blk-cgroup: config options re-arrangement
[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 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
59
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
62
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
66
67 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
68 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
69 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
70
71 #define sample_valid(samples)   ((samples) > 80)
72 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
73
74 /*
75  * Most of our rbtree usage is for sorting with min extraction, so
76  * if we cache the leftmost node we don't have to walk down the tree
77  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
78  * move this into the elevator for the rq sorting as well.
79  */
80 struct cfq_rb_root {
81         struct rb_root rb;
82         struct rb_node *left;
83         unsigned count;
84         unsigned total_weight;
85         u64 min_vdisktime;
86         struct rb_node *active;
87 };
88 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
89                         .count = 0, .min_vdisktime = 0, }
90
91 /*
92  * Per process-grouping structure
93  */
94 struct cfq_queue {
95         /* reference count */
96         atomic_t ref;
97         /* various state flags, see below */
98         unsigned int flags;
99         /* parent cfq_data */
100         struct cfq_data *cfqd;
101         /* service_tree member */
102         struct rb_node rb_node;
103         /* service_tree key */
104         unsigned long rb_key;
105         /* prio tree member */
106         struct rb_node p_node;
107         /* prio tree root we belong to, if any */
108         struct rb_root *p_root;
109         /* sorted list of pending requests */
110         struct rb_root sort_list;
111         /* if fifo isn't expired, next request to serve */
112         struct request *next_rq;
113         /* requests queued in sort_list */
114         int queued[2];
115         /* currently allocated requests */
116         int allocated[2];
117         /* fifo list of requests in sort_list */
118         struct list_head fifo;
119
120         /* time when queue got scheduled in to dispatch first request. */
121         unsigned long dispatch_start;
122         unsigned int allocated_slice;
123         unsigned int slice_dispatch;
124         /* time when first request from queue completed and slice started. */
125         unsigned long slice_start;
126         unsigned long slice_end;
127         long slice_resid;
128
129         /* pending metadata requests */
130         int meta_pending;
131         /* number of requests that are on the dispatch list or inside driver */
132         int dispatched;
133
134         /* io prio of this group */
135         unsigned short ioprio, org_ioprio;
136         unsigned short ioprio_class, org_ioprio_class;
137
138         pid_t pid;
139
140         u32 seek_history;
141         sector_t last_request_pos;
142
143         struct cfq_rb_root *service_tree;
144         struct cfq_queue *new_cfqq;
145         struct cfq_group *cfqg;
146         struct cfq_group *orig_cfqg;
147 };
148
149 /*
150  * First index in the service_trees.
151  * IDLE is handled separately, so it has negative index
152  */
153 enum wl_prio_t {
154         BE_WORKLOAD = 0,
155         RT_WORKLOAD = 1,
156         IDLE_WORKLOAD = 2,
157 };
158
159 /*
160  * Second index in the service_trees.
161  */
162 enum wl_type_t {
163         ASYNC_WORKLOAD = 0,
164         SYNC_NOIDLE_WORKLOAD = 1,
165         SYNC_WORKLOAD = 2
166 };
167
168 /* This is per cgroup per device grouping structure */
169 struct cfq_group {
170         /* group service_tree member */
171         struct rb_node rb_node;
172
173         /* group service_tree key */
174         u64 vdisktime;
175         unsigned int weight;
176         bool on_st;
177
178         /* number of cfqq currently on this group */
179         int nr_cfqq;
180
181         /* Per group busy queus average. Useful for workload slice calc. */
182         unsigned int busy_queues_avg[2];
183         /*
184          * rr lists of queues with requests, onle rr for each priority class.
185          * Counts are embedded in the cfq_rb_root
186          */
187         struct cfq_rb_root service_trees[2][3];
188         struct cfq_rb_root service_tree_idle;
189
190         unsigned long saved_workload_slice;
191         enum wl_type_t saved_workload;
192         enum wl_prio_t saved_serving_prio;
193         struct blkio_group blkg;
194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
195         struct hlist_node cfqd_node;
196         atomic_t ref;
197 #endif
198 };
199
200 /*
201  * Per block device queue structure
202  */
203 struct cfq_data {
204         struct request_queue *queue;
205         /* Root service tree for cfq_groups */
206         struct cfq_rb_root grp_service_tree;
207         struct cfq_group root_group;
208
209         /*
210          * The priority currently being served
211          */
212         enum wl_prio_t serving_prio;
213         enum wl_type_t serving_type;
214         unsigned long workload_expires;
215         struct cfq_group *serving_group;
216         bool noidle_tree_requires_idle;
217
218         /*
219          * Each priority tree is sorted by next_request position.  These
220          * trees are used when determining if two or more queues are
221          * interleaving requests (see cfq_close_cooperator).
222          */
223         struct rb_root prio_trees[CFQ_PRIO_LISTS];
224
225         unsigned int busy_queues;
226
227         int rq_in_driver;
228         int rq_in_flight[2];
229
230         /*
231          * queue-depth detection
232          */
233         int rq_queued;
234         int hw_tag;
235         /*
236          * hw_tag can be
237          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
239          *  0 => no NCQ
240          */
241         int hw_tag_est_depth;
242         unsigned int hw_tag_samples;
243
244         /*
245          * idle window management
246          */
247         struct timer_list idle_slice_timer;
248         struct work_struct unplug_work;
249
250         struct cfq_queue *active_queue;
251         struct cfq_io_context *active_cic;
252
253         /*
254          * async queue for each priority case
255          */
256         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
257         struct cfq_queue *async_idle_cfqq;
258
259         sector_t last_position;
260
261         /*
262          * tunables, see top of file
263          */
264         unsigned int cfq_quantum;
265         unsigned int cfq_fifo_expire[2];
266         unsigned int cfq_back_penalty;
267         unsigned int cfq_back_max;
268         unsigned int cfq_slice[2];
269         unsigned int cfq_slice_async_rq;
270         unsigned int cfq_slice_idle;
271         unsigned int cfq_latency;
272         unsigned int cfq_group_isolation;
273
274         struct list_head cic_list;
275
276         /*
277          * Fallback dummy cfqq for extreme OOM conditions
278          */
279         struct cfq_queue oom_cfqq;
280
281         unsigned long last_delayed_sync;
282
283         /* List of cfq groups being managed on this device*/
284         struct hlist_head cfqg_list;
285         struct rcu_head rcu;
286 };
287
288 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
289
290 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
291                                             enum wl_prio_t prio,
292                                             enum wl_type_t type)
293 {
294         if (!cfqg)
295                 return NULL;
296
297         if (prio == IDLE_WORKLOAD)
298                 return &cfqg->service_tree_idle;
299
300         return &cfqg->service_trees[prio][type];
301 }
302
303 enum cfqq_state_flags {
304         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
305         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
306         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
307         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
308         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
309         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
310         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
311         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
312         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
313         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
314         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
315         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
316         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
317 };
318
319 #define CFQ_CFQQ_FNS(name)                                              \
320 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
321 {                                                                       \
322         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
323 }                                                                       \
324 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
325 {                                                                       \
326         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
327 }                                                                       \
328 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
329 {                                                                       \
330         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
331 }
332
333 CFQ_CFQQ_FNS(on_rr);
334 CFQ_CFQQ_FNS(wait_request);
335 CFQ_CFQQ_FNS(must_dispatch);
336 CFQ_CFQQ_FNS(must_alloc_slice);
337 CFQ_CFQQ_FNS(fifo_expire);
338 CFQ_CFQQ_FNS(idle_window);
339 CFQ_CFQQ_FNS(prio_changed);
340 CFQ_CFQQ_FNS(slice_new);
341 CFQ_CFQQ_FNS(sync);
342 CFQ_CFQQ_FNS(coop);
343 CFQ_CFQQ_FNS(split_coop);
344 CFQ_CFQQ_FNS(deep);
345 CFQ_CFQQ_FNS(wait_busy);
346 #undef CFQ_CFQQ_FNS
347
348 #ifdef CONFIG_CFQ_GROUP_IOSCHED
349 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
350         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
353
354 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
355         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
356                                 blkg_path(&(cfqg)->blkg), ##args);      \
357
358 #else
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
360         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
361 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
362 #endif
363 #define cfq_log(cfqd, fmt, args...)     \
364         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
365
366 /* Traverses through cfq group service trees */
367 #define for_each_cfqg_st(cfqg, i, j, st) \
368         for (i = 0; i <= IDLE_WORKLOAD; i++) \
369                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370                         : &cfqg->service_tree_idle; \
371                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372                         (i == IDLE_WORKLOAD && j == 0); \
373                         j++, st = i < IDLE_WORKLOAD ? \
374                         &cfqg->service_trees[i][j]: NULL) \
375
376
377 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
378 {
379         if (cfq_class_idle(cfqq))
380                 return IDLE_WORKLOAD;
381         if (cfq_class_rt(cfqq))
382                 return RT_WORKLOAD;
383         return BE_WORKLOAD;
384 }
385
386
387 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
388 {
389         if (!cfq_cfqq_sync(cfqq))
390                 return ASYNC_WORKLOAD;
391         if (!cfq_cfqq_idle_window(cfqq))
392                 return SYNC_NOIDLE_WORKLOAD;
393         return SYNC_WORKLOAD;
394 }
395
396 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
397                                         struct cfq_data *cfqd,
398                                         struct cfq_group *cfqg)
399 {
400         if (wl == IDLE_WORKLOAD)
401                 return cfqg->service_tree_idle.count;
402
403         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
404                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
405                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
406 }
407
408 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
409                                         struct cfq_group *cfqg)
410 {
411         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
412                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
413 }
414
415 static void cfq_dispatch_insert(struct request_queue *, struct request *);
416 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
417                                        struct io_context *, gfp_t);
418 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
419                                                 struct io_context *);
420
421 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
422                                             bool is_sync)
423 {
424         return cic->cfqq[is_sync];
425 }
426
427 static inline void cic_set_cfqq(struct cfq_io_context *cic,
428                                 struct cfq_queue *cfqq, bool is_sync)
429 {
430         cic->cfqq[is_sync] = cfqq;
431 }
432
433 /*
434  * We regard a request as SYNC, if it's either a read or has the SYNC bit
435  * set (in which case it could also be direct WRITE).
436  */
437 static inline bool cfq_bio_sync(struct bio *bio)
438 {
439         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
440 }
441
442 /*
443  * scheduler run of queue, if there are requests pending and no one in the
444  * driver that will restart queueing
445  */
446 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
447 {
448         if (cfqd->busy_queues) {
449                 cfq_log(cfqd, "schedule dispatch");
450                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
451         }
452 }
453
454 static int cfq_queue_empty(struct request_queue *q)
455 {
456         struct cfq_data *cfqd = q->elevator->elevator_data;
457
458         return !cfqd->rq_queued;
459 }
460
461 /*
462  * Scale schedule slice based on io priority. Use the sync time slice only
463  * if a queue is marked sync and has sync io queued. A sync queue with async
464  * io only, should not get full sync slice length.
465  */
466 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
467                                  unsigned short prio)
468 {
469         const int base_slice = cfqd->cfq_slice[sync];
470
471         WARN_ON(prio >= IOPRIO_BE_NR);
472
473         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
474 }
475
476 static inline int
477 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
478 {
479         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
480 }
481
482 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
483 {
484         u64 d = delta << CFQ_SERVICE_SHIFT;
485
486         d = d * BLKIO_WEIGHT_DEFAULT;
487         do_div(d, cfqg->weight);
488         return d;
489 }
490
491 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
492 {
493         s64 delta = (s64)(vdisktime - min_vdisktime);
494         if (delta > 0)
495                 min_vdisktime = vdisktime;
496
497         return min_vdisktime;
498 }
499
500 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
501 {
502         s64 delta = (s64)(vdisktime - min_vdisktime);
503         if (delta < 0)
504                 min_vdisktime = vdisktime;
505
506         return min_vdisktime;
507 }
508
509 static void update_min_vdisktime(struct cfq_rb_root *st)
510 {
511         u64 vdisktime = st->min_vdisktime;
512         struct cfq_group *cfqg;
513
514         if (st->active) {
515                 cfqg = rb_entry_cfqg(st->active);
516                 vdisktime = cfqg->vdisktime;
517         }
518
519         if (st->left) {
520                 cfqg = rb_entry_cfqg(st->left);
521                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
522         }
523
524         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
525 }
526
527 /*
528  * get averaged number of queues of RT/BE priority.
529  * average is updated, with a formula that gives more weight to higher numbers,
530  * to quickly follows sudden increases and decrease slowly
531  */
532
533 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
534                                         struct cfq_group *cfqg, bool rt)
535 {
536         unsigned min_q, max_q;
537         unsigned mult  = cfq_hist_divisor - 1;
538         unsigned round = cfq_hist_divisor / 2;
539         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
540
541         min_q = min(cfqg->busy_queues_avg[rt], busy);
542         max_q = max(cfqg->busy_queues_avg[rt], busy);
543         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
544                 cfq_hist_divisor;
545         return cfqg->busy_queues_avg[rt];
546 }
547
548 static inline unsigned
549 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
550 {
551         struct cfq_rb_root *st = &cfqd->grp_service_tree;
552
553         return cfq_target_latency * cfqg->weight / st->total_weight;
554 }
555
556 static inline void
557 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
558 {
559         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
560         if (cfqd->cfq_latency) {
561                 /*
562                  * interested queues (we consider only the ones with the same
563                  * priority class in the cfq group)
564                  */
565                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
566                                                 cfq_class_rt(cfqq));
567                 unsigned sync_slice = cfqd->cfq_slice[1];
568                 unsigned expect_latency = sync_slice * iq;
569                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
570
571                 if (expect_latency > group_slice) {
572                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
573                         /* scale low_slice according to IO priority
574                          * and sync vs async */
575                         unsigned low_slice =
576                                 min(slice, base_low_slice * slice / sync_slice);
577                         /* the adapted slice value is scaled to fit all iqs
578                          * into the target latency */
579                         slice = max(slice * group_slice / expect_latency,
580                                     low_slice);
581                 }
582         }
583         cfqq->slice_start = jiffies;
584         cfqq->slice_end = jiffies + slice;
585         cfqq->allocated_slice = slice;
586         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
587 }
588
589 /*
590  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591  * isn't valid until the first request from the dispatch is activated
592  * and the slice time set.
