blkio: Add more debug-only per-cgroup stats
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
17
18 /*
19  * tunables
20  */
21 /* max queue in one round of service */
22 static const int cfq_quantum = 8;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
34
35 /*
36  * offset from end of service tree
37  */
38 #define CFQ_IDLE_DELAY          (HZ / 5)
39
40 /*
41  * below this threshold, we consider thinktime immediate
42  */
43 #define CFQ_MIN_TT              (2)
44
45 #define CFQ_SLICE_SCALE         (5)
46 #define CFQ_HW_QUEUE_MIN        (5)
47 #define CFQ_SERVICE_SHIFT       12
48
49 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
50 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
51 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
52 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
53
54 #define RQ_CIC(rq)              \
55         ((struct cfq_io_context *) (rq)->elevator_private)
56 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
57
58 static struct kmem_cache *cfq_pool;
59 static struct kmem_cache *cfq_ioc_pool;
60
61 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
62 static struct completion *ioc_gone;
63 static DEFINE_SPINLOCK(ioc_gone_lock);
64
65 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68
69 #define sample_valid(samples)   ((samples) > 80)
70 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
71
72 /*
73  * Most of our rbtree usage is for sorting with min extraction, so
74  * if we cache the leftmost node we don't have to walk down the tree
75  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76  * move this into the elevator for the rq sorting as well.
77  */
78 struct cfq_rb_root {
79         struct rb_root rb;
80         struct rb_node *left;
81         unsigned count;
82         unsigned total_weight;
83         u64 min_vdisktime;
84         struct rb_node *active;
85 };
86 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87                         .count = 0, .min_vdisktime = 0, }
88
89 /*
90  * Per process-grouping structure
91  */
92 struct cfq_queue {
93         /* reference count */
94         atomic_t ref;
95         /* various state flags, see below */
96         unsigned int flags;
97         /* parent cfq_data */
98         struct cfq_data *cfqd;
99         /* service_tree member */
100         struct rb_node rb_node;
101         /* service_tree key */
102         unsigned long rb_key;
103         /* prio tree member */
104         struct rb_node p_node;
105         /* prio tree root we belong to, if any */
106         struct rb_root *p_root;
107         /* sorted list of pending requests */
108         struct rb_root sort_list;
109         /* if fifo isn't expired, next request to serve */
110         struct request *next_rq;
111         /* requests queued in sort_list */
112         int queued[2];
113         /* currently allocated requests */
114         int allocated[2];
115         /* fifo list of requests in sort_list */
116         struct list_head fifo;
117
118         /* time when queue got scheduled in to dispatch first request. */
119         unsigned long dispatch_start;
120         unsigned int allocated_slice;
121         unsigned int slice_dispatch;
122         /* time when first request from queue completed and slice started. */
123         unsigned long slice_start;
124         unsigned long slice_end;
125         long slice_resid;
126
127         /* pending metadata requests */
128         int meta_pending;
129         /* number of requests that are on the dispatch list or inside driver */
130         int dispatched;
131
132         /* io prio of this group */
133         unsigned short ioprio, org_ioprio;
134         unsigned short ioprio_class, org_ioprio_class;
135
136         pid_t pid;
137
138         u32 seek_history;
139         sector_t last_request_pos;
140
141         struct cfq_rb_root *service_tree;
142         struct cfq_queue *new_cfqq;
143         struct cfq_group *cfqg;
144         struct cfq_group *orig_cfqg;
145 };
146
147 /*
148  * First index in the service_trees.
149  * IDLE is handled separately, so it has negative index
150  */
151 enum wl_prio_t {
152         BE_WORKLOAD = 0,
153         RT_WORKLOAD = 1,
154         IDLE_WORKLOAD = 2,
155 };
156
157 /*
158  * Second index in the service_trees.
159  */
160 enum wl_type_t {
161         ASYNC_WORKLOAD = 0,
162         SYNC_NOIDLE_WORKLOAD = 1,
163         SYNC_WORKLOAD = 2
164 };
165
166 /* This is per cgroup per device grouping structure */
167 struct cfq_group {
168         /* group service_tree member */
169         struct rb_node rb_node;
170
171         /* group service_tree key */
172         u64 vdisktime;
173         unsigned int weight;
174         bool on_st;
175
176         /* number of cfqq currently on this group */
177         int nr_cfqq;
178
179         /* Per group busy queus average. Useful for workload slice calc. */
180         unsigned int busy_queues_avg[2];
181         /*
182          * rr lists of queues with requests, onle rr for each priority class.
183          * Counts are embedded in the cfq_rb_root
184          */
185         struct cfq_rb_root service_trees[2][3];
186         struct cfq_rb_root service_tree_idle;
187
188         unsigned long saved_workload_slice;
189         enum wl_type_t saved_workload;
190         enum wl_prio_t saved_serving_prio;
191         struct blkio_group blkg;
192 #ifdef CONFIG_CFQ_GROUP_IOSCHED
193         struct hlist_node cfqd_node;
194         atomic_t ref;
195 #endif
196 };
197
198 /*
199  * Per block device queue structure
200  */
201 struct cfq_data {
202         struct request_queue *queue;
203         /* Root service tree for cfq_groups */
204         struct cfq_rb_root grp_service_tree;
205         struct cfq_group root_group;
206
207         /*
208          * The priority currently being served
209          */
210         enum wl_prio_t serving_prio;
211         enum wl_type_t serving_type;
212         unsigned long workload_expires;
213         struct cfq_group *serving_group;
214         bool noidle_tree_requires_idle;
215
216         /*
217          * Each priority tree is sorted by next_request position.  These
218          * trees are used when determining if two or more queues are
219          * interleaving requests (see cfq_close_cooperator).
220          */
221         struct rb_root prio_trees[CFQ_PRIO_LISTS];
222
223         unsigned int busy_queues;
224
225         int rq_in_driver;
226         int rq_in_flight[2];
227
228         /*
229          * queue-depth detection
230          */
231         int rq_queued;
232         int hw_tag;
233         /*
234          * hw_tag can be
235          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
236          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
237          *  0 => no NCQ
238          */
239         int hw_tag_est_depth;
240         unsigned int hw_tag_samples;
241
242         /*
243          * idle window management
244          */
245         struct timer_list idle_slice_timer;
246         struct work_struct unplug_work;
247
248         struct cfq_queue *active_queue;
249         struct cfq_io_context *active_cic;
250
251         /*
252          * async queue for each priority case
253          */
254         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
255         struct cfq_queue *async_idle_cfqq;
256
257         sector_t last_position;
258
259         /*
260          * tunables, see top of file
261          */
262         unsigned int cfq_quantum;
263         unsigned int cfq_fifo_expire[2];
264         unsigned int cfq_back_penalty;
265         unsigned int cfq_back_max;
266         unsigned int cfq_slice[2];
267         unsigned int cfq_slice_async_rq;
268         unsigned int cfq_slice_idle;
269         unsigned int cfq_latency;
270         unsigned int cfq_group_isolation;
271
272         struct list_head cic_list;
273
274         /*
275          * Fallback dummy cfqq for extreme OOM conditions
276          */
277         struct cfq_queue oom_cfqq;
278
279         unsigned long last_delayed_sync;
280
281         /* List of cfq groups being managed on this device*/
282         struct hlist_head cfqg_list;
283         struct rcu_head rcu;
284 };
285
286 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
287
288 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
289                                             enum wl_prio_t prio,
290                                             enum wl_type_t type)
291 {
292         if (!cfqg)
293                 return NULL;
294
295         if (prio == IDLE_WORKLOAD)
296                 return &cfqg->service_tree_idle;
297
298         return &cfqg->service_trees[prio][type];
299 }
300
301 enum cfqq_state_flags {
302         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
303         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
304         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
305         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
306         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
307         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
308         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
309         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
310         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
311         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
312         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
313         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
314         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
315 };
316
317 #define CFQ_CFQQ_FNS(name)                                              \
318 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
319 {                                                                       \
320         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
321 }                                                                       \
322 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
323 {                                                                       \
324         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
325 }                                                                       \
326 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
327 {                                                                       \
328         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
329 }
330
331 CFQ_CFQQ_FNS(on_rr);
332 CFQ_CFQQ_FNS(wait_request);
333 CFQ_CFQQ_FNS(must_dispatch);
334 CFQ_CFQQ_FNS(must_alloc_slice);
335 CFQ_CFQQ_FNS(fifo_expire);
336 CFQ_CFQQ_FNS(idle_window);
337 CFQ_CFQQ_FNS(prio_changed);
338 CFQ_CFQQ_FNS(slice_new);
339 CFQ_CFQQ_FNS(sync);
340 CFQ_CFQQ_FNS(coop);
341 CFQ_CFQQ_FNS(split_coop);
342 CFQ_CFQQ_FNS(deep);
343 CFQ_CFQQ_FNS(wait_busy);
344 #undef CFQ_CFQQ_FNS
345
346 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
