cfqd->hw_tag = -1;
3741         /*
3742          * we optimistically start assuming sync ops weren't delayed in last
3743          * second, in order to have larger depth for async operations.
3744          */
3745         cfqd->last_delayed_sync = jiffies - HZ;
3746         INIT_RCU_HEAD(&cfqd->rcu);
3747         return cfqd;
3748 }
3749
3750 static void cfq_slab_kill(void)
3751 {
3752         /*
3753          * Caller already ensured that pending RCU callbacks are completed,
3754          * so we should have no busy allocations at this point.
3755          */
3756         if (cfq_pool)
3757                 kmem_cache_destroy(cfq_pool);
3758         if (cfq_ioc_pool)
3759                 kmem_cache_destroy(cfq_ioc_pool);
3760 }
3761
3762 static int __init cfq_slab_setup(void)
3763 {
3764         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3765         if (!cfq_pool)
3766                 goto fail;
3767
3768         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3769         if (!cfq_ioc_pool)
3770                 goto fail;
3771
3772         return 0;
3773 fail:
3774         cfq_slab_kill();
3775         return -ENOMEM;
3776 }
3777
3778 /*
3779  * sysfs parts below -->
3780  */
3781 static ssize_t
3782 cfq_var_show(unsigned int var, char *page)
3783 {
3784         return sprintf(page, "%d\n", var);
3785 }
3786
3787 static ssize_t
3788 cfq_var_store(unsigned int *var, const char *page, size_t count)
3789 {
3790         char *p = (char *) page;
3791
3792         *var = simple_strtoul(p, &p, 10);
3793         return count;
3794 }
3795
3796 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3797 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3798 {                                                                       \
3799         struct cfq_data *cfqd = e->elevator_data;                       \
3800         unsigned int __data = __VAR;                                    \
3801         if (__CONV)                                                     \
3802                 __data = jiffies_to_msecs(__data);                      \
3803         return cfq_var_show(__data, (page));                            \
3804 }
3805 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3806 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3807 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3808 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3809 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3810 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3811 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3812 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3813 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3814 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3815 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3816 #undef SHOW_FUNCTION
3817
3818 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3819 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3820 {                                                                       \
3821         struct cfq_data *cfqd = e->elevator_data;                       \
3822         unsigned int __data;                                            \
3823         int ret = cfq_var_store(&__data, (page), count);                \
3824         if (__data < (MIN))                                             \
3825                 __data = (MIN);                                         \
3826         else if (__data > (MAX))                                        \
3827                 __data = (MAX);                                         \
3828         if (__CONV)                                                     \
3829                 *(__PTR) = msecs_to_jiffies(__data);                    \
3830         else                                                            \
3831                 *(__PTR) = __data;                                      \
3832         return ret;                                                     \
3833 }
3834 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3835 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3836                 UINT_MAX, 1);
3837 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3838                 UINT_MAX, 1);
3839 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3840 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3841                 UINT_MAX, 0);
3842 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3843 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3844 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3845 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3846                 UINT_MAX, 0);
3847 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3848 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3849 #undef STORE_FUNCTION
3850
3851 #define CFQ_ATTR(name) \
3852         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3853
3854 static struct elv_fs_entry cfq_attrs[] = {
3855         CFQ_ATTR(quantum),
3856         CFQ_ATTR(fifo_expire_sync),
3857         CFQ_ATTR(fifo_expire_async),
3858         CFQ_ATTR(back_seek_max),
3859         CFQ_ATTR(back_seek_penalty),
3860         CFQ_ATTR(slice_sync),
3861         CFQ_ATTR(slice_async),
3862         CFQ_ATTR(slice_async_rq),
3863         CFQ_ATTR(slice_idle),
3864         CFQ_ATTR(low_latency),
3865         CFQ_ATTR(group_isolation),
3866         __ATTR_NULL
3867 };
3868
3869 static struct elevator_type iosched_cfq = {
3870         .ops = {
3871                 .elevator_merge_fn =            cfq_merge,
3872                 .elevator_merged_fn =           cfq_merged_request,
3873                 .elevator_merge_req_fn =        cfq_merged_requests,
3874                 .elevator_allow_merge_fn =      cfq_allow_merge,
3875                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3876                 .elevator_add_req_fn =          cfq_insert_request,
3877                 .elevator_activate_req_fn =     cfq_activate_request,
3878                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3879                 .elevator_queue_empty_fn =      cfq_queue_empty,
3880                 .elevator_completed_req_fn =    cfq_completed_request,
3881                 .elevator_former_req_fn =       elv_rb_former_request,
3882                 .elevator_latter_req_fn =       elv_rb_latter_request,
3883                 .elevator_set_req_fn =          cfq_set_request,
3884                 .elevator_put_req_fn =          cfq_put_request,
3885                 .elevator_may_queue_fn =        cfq_may_queue,
3886                 .elevator_init_fn =             cfq_init_queue,
3887                 .elevator_exit_fn =             cfq_exit_queue,
3888                 .trim =                         cfq_free_io_context,
3889         },
3890         .elevator_attrs =       cfq_attrs,
3891         .elevator_name =        "cfq",
3892         .elevator_owner =       THIS_MODULE,
3893 };
3894
3895 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3896 static struct blkio_policy_type blkio_policy_cfq = {
3897         .ops = {
3898                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3899                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3900         },
3901 };
3902 #else
3903 static struct blkio_policy_type blkio_policy_cfq;
3904 #endif
3905
3906 static int __init cfq_init(void)
3907 {
3908         /*
3909          * could be 0 on HZ < 1000 setups
3910          */
3911         if (!cfq_slice_async)
3912                 cfq_slice_async = 1;
3913         if (!cfq_slice_idle)
3914                 cfq_slice_idle = 1;
3915
3916         if (cfq_slab_setup())
3917                 return -ENOMEM;
3918
3919         elv_register(&iosched_cfq);
3920         blkio_policy_register(&blkio_policy_cfq);
3921
3922         return 0;
3923 }
3924
3925 static void __exit cfq_exit(void)
3926 {
3927         DECLARE_COMPLETION_ONSTACK(all_gone);
3928         blkio_policy_unregister(&blkio_policy_cfq);
3929         elv_unregister(&iosched_cfq);
3930         ioc_gone = &all_gone;
3931         /* ioc_gone's update must be visible before reading ioc_count */
3932         smp_wmb();
3933
3934         /*
3935          * this also protects us from entering cfq_slab_kill() with
3936          * pending RCU callbacks
3937          */
3938         if (elv_ioc_count_read(cfq_ioc_count))
3939                 wait_for_completion(&all_gone);
3940         cfq_slab_kill();
3941 }
3942
3943 module_init(cfq_init);
3944 module_exit(cfq_exit);
3945
3946 MODULE_AUTHOR("Jens Axboe");
3947 MODULE_LICENSE("GPL");
3948 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");