593  */
594 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
595 {
596         if (cfq_cfqq_slice_new(cfqq))
597                 return 0;
598         if (time_before(jiffies, cfqq->slice_end))
599                 return 0;
600
601         return 1;
602 }
603
604 /*
605  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
606  * We choose the request that is closest to the head right now. Distance
607  * behind the head is penalized and only allowed to a certain extent.
608  */
609 static struct request *
610 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
611 {
612         sector_t s1, s2, d1 = 0, d2 = 0;
613         unsigned long back_max;
614 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
615 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
616         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
617
618         if (rq1 == NULL || rq1 == rq2)
619                 return rq2;
620         if (rq2 == NULL)
621                 return rq1;
622
623         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
624                 return rq1;
625         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
626                 return rq2;
627         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
628                 return rq1;
629         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
630                 return rq2;
631
632         s1 = blk_rq_pos(rq1);
633         s2 = blk_rq_pos(rq2);
634
635         /*
636          * by definition, 1KiB is 2 sectors
637          */
638         back_max = cfqd->cfq_back_max * 2;
639
640         /*
641          * Strict one way elevator _except_ in the case where we allow
642          * short backward seeks which are biased as twice the cost of a
643          * similar forward seek.
644          */
645         if (s1 >= last)
646                 d1 = s1 - last;
647         else if (s1 + back_max >= last)
648                 d1 = (last - s1) * cfqd->cfq_back_penalty;
649         else
650                 wrap |= CFQ_RQ1_WRAP;
651
652         if (s2 >= last)
653                 d2 = s2 - last;
654         else if (s2 + back_max >= last)
655                 d2 = (last - s2) * cfqd->cfq_back_penalty;
656         else
657                 wrap |= CFQ_RQ2_WRAP;
658
659         /* Found required data */
660
661         /*
662          * By doing switch() on the bit mask "wrap" we avoid having to
663          * check two variables for all permutations: --> faster!
664          */
665         switch (wrap) {
666         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
667                 if (d1 < d2)
668                         return rq1;
669                 else if (d2 < d1)
670                         return rq2;
671                 else {
672                         if (s1 >= s2)
673                                 return rq1;
674                         else
675                                 return rq2;
676                 }
677
678         case CFQ_RQ2_WRAP:
679                 return rq1;
680         case CFQ_RQ1_WRAP:
681                 return rq2;
682         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
683         default:
684                 /*
685                  * Since both rqs are wrapped,
686                  * start with the one that's further behind head
687                  * (--> only *one* back seek required),
688                  * since back seek takes more time than forward.
689                  */
690                 if (s1 <= s2)
691                         return rq1;
692                 else
693                         return rq2;
694         }
695 }
696
697 /*
698  * The below is leftmost cache rbtree addon
699  */
700 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
701 {
702         /* Service tree is empty */
703         if (!root->count)
704                 return NULL;
705
706         if (!root->left)
707                 root->left = rb_first(&root->rb);
708
709         if (root->left)
710                 return rb_entry(root->left, struct cfq_queue, rb_node);
711
712         return NULL;
713 }
714
715 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
716 {
717         if (!root->left)
718                 root->left = rb_first(&root->rb);
719
720         if (root->left)
721                 return rb_entry_cfqg(root->left);
722
723         return NULL;
724 }
725
726 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
727 {
728         rb_erase(n, root);
729         RB_CLEAR_NODE(n);
730 }
731
732 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
733 {
734         if (root->left == n)
735                 root->left = NULL;
736         rb_erase_init(n, &root->rb);
737         --root->count;
738 }
739
740 /*
741  * would be nice to take fifo expire time into account as well
742  */
743 static struct request *
744 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745                   struct request *last)
746 {
747         struct rb_node *rbnext = rb_next(&last->rb_node);
748         struct rb_node *rbprev = rb_prev(&last->rb_node);
749         struct request *next = NULL, *prev = NULL;
750
751         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
752
753         if (rbprev)
754                 prev = rb_entry_rq(rbprev);
755
756         if (rbnext)
757                 next = rb_entry_rq(rbnext);
758         else {
759                 rbnext = rb_first(&cfqq->sort_list);
760                 if (rbnext && rbnext != &last->rb_node)
761                         next = rb_entry_rq(rbnext);
762         }
763
764         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
765 }
766
767 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
768                                       struct cfq_queue *cfqq)
769 {
770         /*
771          * just an approximation, should be ok.
772          */
773         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
774                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
775 }
776
777 static inline s64
778 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
779 {
780         return cfqg->vdisktime - st->min_vdisktime;
781 }
782
783 static void
784 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
785 {
786         struct rb_node **node = &st->rb.rb_node;
787         struct rb_node *parent = NULL;
788         struct cfq_group *__cfqg;
789         s64 key = cfqg_key(st, cfqg);
790         int left = 1;
791
792         while (*node != NULL) {
793                 parent = *node;
794                 __cfqg = rb_entry_cfqg(parent);
795
796                 if (key < cfqg_key(st, __cfqg))
797                         node = &parent->rb_left;
798                 else {
799                         node = &parent->rb_right;
800                         left = 0;
801                 }
802         }
803
804         if (left)
805                 st->left = &cfqg->rb_node;
806
807         rb_link_node(&cfqg->rb_node, parent, node);
808         rb_insert_color(&cfqg->rb_node, &st->rb);
809 }
810
811 static void
812 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
813 {
814         struct cfq_rb_root *st = &cfqd->grp_service_tree;
815         struct cfq_group *__cfqg;
816         struct rb_node *n;
817
818         cfqg->nr_cfqq++;
819         if (cfqg->on_st)
820                 return;
821
822         /*
823          * Currently put the group at the end. Later implement something
824          * so that groups get lesser vtime based on their weights, so that
825          * if group does not loose all if it was not continously backlogged.
826          */
827         n = rb_last(&st->rb);
828         if (n) {
829                 __cfqg = rb_entry_cfqg(n);
830                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
831         } else
832                 cfqg->vdisktime = st->min_vdisktime;
833
834         __cfq_group_service_tree_add(st, cfqg);
835         cfqg->on_st = true;
836         st->total_weight += cfqg->weight;
837 }
838
839 static void
840 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
841 {
842         struct cfq_rb_root *st = &cfqd->grp_service_tree;
843
844         if (st->active == &cfqg->rb_node)
845                 st->active = NULL;
846
847         BUG_ON(cfqg->nr_cfqq < 1);
848         cfqg->nr_cfqq--;
849
850         /* If there are other cfq queues under this group, don't delete it */
851         if (cfqg->nr_cfqq)
852                 return;
853
854         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
855         cfqg->on_st = false;
856         st->total_weight -= cfqg->weight;
857         if (!RB_EMPTY_NODE(&cfqg->rb_node))
858                 cfq_rb_erase(&cfqg->rb_node, st);
859         cfqg->saved_workload_slice = 0;
860         blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
861 }
862
863 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
864 {
865         unsigned int slice_used;
866
867         /*
868          * Queue got expired before even a single request completed or
869          * got expired immediately after first request completion.
870          */
871         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
872                 /*
873                  * Also charge the seek time incurred to the group, otherwise
874                  * if there are mutiple queues in the group, each can dispatch
875                  * a single request on seeky media and cause lots of seek time
876                  * and group will never know it.
877                  */
878                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
879                                         1);
880         } else {
881                 slice_used = jiffies - cfqq->slice_start;
882                 if (slice_used > cfqq->allocated_slice)
883                         slice_used = cfqq->allocated_slice;
884         }
885
886         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
887         return slice_used;
888 }
889
890 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
891                                 struct cfq_queue *cfqq)
892 {
893         struct cfq_rb_root *st = &cfqd->grp_service_tree;
894         unsigned int used_sl, charge_sl;
895         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
896                         - cfqg->service_tree_idle.count;
897
898         BUG_ON(nr_sync < 0);
899         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
900
901         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
902                 charge_sl = cfqq->allocated_slice;
903
904         /* Can't update vdisktime while group is on service tree */
905         cfq_rb_erase(&cfqg->rb_node, st);
906         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
907         __cfq_group_service_tree_add(st, cfqg);
908
909         /* This group is being expired. Save the context */
910         if (time_after(cfqd->workload_expires, jiffies)) {
911                 cfqg->saved_workload_slice = cfqd->workload_expires
912                                                 - jiffies;
913                 cfqg->saved_workload = cfqd->serving_type;
914                 cfqg->saved_serving_prio = cfqd->serving_prio;
915         } else
916                 cfqg->saved_workload_slice = 0;
917
918         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
919                                         st->min_vdisktime);
920         blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
921         blkiocg_set_start_empty_time(&cfqg->blkg);
922 }
923
924 #ifdef CONFIG_CFQ_GROUP_IOSCHED
925 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
926 {
927         if (blkg)
928                 return container_of(blkg, struct cfq_group, blkg);
929         return NULL;
930 }
931
932 void
933 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
934 {
935         cfqg_of_blkg(blkg)->weight = weight;
936 }
937
938 static struct cfq_group *
939 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
940 {
941         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
942         struct cfq_group *cfqg = NULL;
943         void *key = cfqd;
944         int i, j;
945         struct cfq_rb_root *st;
946         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
947         unsigned int major, minor;
948
949         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
950         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
951                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
952                 cfqg->blkg.dev = MKDEV(major, minor);
953                 goto done;
954         }
955         if (cfqg || !create)
956                 goto done;
957
958         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
959         if (!cfqg)
960                 goto done;
961
962         for_each_cfqg_st(cfqg, i, j, st)
963                 *st = CFQ_RB_ROOT;
964         RB_CLEAR_NODE(&cfqg->rb_node);
965
966         /*
967          * Take the initial reference that will be released on destroy
968          * This can be thought of a joint reference by cgroup and
969          * elevator which will be dropped by either elevator exit
970          * or cgroup deletion path depending on who is exiting first.
971          */
972         atomic_set(&cfqg->ref, 1);
973
974         /* Add group onto cgroup list */
975         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
976         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
977                                         MKDEV(major, minor));
978         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
979
980         /* Add group on cfqd list */
981         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
982
983 done:
984         return cfqg;
985 }
986
987 /*
988  * Search for the cfq group current task belongs to. If create = 1, then also
989  * create the cfq group if it does not exist. request_queue lock must be held.
990  */
991 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
992 {
993         struct cgroup *cgroup;
994         struct cfq_group *cfqg = NULL;
995
996         rcu_read_lock();
997         cgroup = task_cgroup(current, blkio_subsys_id);
998         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
999         if (!cfqg && create)
1000                 cfqg = &cfqd->root_group;
1001         rcu_read_unlock();
1002         return cfqg;
1003 }
1004
1005 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1006 {
1007         atomic_inc(&cfqg->ref);
1008         return cfqg;
1009 }
1010
1011 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1012 {
1013         /* Currently, all async queues are mapped to root group */
1014         if (!cfq_cfqq_sync(cfqq))
1015                 cfqg = &cfqq->cfqd->root_group;
1016
1017         cfqq->cfqg = cfqg;
1018         /* cfqq reference on cfqg */
1019         atomic_inc(&cfqq->cfqg->ref);
1020 }
1021
1022 static void cfq_put_cfqg(struct cfq_group *cfqg)
1023 {
1024         struct cfq_rb_root *st;
1025         int i, j;
1026
1027         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1028         if (!atomic_dec_and_test(&cfqg->ref))
1029                 return;
1030         for_each_cfqg_st(cfqg, i, j, st)
1031                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1032         kfree(cfqg);
1033 }
1034
1035 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1036 {
1037         /* Something wrong if we are trying to remove same group twice */
1038         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1039
1040         hlist_del_init(&cfqg->cfqd_node);
1041
1042         /*
1043          * Put the reference taken at the time of creation so that when all
1044          * queues are gone, group can be destroyed.
1045          */
1046         cfq_put_cfqg(cfqg);
1047 }
1048
1049 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1050 {
1051         struct hlist_node *pos, *n;
1052         struct cfq_group *cfqg;
1053
1054         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1055                 /*
1056                  * If cgroup removal path got to blk_group first and removed
1057                  * it from cgroup list, then it will take care of destroying
1058                  * cfqg also.
1059                  */
1060                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1061                         cfq_destroy_cfqg(cfqd, cfqg);
1062         }
1063 }
1064
1065 /*
1066  * Blk cgroup controller notification saying that blkio_group object is being
1067  * delinked as associated cgroup object is going away. That also means that
1068  * no new IO will come in this group. So get rid of this group as soon as
1069  * any pending IO in the group is finished.
1070  *
1071  * This function is called under rcu_read_lock(). key is the rcu protected
1072  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1073  * read lock.
1074  *
1075  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1076  * it should not be NULL as even if elevator was exiting, cgroup deltion
1077  * path got to it first.
1078  */
1079 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1080 {
1081         unsigned long  flags;
1082         struct cfq_data *cfqd = key;
1083
1084         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1085         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1086         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1087 }
1088
1089 #else /* GROUP_IOSCHED */
1090 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1091 {
1092         return &cfqd->root_group;
1093 }
1094
1095 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1096 {
1097         return NULL;
1098 }
1099
1100 static inline void
1101 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1102         cfqq->cfqg = cfqg;
1103 }
1104
1105 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1106 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1107
1108 #endif /* GROUP_IOSCHED */
1109
1110 /*
1111  * The cfqd->service_trees holds all pending cfq_queue's that have
1112  * requests waiting to be processed. It is sorted in the order that
1113  * we will service the queues.