347 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
348         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
349                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
350                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
351
352 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
353         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
354                                 blkg_path(&(cfqg)->blkg), ##args);      \
355
356 #else
357 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
358         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
359 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
360 #endif
361 #define cfq_log(cfqd, fmt, args...)     \
362         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
363
364 /* Traverses through cfq group service trees */
365 #define for_each_cfqg_st(cfqg, i, j, st) \
366         for (i = 0; i <= IDLE_WORKLOAD; i++) \
367                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
368                         : &cfqg->service_tree_idle; \
369                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
370                         (i == IDLE_WORKLOAD && j == 0); \
371                         j++, st = i < IDLE_WORKLOAD ? \
372                         &cfqg->service_trees[i][j]: NULL) \
373
374
375 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
376 {
377         if (cfq_class_idle(cfqq))
378                 return IDLE_WORKLOAD;
379         if (cfq_class_rt(cfqq))
380                 return RT_WORKLOAD;
381         return BE_WORKLOAD;
382 }
383
384
385 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
386 {
387         if (!cfq_cfqq_sync(cfqq))
388                 return ASYNC_WORKLOAD;
389         if (!cfq_cfqq_idle_window(cfqq))
390                 return SYNC_NOIDLE_WORKLOAD;
391         return SYNC_WORKLOAD;
392 }
393
394 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
395                                         struct cfq_data *cfqd,
396                                         struct cfq_group *cfqg)
397 {
398         if (wl == IDLE_WORKLOAD)
399                 return cfqg->service_tree_idle.count;
400
401         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
402                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
403                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
404 }
405
406 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
407                                         struct cfq_group *cfqg)
408 {
409         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
410                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
411 }
412
413 static void cfq_dispatch_insert(struct request_queue *, struct request *);
414 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
415                                        struct io_context *, gfp_t);
416 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
417                                                 struct io_context *);
418
419 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
420                                             bool is_sync)
421 {
422         return cic->cfqq[is_sync];
423 }
424
425 static inline void cic_set_cfqq(struct cfq_io_context *cic,
426                                 struct cfq_queue *cfqq, bool is_sync)
427 {
428         cic->cfqq[is_sync] = cfqq;
429 }
430
431 /*
432  * We regard a request as SYNC, if it's either a read or has the SYNC bit
433  * set (in which case it could also be direct WRITE).
434  */
435 static inline bool cfq_bio_sync(struct bio *bio)
436 {
437         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
438 }
439
440 /*
441  * scheduler run of queue, if there are requests pending and no one in the
442  * driver that will restart queueing
443  */
444 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
445 {
446         if (cfqd->busy_queues) {
447                 cfq_log(cfqd, "schedule dispatch");
448                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
449         }
450 }
451
452 static int cfq_queue_empty(struct request_queue *q)
453 {
454         struct cfq_data *cfqd = q->elevator->elevator_data;
455
456         return !cfqd->rq_queued;
457 }
458
459 /*
460  * Scale schedule slice based on io priority. Use the sync time slice only
461  * if a queue is marked sync and has sync io queued. A sync queue with async
462  * io only, should not get full sync slice length.
463  */
464 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
465                                  unsigned short prio)
466 {
467         const int base_slice = cfqd->cfq_slice[sync];
468
469         WARN_ON(prio >= IOPRIO_BE_NR);
470
471         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
472 }
473
474 static inline int
475 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
476 {
477         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
478 }
479
480 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
481 {
482         u64 d = delta << CFQ_SERVICE_SHIFT;
483
484         d = d * BLKIO_WEIGHT_DEFAULT;
485         do_div(d, cfqg->weight);
486         return d;
487 }
488
489 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
490 {
491         s64 delta = (s64)(vdisktime - min_vdisktime);
492         if (delta > 0)
493                 min_vdisktime = vdisktime;
494
495         return min_vdisktime;
496 }
497
498 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
499 {
500         s64 delta = (s64)(vdisktime - min_vdisktime);
501         if (delta < 0)
502                 min_vdisktime = vdisktime;
503
504         return min_vdisktime;
505 }
506
507 static void update_min_vdisktime(struct cfq_rb_root *st)
508 {
509         u64 vdisktime = st->min_vdisktime;
510         struct cfq_group *cfqg;
511
512         if (st->active) {
513                 cfqg = rb_entry_cfqg(st->active);
514                 vdisktime = cfqg->vdisktime;
515         }
516
517         if (st->left) {
518                 cfqg = rb_entry_cfqg(st->left);
519                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
520         }
521
522         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
523 }
524
525 /*
526  * get averaged number of queues of RT/BE priority.
527  * average is updated, with a formula that gives more weight to higher numbers,
528  * to quickly follows sudden increases and decrease slowly
529  */
530
531 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
532                                         struct cfq_group *cfqg, bool rt)
533 {
534         unsigned min_q, max_q;
535         unsigned mult  = cfq_hist_divisor - 1;
536         unsigned round = cfq_hist_divisor / 2;
537         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
538
539         min_q = min(cfqg->busy_queues_avg[rt], busy);
540         max_q = max(cfqg->busy_queues_avg[rt], busy);
541         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
542                 cfq_hist_divisor;
543         return cfqg->busy_queues_avg[rt];
544 }
545
546 static inline unsigned
547 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
548 {
549         struct cfq_rb_root *st = &cfqd->grp_service_tree;
550
551         return cfq_target_latency * cfqg->weight / st->total_weight;
552 }
553
554 static inline void
555 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
556 {
557         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
558         if (cfqd->cfq_latency) {
559                 /*
560                  * interested queues (we consider only the ones with the same
561                  * priority class in the cfq group)
562                  */
563                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
564                                                 cfq_class_rt(cfqq));
565                 unsigned sync_slice = cfqd->cfq_slice[1];
566                 unsigned expect_latency = sync_slice * iq;
567                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
568
569                 if (expect_latency > group_slice) {
570                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
571                         /* scale low_slice according to IO priority
572                          * and sync vs async */
573                         unsigned low_slice =
574                                 min(slice, base_low_slice * slice / sync_slice);
575                         /* the adapted slice value is scaled to fit all iqs
576                          * into the target latency */
577                         slice = max(slice * group_slice / expect_latency,
578                                     low_slice);
579                 }
580         }
581         cfqq->slice_start = jiffies;
582         cfqq->slice_end = jiffies + slice;
583         cfqq->allocated_slice = slice;
584         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
585 }
586
587 /*
588  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
589  * isn't valid until the first request from the dispatch is activated
590  * and the slice time set.
591  */
592 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
593 {
594         if (cfq_cfqq_slice_new(cfqq))
595                 return 0;
596         if (time_before(jiffies, cfqq->slice_end))
597                 return 0;
598
599         return 1;
600 }
601
602 /*
603  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
604  * We choose the request that is closest to the head right now. Distance
605  * behind the head is penalized and only allowed to a certain extent.
606  */
607 static struct request *
608 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
609 {
610         sector_t s1, s2, d1 = 0, d2 = 0;
611         unsigned long back_max;
612 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
613 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
614         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
615
616         if (rq1 == NULL || rq1 == rq2)
617                 return rq2;
618         if (rq2 == NULL)
619                 return rq1;
620
621         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
622                 return rq1;
623         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
624                 return rq2;
625         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
626                 return rq1;
627         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
628                 return rq2;
629
630         s1 = blk_rq_pos(rq1);
631         s2 = blk_rq_pos(rq2);
632
633         /*
634          * by definition, 1KiB is 2 sectors
635          */
636         back_max = cfqd->cfq_back_max * 2;
637
638         /*
639          * Strict one way elevator _except_ in the case where we allow
640          * short backward seeks which are biased as twice the cost of a
641          * similar forward seek.