1114  */
1115 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1116                                  bool add_front)
1117 {
1118         struct rb_node **p, *parent;
1119         struct cfq_queue *__cfqq;
1120         unsigned long rb_key;
1121         struct cfq_rb_root *service_tree;
1122         int left;
1123         int new_cfqq = 1;
1124         int group_changed = 0;
1125
1126 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1127         if (!cfqd->cfq_group_isolation
1128             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1129             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1130                 /* Move this cfq to root group */
1131                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1132                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1133                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1134                 cfqq->orig_cfqg = cfqq->cfqg;
1135                 cfqq->cfqg = &cfqd->root_group;
1136                 atomic_inc(&cfqd->root_group.ref);
1137                 group_changed = 1;
1138         } else if (!cfqd->cfq_group_isolation
1139                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1140                 /* cfqq is sequential now needs to go to its original group */
1141                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1142                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1143                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1144                 cfq_put_cfqg(cfqq->cfqg);
1145                 cfqq->cfqg = cfqq->orig_cfqg;
1146                 cfqq->orig_cfqg = NULL;
1147                 group_changed = 1;
1148                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1149         }
1150 #endif
1151
1152         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1153                                                 cfqq_type(cfqq));
1154         if (cfq_class_idle(cfqq)) {
1155                 rb_key = CFQ_IDLE_DELAY;
1156                 parent = rb_last(&service_tree->rb);
1157                 if (parent && parent != &cfqq->rb_node) {
1158                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1159                         rb_key += __cfqq->rb_key;
1160                 } else
1161                         rb_key += jiffies;
1162         } else if (!add_front) {
1163                 /*
1164                  * Get our rb key offset. Subtract any residual slice
1165                  * value carried from last service. A negative resid
1166                  * count indicates slice overrun, and this should position
1167                  * the next service time further away in the tree.
1168                  */
1169                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1170                 rb_key -= cfqq->slice_resid;
1171                 cfqq->slice_resid = 0;
1172         } else {
1173                 rb_key = -HZ;
1174                 __cfqq = cfq_rb_first(service_tree);
1175                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1176         }
1177
1178         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1179                 new_cfqq = 0;
1180                 /*
1181                  * same position, nothing more to do
1182                  */
1183                 if (rb_key == cfqq->rb_key &&
1184                     cfqq->service_tree == service_tree)
1185                         return;
1186
1187                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1188                 cfqq->service_tree = NULL;
1189         }
1190
1191         left = 1;
1192         parent = NULL;
1193         cfqq->service_tree = service_tree;
1194         p = &service_tree->rb.rb_node;
1195         while (*p) {
1196                 struct rb_node **n;
1197
1198                 parent = *p;
1199                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1200
1201                 /*
1202                  * sort by key, that represents service time.
1203                  */
1204                 if (time_before(rb_key, __cfqq->rb_key))
1205                         n = &(*p)->rb_left;
1206                 else {
1207                         n = &(*p)->rb_right;
1208                         left = 0;
1209                 }
1210
1211                 p = n;
1212         }
1213
1214         if (left)
1215                 service_tree->left = &cfqq->rb_node;
1216
1217         cfqq->rb_key = rb_key;
1218         rb_link_node(&cfqq->rb_node, parent, p);
1219         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1220         service_tree->count++;
1221         if ((add_front || !new_cfqq) && !group_changed)
1222                 return;
1223         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1224 }
1225
1226 static struct cfq_queue *
1227 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1228                      sector_t sector, struct rb_node **ret_parent,
1229                      struct rb_node ***rb_link)
1230 {
1231         struct rb_node **p, *parent;
1232         struct cfq_queue *cfqq = NULL;
1233
1234         parent = NULL;
1235         p = &root->rb_node;
1236         while (*p) {
1237                 struct rb_node **n;
1238
1239                 parent = *p;
1240                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1241
1242                 /*
1243                  * Sort strictly based on sector.  Smallest to the left,
1244                  * largest to the right.
1245                  */
1246                 if (sector > blk_rq_pos(cfqq->next_rq))
1247                         n = &(*p)->rb_right;
1248                 else if (sector < blk_rq_pos(cfqq->next_rq))
1249                         n = &(*p)->rb_left;
1250                 else
1251                         break;
1252                 p = n;
1253                 cfqq = NULL;
1254         }
1255
1256         *ret_parent = parent;
1257         if (rb_link)
1258                 *rb_link = p;
1259         return cfqq;
1260 }
1261
1262 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1263 {
1264         struct rb_node **p, *parent;
1265         struct cfq_queue *__cfqq;
1266
1267         if (cfqq->p_root) {
1268                 rb_erase(&cfqq->p_node, cfqq->p_root);
1269                 cfqq->p_root = NULL;
1270         }
1271
1272         if (cfq_class_idle(cfqq))
1273                 return;
1274         if (!cfqq->next_rq)
1275                 return;
1276
1277         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1278         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1279                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1280         if (!__cfqq) {
1281                 rb_link_node(&cfqq->p_node, parent, p);
1282                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1283         } else
1284                 cfqq->p_root = NULL;
1285 }
1286
1287 /*
1288  * Update cfqq's position in the service tree.
1289  */
1290 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1291 {
1292         /*
1293          * Resorting requires the cfqq to be on the RR list already.
1294          */
1295         if (cfq_cfqq_on_rr(cfqq)) {
1296                 cfq_service_tree_add(cfqd, cfqq, 0);
1297                 cfq_prio_tree_add(cfqd, cfqq);
1298         }
1299 }
1300
1301 /*
1302  * add to busy list of queues for service, trying to be fair in ordering
1303  * the pending list according to last request service
1304  */
1305 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1306 {
1307         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1308         BUG_ON(cfq_cfqq_on_rr(cfqq));
1309         cfq_mark_cfqq_on_rr(cfqq);
1310         cfqd->busy_queues++;
1311
1312         cfq_resort_rr_list(cfqd, cfqq);
1313 }
1314
1315 /*
1316  * Called when the cfqq no longer has requests pending, remove it from
1317  * the service tree.
1318  */
1319 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1320 {
1321         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1322         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1323         cfq_clear_cfqq_on_rr(cfqq);
1324
1325         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1326                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1327                 cfqq->service_tree = NULL;
1328         }
1329         if (cfqq->p_root) {
1330                 rb_erase(&cfqq->p_node, cfqq->p_root);
1331                 cfqq->p_root = NULL;
1332         }
1333
1334         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1335         BUG_ON(!cfqd->busy_queues);
1336         cfqd->busy_queues--;
1337 }
1338
1339 /*
1340  * rb tree support functions
1341  */
1342 static void cfq_del_rq_rb(struct request *rq)
1343 {
1344         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1345         const int sync = rq_is_sync(rq);
1346
1347         BUG_ON(!cfqq->queued[sync]);
1348         cfqq->queued[sync]--;
1349
1350         elv_rb_del(&cfqq->sort_list, rq);
1351
1352         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1353                 /*
1354                  * Queue will be deleted from service tree when we actually
1355                  * expire it later. Right now just remove it from prio tree
1356                  * as it is empty.
1357                  */
1358                 if (cfqq->p_root) {
1359                         rb_erase(&cfqq->p_node, cfqq->p_root);
1360                         cfqq->p_root = NULL;
1361                 }
1362         }
1363 }
1364
1365 static void cfq_add_rq_rb(struct request *rq)
1366 {
1367         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1368         struct cfq_data *cfqd = cfqq->cfqd;
1369         struct request *__alias, *prev;
1370
1371         cfqq->queued[rq_is_sync(rq)]++;
1372
1373         /*
1374          * looks a little odd, but the first insert might return an alias.
1375          * if that happens, put the alias on the dispatch list
1376          */
1377         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1378                 cfq_dispatch_insert(cfqd->queue, __alias);
1379
1380         if (!cfq_cfqq_on_rr(cfqq))
1381                 cfq_add_cfqq_rr(cfqd, cfqq);
1382
1383         /*
1384          * check if this request is a better next-serve candidate
1385          */
1386         prev = cfqq->next_rq;
1387         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1388
1389         /*
1390          * adjust priority tree position, if ->next_rq changes
1391          */
1392         if (prev != cfqq->next_rq)
1393                 cfq_prio_tree_add(cfqd, cfqq);
1394
1395         BUG_ON(!cfqq->next_rq);
1396 }
1397
1398 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1399 {
1400         elv_rb_del(&cfqq->sort_list, rq);
1401         cfqq->queued[rq_is_sync(rq)]--;
1402         blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1403                                                 rq_is_sync(rq));
1404         cfq_add_rq_rb(rq);
1405         blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1406                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1407                         rq_is_sync(rq));
1408 }
1409
1410 static struct request *
1411 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1412 {
1413         struct task_struct *tsk = current;
1414         struct cfq_io_context *cic;
1415         struct cfq_queue *cfqq;
1416
1417         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1418         if (!cic)
1419                 return NULL;
1420
1421         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1422         if (cfqq) {
1423                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1424
1425                 return elv_rb_find(&cfqq->sort_list, sector);
1426         }
1427
1428         return NULL;
1429 }
1430
1431 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1432 {
1433         struct cfq_data *cfqd = q->elevator->elevator_data;
1434
1435         cfqd->rq_in_driver++;
1436         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1437                                                 cfqd->rq_in_driver);
1438
1439         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1440 }
1441
1442 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1443 {
1444         struct cfq_data *cfqd = q->elevator->elevator_data;
1445
1446         WARN_ON(!cfqd->rq_in_driver);
1447         cfqd->rq_in_driver--;
1448         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1449                                                 cfqd->rq_in_driver);
1450 }
1451
1452 static void cfq_remove_request(struct request *rq)
1453 {
1454         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1455
1456         if (cfqq->next_rq == rq)
1457                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1458
1459         list_del_init(&rq->queuelist);
1460         cfq_del_rq_rb(rq);
1461
1462         cfqq->cfqd->rq_queued--;
1463         blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1464                                                 rq_is_sync(rq));
1465         if (rq_is_meta(rq)) {
1466                 WARN_ON(!cfqq->meta_pending);
1467                 cfqq->meta_pending--;
1468         }
1469 }
1470
1471 static int cfq_merge(struct request_queue *q, struct request **req,
1472                      struct bio *bio)
1473 {
1474         struct cfq_data *cfqd = q->elevator->elevator_data;
1475         struct request *__rq;
1476
1477         __rq = cfq_find_rq_fmerge(cfqd, bio);
1478         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1479                 *req = __rq;
1480                 return ELEVATOR_FRONT_MERGE;
1481         }
1482
1483         return ELEVATOR_NO_MERGE;
1484 }
1485
1486 static void cfq_merged_request(struct request_queue *q, struct request *req,
1487                                int type)
1488 {
1489         if (type == ELEVATOR_FRONT_MERGE) {
1490                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1491
1492                 cfq_reposition_rq_rb(cfqq, req);
1493         }
1494 }
1495
1496 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1497                                 struct bio *bio)
1498 {
1499         blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, bio_data_dir(bio),
1500                                         cfq_bio_sync(bio));
1501 }
1502
1503 static void
1504 cfq_merged_requests(struct request_queue *q, struct request *rq,
1505                     struct request *next)
1506 {
1507         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1508         /*
1509          * reposition in fifo if next is older than rq
1510          */
1511         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1512             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1513                 list_move(&rq->queuelist, &next->queuelist);
1514                 rq_set_fifo_time(rq, rq_fifo_time(next));
1515         }
1516
1517         if (cfqq->next_rq == next)
1518                 cfqq->next_rq = rq;
1519         cfq_remove_request(next);
1520         blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(next),
1521                                         rq_is_sync(next));
1522 }
1523
1524 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1525                            struct bio *bio)
1526 {
1527         struct cfq_data *cfqd = q->elevator->elevator_data;
1528         struct cfq_io_context *cic;
1529         struct cfq_queue *cfqq;
1530
1531         /*
1532          * Disallow merge of a sync bio into an async request.
1533          */
1534         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1535                 return false;
1536
1537         /*
1538          * Lookup the cfqq that this bio will be queued with. Allow
1539          * merge only if rq is queued there.
1540          */
1541         cic = cfq_cic_lookup(cfqd, current->io_context);
1542         if (!cic)
1543                 return false;
1544
1545         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1546         return cfqq == RQ_CFQQ(rq);
1547 }
1548
1549 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1550 {
1551         del_timer(&cfqd->idle_slice_timer);
1552         blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1553 }
1554
1555 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1556                                    struct cfq_queue *cfqq)
1557 {
1558         if (cfqq) {
1559                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1560                                 cfqd->serving_prio, cfqd->serving_type);
1561                 blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1562                 cfqq->slice_start = 0;
1563                 cfqq->dispatch_start = jiffies;
1564                 cfqq->allocated_slice = 0;
1565                 cfqq->slice_end = 0;
1566                 cfqq->slice_dispatch = 0;
1567
1568                 cfq_clear_cfqq_wait_request(cfqq);
1569                 cfq_clear_cfqq_must_dispatch(cfqq);
1570                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1571                 cfq_clear_cfqq_fifo_expire(cfqq);
1572                 cfq_mark_cfqq_slice_new(cfqq);
1573
1574                 cfq_del_timer(cfqd, cfqq);
1575         }
1576
1577         cfqd->active_queue = cfqq;
1578 }
1579
1580 /*
1581  * current cfqq expired its slice (or was too idle), select new one
1582  */
1583 static void
1584 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1585                     bool timed_out)
1586 {
1587         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1588
1589         if (cfq_cfqq_wait_request(cfqq))
1590                 cfq_del_timer(cfqd, cfqq);
1591
1592         cfq_clear_cfqq_wait_request(cfqq);
1593         cfq_clear_cfqq_wait_busy(cfqq);
1594
1595         /*
1596          * If this cfqq is shared between multiple processes, check to
1597          * make sure that those processes are still issuing I/Os within
1598          * the mean seek distance.  If not, it may be time to break the
1599          * queues apart again.
1600          */
1601         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1602                 cfq_mark_cfqq_split_coop(cfqq);
1603
1604         /*
1605          * store what was left of this slice, if the queue idled/timed out
1606          */
1607         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1608                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1609                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1610         }
1611
1612         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1613
1614         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1615                 cfq_del_cfqq_rr(cfqd, cfqq);
1616
1617         cfq_resort_rr_list(cfqd, cfqq);
1618
1619         if (cfqq == cfqd->active_queue)
1620                 cfqd->active_queue = NULL;
1621
1622         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1623                 cfqd->grp_service_tree.active = NULL;
1624
1625         if (cfqd->active_cic) {
1626                 put_io_context(cfqd->active_cic->ioc);
1627                 cfqd->active_cic = NULL;
1628         }
1629 }
1630
1631 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1632 {
1633         struct cfq_queue *cfqq = cfqd->active_queue;
1634
1635         if (cfqq)
1636                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1637 }
1638
1639 /*
1640  * Get next queue for service. Unless we have a queue preemption,
1641  * we'll simply select the first cfqq in the service tree.
1642  */
1643 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1644 {
1645         struct cfq_rb_root *service_tree =
1646                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1647                                         cfqd->serving_type);
1648
1649         if (!cfqd->rq_queued)
1650                 return NULL;
1651
1652         /* There is nothing to dispatch */
1653         if (!service_tree)
1654                 return NULL;
1655         if (RB_EMPTY_ROOT(&service_tree->rb))
1656                 return NULL;
1657         return cfq_rb_first(service_tree);
1658 }
1659
1660 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1661 {
1662         struct cfq_group *cfqg;
1663         struct cfq_queue *cfqq;
1664         int i, j;
1665         struct cfq_rb_root *st;
1666
1667         if (!cfqd->rq_queued)
1668                 return NULL;
1669
1670         cfqg = cfq_get_next_cfqg(cfqd);
1671         if (!cfqg)
1672                 return NULL;
1673
1674         for_each_cfqg_st(cfqg, i, j, st)
1675                 if ((cfqq = cfq_rb_first(st)) != NULL)
1676                         return cfqq;
1677         return NULL;
1678 }
1679
1680 /*
1681  * Get and set a new active queue for service.
1682  */
1683 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1684                                               struct cfq_queue *cfqq)
1685 {
1686         if (!cfqq)
1687                 cfqq = cfq_get_next_queue(cfqd);
1688
1689         __cfq_set_active_queue(cfqd, cfqq);
1690         return cfqq;
1691 }
1692
1693 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1694                                           struct request *rq)
1695 {
1696         if (blk_rq_pos(rq) >= cfqd->last_position)
1697                 return blk_rq_pos(rq) - cfqd->last_position;
1698         else
1699                 return cfqd->last_position - blk_rq_pos(rq);
1700 }
1701
1702 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1703                                struct request *rq)
1704 {
1705         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1706 }
1707
1708 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1709                                     struct cfq_queue *cur_cfqq)
1710 {
1711         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1712         struct rb_node *parent, *node;
1713         struct cfq_queue *__cfqq;
1714         sector_t sector = cfqd->last_position;
1715
1716         if (RB_EMPTY_ROOT(root))
1717                 return NULL;
1718
1719         /*
1720          * First, if we find a request starting at the end of the last
1721          * request, choose it.
1722          */
1723         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1724         if (__cfqq)
1725                 return __cfqq;
1726
1727         /*
1728          * If the exact sector wasn't found, the parent of the NULL leaf
1729          * will contain the closest sector.
1730          */
1731         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1732         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1733                 return __cfqq;
1734
1735         if (blk_rq_pos(__cfqq->next_rq) < sector)
1736                 node = rb_next(&__cfqq->p_node);
1737         else
1738                 node = rb_prev(&__cfqq->p_node);
1739         if (!node)
1740                 return NULL;
1741
1742         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1743         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1744                 return __cfqq;
1745
1746         return NULL;
1747 }
1748
1749 /*
1750  * cfqd - obvious
1751  * cur_cfqq - passed in so that we don't decide that the current queue is
1752  *            closely cooperating with itself.
1753  *
1754  * So, basically we're assuming that that cur_cfqq has dispatched at least
1755  * one request, and that cfqd->last_position reflects a position on the disk
1756  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1757  * assumption.
1758  */
1759 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1760                                               struct cfq_queue *cur_cfqq)
1761 {
1762         struct cfq_queue *cfqq;
1763
1764         if (cfq_class_idle(cur_cfqq))
1765                 return NULL;
1766         if (!cfq_cfqq_sync(cur_cfqq))
1767                 return NULL;
1768         if (CFQQ_SEEKY(cur_cfqq))
1769                 return NULL;
1770
1771         /*
1772          * Don't search priority tree if it's the only queue in the group.
1773          */
1774         if (cur_cfqq->cfqg->nr_cfqq == 1)
1775                 return NULL;
1776
1777         /*
1778          * We should notice if some of the queues are cooperating, eg
1779          * working closely on the same area of the disk. In that case,
1780          * we can group them together and don't waste time idling.
1781          */
1782         cfqq = cfqq_close(cfqd, cur_cfqq);
1783         if (!cfqq)
1784                 return NULL;
1785
1786         /* If new queue belongs to different cfq_group, don't choose it */
1787         if (cur_cfqq->cfqg != cfqq->cfqg)
1788                 return NULL;
1789
1790         /*
1791          * It only makes sense to merge sync queues.
1792          */
1793         if (!cfq_cfqq_sync(cfqq))
1794                 return NULL;
1795         if (CFQQ_SEEKY(cfqq))
1796                 return NULL;
1797
1798         /*
1799          * Do not merge queues of different priority classes
1800          */
1801         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1802                 return NULL;
1803
1804         return cfqq;
1805 }
1806
1807 /*
1808  * Determine whether we should enforce idle window for this queue.
1809  */
1810
1811 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1812 {
1813         enum wl_prio_t prio = cfqq_prio(cfqq);
1814         struct cfq_rb_root *service_tree = cfqq->service_tree;
1815
1816         BUG_ON(!service_tree);
1817         BUG_ON(!service_tree->count);
1818
1819         /* We never do for idle class queues. */
1820         if (prio == IDLE_WORKLOAD)
1821                 return false;
1822
1823         /* We do for queues that were marked with idle window flag. */
1824         if (cfq_cfqq_idle_window(cfqq) &&
1825            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1826                 return true;
1827
1828         /*
1829          * Otherwise, we do only if they are the last ones
1830          * in their service tree.
1831          */
1832         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1833                 return 1;
1834         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1835                         service_tree->count);
1836         return 0;
1837 }
1838
1839 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1840 {
1841         struct cfq_queue *cfqq = cfqd->active_queue;
1842         struct cfq_io_context *cic;
1843         unsigned long sl;
1844
1845         /*
1846          * SSD device without seek penalty, disable idling. But only do so
1847          * for devices that support queuing, otherwise we still have a problem
1848          * with sync vs async workloads.
1849          */
1850         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1851                 return;
1852
1853         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1854         WARN_ON(cfq_cfqq_slice_new(cfqq));
1855
1856         /*
1857          * idle is disabled, either manually or by past process history
1858          */
1859         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1860                 return;
1861
1862         /*
1863          * still active requests from this queue, don't idle
1864          */
1865         if (cfqq->dispatched)
1866                 return;
1867
1868         /*
1869          * task has exited, don't wait
1870          */
1871         cic = cfqd->active_cic;
1872         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1873                 return;
1874
1875         /*
1876          * If our average think time is larger than the remaining time
1877          * slice, then don't idle. This avoids overrunning the allotted
1878          * time slice.
1879          */
1880         if (sample_valid(cic->ttime_samples) &&
1881             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1882                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1883                                 cic->ttime_mean);
1884                 return;
1885         }
1886
1887         cfq_mark_cfqq_wait_request(cfqq);
1888
1889         sl = cfqd->cfq_slice_idle;
1890
1891         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1892         blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1893         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1894 }
1895
1896 /*
1897  * Move request from internal lists to the request queue dispatch list.
1898  */
1899 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1900 {
1901         struct cfq_data *cfqd = q->elevator->elevator_data;
1902         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1903
1904         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1905
1906         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1907         cfq_remove_request(rq);
1908         cfqq->dispatched++;
1909         elv_dispatch_sort(q, rq);
1910
1911         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1912         blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1913                                         rq_data_dir(rq), rq_is_sync(rq));
1914 }
1915
1916 /*
1917  * return expired entry, or NULL to just start from scratch in rbtree
1918  */
1919 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1920 {
1921         struct request *rq = NULL;
1922
1923         if (cfq_cfqq_fifo_expire(cfqq))
1924                 return NULL;
1925
1926         cfq_mark_cfqq_fifo_expire(cfqq);
1927
1928         if (list_empty(&cfqq->fifo))
1929                 return NULL;
1930
1931         rq = rq_entry_fifo(cfqq->fifo.next);
1932         if (time_before(jiffies, rq_fifo_time(rq)))
1933                 rq = NULL;
1934
1935         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1936         return rq;
1937 }
1938
1939 static inline int
1940 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1941 {
1942         const int base_rq = cfqd->cfq_slice_async_rq;
1943
1944         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1945
1946         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1947 }
1948
1949 /*
1950  * Must be called with the queue_lock held.
1951  */
1952 static int cfqq_process_refs(struct cfq_queue *cfqq)
1953 {
1954         int process_refs, io_refs;
1955
1956         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1957         process_refs = atomic_read(&cfqq->ref) - io_refs;
1958         BUG_ON(process_refs < 0);
1959         return process_refs;
1960 }
1961
1962 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1963 {
1964         int process_refs, new_process_refs;
1965         struct cfq_queue *__cfqq;
1966
1967         /* Avoid a circular list and skip interim queue merges */
1968         while ((__cfqq = new_cfqq->new_cfqq)) {
1969                 if (__cfqq == cfqq)
1970                         return;
1971                 new_cfqq = __cfqq;
1972         }
1973
1974         process_refs = cfqq_process_refs(cfqq);
1975         /*
1976          * If the process for the cfqq has gone away, there is no
1977          * sense in merging the queues.
1978          */
1979         if (process_refs == 0)
1980                 return;
1981
1982         /*
1983          * Merge in the direction of the lesser amount of work.
1984          */
1985         new_process_refs = cfqq_process_refs(new_cfqq);
1986         if (new_process_refs >= process_refs) {
1987                 cfqq->new_cfqq = new_cfqq;
1988                 atomic_add(process_refs, &new_cfqq->ref);
1989         } else {
1990                 new_cfqq->new_cfqq = cfqq;
1991                 atomic_add(new_process_refs, &cfqq->ref);
1992         }
1993 }
1994
1995 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1996                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1997 {
1998         struct cfq_queue *queue;
1999         int i;
2000         bool key_valid = false;
2001         unsigned long lowest_key = 0;
2002         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2003
2004         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2005                 /* select the one with lowest rb_key */
2006                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2007                 if (queue &&
2008                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2009                         lowest_key = queue->rb_key;
2010                         cur_best = i;
2011                         key_valid = true;
2012                 }
2013         }
2014
2015         return cur_best;
2016 }
2017
2018 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2019 {
2020         unsigned slice;
2021         unsigned count;
2022         struct cfq_rb_root *st;
2023         unsigned group_slice;
2024
2025         if (!cfqg) {
2026                 cfqd->serving_prio = IDLE_WORKLOAD;
2027                 cfqd->workload_expires = jiffies + 1;
2028                 return;
2029         }
2030
2031         /* Choose next priority. RT > BE > IDLE */
2032         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2033                 cfqd->serving_prio = RT_WORKLOAD;
2034         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2035                 cfqd->serving_prio = BE_WORKLOAD;
2036         else {
2037                 cfqd->serving_prio = IDLE_WORKLOAD;
2038                 cfqd->workload_expires = jiffies + 1;
2039                 return;
2040         }
2041
2042         /*
2043          * For RT and BE, we have to choose also the type
2044          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2045          * expiration time
2046          */
2047         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2048         count = st->count;
2049
2050         /*
2051          * check workload expiration, and that we still have other queues ready
2052          */
2053         if (count && !time_after(jiffies, cfqd->workload_expires))
2054                 return;
2055
2056         /* otherwise select new workload type */
2057         cfqd->serving_type =
2058                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2059         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2060         count = st->count;
2061
2062         /*
2063          * the workload slice is computed as a fraction of target latency
2064          * proportional to the number of queues in that workload, over
2065          * all the queues in the same priority class
2066          */
2067         group_slice = cfq_group_slice(cfqd, cfqg);
2068
2069         slice = group_slice * count /
2070                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2071                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2072
2073         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2074                 unsigned int tmp;
2075
2076                 /*
2077                  * Async queues are currently system wide. Just taking
2078                  * proportion of queues with-in same group will lead to higher
2079                  * async ratio system wide as generally root group is going
2080                  * to have higher weight. A more accurate thing would be to
2081                  * calculate system wide asnc/sync ratio.
2082                  */
2083                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2084                 tmp = tmp/cfqd->busy_queues;
2085                 slice = min_t(unsigned, slice, tmp);
2086
2087                 /* async workload slice is scaled down according to
2088                  * the sync/async slice ratio. */
2089                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2090         } else
2091                 /* sync workload slice is at least 2 * cfq_slice_idle */
2092                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2093
2094         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2095         cfq_log(cfqd, "workload slice:%d", slice);
2096         cfqd->workload_expires = jiffies + slice;
2097         cfqd->noidle_tree_requires_idle = false;
2098 }
2099
2100 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2101 {
2102         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2103         struct cfq_group *cfqg;
2104
2105         if (RB_EMPTY_ROOT(&st->rb))
2106                 return NULL;
2107         cfqg = cfq_rb_first_group(st);
2108         st->active = &cfqg->rb_node;
2109         update_min_vdisktime(st);
2110         return cfqg;
2111 }
2112
2113 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2114 {
2115         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2116
2117         cfqd->serving_group = cfqg;
2118
2119         /* Restore the workload type data */
2120         if (cfqg->saved_workload_slice) {
2121                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2122                 cfqd->serving_type = cfqg->saved_workload;
2123                 cfqd->serving_prio = cfqg->saved_serving_prio;
2124         } else
2125                 cfqd->workload_expires = jiffies - 1;
2126
2127         choose_service_tree(cfqd, cfqg);
2128 }
2129
2130 /*
2131  * Select a queue for service. If we have a current active queue,
2132  * check whether to continue servicing it, or retrieve and set a new one.