642          */
643         if (s1 >= last)
644                 d1 = s1 - last;
645         else if (s1 + back_max >= last)
646                 d1 = (last - s1) * cfqd->cfq_back_penalty;
647         else
648                 wrap |= CFQ_RQ1_WRAP;
649
650         if (s2 >= last)
651                 d2 = s2 - last;
652         else if (s2 + back_max >= last)
653                 d2 = (last - s2) * cfqd->cfq_back_penalty;
654         else
655                 wrap |= CFQ_RQ2_WRAP;
656
657         /* Found required data */
658
659         /*
660          * By doing switch() on the bit mask "wrap" we avoid having to
661          * check two variables for all permutations: --> faster!
662          */
663         switch (wrap) {
664         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
665                 if (d1 < d2)
666                         return rq1;
667                 else if (d2 < d1)
668                         return rq2;
669                 else {
670                         if (s1 >= s2)
671                                 return rq1;
672                         else
673                                 return rq2;
674                 }
675
676         case CFQ_RQ2_WRAP:
677                 return rq1;
678         case CFQ_RQ1_WRAP:
679                 return rq2;
680         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
681         default:
682                 /*
683                  * Since both rqs are wrapped,
684                  * start with the one that's further behind head
685                  * (--> only *one* back seek required),
686                  * since back seek takes more time than forward.
687                  */
688                 if (s1 <= s2)
689                         return rq1;
690                 else
691                         return rq2;
692         }
693 }
694
695 /*
696  * The below is leftmost cache rbtree addon
697  */
698 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
699 {
700         /* Service tree is empty */
701         if (!root->count)
702                 return NULL;
703
704         if (!root->left)
705                 root->left = rb_first(&root->rb);
706
707         if (root->left)
708                 return rb_entry(root->left, struct cfq_queue, rb_node);
709
710         return NULL;
711 }
712
713 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
714 {
715         if (!root->left)
716                 root->left = rb_first(&root->rb);
717
718         if (root->left)
719                 return rb_entry_cfqg(root->left);
720
721         return NULL;
722 }
723
724 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
725 {
726         rb_erase(n, root);
727         RB_CLEAR_NODE(n);
728 }
729
730 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
731 {
732         if (root->left == n)
733                 root->left = NULL;
734         rb_erase_init(n, &root->rb);
735         --root->count;
736 }
737
738 /*
739  * would be nice to take fifo expire time into account as well
740  */
741 static struct request *
742 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
743                   struct request *last)
744 {
745         struct rb_node *rbnext = rb_next(&last->rb_node);
746         struct rb_node *rbprev = rb_prev(&last->rb_node);
747         struct request *next = NULL, *prev = NULL;
748
749         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
750
751         if (rbprev)
752                 prev = rb_entry_rq(rbprev);
753
754         if (rbnext)
755                 next = rb_entry_rq(rbnext);
756         else {
757                 rbnext = rb_first(&cfqq->sort_list);
758                 if (rbnext && rbnext != &last->rb_node)
759                         next = rb_entry_rq(rbnext);
760         }
761
762         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
763 }
764
765 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
766                                       struct cfq_queue *cfqq)
767 {
768         /*
769          * just an approximation, should be ok.
770          */
771         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
772                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
773 }
774
775 static inline s64
776 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
777 {
778         return cfqg->vdisktime - st->min_vdisktime;
779 }
780
781 static void
782 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
783 {
784         struct rb_node **node = &st->rb.rb_node;
785         struct rb_node *parent = NULL;
786         struct cfq_group *__cfqg;
787         s64 key = cfqg_key(st, cfqg);
788         int left = 1;
789
790         while (*node != NULL) {
791                 parent = *node;
792                 __cfqg = rb_entry_cfqg(parent);
793
794                 if (key < cfqg_key(st, __cfqg))
795                         node = &parent->rb_left;
796                 else {
797                         node = &parent->rb_right;
798                         left = 0;
799                 }
800         }
801
802         if (left)
803                 st->left = &cfqg->rb_node;
804
805         rb_link_node(&cfqg->rb_node, parent, node);
806         rb_insert_color(&cfqg->rb_node, &st->rb);
807 }
808
809 static void
810 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
811 {
812         struct cfq_rb_root *st = &cfqd->grp_service_tree;
813         struct cfq_group *__cfqg;
814         struct rb_node *n;
815
816         cfqg->nr_cfqq++;
817         if (cfqg->on_st)
818                 return;
819
820         /*
821          * Currently put the group at the end. Later implement something
822          * so that groups get lesser vtime based on their weights, so that
823          * if group does not loose all if it was not continously backlogged.
824          */
825         n = rb_last(&st->rb);
826         if (n) {
827                 __cfqg = rb_entry_cfqg(n);
828                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
829         } else
830                 cfqg->vdisktime = st->min_vdisktime;
831
832         __cfq_group_service_tree_add(st, cfqg);
833         cfqg->on_st = true;
834         st->total_weight += cfqg->weight;
835 }
836
837 static void
838 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
839 {
840         struct cfq_rb_root *st = &cfqd->grp_service_tree;
841
842         if (st->active == &cfqg->rb_node)
843                 st->active = NULL;
844
845         BUG_ON(cfqg->nr_cfqq < 1);
846         cfqg->nr_cfqq--;
847
848         /* If there are other cfq queues under this group, don't delete it */
849         if (cfqg->nr_cfqq)
850                 return;
851
852         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
853         cfqg->on_st = false;
854         st->total_weight -= cfqg->weight;
855         if (!RB_EMPTY_NODE(&cfqg->rb_node))
856                 cfq_rb_erase(&cfqg->rb_node, st);
857         cfqg->saved_workload_slice = 0;
858         blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
859 }
860
861 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
862 {
863         unsigned int slice_used;
864
865         /*
866          * Queue got expired before even a single request completed or
867          * got expired immediately after first request completion.
868          */
869         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
870                 /*
871                  * Also charge the seek time incurred to the group, otherwise
872                  * if there are mutiple queues in the group, each can dispatch
873                  * a single request on seeky media and cause lots of seek time
874                  * and group will never know it.
875                  */
876                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
877                                         1);
878         } else {
879                 slice_used = jiffies - cfqq->slice_start;
880                 if (slice_used > cfqq->allocated_slice)
881                         slice_used = cfqq->allocated_slice;
882         }
883
884         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
885         return slice_used;
886 }
887
888 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
889                                 struct cfq_queue *cfqq, bool forced)
890 {
891         struct cfq_rb_root *st = &cfqd->grp_service_tree;
892         unsigned int used_sl, charge_sl;
893         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
894                         - cfqg->service_tree_idle.count;
895
896         BUG_ON(nr_sync < 0);
897         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
898
899         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
900                 charge_sl = cfqq->allocated_slice;
901
902         /* Can't update vdisktime while group is on service tree */
903         cfq_rb_erase(&cfqg->rb_node, st);
904         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
905         __cfq_group_service_tree_add(st, cfqg);
906
907         /* This group is being expired. Save the context */
908         if (time_after(cfqd->workload_expires, jiffies)) {
909                 cfqg->saved_workload_slice = cfqd->workload_expires
910                                                 - jiffies;
911                 cfqg->saved_workload = cfqd->serving_type;
912                 cfqg->saved_serving_prio = cfqd->serving_prio;
913         } else
914                 cfqg->saved_workload_slice = 0;
915
916         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
917                                         st->min_vdisktime);
918         blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
919         blkiocg_set_start_empty_time(&cfqg->blkg, forced);
920 }
921
922 #ifdef CONFIG_CFQ_GROUP_IOSCHED
923 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
924 {
925         if (blkg)
926                 return container_of(blkg, struct cfq_group, blkg);
927         return NULL;
928 }
929
930 void
931 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
932 {
933         cfqg_of_blkg(blkg)->weight = weight;
934 }
935
936 static struct cfq_group *
937 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
938 {
939         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
940         struct cfq_group *cfqg = NULL;
941         void *key = cfqd;
942         int i, j;
943         struct cfq_rb_root *st;
944         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
945         unsigned int major, minor;
946
947         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
948         if (cfqg || !create)
949                 goto done;
950
951         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
952         if (!cfqg)
953                 goto done;
954
955         cfqg->weight = blkcg->weight;
956         for_each_cfqg_st(cfqg, i, j, st)
957                 *st = CFQ_RB_ROOT;
958         RB_CLEAR_NODE(&cfqg->rb_node);
959         blkio_group_init(&cfqg->blkg);
960
961         /*
962          * Take the initial reference that will be released on destroy
963          * This can be thought of a joint reference by cgroup and
964          * elevator which will be dropped by either elevator exit
965          * or cgroup deletion path depending on who is exiting first.
966          */
967         atomic_set(&cfqg->ref, 1);
968
969         /* Add group onto cgroup list */
970         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
971         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
972                                         MKDEV(major, minor));
973
974         /* Add group on cfqd list */
975         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
976
977 done:
978         return cfqg;
979 }
980
981 /*
982  * Search for the cfq group current task belongs to. If create = 1, then also
983  * create the cfq group if it does not exist. request_queue lock must be held.
984  */
985 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
986 {
987         struct cgroup *cgroup;
988         struct cfq_group *cfqg = NULL;
989
990         rcu_read_lock();
991         cgroup = task_cgroup(current, blkio_subsys_id);
992         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
993         if (!cfqg && create)
994                 cfqg = &cfqd->root_group;
995         rcu_read_unlock();
996         return cfqg;
997 }
998
999 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1000 {
1001         /* Currently, all async queues are mapped to root group */
1002         if (!cfq_cfqq_sync(cfqq))
1003                 cfqg = &cfqq->cfqd->root_group;
1004
1005         cfqq->cfqg = cfqg;
1006         /* cfqq reference on cfqg */
1007         atomic_inc(&cfqq->cfqg->ref);
1008 }
1009
1010 static void cfq_put_cfqg(struct cfq_group *cfqg)
1011 {
1012         struct cfq_rb_root *st;
1013         int i, j;
1014
1015         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1016         if (!atomic_dec_and_test(&cfqg->ref))
1017                 return;
1018         for_each_cfqg_st(cfqg, i, j, st)
1019                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1020         kfree(cfqg);
1021 }
1022
1023 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1024 {
1025         /* Something wrong if we are trying to remove same group twice */
1026         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1027
1028         hlist_del_init(&cfqg->cfqd_node);
1029
1030         /*
1031          * Put the reference taken at the time of creation so that when all
1032          * queues are gone, group can be destroyed.
1033          */
1034         cfq_put_cfqg(cfqg);
1035 }
1036
1037 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1038 {
1039         struct hlist_node *pos, *n;
1040         struct cfq_group *cfqg;
1041
1042         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1043                 /*
1044                  * If cgroup removal path got to blk_group first and removed
1045                  * it from cgroup list, then it will take care of destroying
1046                  * cfqg also.
1047                  */
1048                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1049                         cfq_destroy_cfqg(cfqd, cfqg);
1050         }
1051 }
1052
1053 /*
1054  * Blk cgroup controller notification saying that blkio_group object is being
1055  * delinked as associated cgroup object is going away. That also means that
1056  * no new IO will come in this group. So get rid of this group as soon as
1057  * any pending IO in the group is finished.
1058  *
1059  * This function is called under rcu_read_lock(). key is the rcu protected
1060  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1061  * read lock.
1062  *
1063  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1064  * it should not be NULL as even if elevator was exiting, cgroup deltion
1065  * path got to it first.
1066  */
1067 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1068 {
1069         unsigned long  flags;
1070         struct cfq_data *cfqd = key;
1071
1072         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1073         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1074         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1075 }
1076
1077 #else /* GROUP_IOSCHED */
1078 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1079 {
1080         return &cfqd->root_group;
1081 }
1082 static inline void
1083 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1084         cfqq->cfqg = cfqg;
1085 }
1086
1087 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1088 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1089
1090 #endif /* GROUP_IOSCHED */
1091
1092 /*
1093  * The cfqd->service_trees holds all pending cfq_queue's that have
1094  * requests waiting to be processed. It is sorted in the order that
1095  * we will service the queues.
1096  */
1097 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1098                                  bool add_front)
1099 {
1100         struct rb_node **p, *parent;
1101         struct cfq_queue *__cfqq;
1102         unsigned long rb_key;
1103         struct cfq_rb_root *service_tree;
1104         int left;
1105         int new_cfqq = 1;
1106         int group_changed = 0;
1107
1108 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1109         if (!cfqd->cfq_group_isolation
1110             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1111             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1112                 /* Move this cfq to root group */
1113                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1114                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1115                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1116                 cfqq->orig_cfqg = cfqq->cfqg;
1117                 cfqq->cfqg = &cfqd->root_group;
1118                 atomic_inc(&cfqd->root_group.ref);
1119                 group_changed = 1;
1120         } else if (!cfqd->cfq_group_isolation
1121                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1122                 /* cfqq is sequential now needs to go to its original group */
1123                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1124                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1125                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1126                 cfq_put_cfqg(cfqq->cfqg);
1127                 cfqq->cfqg = cfqq->orig_cfqg;
1128                 cfqq->orig_cfqg = NULL;
1129                 group_changed = 1;
1130                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1131         }
1132 #endif
1133
1134         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1135                                                 cfqq_type(cfqq));
1136         if (cfq_class_idle(cfqq)) {
1137                 rb_key = CFQ_IDLE_DELAY;
1138                 parent = rb_last(&service_tree->rb);
1139                 if (parent && parent != &cfqq->rb_node) {
1140                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1141                         rb_key += __cfqq->rb_key;
1142                 } else
1143                         rb_key += jiffies;
1144         } else if (!add_front) {
1145                 /*
1146                  * Get our rb key offset. Subtract any residual slice
1147                  * value carried from last service. A negative resid
1148                  * count indicates slice overrun, and this should position
1149                  * the next service time further away in the tree.
1150                  */
1151                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1152                 rb_key -= cfqq->slice_resid;
1153                 cfqq->slice_resid = 0;
1154         } else {
1155                 rb_key = -HZ;
1156                 __cfqq = cfq_rb_first(service_tree);
1157                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1158         }
1159
1160         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1161                 new_cfqq = 0;
1162                 /*
1163                  * same position, nothing more to do
1164                  */
1165                 if (rb_key == cfqq->rb_key &&
1166                     cfqq->service_tree == service_tree)
1167                         return;
1168
1169                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1170                 cfqq->service_tree = NULL;
1171         }
1172
1173         left = 1;
1174         parent = NULL;
1175         cfqq->service_tree = service_tree;
1176         p = &service_tree->rb.rb_node;
1177         while (*p) {
1178                 struct rb_node **n;
1179
1180                 parent = *p;
1181                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1182
1183                 /*
1184                  * sort by key, that represents service time.
1185                  */
1186                 if (time_before(rb_key, __cfqq->rb_key))
1187                         n = &(*p)->rb_left;
1188                 else {
1189                         n = &(*p)->rb_right;
1190                         left = 0;
1191                 }
1192
1193                 p = n;
1194         }
1195
1196         if (left)
1197                 service_tree->left = &cfqq->rb_node;
1198
1199         cfqq->rb_key = rb_key;
1200         rb_link_node(&cfqq->rb_node, parent, p);
1201         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1202         service_tree->count++;
1203         if ((add_front || !new_cfqq) && !group_changed)
1204                 return;
1205         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1206 }
1207
1208 static struct cfq_queue *
1209 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1210                      sector_t sector, struct rb_node **ret_parent,
1211                      struct rb_node ***rb_link)
1212 {
1213         struct rb_node **p, *parent;
1214         struct cfq_queue *cfqq = NULL;
1215
1216         parent = NULL;
1217         p = &root->rb_node;
1218         while (*p) {
1219                 struct rb_node **n;
1220
1221                 parent = *p;
1222                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1223
1224                 /*
1225                  * Sort strictly based on sector.  Smallest to the left,
1226                  * largest to the right.
1227                  */
1228                 if (sector > blk_rq_pos(cfqq->next_rq))
1229                         n = &(*p)->rb_right;
1230                 else if (sector < blk_rq_pos(cfqq->next_rq))
1231                         n = &(*p)->rb_left;
1232                 else
1233                         break;
1234                 p = n;
1235                 cfqq = NULL;
1236         }
1237
1238         *ret_parent = parent;
1239         if (rb_link)
1240                 *rb_link = p;
1241         return cfqq;
1242 }
1243
1244 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1245 {
1246         struct rb_node **p, *parent;
1247         struct cfq_queue *__cfqq;
1248
1249         if (cfqq->p_root) {
1250                 rb_erase(&cfqq->p_node, cfqq->p_root);
1251                 cfqq->p_root = NULL;
1252         }
1253
1254         if (cfq_class_idle(cfqq))
1255                 return;
1256         if (!cfqq->next_rq)
1257                 return;
1258
1259         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1260         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1261                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1262         if (!__cfqq) {
1263                 rb_link_node(&cfqq->p_node, parent, p);
1264                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1265         } else
1266                 cfqq->p_root = NULL;
1267 }
1268
1269 /*
1270  * Update cfqq's position in the service tree.