2133  */
2134 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2135 {
2136         struct cfq_queue *cfqq, *new_cfqq = NULL;
2137
2138         cfqq = cfqd->active_queue;
2139         if (!cfqq)
2140                 goto new_queue;
2141
2142         if (!cfqd->rq_queued)
2143                 return NULL;
2144
2145         /*
2146          * We were waiting for group to get backlogged. Expire the queue
2147          */
2148         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2149                 goto expire;
2150
2151         /*
2152          * The active queue has run out of time, expire it and select new.
2153          */
2154         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2155                 /*
2156                  * If slice had not expired at the completion of last request
2157                  * we might not have turned on wait_busy flag. Don't expire
2158                  * the queue yet. Allow the group to get backlogged.
2159                  *
2160                  * The very fact that we have used the slice, that means we
2161                  * have been idling all along on this queue and it should be
2162                  * ok to wait for this request to complete.
2163                  */
2164                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2165                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2166                         cfqq = NULL;
2167                         goto keep_queue;
2168                 } else
2169                         goto expire;
2170         }
2171
2172         /*
2173          * The active queue has requests and isn't expired, allow it to
2174          * dispatch.
2175          */
2176         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2177                 goto keep_queue;
2178
2179         /*
2180          * If another queue has a request waiting within our mean seek
2181          * distance, let it run.  The expire code will check for close
2182          * cooperators and put the close queue at the front of the service
2183          * tree.  If possible, merge the expiring queue with the new cfqq.
2184          */
2185         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2186         if (new_cfqq) {
2187                 if (!cfqq->new_cfqq)
2188                         cfq_setup_merge(cfqq, new_cfqq);
2189                 goto expire;
2190         }
2191
2192         /*
2193          * No requests pending. If the active queue still has requests in
2194          * flight or is idling for a new request, allow either of these
2195          * conditions to happen (or time out) before selecting a new queue.
2196          */
2197         if (timer_pending(&cfqd->idle_slice_timer) ||
2198             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2199                 cfqq = NULL;
2200                 goto keep_queue;
2201         }
2202
2203 expire:
2204         cfq_slice_expired(cfqd, 0);
2205 new_queue:
2206         /*
2207          * Current queue expired. Check if we have to switch to a new
2208          * service tree
2209          */
2210         if (!new_cfqq)
2211                 cfq_choose_cfqg(cfqd);
2212
2213         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2214 keep_queue:
2215         return cfqq;
2216 }
2217
2218 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2219 {
2220         int dispatched = 0;
2221
2222         while (cfqq->next_rq) {
2223                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2224                 dispatched++;
2225         }
2226
2227         BUG_ON(!list_empty(&cfqq->fifo));
2228
2229         /* By default cfqq is not expired if it is empty. Do it explicitly */
2230         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2231         return dispatched;
2232 }
2233
2234 /*
2235  * Drain our current requests. Used for barriers and when switching
2236  * io schedulers on-the-fly.
2237  */
2238 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2239 {
2240         struct cfq_queue *cfqq;
2241         int dispatched = 0;
2242
2243         /* Expire the timeslice of the current active queue first */
2244         cfq_slice_expired(cfqd, 0);
2245         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2246                 __cfq_set_active_queue(cfqd, cfqq);
2247                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2248         }
2249
2250         BUG_ON(cfqd->busy_queues);
2251
2252         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2253         return dispatched;
2254 }
2255
2256 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2257         struct cfq_queue *cfqq)
2258 {
2259         /* the queue hasn't finished any request, can't estimate */
2260         if (cfq_cfqq_slice_new(cfqq))
2261                 return 1;
2262         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2263                 cfqq->slice_end))
2264                 return 1;
2265
2266         return 0;
2267 }
2268
2269 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2270 {
2271         unsigned int max_dispatch;
2272
2273         /*
2274          * Drain async requests before we start sync IO
2275          */
2276         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2277                 return false;
2278
2279         /*
2280          * If this is an async queue and we have sync IO in flight, let it wait
2281          */
2282         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2283                 return false;
2284
2285         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2286         if (cfq_class_idle(cfqq))
2287                 max_dispatch = 1;
2288
2289         /*
2290          * Does this cfqq already have too much IO in flight?
2291          */
2292         if (cfqq->dispatched >= max_dispatch) {
2293                 /*
2294                  * idle queue must always only have a single IO in flight
2295                  */
2296                 if (cfq_class_idle(cfqq))
2297                         return false;
2298
2299                 /*
2300                  * We have other queues, don't allow more IO from this one
2301                  */
2302                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2303                         return false;
2304
2305                 /*
2306                  * Sole queue user, no limit
2307                  */
2308                 if (cfqd->busy_queues == 1)
2309                         max_dispatch = -1;
2310                 else
2311                         /*
2312                          * Normally we start throttling cfqq when cfq_quantum/2
2313                          * requests have been dispatched. But we can drive
2314                          * deeper queue depths at the beginning of slice
2315                          * subjected to upper limit of cfq_quantum.
2316                          * */
2317                         max_dispatch = cfqd->cfq_quantum;
2318         }
2319
2320         /*
2321          * Async queues must wait a bit before being allowed dispatch.
2322          * We also ramp up the dispatch depth gradually for async IO,
2323          * based on the last sync IO we serviced
2324          */
2325         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2326                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2327                 unsigned int depth;
2328
2329                 depth = last_sync / cfqd->cfq_slice[1];
2330                 if (!depth && !cfqq->dispatched)
2331                         depth = 1;
2332                 if (depth < max_dispatch)
2333                         max_dispatch = depth;
2334         }
2335
2336         /*
2337          * If we're below the current max, allow a dispatch
2338          */
2339         return cfqq->dispatched < max_dispatch;
2340 }
2341
2342 /*
2343  * Dispatch a request from cfqq, moving them to the request queue
2344  * dispatch list.
2345  */
2346 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2347 {
2348         struct request *rq;
2349
2350         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2351
2352         if (!cfq_may_dispatch(cfqd, cfqq))
2353                 return false;
2354
2355         /*
2356          * follow expired path, else get first next available
2357          */
2358         rq = cfq_check_fifo(cfqq);
2359         if (!rq)
2360                 rq = cfqq->next_rq;
2361
2362         /*
2363          * insert request into driver dispatch list
2364          */
2365         cfq_dispatch_insert(cfqd->queue, rq);
2366
2367         if (!cfqd->active_cic) {
2368                 struct cfq_io_context *cic = RQ_CIC(rq);
2369
2370                 atomic_long_inc(&cic->ioc->refcount);
2371                 cfqd->active_cic = cic;
2372         }
2373
2374         return true;
2375 }
2376
2377 /*
2378  * Find the cfqq that we need to service and move a request from that to the
2379  * dispatch list
2380  */
2381 static int cfq_dispatch_requests(struct request_queue *q, int force)
2382 {
2383         struct cfq_data *cfqd = q->elevator->elevator_data;
2384         struct cfq_queue *cfqq;
2385
2386         if (!cfqd->busy_queues)
2387                 return 0;
2388
2389         if (unlikely(force))
2390                 return cfq_forced_dispatch(cfqd);
2391
2392         cfqq = cfq_select_queue(cfqd);
2393         if (!cfqq)
2394                 return 0;
2395
2396         /*
2397          * Dispatch a request from this cfqq, if it is allowed
2398          */
2399         if (!cfq_dispatch_request(cfqd, cfqq))
2400                 return 0;
2401
2402         cfqq->slice_dispatch++;
2403         cfq_clear_cfqq_must_dispatch(cfqq);
2404
2405         /*
2406          * expire an async queue immediately if it has used up its slice. idle
2407          * queue always expire after 1 dispatch round.
2408          */
2409         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2410             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2411             cfq_class_idle(cfqq))) {
2412                 cfqq->slice_end = jiffies + 1;
2413                 cfq_slice_expired(cfqd, 0);
2414         }
2415
2416         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2417         return 1;
2418 }
2419
2420 /*
2421  * task holds one reference to the queue, dropped when task exits. each rq
2422  * in-flight on this queue also holds a reference, dropped when rq is freed.
2423  *
2424  * Each cfq queue took a reference on the parent group. Drop it now.
2425  * queue lock must be held here.
2426  */
2427 static void cfq_put_queue(struct cfq_queue *cfqq)
2428 {
2429         struct cfq_data *cfqd = cfqq->cfqd;
2430         struct cfq_group *cfqg, *orig_cfqg;
2431
2432         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2433
2434         if (!atomic_dec_and_test(&cfqq->ref))
2435                 return;
2436
2437         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2438         BUG_ON(rb_first(&cfqq->sort_list));
2439         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2440         cfqg = cfqq->cfqg;
2441         orig_cfqg = cfqq->orig_cfqg;
2442
2443         if (unlikely(cfqd->active_queue == cfqq)) {
2444                 __cfq_slice_expired(cfqd, cfqq, 0);
2445                 cfq_schedule_dispatch(cfqd);
2446         }
2447
2448         BUG_ON(cfq_cfqq_on_rr(cfqq));
2449         kmem_cache_free(cfq_pool, cfqq);
2450         cfq_put_cfqg(cfqg);
2451         if (orig_cfqg)
2452                 cfq_put_cfqg(orig_cfqg);
2453 }
2454
2455 /*
2456  * Must always be called with the rcu_read_lock() held
2457  */
2458 static void
2459 __call_for_each_cic(struct io_context *ioc,
2460                     void (*func)(struct io_context *, struct cfq_io_context *))
2461 {
2462         struct cfq_io_context *cic;
2463         struct hlist_node *n;
2464
2465         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2466                 func(ioc, cic);
2467 }
2468
2469 /*
2470  * Call func for each cic attached to this ioc.
2471  */
2472 static void
2473 call_for_each_cic(struct io_context *ioc,
2474                   void (*func)(struct io_context *, struct cfq_io_context *))
2475 {
2476         rcu_read_lock();
2477         __call_for_each_cic(ioc, func);
2478         rcu_read_unlock();
2479 }
2480
2481 static void cfq_cic_free_rcu(struct rcu_head *head)
2482 {
2483         struct cfq_io_context *cic;
2484
2485         cic = container_of(head, struct cfq_io_context, rcu_head);
2486
2487         kmem_cache_free(cfq_ioc_pool, cic);
2488         elv_ioc_count_dec(cfq_ioc_count);
2489
2490         if (ioc_gone) {
2491                 /*
2492                  * CFQ scheduler is exiting, grab exit lock and check
2493                  * the pending io context count. If it hits zero,
2494                  * complete ioc_gone and set it back to NULL
2495                  */
2496                 spin_lock(&ioc_gone_lock);
2497                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2498                         complete(ioc_gone);
2499                         ioc_gone = NULL;
2500                 }
2501                 spin_unlock(&ioc_gone_lock);
2502         }
2503 }
2504
2505 static void cfq_cic_free(struct cfq_io_context *cic)
2506 {
2507         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2508 }
2509
2510 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2511 {
2512         unsigned long flags;
2513
2514         BUG_ON(!cic->dead_key);
2515
2516         spin_lock_irqsave(&ioc->lock, flags);
2517         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2518         hlist_del_rcu(&cic->cic_list);
2519         spin_unlock_irqrestore(&ioc->lock, flags);
2520
2521         cfq_cic_free(cic);
2522 }
2523
2524 /*
2525  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2526  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2527  * and ->trim() which is called with the task lock held
2528  */
2529 static void cfq_free_io_context(struct io_context *ioc)
2530 {
2531         /*
2532          * ioc->refcount is zero here, or we are called from elv_unregister(),
2533          * so no more cic's are allowed to be linked into this ioc.  So it
2534          * should be ok to iterate over the known list, we will see all cic's
2535          * since no new ones are added.
2536          */
2537         __call_for_each_cic(ioc, cic_free_func);
2538 }
2539
2540 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2541 {
2542         struct cfq_queue *__cfqq, *next;
2543
2544         if (unlikely(cfqq == cfqd->active_queue)) {
2545                 __cfq_slice_expired(cfqd, cfqq, 0);
2546                 cfq_schedule_dispatch(cfqd);
2547         }
2548
2549         /*
2550          * If this queue was scheduled to merge with another queue, be
2551          * sure to drop the reference taken on that queue (and others in
2552          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2553          */
2554         __cfqq = cfqq->new_cfqq;
2555         while (__cfqq) {
2556                 if (__cfqq == cfqq) {
2557                         WARN(1, "cfqq->new_cfqq loop detected\n");
2558                         break;
2559                 }
2560                 next = __cfqq->new_cfqq;
2561                 cfq_put_queue(__cfqq);
2562                 __cfqq = next;
2563         }
2564
2565         cfq_put_queue(cfqq);
2566 }
2567
2568 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2569                                          struct cfq_io_context *cic)
2570 {
2571         struct io_context *ioc = cic->ioc;
2572
2573         list_del_init(&cic->queue_list);
2574
2575         /*
2576          * Make sure key == NULL is seen for dead queues
2577          */
2578         smp_wmb();
2579         cic->dead_key = (unsigned long) cic->key;
2580         cic->key = NULL;
2581
2582         if (ioc->ioc_data == cic)
2583                 rcu_assign_pointer(ioc->ioc_data, NULL);
2584
2585         if (cic->cfqq[BLK_RW_ASYNC]) {
2586                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2587                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2588         }
2589
2590         if (cic->cfqq[BLK_RW_SYNC]) {
2591                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2592                 cic->cfqq[BLK_RW_SYNC] = NULL;
2593         }
2594 }
2595
2596 static void cfq_exit_single_io_context(struct io_context *ioc,
2597                                        struct cfq_io_context *cic)
2598 {
2599         struct cfq_data *cfqd = cic->key;
2600
2601         if (cfqd) {
2602                 struct request_queue *q = cfqd->queue;
2603                 unsigned long flags;
2604
2605                 spin_lock_irqsave(q->queue_lock, flags);
2606
2607                 /*
2608                  * Ensure we get a fresh copy of the ->key to prevent
2609                  * race between exiting task and queue
2610                  */
2611                 smp_read_barrier_depends();
2612                 if (cic->key)
2613                         __cfq_exit_single_io_context(cfqd, cic);
2614
2615                 spin_unlock_irqrestore(q->queue_lock, flags);
2616         }
2617 }
2618
2619 /*
2620  * The process that ioc belongs to has exited, we need to clean up
2621  * and put the internal structures we have that belongs to that process.