1271  */
1272 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1273 {
1274         /*
1275          * Resorting requires the cfqq to be on the RR list already.
1276          */
1277         if (cfq_cfqq_on_rr(cfqq)) {
1278                 cfq_service_tree_add(cfqd, cfqq, 0);
1279                 cfq_prio_tree_add(cfqd, cfqq);
1280         }
1281 }
1282
1283 /*
1284  * add to busy list of queues for service, trying to be fair in ordering
1285  * the pending list according to last request service
1286  */
1287 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1288 {
1289         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1290         BUG_ON(cfq_cfqq_on_rr(cfqq));
1291         cfq_mark_cfqq_on_rr(cfqq);
1292         cfqd->busy_queues++;
1293
1294         cfq_resort_rr_list(cfqd, cfqq);
1295 }
1296
1297 /*
1298  * Called when the cfqq no longer has requests pending, remove it from
1299  * the service tree.
1300  */
1301 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1302 {
1303         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1304         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1305         cfq_clear_cfqq_on_rr(cfqq);
1306
1307         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1308                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1309                 cfqq->service_tree = NULL;
1310         }
1311         if (cfqq->p_root) {
1312                 rb_erase(&cfqq->p_node, cfqq->p_root);
1313                 cfqq->p_root = NULL;
1314         }
1315
1316         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1317         BUG_ON(!cfqd->busy_queues);
1318         cfqd->busy_queues--;
1319 }
1320
1321 /*
1322  * rb tree support functions
1323  */
1324 static void cfq_del_rq_rb(struct request *rq)
1325 {
1326         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1327         const int sync = rq_is_sync(rq);
1328
1329         BUG_ON(!cfqq->queued[sync]);
1330         cfqq->queued[sync]--;
1331
1332         elv_rb_del(&cfqq->sort_list, rq);
1333
1334         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1335                 /*
1336                  * Queue will be deleted from service tree when we actually
1337                  * expire it later. Right now just remove it from prio tree
1338                  * as it is empty.
1339                  */
1340                 if (cfqq->p_root) {
1341                         rb_erase(&cfqq->p_node, cfqq->p_root);
1342                         cfqq->p_root = NULL;
1343                 }
1344         }
1345 }
1346
1347 static void cfq_add_rq_rb(struct request *rq)
1348 {
1349         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1350         struct cfq_data *cfqd = cfqq->cfqd;
1351         struct request *__alias, *prev;
1352
1353         cfqq->queued[rq_is_sync(rq)]++;
1354
1355         /*
1356          * looks a little odd, but the first insert might return an alias.
1357          * if that happens, put the alias on the dispatch list
1358          */
1359         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1360                 cfq_dispatch_insert(cfqd->queue, __alias);
1361
1362         if (!cfq_cfqq_on_rr(cfqq))
1363                 cfq_add_cfqq_rr(cfqd, cfqq);
1364
1365         /*
1366          * check if this request is a better next-serve candidate
1367          */
1368         prev = cfqq->next_rq;
1369         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1370
1371         /*
1372          * adjust priority tree position, if ->next_rq changes
1373          */
1374         if (prev != cfqq->next_rq)
1375                 cfq_prio_tree_add(cfqd, cfqq);
1376
1377         BUG_ON(!cfqq->next_rq);
1378 }
1379
1380 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1381 {
1382         elv_rb_del(&cfqq->sort_list, rq);
1383         cfqq->queued[rq_is_sync(rq)]--;
1384         blkiocg_update_request_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1385                                                 rq_is_sync(rq));
1386         cfq_add_rq_rb(rq);
1387         blkiocg_update_request_add_stats(
1388                         &cfqq->cfqg->blkg, &cfqq->cfqd->serving_group->blkg,
1389                         rq_data_dir(rq), rq_is_sync(rq));
1390 }
1391
1392 static struct request *
1393 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1394 {
1395         struct task_struct *tsk = current;
1396         struct cfq_io_context *cic;
1397         struct cfq_queue *cfqq;
1398
1399         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1400         if (!cic)
1401                 return NULL;
1402
1403         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1404         if (cfqq) {
1405                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1406
1407                 return elv_rb_find(&cfqq->sort_list, sector);
1408         }
1409
1410         return NULL;
1411 }
1412
1413 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1414 {
1415         struct cfq_data *cfqd = q->elevator->elevator_data;
1416
1417         cfqd->rq_in_driver++;
1418         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1419                                                 cfqd->rq_in_driver);
1420
1421         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1422 }
1423
1424 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1425 {
1426         struct cfq_data *cfqd = q->elevator->elevator_data;
1427
1428         WARN_ON(!cfqd->rq_in_driver);
1429         cfqd->rq_in_driver--;
1430         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1431                                                 cfqd->rq_in_driver);
1432 }
1433
1434 static void cfq_remove_request(struct request *rq)
1435 {
1436         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1437
1438         if (cfqq->next_rq == rq)
1439                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1440
1441         list_del_init(&rq->queuelist);
1442         cfq_del_rq_rb(rq);
1443
1444         cfqq->cfqd->rq_queued--;
1445         blkiocg_update_request_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1446                                                 rq_is_sync(rq));
1447         if (rq_is_meta(rq)) {
1448                 WARN_ON(!cfqq->meta_pending);
1449                 cfqq->meta_pending--;
1450         }
1451 }
1452
1453 static int cfq_merge(struct request_queue *q, struct request **req,
1454                      struct bio *bio)
1455 {
1456         struct cfq_data *cfqd = q->elevator->elevator_data;
1457         struct request *__rq;
1458
1459         __rq = cfq_find_rq_fmerge(cfqd, bio);
1460         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1461                 *req = __rq;
1462                 return ELEVATOR_FRONT_MERGE;
1463         }
1464
1465         return ELEVATOR_NO_MERGE;
1466 }
1467
1468 static void cfq_merged_request(struct request_queue *q, struct request *req,
1469                                int type)
1470 {
1471         if (type == ELEVATOR_FRONT_MERGE) {
1472                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1473
1474                 cfq_reposition_rq_rb(cfqq, req);
1475         }
1476 }
1477
1478 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1479                                 struct bio *bio)
1480 {
1481         struct cfq_queue *cfqq = RQ_CFQQ(req);
1482         blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, bio_data_dir(bio),
1483                                         cfq_bio_sync(bio));
1484 }
1485
1486 static void
1487 cfq_merged_requests(struct request_queue *q, struct request *rq,
1488                     struct request *next)
1489 {
1490         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1491         /*
1492          * reposition in fifo if next is older than rq
1493          */
1494         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1495             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1496                 list_move(&rq->queuelist, &next->queuelist);
1497                 rq_set_fifo_time(rq, rq_fifo_time(next));
1498         }
1499
1500         if (cfqq->next_rq == next)
1501                 cfqq->next_rq = rq;
1502         cfq_remove_request(next);
1503         blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, rq_data_dir(next),
1504                                         rq_is_sync(next));
1505 }
1506
1507 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1508                            struct bio *bio)
1509 {
1510         struct cfq_data *cfqd = q->elevator->elevator_data;
1511         struct cfq_io_context *cic;
1512         struct cfq_queue *cfqq;
1513
1514         /*
1515          * Disallow merge of a sync bio into an async request.
1516          */
1517         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1518                 return false;
1519
1520         /*
1521          * Lookup the cfqq that this bio will be queued with. Allow
1522          * merge only if rq is queued there.
1523          */
1524         cic = cfq_cic_lookup(cfqd, current->io_context);
1525         if (!cic)
1526                 return false;
1527
1528         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1529         return cfqq == RQ_CFQQ(rq);
1530 }
1531
1532 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1533 {
1534         del_timer(&cfqd->idle_slice_timer);
1535         blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1536 }
1537
1538 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1539                                    struct cfq_queue *cfqq)
1540 {
1541         if (cfqq) {
1542                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1543                                 cfqd->serving_prio, cfqd->serving_type);
1544                 blkiocg_update_set_active_queue_stats(&cfqq->cfqg->blkg);
1545                 cfqq->slice_start = 0;
1546                 cfqq->dispatch_start = jiffies;
1547                 cfqq->allocated_slice = 0;
1548                 cfqq->slice_end = 0;
1549                 cfqq->slice_dispatch = 0;
1550
1551                 cfq_clear_cfqq_wait_request(cfqq);
1552                 cfq_clear_cfqq_must_dispatch(cfqq);
1553                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1554                 cfq_clear_cfqq_fifo_expire(cfqq);
1555                 cfq_mark_cfqq_slice_new(cfqq);
1556
1557                 cfq_del_timer(cfqd, cfqq);
1558         }
1559
1560         cfqd->active_queue = cfqq;
1561 }
1562
1563 /*
1564  * current cfqq expired its slice (or was too idle), select new one
1565  */
1566 static void
1567 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1568                     bool timed_out, bool forced)
1569 {
1570         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1571
1572         if (cfq_cfqq_wait_request(cfqq))
1573                 cfq_del_timer(cfqd, cfqq);
1574
1575         cfq_clear_cfqq_wait_request(cfqq);
1576         cfq_clear_cfqq_wait_busy(cfqq);
1577
1578         /*
1579          * If this cfqq is shared between multiple processes, check to
1580          * make sure that those processes are still issuing I/Os within
1581          * the mean seek distance.  If not, it may be time to break the
1582          * queues apart again.
1583          */
1584         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1585                 cfq_mark_cfqq_split_coop(cfqq);
1586
1587         /*
1588          * store what was left of this slice, if the queue idled/timed out
1589          */
1590         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1591                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1592                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1593         }
1594
1595         cfq_group_served(cfqd, cfqq->cfqg, cfqq, forced);
1596
1597         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1598                 cfq_del_cfqq_rr(cfqd, cfqq);
1599
1600         cfq_resort_rr_list(cfqd, cfqq);
1601
1602         if (cfqq == cfqd->active_queue)
1603                 cfqd->active_queue = NULL;
1604
1605         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1606                 cfqd->grp_service_tree.active = NULL;
1607
1608         if (cfqd->active_cic) {
1609                 put_io_context(cfqd->active_cic->ioc);
1610                 cfqd->active_cic = NULL;
1611         }
1612 }
1613
1614 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out,
1615                                         bool forced)
1616 {
1617         struct cfq_queue *cfqq = cfqd->active_queue;
1618
1619         if (cfqq)
1620                 __cfq_slice_expired(cfqd, cfqq, timed_out, forced);
1621 }
1622
1623 /*
1624  * Get next queue for service. Unless we have a queue preemption,
1625  * we'll simply select the first cfqq in the service tree.
1626  */
1627 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1628 {
1629         struct cfq_rb_root *service_tree =
1630                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1631                                         cfqd->serving_type);
1632
1633         if (!cfqd->rq_queued)
1634                 return NULL;
1635
1636         /* There is nothing to dispatch */
1637         if (!service_tree)
1638                 return NULL;
1639         if (RB_EMPTY_ROOT(&service_tree->rb))
1640                 return NULL;
1641         return cfq_rb_first(service_tree);
1642 }
1643
1644 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1645 {
1646         struct cfq_group *cfqg;
1647         struct cfq_queue *cfqq;
1648         int i, j;
1649         struct cfq_rb_root *st;
1650
1651         if (!cfqd->rq_queued)
1652                 return NULL;
1653
1654         cfqg = cfq_get_next_cfqg(cfqd);
1655         if (!cfqg)
1656                 return NULL;
1657
1658         for_each_cfqg_st(cfqg, i, j, st)
1659                 if ((cfqq = cfq_rb_first(st)) != NULL)
1660                         return cfqq;
1661         return NULL;
1662 }
1663
1664 /*
1665  * Get and set a new active queue for service.
1666  */
1667 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1668                                               struct cfq_queue *cfqq)
1669 {
1670         if (!cfqq)
1671                 cfqq = cfq_get_next_queue(cfqd);
1672
1673         __cfq_set_active_queue(cfqd, cfqq);
1674         return cfqq;
1675 }
1676
1677 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1678                                           struct request *rq)
1679 {
1680         if (blk_rq_pos(rq) >= cfqd->last_position)
1681                 return blk_rq_pos(rq) - cfqd->last_position;
1682         else
1683                 return cfqd->last_position - blk_rq_pos(rq);
1684 }
1685
1686 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1687                                struct request *rq)
1688 {
1689         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1690 }
1691
1692 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1693                                     struct cfq_queue *cur_cfqq)
1694 {
1695         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1696         struct rb_node *parent, *node;
1697         struct cfq_queue *__cfqq;
1698         sector_t sector = cfqd->last_position;
1699
1700         if (RB_EMPTY_ROOT(root))
1701                 return NULL;
1702
1703         /*
1704          * First, if we find a request starting at the end of the last
1705          * request, choose it.
1706          */
1707         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1708         if (__cfqq)
1709                 return __cfqq;
1710
1711         /*
1712          * If the exact sector wasn't found, the parent of the NULL leaf
1713          * will contain the closest sector.
1714          */
1715         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1716         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1717                 return __cfqq;
1718
1719         if (blk_rq_pos(__cfqq->next_rq) < sector)
1720                 node = rb_next(&__cfqq->p_node);
1721         else
1722                 node = rb_prev(&__cfqq->p_node);
1723         if (!node)
1724                 return NULL;
1725
1726         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1727         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1728                 return __cfqq;
1729
1730         return NULL;
1731 }
1732
1733 /*
1734  * cfqd - obvious
1735  * cur_cfqq - passed in so that we don't decide that the current queue is
1736  *            closely cooperating with itself.
1737  *
1738  * So, basically we're assuming that that cur_cfqq has dispatched at least
1739  * one request, and that cfqd->last_position reflects a position on the disk
1740  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1741  * assumption.
1742  */
1743 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1744                                               struct cfq_queue *cur_cfqq)
1745 {
1746         struct cfq_queue *cfqq;
1747
1748         if (cfq_class_idle(cur_cfqq))
1749                 return NULL;
1750         if (!cfq_cfqq_sync(cur_cfqq))
1751                 return NULL;
1752         if (CFQQ_SEEKY(cur_cfqq))
1753                 return NULL;
1754
1755         /*
1756          * Don't search priority tree if it's the only queue in the group.
1757          */
1758         if (cur_cfqq->cfqg->nr_cfqq == 1)
1759                 return NULL;
1760
1761         /*
1762          * We should notice if some of the queues are cooperating, eg
1763          * working closely on the same area of the disk. In that case,
1764          * we can group them together and don't waste time idling.
1765          */
1766         cfqq = cfqq_close(cfqd, cur_cfqq);
1767         if (!cfqq)
1768                 return NULL;
1769
1770         /* If new queue belongs to different cfq_group, don't choose it */
1771         if (cur_cfqq->cfqg != cfqq->cfqg)
1772                 return NULL;
1773
1774         /*
1775          * It only makes sense to merge sync queues.
1776          */
1777         if (!cfq_cfqq_sync(cfqq))
1778                 return NULL;
1779         if (CFQQ_SEEKY(cfqq))
1780                 return NULL;
1781
1782         /*
1783          * Do not merge queues of different priority classes
1784          */
1785         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1786                 return NULL;
1787
1788         return cfqq;
1789 }
1790
1791 /*
1792  * Determine whether we should enforce idle window for this queue.
1793  */
1794
1795 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1796 {
1797         enum wl_prio_t prio = cfqq_prio(cfqq);
1798         struct cfq_rb_root *service_tree = cfqq->service_tree;
1799
1800         BUG_ON(!service_tree);
1801         BUG_ON(!service_tree->count);
1802
1803         /* We never do for idle class queues. */
1804         if (prio == IDLE_WORKLOAD)
1805                 return false;
1806
1807         /* We do for queues that were marked with idle window flag. */
1808         if (cfq_cfqq_idle_window(cfqq) &&
1809            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1810                 return true;
1811
1812         /*
1813          * Otherwise, we do only if they are the last ones
1814          * in their service tree.
1815          */
1816         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1817                 return 1;
1818         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1819                         service_tree->count);
1820         return 0;
1821 }
1822
1823 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1824 {
1825         struct cfq_queue *cfqq = cfqd->active_queue;
1826         struct cfq_io_context *cic;
1827         unsigned long sl;
1828
1829         /*
1830          * SSD device without seek penalty, disable idling. But only do so
1831          * for devices that support queuing, otherwise we still have a problem
1832          * with sync vs async workloads.
1833          */
1834         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1835                 return;
1836
1837         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1838         WARN_ON(cfq_cfqq_slice_new(cfqq));
1839
1840         /*
1841          * idle is disabled, either manually or by past process history
1842          */
1843         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1844                 return;
1845
1846         /*
1847          * still active requests from this queue, don't idle
1848          */
1849         if (cfqq->dispatched)
1850                 return;
1851
1852         /*
1853          * task has exited, don't wait
1854          */
1855         cic = cfqd->active_cic;
1856         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1857                 return;
1858
1859         /*
1860          * If our average think time is larger than the remaining time
1861          * slice, then don't idle. This avoids overrunning the allotted
1862          * time slice.