2622  */
2623 static void cfq_exit_io_context(struct io_context *ioc)
2624 {
2625         call_for_each_cic(ioc, cfq_exit_single_io_context);
2626 }
2627
2628 static struct cfq_io_context *
2629 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2630 {
2631         struct cfq_io_context *cic;
2632
2633         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2634                                                         cfqd->queue->node);
2635         if (cic) {
2636                 cic->last_end_request = jiffies;
2637                 INIT_LIST_HEAD(&cic->queue_list);
2638                 INIT_HLIST_NODE(&cic->cic_list);
2639                 cic->dtor = cfq_free_io_context;
2640                 cic->exit = cfq_exit_io_context;
2641                 elv_ioc_count_inc(cfq_ioc_count);
2642         }
2643
2644         return cic;
2645 }
2646
2647 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2648 {
2649         struct task_struct *tsk = current;
2650         int ioprio_class;
2651
2652         if (!cfq_cfqq_prio_changed(cfqq))
2653                 return;
2654
2655         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2656         switch (ioprio_class) {
2657         default:
2658                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2659         case IOPRIO_CLASS_NONE:
2660                 /*
2661                  * no prio set, inherit CPU scheduling settings
2662                  */
2663                 cfqq->ioprio = task_nice_ioprio(tsk);
2664                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2665                 break;
2666         case IOPRIO_CLASS_RT:
2667                 cfqq->ioprio = task_ioprio(ioc);
2668                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2669                 break;
2670         case IOPRIO_CLASS_BE:
2671                 cfqq->ioprio = task_ioprio(ioc);
2672                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2673                 break;
2674         case IOPRIO_CLASS_IDLE:
2675                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2676                 cfqq->ioprio = 7;
2677                 cfq_clear_cfqq_idle_window(cfqq);
2678                 break;
2679         }
2680
2681         /*
2682          * keep track of original prio settings in case we have to temporarily
2683          * elevate the priority of this queue
2684          */
2685         cfqq->org_ioprio = cfqq->ioprio;
2686         cfqq->org_ioprio_class = cfqq->ioprio_class;
2687         cfq_clear_cfqq_prio_changed(cfqq);
2688 }
2689
2690 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2691 {
2692         struct cfq_data *cfqd = cic->key;
2693         struct cfq_queue *cfqq;
2694         unsigned long flags;
2695
2696         if (unlikely(!cfqd))
2697                 return;
2698
2699         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2700
2701         cfqq = cic->cfqq[BLK_RW_ASYNC];
2702         if (cfqq) {
2703                 struct cfq_queue *new_cfqq;
2704                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2705                                                 GFP_ATOMIC);
2706                 if (new_cfqq) {
2707                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2708                         cfq_put_queue(cfqq);
2709                 }
2710         }
2711
2712         cfqq = cic->cfqq[BLK_RW_SYNC];
2713         if (cfqq)
2714                 cfq_mark_cfqq_prio_changed(cfqq);
2715
2716         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2717 }
2718
2719 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2720 {
2721         call_for_each_cic(ioc, changed_ioprio);
2722         ioc->ioprio_changed = 0;
2723 }
2724
2725 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2726                           pid_t pid, bool is_sync)
2727 {
2728         RB_CLEAR_NODE(&cfqq->rb_node);
2729         RB_CLEAR_NODE(&cfqq->p_node);
2730         INIT_LIST_HEAD(&cfqq->fifo);
2731
2732         atomic_set(&cfqq->ref, 0);
2733         cfqq->cfqd = cfqd;
2734
2735         cfq_mark_cfqq_prio_changed(cfqq);
2736
2737         if (is_sync) {
2738                 if (!cfq_class_idle(cfqq))
2739                         cfq_mark_cfqq_idle_window(cfqq);
2740                 cfq_mark_cfqq_sync(cfqq);
2741         }
2742         cfqq->pid = pid;
2743 }
2744
2745 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2746 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2747 {
2748         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2749         struct cfq_data *cfqd = cic->key;
2750         unsigned long flags;
2751         struct request_queue *q;
2752
2753         if (unlikely(!cfqd))
2754                 return;
2755
2756         q = cfqd->queue;
2757
2758         spin_lock_irqsave(q->queue_lock, flags);
2759
2760         if (sync_cfqq) {
2761                 /*
2762                  * Drop reference to sync queue. A new sync queue will be
2763                  * assigned in new group upon arrival of a fresh request.
2764                  */
2765                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2766                 cic_set_cfqq(cic, NULL, 1);
2767                 cfq_put_queue(sync_cfqq);
2768         }
2769
2770         spin_unlock_irqrestore(q->queue_lock, flags);
2771 }
2772
2773 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2774 {
2775         call_for_each_cic(ioc, changed_cgroup);
2776         ioc->cgroup_changed = 0;
2777 }
2778 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2779
2780 static struct cfq_queue *
2781 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2782                      struct io_context *ioc, gfp_t gfp_mask)
2783 {
2784         struct cfq_queue *cfqq, *new_cfqq = NULL;
2785         struct cfq_io_context *cic;
2786         struct cfq_group *cfqg;
2787
2788 retry:
2789         cfqg = cfq_get_cfqg(cfqd, 1);
2790         cic = cfq_cic_lookup(cfqd, ioc);
2791         /* cic always exists here */
2792         cfqq = cic_to_cfqq(cic, is_sync);
2793
2794         /*
2795          * Always try a new alloc if we fell back to the OOM cfqq
2796          * originally, since it should just be a temporary situation.
2797          */
2798         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2799                 cfqq = NULL;
2800                 if (new_cfqq) {
2801                         cfqq = new_cfqq;
2802                         new_cfqq = NULL;
2803                 } else if (gfp_mask & __GFP_WAIT) {
2804                         spin_unlock_irq(cfqd->queue->queue_lock);
2805                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2806                                         gfp_mask | __GFP_ZERO,
2807                                         cfqd->queue->node);
2808                         spin_lock_irq(cfqd->queue->queue_lock);
2809                         if (new_cfqq)
2810                                 goto retry;
2811                 } else {
2812                         cfqq = kmem_cache_alloc_node(cfq_pool,
2813                                         gfp_mask | __GFP_ZERO,
2814                                         cfqd->queue->node);
2815                 }
2816
2817                 if (cfqq) {
2818                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2819                         cfq_init_prio_data(cfqq, ioc);
2820                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2821                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2822                 } else
2823                         cfqq = &cfqd->oom_cfqq;
2824         }
2825
2826         if (new_cfqq)
2827                 kmem_cache_free(cfq_pool, new_cfqq);
2828
2829         return cfqq;
2830 }
2831
2832 static struct cfq_queue **
2833 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2834 {
2835         switch (ioprio_class) {
2836         case IOPRIO_CLASS_RT:
2837                 return &cfqd->async_cfqq[0][ioprio];
2838         case IOPRIO_CLASS_BE:
2839                 return &cfqd->async_cfqq[1][ioprio];
2840         case IOPRIO_CLASS_IDLE:
2841                 return &cfqd->async_idle_cfqq;
2842         default:
2843                 BUG();
2844         }
2845 }
2846
2847 static struct cfq_queue *
2848 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2849               gfp_t gfp_mask)
2850 {
2851         const int ioprio = task_ioprio(ioc);
2852         const int ioprio_class = task_ioprio_class(ioc);
2853         struct cfq_queue **async_cfqq = NULL;
2854         struct cfq_queue *cfqq = NULL;
2855
2856         if (!is_sync) {
2857                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2858                 cfqq = *async_cfqq;
2859         }
2860
2861         if (!cfqq)
2862                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2863
2864         /*
2865          * pin the queue now that it's allocated, scheduler exit will prune it
2866          */
2867         if (!is_sync && !(*async_cfqq)) {
2868                 atomic_inc(&cfqq->ref);
2869                 *async_cfqq = cfqq;
2870         }
2871
2872         atomic_inc(&cfqq->ref);
2873         return cfqq;
2874 }
2875
2876 /*
2877  * We drop cfq io contexts lazily, so we may find a dead one.
2878  */
2879 static void
2880 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2881                   struct cfq_io_context *cic)
2882 {
2883         unsigned long flags;
2884
2885         WARN_ON(!list_empty(&cic->queue_list));
2886
2887         spin_lock_irqsave(&ioc->lock, flags);
2888
2889         BUG_ON(ioc->ioc_data == cic);
2890
2891         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2892         hlist_del_rcu(&cic->cic_list);
2893         spin_unlock_irqrestore(&ioc->lock, flags);
2894
2895         cfq_cic_free(cic);
2896 }
2897
2898 static struct cfq_io_context *
2899 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2900 {
2901         struct cfq_io_context *cic;
2902         unsigned long flags;
2903         void *k;
2904
2905         if (unlikely(!ioc))
2906                 return NULL;
2907
2908         rcu_read_lock();
2909
2910         /*
2911          * we maintain a last-hit cache, to avoid browsing over the tree
2912          */
2913         cic = rcu_dereference(ioc->ioc_data);
2914         if (cic && cic->key == cfqd) {
2915                 rcu_read_unlock();
2916                 return cic;
2917         }
2918
2919         do {
2920                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2921                 rcu_read_unlock();
2922                 if (!cic)
2923                         break;
2924                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2925                 k = cic->key;
2926                 if (unlikely(!k)) {
2927                         cfq_drop_dead_cic(cfqd, ioc, cic);
2928                         rcu_read_lock();
2929                         continue;
2930                 }
2931
2932                 spin_lock_irqsave(&ioc->lock, flags);
2933                 rcu_assign_pointer(ioc->ioc_data, cic);
2934                 spin_unlock_irqrestore(&ioc->lock, flags);
2935                 break;
2936         } while (1);
2937
2938         return cic;
2939 }
2940
2941 /*
2942  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2943  * the process specific cfq io context when entered from the block layer.
2944  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2945  */
2946 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2947                         struct cfq_io_context *cic, gfp_t gfp_mask)
2948 {
2949         unsigned long flags;
2950         int ret;
2951
2952         ret = radix_tree_preload(gfp_mask);
2953         if (!ret) {
2954                 cic->ioc = ioc;
2955                 cic->key = cfqd;
2956
2957                 spin_lock_irqsave(&ioc->lock, flags);
2958                 ret = radix_tree_insert(&ioc->radix_root,
2959                                                 (unsigned long) cfqd, cic);
2960                 if (!ret)
2961                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2962                 spin_unlock_irqrestore(&ioc->lock, flags);
2963
2964                 radix_tree_preload_end();
2965
2966                 if (!ret) {
2967                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2968                         list_add(&cic->queue_list, &cfqd->cic_list);
2969                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2970                 }
2971         }
2972
2973         if (ret)
2974                 printk(KERN_ERR "cfq: cic link failed!\n");
2975
2976         return ret;
2977 }
2978
2979 /*
2980  * Setup general io context and cfq io context. There can be several cfq
2981  * io contexts per general io context, if this process is doing io to more
2982  * than one device managed by cfq.
2983  */
2984 static struct cfq_io_context *
2985 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2986 {
2987         struct io_context *ioc = NULL;
2988         struct cfq_io_context *cic;
2989
2990         might_sleep_if(gfp_mask & __GFP_WAIT);
2991
2992         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2993         if (!ioc)
2994                 return NULL;
2995
2996         cic = cfq_cic_lookup(cfqd, ioc);
2997         if (cic)
2998                 goto out;
2999
3000         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3001         if (cic == NULL)
3002                 goto err;
3003
3004         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3005                 goto err_free;
3006
3007 out:
3008         smp_read_barrier_depends();
3009         if (unlikely(ioc->ioprio_changed))
3010                 cfq_ioc_set_ioprio(ioc);
3011
3012 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3013         if (unlikely(ioc->cgroup_changed))
3014                 cfq_ioc_set_cgroup(ioc);
3015 #endif
3016         return cic;
3017 err_free:
3018         cfq_cic_free(cic);
3019 err:
3020         put_io_context(ioc);
3021         return NULL;
3022 }
3023
3024 static void
3025 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3026 {
3027         unsigned long elapsed = jiffies - cic->last_end_request;
3028         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3029
3030         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3031         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3032         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3033 }
3034
3035 static void
3036 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3037                        struct request *rq)
3038 {
3039         sector_t sdist = 0;
3040         sector_t n_sec = blk_rq_sectors(rq);
3041         if (cfqq->last_request_pos) {
3042                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3043                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3044                 else
3045                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3046         }
3047
3048         cfqq->seek_history <<= 1;
3049         if (blk_queue_nonrot(cfqd->queue))
3050                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3051         else
3052                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3053 }
3054
3055 /*
3056  * Disable idle window if the process thinks too long or seeks so much that
3057  * it doesn't matter
3058  */
3059 static void
3060 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3061                        struct cfq_io_context *cic)
3062 {
3063         int old_idle, enable_idle;
3064
3065         /*
3066          * Don't idle for async or idle io prio class
3067          */
3068         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3069                 return;
3070
3071         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3072
3073         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3074                 cfq_mark_cfqq_deep(cfqq);
3075
3076         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3077             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3078                 enable_idle = 0;
3079         else if (sample_valid(cic->ttime_samples)) {
3080                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3081                         enable_idle = 0;
3082                 else
3083                         enable_idle = 1;
3084         }
3085
3086         if (old_idle != enable_idle) {
3087                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3088                 if (enable_idle)
3089                         cfq_mark_cfqq_idle_window(cfqq);
3090                 else
3091                         cfq_clear_cfqq_idle_window(cfqq);
3092         }
3093 }
3094
3095 /*
3096  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3097  * no or if we aren't sure, a 1 will cause a preempt.