1863          */
1864         if (sample_valid(cic->ttime_samples) &&
1865             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1866                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1867                                 cic->ttime_mean);
1868                 return;
1869         }
1870
1871         cfq_mark_cfqq_wait_request(cfqq);
1872
1873         sl = cfqd->cfq_slice_idle;
1874
1875         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1876         blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1877         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1878 }
1879
1880 /*
1881  * Move request from internal lists to the request queue dispatch list.
1882  */
1883 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1884 {
1885         struct cfq_data *cfqd = q->elevator->elevator_data;
1886         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1887
1888         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1889
1890         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1891         cfq_remove_request(rq);
1892         cfqq->dispatched++;
1893         elv_dispatch_sort(q, rq);
1894
1895         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1896         blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1897                                         rq_data_dir(rq), rq_is_sync(rq));
1898 }
1899
1900 /*
1901  * return expired entry, or NULL to just start from scratch in rbtree
1902  */
1903 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1904 {
1905         struct request *rq = NULL;
1906
1907         if (cfq_cfqq_fifo_expire(cfqq))
1908                 return NULL;
1909
1910         cfq_mark_cfqq_fifo_expire(cfqq);
1911
1912         if (list_empty(&cfqq->fifo))
1913                 return NULL;
1914
1915         rq = rq_entry_fifo(cfqq->fifo.next);
1916         if (time_before(jiffies, rq_fifo_time(rq)))
1917                 rq = NULL;
1918
1919         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1920         return rq;
1921 }
1922
1923 static inline int
1924 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1925 {
1926         const int base_rq = cfqd->cfq_slice_async_rq;
1927
1928         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1929
1930         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1931 }
1932
1933 /*
1934  * Must be called with the queue_lock held.
1935  */
1936 static int cfqq_process_refs(struct cfq_queue *cfqq)
1937 {
1938         int process_refs, io_refs;
1939
1940         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1941         process_refs = atomic_read(&cfqq->ref) - io_refs;
1942         BUG_ON(process_refs < 0);
1943         return process_refs;
1944 }
1945
1946 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1947 {
1948         int process_refs, new_process_refs;
1949         struct cfq_queue *__cfqq;
1950
1951         /* Avoid a circular list and skip interim queue merges */
1952         while ((__cfqq = new_cfqq->new_cfqq)) {
1953                 if (__cfqq == cfqq)
1954                         return;
1955                 new_cfqq = __cfqq;
1956         }
1957
1958         process_refs = cfqq_process_refs(cfqq);
1959         /*
1960          * If the process for the cfqq has gone away, there is no
1961          * sense in merging the queues.
1962          */
1963         if (process_refs == 0)
1964                 return;
1965
1966         /*
1967          * Merge in the direction of the lesser amount of work.
1968          */
1969         new_process_refs = cfqq_process_refs(new_cfqq);
1970         if (new_process_refs >= process_refs) {
1971                 cfqq->new_cfqq = new_cfqq;
1972                 atomic_add(process_refs, &new_cfqq->ref);
1973         } else {
1974                 new_cfqq->new_cfqq = cfqq;
1975                 atomic_add(new_process_refs, &cfqq->ref);
1976         }
1977 }
1978
1979 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1980                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1981 {
1982         struct cfq_queue *queue;
1983         int i;
1984         bool key_valid = false;
1985         unsigned long lowest_key = 0;
1986         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1987
1988         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1989                 /* select the one with lowest rb_key */
1990                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1991                 if (queue &&
1992                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1993                         lowest_key = queue->rb_key;
1994                         cur_best = i;
1995                         key_valid = true;
1996                 }
1997         }
1998
1999         return cur_best;
2000 }
2001
2002 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2003 {
2004         unsigned slice;
2005         unsigned count;
2006         struct cfq_rb_root *st;
2007         unsigned group_slice;
2008
2009         if (!cfqg) {
2010                 cfqd->serving_prio = IDLE_WORKLOAD;
2011                 cfqd->workload_expires = jiffies + 1;
2012                 return;
2013         }
2014
2015         /* Choose next priority. RT > BE > IDLE */
2016         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2017                 cfqd->serving_prio = RT_WORKLOAD;
2018         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2019                 cfqd->serving_prio = BE_WORKLOAD;
2020         else {
2021                 cfqd->serving_prio = IDLE_WORKLOAD;
2022                 cfqd->workload_expires = jiffies + 1;
2023                 return;
2024         }
2025
2026         /*
2027          * For RT and BE, we have to choose also the type
2028          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2029          * expiration time
2030          */
2031         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2032         count = st->count;
2033
2034         /*
2035          * check workload expiration, and that we still have other queues ready
2036          */
2037         if (count && !time_after(jiffies, cfqd->workload_expires))
2038                 return;
2039
2040         /* otherwise select new workload type */
2041         cfqd->serving_type =
2042                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2043         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2044         count = st->count;
2045
2046         /*
2047          * the workload slice is computed as a fraction of target latency
2048          * proportional to the number of queues in that workload, over
2049          * all the queues in the same priority class
2050          */
2051         group_slice = cfq_group_slice(cfqd, cfqg);
2052
2053         slice = group_slice * count /
2054                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2055                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2056
2057         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2058                 unsigned int tmp;
2059
2060                 /*
2061                  * Async queues are currently system wide. Just taking
2062                  * proportion of queues with-in same group will lead to higher
2063                  * async ratio system wide as generally root group is going
2064                  * to have higher weight. A more accurate thing would be to
2065                  * calculate system wide asnc/sync ratio.
2066                  */
2067                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2068                 tmp = tmp/cfqd->busy_queues;
2069                 slice = min_t(unsigned, slice, tmp);
2070
2071                 /* async workload slice is scaled down according to
2072                  * the sync/async slice ratio. */
2073                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2074         } else
2075                 /* sync workload slice is at least 2 * cfq_slice_idle */
2076                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2077
2078         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2079         cfq_log(cfqd, "workload slice:%d", slice);
2080         cfqd->workload_expires = jiffies + slice;
2081         cfqd->noidle_tree_requires_idle = false;
2082 }
2083
2084 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2085 {
2086         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2087         struct cfq_group *cfqg;
2088
2089         if (RB_EMPTY_ROOT(&st->rb))
2090                 return NULL;
2091         cfqg = cfq_rb_first_group(st);
2092         st->active = &cfqg->rb_node;
2093         update_min_vdisktime(st);
2094         return cfqg;
2095 }
2096
2097 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2098 {
2099         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2100
2101         cfqd->serving_group = cfqg;
2102
2103         /* Restore the workload type data */
2104         if (cfqg->saved_workload_slice) {
2105                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2106                 cfqd->serving_type = cfqg->saved_workload;
2107                 cfqd->serving_prio = cfqg->saved_serving_prio;
2108         } else
2109                 cfqd->workload_expires = jiffies - 1;
2110
2111         choose_service_tree(cfqd, cfqg);
2112 }
2113
2114 /*
2115  * Select a queue for service. If we have a current active queue,
2116  * check whether to continue servicing it, or retrieve and set a new one.
2117  */
2118 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2119 {
2120         struct cfq_queue *cfqq, *new_cfqq = NULL;
2121
2122         cfqq = cfqd->active_queue;
2123         if (!cfqq)
2124                 goto new_queue;
2125
2126         if (!cfqd->rq_queued)
2127                 return NULL;
2128
2129         /*
2130          * We were waiting for group to get backlogged. Expire the queue
2131          */
2132         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2133                 goto expire;
2134
2135         /*
2136          * The active queue has run out of time, expire it and select new.
2137          */
2138         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2139                 /*
2140                  * If slice had not expired at the completion of last request
2141                  * we might not have turned on wait_busy flag. Don't expire
2142                  * the queue yet. Allow the group to get backlogged.
2143                  *
2144                  * The very fact that we have used the slice, that means we
2145                  * have been idling all along on this queue and it should be
2146                  * ok to wait for this request to complete.
2147                  */
2148                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2149                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2150                         cfqq = NULL;
2151                         goto keep_queue;
2152                 } else
2153                         goto expire;
2154         }
2155
2156         /*
2157          * The active queue has requests and isn't expired, allow it to
2158          * dispatch.