3098  */
3099 static bool
3100 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3101                    struct request *rq)
3102 {
3103         struct cfq_queue *cfqq;
3104
3105         cfqq = cfqd->active_queue;
3106         if (!cfqq)
3107                 return false;
3108
3109         if (cfq_class_idle(new_cfqq))
3110                 return false;
3111
3112         if (cfq_class_idle(cfqq))
3113                 return true;
3114
3115         /*
3116          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3117          */
3118         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3119                 return false;
3120
3121         /*
3122          * if the new request is sync, but the currently running queue is
3123          * not, let the sync request have priority.
3124          */
3125         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3126                 return true;
3127
3128         if (new_cfqq->cfqg != cfqq->cfqg)
3129                 return false;
3130
3131         if (cfq_slice_used(cfqq))
3132                 return true;
3133
3134         /* Allow preemption only if we are idling on sync-noidle tree */
3135         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3136             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3137             new_cfqq->service_tree->count == 2 &&
3138             RB_EMPTY_ROOT(&cfqq->sort_list))
3139                 return true;
3140
3141         /*
3142          * So both queues are sync. Let the new request get disk time if
3143          * it's a metadata request and the current queue is doing regular IO.
3144          */
3145         if (rq_is_meta(rq) && !cfqq->meta_pending)
3146                 return true;
3147
3148         /*
3149          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3150          */
3151         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3152                 return true;
3153
3154         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3155                 return false;
3156
3157         /*
3158          * if this request is as-good as one we would expect from the
3159          * current cfqq, let it preempt
3160          */
3161         if (cfq_rq_close(cfqd, cfqq, rq))
3162                 return true;
3163
3164         return false;
3165 }
3166
3167 /*
3168  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3169  * let it have half of its nominal slice.
3170  */
3171 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3172 {
3173         cfq_log_cfqq(cfqd, cfqq, "preempt");
3174         cfq_slice_expired(cfqd, 1);
3175
3176         /*
3177          * Put the new queue at the front of the of the current list,
3178          * so we know that it will be selected next.
3179          */
3180         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3181
3182         cfq_service_tree_add(cfqd, cfqq, 1);
3183
3184         cfqq->slice_end = 0;
3185         cfq_mark_cfqq_slice_new(cfqq);
3186 }
3187
3188 /*
3189  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3190  * something we should do about it
3191  */
3192 static void
3193 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3194                 struct request *rq)
3195 {
3196         struct cfq_io_context *cic = RQ_CIC(rq);
3197
3198         cfqd->rq_queued++;
3199         if (rq_is_meta(rq))
3200                 cfqq->meta_pending++;
3201
3202         cfq_update_io_thinktime(cfqd, cic);
3203         cfq_update_io_seektime(cfqd, cfqq, rq);
3204         cfq_update_idle_window(cfqd, cfqq, cic);
3205
3206         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3207
3208         if (cfqq == cfqd->active_queue) {
3209                 /*
3210                  * Remember that we saw a request from this process, but
3211                  * don't start queuing just yet. Otherwise we risk seeing lots
3212                  * of tiny requests, because we disrupt the normal plugging
3213                  * and merging. If the request is already larger than a single
3214                  * page, let it rip immediately. For that case we assume that
3215                  * merging is already done. Ditto for a busy system that
3216                  * has other work pending, don't risk delaying until the
3217                  * idle timer unplug to continue working.
3218                  */
3219                 if (cfq_cfqq_wait_request(cfqq)) {
3220                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3221                             cfqd->busy_queues > 1) {
3222                                 cfq_del_timer(cfqd, cfqq);
3223                                 cfq_clear_cfqq_wait_request(cfqq);
3224                                 __blk_run_queue(cfqd->queue);
3225                         } else {
3226                                 blkiocg_update_idle_time_stats(
3227                                                 &cfqq->cfqg->blkg);
3228                                 cfq_mark_cfqq_must_dispatch(cfqq);
3229                         }
3230                 }
3231         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3232                 /*
3233                  * not the active queue - expire current slice if it is
3234                  * idle and has expired it's mean thinktime or this new queue
3235                  * has some old slice time left and is of higher priority or
3236                  * this new queue is RT and the current one is BE
3237                  */
3238                 cfq_preempt_queue(cfqd, cfqq);
3239                 __blk_run_queue(cfqd->queue);
3240         }
3241 }
3242
3243 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3244 {
3245         struct cfq_data *cfqd = q->elevator->elevator_data;
3246         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3247
3248         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3249         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3250
3251         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3252         list_add_tail(&rq->queuelist, &cfqq->fifo);
3253         cfq_add_rq_rb(rq);
3254         blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3255                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3256                         rq_is_sync(rq));
3257         cfq_rq_enqueued(cfqd, cfqq, rq);
3258 }
3259
3260 /*
3261  * Update hw_tag based on peak queue depth over 50 samples under
3262  * sufficient load.
3263  */
3264 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3265 {
3266         struct cfq_queue *cfqq = cfqd->active_queue;
3267
3268         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3269                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3270
3271         if (cfqd->hw_tag == 1)
3272                 return;
3273
3274         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3275             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3276                 return;
3277
3278         /*
3279          * If active queue hasn't enough requests and can idle, cfq might not
3280          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3281          * case
3282          */
3283         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3284             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3285             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3286                 return;
3287
3288         if (cfqd->hw_tag_samples++ < 50)
3289                 return;
3290
3291         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3292                 cfqd->hw_tag = 1;
3293         else
3294                 cfqd->hw_tag = 0;
3295 }
3296
3297 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3298 {
3299         struct cfq_io_context *cic = cfqd->active_cic;
3300
3301         /* If there are other queues in the group, don't wait */
3302         if (cfqq->cfqg->nr_cfqq > 1)
3303                 return false;
3304
3305         if (cfq_slice_used(cfqq))
3306                 return true;
3307
3308         /* if slice left is less than think time, wait busy */
3309         if (cic && sample_valid(cic->ttime_samples)
3310             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3311                 return true;
3312
3313         /*
3314          * If think times is less than a jiffy than ttime_mean=0 and above
3315          * will not be true. It might happen that slice has not expired yet
3316          * but will expire soon (4-5 ns) during select_queue(). To cover the
3317          * case where think time is less than a jiffy, mark the queue wait
3318          * busy if only 1 jiffy is left in the slice.
3319          */
3320         if (cfqq->slice_end - jiffies == 1)
3321                 return true;
3322
3323         return false;
3324 }
3325
3326 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3327 {
3328         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3329         struct cfq_data *cfqd = cfqq->cfqd;
3330         const int sync = rq_is_sync(rq);
3331         unsigned long now;
3332
3333         now = jiffies;
3334         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3335
3336         cfq_update_hw_tag(cfqd);
3337
3338         WARN_ON(!cfqd->rq_in_driver);
3339         WARN_ON(!cfqq->dispatched);
3340         cfqd->rq_in_driver--;
3341         cfqq->dispatched--;
3342         blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3343                         rq_io_start_time_ns(rq), rq_data_dir(rq),
3344                         rq_is_sync(rq));
3345
3346         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3347
3348         if (sync) {
3349                 RQ_CIC(rq)->last_end_request = now;
3350                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3351                         cfqd->last_delayed_sync = now;
3352         }
3353
3354         /*
3355          * If this is the active queue, check if it needs to be expired,
3356          * or if we want to idle in case it has no pending requests.
3357          */
3358         if (cfqd->active_queue == cfqq) {
3359                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3360
3361                 if (cfq_cfqq_slice_new(cfqq)) {
3362                         cfq_set_prio_slice(cfqd, cfqq);
3363                         cfq_clear_cfqq_slice_new(cfqq);
3364                 }
3365
3366                 /*
3367                  * Should we wait for next request to come in before we expire
3368                  * the queue.
3369                  */
3370                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3371                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3372                         cfq_mark_cfqq_wait_busy(cfqq);
3373                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3374                 }
3375
3376                 /*
3377                  * Idling is not enabled on:
3378                  * - expired queues
3379                  * - idle-priority queues
3380                  * - async queues
3381                  * - queues with still some requests queued
3382                  * - when there is a close cooperator
3383                  */
3384                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3385                         cfq_slice_expired(cfqd, 1);
3386                 else if (sync && cfqq_empty &&
3387                          !cfq_close_cooperator(cfqd, cfqq)) {
3388                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3389                         /*
3390                          * Idling is enabled for SYNC_WORKLOAD.
3391                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3392                          * only if we processed at least one !rq_noidle request
3393                          */
3394                         if (cfqd->serving_type == SYNC_WORKLOAD
3395                             || cfqd->noidle_tree_requires_idle
3396                             || cfqq->cfqg->nr_cfqq == 1)
3397                                 cfq_arm_slice_timer(cfqd);
3398                 }
3399         }
3400
3401         if (!cfqd->rq_in_driver)
3402                 cfq_schedule_dispatch(cfqd);
3403 }
3404
3405 /*
3406  * we temporarily boost lower priority queues if they are holding fs exclusive
3407  * resources. they are boosted to normal prio (CLASS_BE/4)
3408  */
3409 static void cfq_prio_boost(struct cfq_queue *cfqq)
3410 {
3411         if (has_fs_excl()) {
3412                 /*
3413                  * boost idle prio on transactions that would lock out other
3414                  * users of the filesystem
3415                  */
3416                 if (cfq_class_idle(cfqq))
3417                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3418                 if (cfqq->ioprio > IOPRIO_NORM)
3419                         cfqq->ioprio = IOPRIO_NORM;
3420         } else {
3421                 /*
3422                  * unboost the queue (if needed)
3423                  */
3424                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3425                 cfqq->ioprio = cfqq->org_ioprio;
3426         }
3427 }
3428
3429 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3430 {
3431         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3432                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3433                 return ELV_MQUEUE_MUST;
3434         }
3435
3436         return ELV_MQUEUE_MAY;
3437 }
3438
3439 static int cfq_may_queue(struct request_queue *q, int rw)
3440 {
3441         struct cfq_data *cfqd = q->elevator->elevator_data;
3442         struct task_struct *tsk = current;
3443         struct cfq_io_context *cic;
3444         struct cfq_queue *cfqq;
3445
3446         /*
3447          * don't force setup of a queue from here, as a call to may_queue
3448          * does not necessarily imply that a request actually will be queued.
3449          * so just lookup a possibly existing queue, or return 'may queue'
3450          * if that fails
3451          */
3452         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3453         if (!cic)
3454                 return ELV_MQUEUE_MAY;
3455
3456         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3457         if (cfqq) {
3458                 cfq_init_prio_data(cfqq, cic->ioc);
3459                 cfq_prio_boost(cfqq);
3460
3461                 return __cfq_may_queue(cfqq);
3462         }
3463
3464         return ELV_MQUEUE_MAY;
3465 }
3466
3467 /*
3468  * queue lock held here
3469  */
3470 static void cfq_put_request(struct request *rq)
3471 {
3472         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3473
3474         if (cfqq) {
3475                 const int rw = rq_data_dir(rq);
3476
3477                 BUG_ON(!cfqq->allocated[rw]);
3478                 cfqq->allocated[rw]--;
3479
3480                 put_io_context(RQ_CIC(rq)->ioc);
3481
3482                 rq->elevator_private = NULL;
3483                 rq->elevator_private2 = NULL;
3484
3485                 /* Put down rq reference on cfqg */
3486                 cfq_put_cfqg(RQ_CFQG(rq));
3487                 rq->elevator_private3 = NULL;
3488
3489                 cfq_put_queue(cfqq);
3490         }
3491 }
3492
3493 static struct cfq_queue *
3494 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3495                 struct cfq_queue *cfqq)
3496 {
3497         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3498         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3499         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3500         cfq_put_queue(cfqq);
3501         return cic_to_cfqq(cic, 1);
3502 }
3503
3504 /*
3505  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3506  * was the last process referring to said cfqq.
3507  */
3508 static struct cfq_queue *
3509 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3510 {
3511         if (cfqq_process_refs(cfqq) == 1) {
3512                 cfqq->pid = current->pid;
3513                 cfq_clear_cfqq_coop(cfqq);
3514                 cfq_clear_cfqq_split_coop(cfqq);
3515                 return cfqq;
3516         }
3517
3518         cic_set_cfqq(cic, NULL, 1);
3519         cfq_put_queue(cfqq);
3520         return NULL;
3521 }
3522 /*
3523  * Allocate cfq data structures associated with this request.
3524  */
3525 static int
3526 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3527 {
3528         struct cfq_data *cfqd = q->elevator->elevator_data;
3529         struct cfq_io_context *cic;
3530         const int rw = rq_data_dir(rq);
3531         const bool is_sync = rq_is_sync(rq);
3532         struct cfq_queue *cfqq;
3533         unsigned long flags;
3534
3535         might_sleep_if(gfp_mask & __GFP_WAIT);
3536
3537         cic = cfq_get_io_context(cfqd, gfp_mask);
3538
3539         spin_lock_irqsave(q->queue_lock, flags);
3540
3541         if (!cic)
3542                 goto queue_fail;
3543
3544 new_queue:
3545         cfqq = cic_to_cfqq(cic, is_sync);
3546         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3547                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3548                 cic_set_cfqq(cic, cfqq, is_sync);
3549         } else {
3550                 /*
3551                  * If the queue was seeky for too long, break it apart.
3552                  */
3553                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3554                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3555                         cfqq = split_cfqq(cic, cfqq);
3556                         if (!cfqq)
3557                                 goto new_queue;
3558                 }
3559
3560                 /*
3561                  * Check to see if this queue is scheduled to merge with
3562                  * another, closely cooperating queue.  The merging of
3563                  * queues happens here as it must be done in process context.
3564                  * The reference on new_cfqq was taken in merge_cfqqs.
3565                  */
3566                 if (cfqq->new_cfqq)
3567                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3568         }
3569
3570         cfqq->allocated[rw]++;
3571         atomic_inc(&cfqq->ref);
3572
3573         spin_unlock_irqrestore(q->queue_lock, flags);
3574
3575         rq->elevator_private = cic;
3576         rq->elevator_private2 = cfqq;
3577         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3578         return 0;
3579
3580 queue_fail:
3581         if (cic)
3582                 put_io_context(cic->ioc);
3583
3584         cfq_schedule_dispatch(cfqd);
3585         spin_unlock_irqrestore(q->queue_lock, flags);
3586         cfq_log(cfqd, "set_request fail");
3587         return 1;
3588 }
3589
3590 static void cfq_kick_queue(struct work_struct *work)
3591 {
3592         struct cfq_data *cfqd =
3593                 container_of(work, struct cfq_data, unplug_work);
3594         struct request_queue *q = cfqd->queue;
3595
3596         spin_lock_irq(q->queue_lock);
3597         __blk_run_queue(cfqd->queue);
3598         spin_unlock_irq(q->queue_lock);
3599 }
3600
3601 /*
3602  * Timer running if the active_queue is currently idling inside its time slice
3603  */
3604 static void cfq_idle_slice_timer(unsigned long data)
3605 {
3606         struct cfq_data *cfqd = (struct cfq_data *) data;
3607         struct cfq_queue *cfqq;
3608         unsigned long flags;
3609         int timed_out = 1;
3610
3611         cfq_log(cfqd, "idle timer fired");
3612
3613         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3614
3615         cfqq = cfqd->active_queue;
3616         if (cfqq) {
3617                 timed_out = 0;
3618
3619                 /*
3620                  * We saw a request before the queue expired, let it through
3621                  */
3622                 if (cfq_cfqq_must_dispatch(cfqq))
3623                         goto out_kick;
3624
3625                 /*
3626                  * expired
3627                  */
3628                 if (cfq_slice_used(cfqq))
3629                         goto expire;
3630
3631                 /*
3632                  * only expire and reinvoke request handler, if there are
3633                  * other queues with pending requests
3634                  */
3635                 if (!cfqd->busy_queues)
3636                         goto out_cont;
3637
3638                 /*
3639                  * not expired and it has a request pending, let it dispatch
3640                  */
3641                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3642                         goto out_kick;
3643
3644                 /*
3645                  * Queue depth flag is reset only when the idle didn't succeed
3646                  */
3647                 cfq_clear_cfqq_deep(cfqq);
3648         }
3649 expire:
3650         cfq_slice_expired(cfqd, timed_out);
3651 out_kick:
3652         cfq_schedule_dispatch(cfqd);
3653 out_cont:
3654         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3655 }
3656
3657 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3658 {
3659         del_timer_sync(&cfqd->idle_slice_timer);
3660         cancel_work_sync(&cfqd->unplug_work);
3661 }
3662
3663 static void cfq_put_async_queues(struct cfq_data *cfqd)
3664 {
3665         int i;
3666
3667         for (i = 0; i < IOPRIO_BE_NR; i++) {
3668                 if (cfqd->async_cfqq[0][i])
3669                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3670                 if (cfqd->async_cfqq[1][i])
3671                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3672         }
3673
3674         if (cfqd->async_idle_cfqq)
3675                 cfq_put_queue(cfqd->async_idle_cfqq);
3676 }
3677
3678 static void cfq_cfqd_free(struct rcu_head *head)
3679 {
3680         kfree(container_of(head, struct cfq_data, rcu));
3681 }
3682
3683 static void cfq_exit_queue(struct elevator_queue *e)
3684 {
3685         struct cfq_data *cfqd = e->elevator_data;
3686         struct request_queue *q = cfqd->queue;
3687
3688         cfq_shutdown_timer_wq(cfqd);
3689
3690         spin_lock_irq(q->queue_lock);
3691
3692         if (cfqd->active_queue)
3693                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3694
3695         while (!list_empty(&cfqd->cic_list)) {
3696                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3697                                                         struct cfq_io_context,
3698                                                         queue_list);
3699
3700                 __cfq_exit_single_io_context(cfqd, cic);
3701         }
3702
3703         cfq_put_async_queues(cfqd);
3704         cfq_release_cfq_groups(cfqd);
3705         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3706
3707         spin_unlock_irq(q->queue_lock);
3708
3709         cfq_shutdown_timer_wq(cfqd);
3710
3711         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3712         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3713 }
3714
3715 static void *cfq_init_queue(struct request_queue *q)
3716 {
3717         struct cfq_data *cfqd;
3718         int i, j;
3719         struct cfq_group *cfqg;
3720         struct cfq_rb_root *st;
3721
3722         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3723         if (!cfqd)
3724                 return NULL;
3725
3726         /* Init root service tree */
3727         cfqd->grp_service_tree = CFQ_RB_ROOT;
3728
3729         /* Init root group */
3730         cfqg = &cfqd->root_group;
3731         for_each_cfqg_st(cfqg, i, j, st)
3732                 *st = CFQ_RB_ROOT;
3733         RB_CLEAR_NODE(&cfqg->rb_node);
3734
3735         /* Give preference to root group over other groups */
3736         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3737
3738 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3739         /*
3740          * Take a reference to root group which we never drop. This is just
3741          * to make sure that cfq_put_cfqg() does not try to kfree root group
3742          */
3743         atomic_set(&cfqg->ref, 1);
3744         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3745                                         0);
3746 #endif
3747         /*
3748          * Not strictly needed (since RB_ROOT just clears the node and we
3749          * zeroed cfqd on alloc), but better be safe in case someone decides
3750          * to add magic to the rb code
3751          */
3752         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3753                 cfqd->prio_trees[i] = RB_ROOT;
3754
3755         /*
3756          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3757          * Grab a permanent reference to it, so that the normal code flow
3758          * will not attempt to free it.
3759          */
3760         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3761         atomic_inc(&cfqd->oom_cfqq.ref);
3762         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3763
3764         INIT_LIST_HEAD(&cfqd->cic_list);
3765
3766         cfqd->queue = q;
3767
3768         init_timer(&cfqd->idle_slice_timer);
3769         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3770         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3771
3772         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3773
3774         cfqd->cfq_quantum = cfq_quantum;
3775         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3776         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3777         cfqd->cfq_back_max = cfq_back_max;
3778         cfqd->cfq_back_penalty = cfq_back_penalty;
3779         cfqd->cfq_slice[0] = cfq_slice_async;
3780         cfqd->cfq_slice[1] = cfq_slice_sync;
3781         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3782         cfqd->cfq_slice_idle = cfq_slice_idle;
3783         cfqd->cfq_latency = 1;
3784         cfqd->cfq_group_isolation = 0;
3785         cfqd->hw_tag = -1;
3786         /*
3787          * we optimistically start assuming sync ops weren't delayed in last
3788          * second, in order to have larger depth for async operations.
3789          */
3790         cfqd->last_delayed_sync = jiffies - HZ;
3791         INIT_RCU_HEAD(&cfqd->rcu);
3792         return cfqd;
3793 }
3794
3795 static void cfq_slab_kill(void)
3796 {
3797         /*
3798          * Caller already ensured that pending RCU callbacks are completed,
3799          * so we should have no busy allocations at this point.
3800          */
3801         if (cfq_pool)
3802                 kmem_cache_destroy(cfq_pool);
3803         if (cfq_ioc_pool)
3804                 kmem_cache_destroy(cfq_ioc_pool);
3805 }
3806
3807 static int __init cfq_slab_setup(void)
3808 {
3809         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3810         if (!cfq_pool)
3811                 goto fail;
3812
3813         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3814         if (!cfq_ioc_pool)
3815                 goto fail;
3816
3817         return 0;
3818 fail:
3819         cfq_slab_kill();
3820         return -ENOMEM;
3821 }
3822
3823 /*
3824  * sysfs parts below -->
3825  */
3826 static ssize_t
3827 cfq_var_show(unsigned int var, char *page)
3828 {
3829         return sprintf(page, "%d\n", var);
3830 }
3831
3832 static ssize_t
3833 cfq_var_store(unsigned int *var, const char *page, size_t count)
3834 {
3835         char *p = (char *) page;
3836
3837         *var = simple_strtoul(p, &p, 10);
3838         return count;
3839 }
3840
3841 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3842 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3843 {                                                                       \
3844         struct cfq_data *cfqd = e->elevator_data;                       \
3845         unsigned int __data = __VAR;                                    \
3846         if (__CONV)                                                     \
3847                 __data = jiffies_to_msecs(__data);                      \
3848         return cfq_var_show(__data, (page));                            \
3849 }
3850 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3851 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3852 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3853 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3854 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3855 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3856 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3857 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3858 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3859 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3860 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3861 #undef SHOW_FUNCTION
3862
3863 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3864 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3865 {                                                                       \
3866         struct cfq_data *cfqd = e->elevator_data;                       \
3867         unsigned int __data;                                            \
3868         int ret = cfq_var_store(&__data, (page), count);                \
3869         if (__data < (MIN))                                             \
3870                 __data = (MIN);                                         \
3871         else if (__data > (MAX))                                        \
3872                 __data = (MAX);                                         \
3873         if (__CONV)                                                     \
3874                 *(__PTR) = msecs_to_jiffies(__data);                    \
3875         else                                                            \
3876                 *(__PTR) = __data;                                      \
3877         return ret;                                                     \
3878 }
3879 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3880 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3881                 UINT_MAX, 1);
3882 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3883                 UINT_MAX, 1);
3884 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3885 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3886                 UINT_MAX, 0);
3887 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3888 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3889 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3890 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3891                 UINT_MAX, 0);
3892 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3893 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3894 #undef STORE_FUNCTION
3895
3896 #define CFQ_ATTR(name) \
3897         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3898
3899 static struct elv_fs_entry cfq_attrs[] = {
3900         CFQ_ATTR(quantum),
3901         CFQ_ATTR(fifo_expire_sync),
3902         CFQ_ATTR(fifo_expire_async),
3903         CFQ_ATTR(back_seek_max),
3904         CFQ_ATTR(back_seek_penalty),
3905         CFQ_ATTR(slice_sync),
3906         CFQ_ATTR(slice_async),
3907         CFQ_ATTR(slice_async_rq),
3908         CFQ_ATTR(slice_idle),
3909         CFQ_ATTR(low_latency),
3910         CFQ_ATTR(group_isolation),
3911         __ATTR_NULL
3912 };
3913
3914 static struct elevator_type iosched_cfq = {
3915         .ops = {
3916                 .elevator_merge_fn =            cfq_merge,
3917                 .elevator_merged_fn =           cfq_merged_request,
3918                 .elevator_merge_req_fn =        cfq_merged_requests,
3919                 .elevator_allow_merge_fn =      cfq_allow_merge,
3920                 .elevator_bio_merged_fn =       cfq_bio_merged,
3921                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3922                 .elevator_add_req_fn =          cfq_insert_request,
3923                 .elevator_activate_req_fn =     cfq_activate_request,
3924                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3925                 .elevator_queue_empty_fn =      cfq_queue_empty,
3926                 .elevator_completed_req_fn =    cfq_completed_request,
3927                 .elevator_former_req_fn =       elv_rb_former_request,
3928                 .elevator_latter_req_fn =       elv_rb_latter_request,
3929                 .elevator_set_req_fn =          cfq_set_request,
3930                 .elevator_put_req_fn =          cfq_put_request,
3931                 .elevator_may_queue_fn =        cfq_may_queue,
3932                 .elevator_init_fn =             cfq_init_queue,
3933                 .elevator_exit_fn =             cfq_exit_queue,
3934                 .trim =                         cfq_free_io_context,
3935         },
3936         .elevator_attrs =       cfq_attrs,
3937         .elevator_name =        "cfq",
3938         .elevator_owner =       THIS_MODULE,
3939 };
3940
3941 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3942 static struct blkio_policy_type blkio_policy_cfq = {
3943         .ops = {
3944                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3945                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3946         },
3947 };
3948 #else
3949 static struct blkio_policy_type blkio_policy_cfq;
3950 #endif
3951
3952 static int __init cfq_init(void)
3953 {
3954         /*
3955          * could be 0 on HZ < 1000 setups
3956          */
3957         if (!cfq_slice_async)
3958                 cfq_slice_async = 1;
3959         if (!cfq_slice_idle)
3960                 cfq_slice_idle = 1;
3961
3962         if (cfq_slab_setup())
3963                 return -ENOMEM;
3964
3965         elv_register(&iosched_cfq);
3966         blkio_policy_register(&blkio_policy_cfq);
3967
3968         return 0;
3969 }
3970
3971 static void __exit cfq_exit(void)
3972 {
3973         DECLARE_COMPLETION_ONSTACK(all_gone);
3974         blkio_policy_unregister(&blkio_policy_cfq);
3975         elv_unregister(&iosched_cfq);
3976         ioc_gone = &all_gone;
3977         /* ioc_gone's update must be visible before reading ioc_count */
3978         smp_wmb();
3979
3980         /*
3981          * this also protects us from entering cfq_slab_kill() with
3982          * pending RCU callbacks
3983          */
3984         if (elv_ioc_count_read(cfq_ioc_count))
3985                 wait_for_completion(&all_gone);
3986         cfq_slab_kill();
3987 }
3988
3989 module_init(cfq_init);
3990 module_exit(cfq_exit);
3991
3992 MODULE_AUTHOR("Jens Axboe");
3993 MODULE_LICENSE("GPL");
3994 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");