2159          */
2160         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2161                 goto keep_queue;
2162
2163         /*
2164          * If another queue has a request waiting within our mean seek
2165          * distance, let it run.  The expire code will check for close
2166          * cooperators and put the close queue at the front of the service
2167          * tree.  If possible, merge the expiring queue with the new cfqq.
2168          */
2169         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2170         if (new_cfqq) {
2171                 if (!cfqq->new_cfqq)
2172                         cfq_setup_merge(cfqq, new_cfqq);
2173                 goto expire;
2174         }
2175
2176         /*
2177          * No requests pending. If the active queue still has requests in
2178          * flight or is idling for a new request, allow either of these
2179          * conditions to happen (or time out) before selecting a new queue.
2180          */
2181         if (timer_pending(&cfqd->idle_slice_timer) ||
2182             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2183                 cfqq = NULL;
2184                 goto keep_queue;
2185         }
2186
2187 expire:
2188         cfq_slice_expired(cfqd, 0, false);
2189 new_queue:
2190         /*
2191          * Current queue expired. Check if we have to switch to a new
2192          * service tree
2193          */
2194         if (!new_cfqq)
2195                 cfq_choose_cfqg(cfqd);
2196
2197         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2198 keep_queue:
2199         return cfqq;
2200 }
2201
2202 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2203 {
2204         int dispatched = 0;
2205
2206         while (cfqq->next_rq) {
2207                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2208                 dispatched++;
2209         }
2210
2211         BUG_ON(!list_empty(&cfqq->fifo));
2212
2213         /* By default cfqq is not expired if it is empty. Do it explicitly */
2214         __cfq_slice_expired(cfqq->cfqd, cfqq, 0, true);
2215         return dispatched;
2216 }
2217
2218 /*
2219  * Drain our current requests. Used for barriers and when switching
2220  * io schedulers on-the-fly.
2221  */
2222 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2223 {
2224         struct cfq_queue *cfqq;
2225         int dispatched = 0;
2226
2227         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2228                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2229
2230         cfq_slice_expired(cfqd, 0, true);
2231         BUG_ON(cfqd->busy_queues);
2232
2233         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2234         return dispatched;
2235 }
2236
2237 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2238         struct cfq_queue *cfqq)
2239 {
2240         /* the queue hasn't finished any request, can't estimate */
2241         if (cfq_cfqq_slice_new(cfqq))
2242                 return 1;
2243         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2244                 cfqq->slice_end))
2245                 return 1;
2246
2247         return 0;
2248 }
2249
2250 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2251 {
2252         unsigned int max_dispatch;
2253
2254         /*
2255          * Drain async requests before we start sync IO
2256          */
2257         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2258                 return false;
2259
2260         /*
2261          * If this is an async queue and we have sync IO in flight, let it wait
2262          */
2263         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2264                 return false;
2265
2266         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2267         if (cfq_class_idle(cfqq))
2268                 max_dispatch = 1;
2269
2270         /*
2271          * Does this cfqq already have too much IO in flight?
2272          */
2273         if (cfqq->dispatched >= max_dispatch) {
2274                 /*
2275                  * idle queue must always only have a single IO in flight
2276                  */
2277                 if (cfq_class_idle(cfqq))
2278                         return false;
2279
2280                 /*
2281                  * We have other queues, don't allow more IO from this one
2282                  */
2283                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2284                         return false;
2285
2286                 /*
2287                  * Sole queue user, no limit
2288                  */
2289                 if (cfqd->busy_queues == 1)
2290                         max_dispatch = -1;
2291                 else
2292                         /*
2293                          * Normally we start throttling cfqq when cfq_quantum/2
2294                          * requests have been dispatched. But we can drive
2295                          * deeper queue depths at the beginning of slice
2296                          * subjected to upper limit of cfq_quantum.
2297                          * */
2298                         max_dispatch = cfqd->cfq_quantum;
2299         }
2300
2301         /*
2302          * Async queues must wait a bit before being allowed dispatch.
2303          * We also ramp up the dispatch depth gradually for async IO,
2304          * based on the last sync IO we serviced
2305          */
2306         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2307                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2308                 unsigned int depth;
2309
2310                 depth = last_sync / cfqd->cfq_slice[1];
2311                 if (!depth && !cfqq->dispatched)
2312                         depth = 1;
2313                 if (depth < max_dispatch)
2314                         max_dispatch = depth;
2315         }
2316
2317         /*
2318          * If we're below the current max, allow a dispatch
2319          */
2320         return cfqq->dispatched < max_dispatch;
2321 }
2322
2323 /*
2324  * Dispatch a request from cfqq, moving them to the request queue
2325  * dispatch list.
2326  */
2327 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2328 {
2329         struct request *rq;
2330
2331         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2332
2333         if (!cfq_may_dispatch(cfqd, cfqq))
2334                 return false;
2335
2336         /*
2337          * follow expired path, else get first next available
2338          */
2339         rq = cfq_check_fifo(cfqq);
2340         if (!rq)
2341                 rq = cfqq->next_rq;
2342
2343         /*
2344          * insert request into driver dispatch list
2345          */
2346         cfq_dispatch_insert(cfqd->queue, rq);
2347
2348         if (!cfqd->active_cic) {
2349                 struct cfq_io_context *cic = RQ_CIC(rq);
2350
2351                 atomic_long_inc(&cic->ioc->refcount);
2352                 cfqd->active_cic = cic;
2353         }
2354
2355         return true;
2356 }
2357
2358 /*
2359  * Find the cfqq that we need to service and move a request from that to the
2360  * dispatch list
2361  */
2362 static int cfq_dispatch_requests(struct request_queue *q, int force)
2363 {
2364         struct cfq_data *cfqd = q->elevator->elevator_data;
2365         struct cfq_queue *cfqq;
2366
2367         if (!cfqd->busy_queues)
2368                 return 0;
2369
2370         if (unlikely(force))
2371                 return cfq_forced_dispatch(cfqd);
2372
2373         cfqq = cfq_select_queue(cfqd);
2374         if (!cfqq)
2375                 return 0;
2376
2377         /*
2378          * Dispatch a request from this cfqq, if it is allowed
2379          */
2380         if (!cfq_dispatch_request(cfqd, cfqq))
2381                 return 0;
2382
2383         cfqq->slice_dispatch++;
2384         cfq_clear_cfqq_must_dispatch(cfqq);
2385
2386         /*
2387          * expire an async queue immediately if it has used up its slice. idle
2388          * queue always expire after 1 dispatch round.
2389          */
2390         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2391             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2392             cfq_class_idle(cfqq))) {
2393                 cfqq->slice_end = jiffies + 1;
2394                 cfq_slice_expired(cfqd, 0, false);
2395         }
2396
2397         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2398         /*
2399          * This is needed since we don't exactly match the mod_timer() and
2400          * del_timer() calls in CFQ.
2401          */
2402         blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
2403         return 1;
2404 }
2405
2406 /*
2407  * task holds one reference to the queue, dropped when task exits. each rq
2408  * in-flight on this queue also holds a reference, dropped when rq is freed.
2409  *
2410  * Each cfq queue took a reference on the parent group. Drop it now.
2411  * queue lock must be held here.
2412  */
2413 static void cfq_put_queue(struct cfq_queue *cfqq)
2414 {
2415         struct cfq_data *cfqd = cfqq->cfqd;
2416         struct cfq_group *cfqg, *orig_cfqg;
2417
2418         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2419
2420         if (!atomic_dec_and_test(&cfqq->ref))
2421                 return;
2422
2423         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2424         BUG_ON(rb_first(&cfqq->sort_list));
2425         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2426         cfqg = cfqq->cfqg;
2427         orig_cfqg = cfqq->orig_cfqg;
2428
2429         if (unlikely(cfqd->active_queue == cfqq)) {
2430                 __cfq_slice_expired(cfqd, cfqq, 0, false);
2431                 cfq_schedule_dispatch(cfqd);
2432         }
2433
2434         BUG_ON(cfq_cfqq_on_rr(cfqq));
2435         kmem_cache_free(cfq_pool, cfqq);
2436         cfq_put_cfqg(cfqg);
2437         if (orig_cfqg)
2438                 cfq_put_cfqg(orig_cfqg);
2439 }
2440
2441 /*
2442  * Must always be called with the rcu_read_lock() held
2443  */
2444 static void
2445 __call_for_each_cic(struct io_context *ioc,
2446                     void (*func)(struct io_context *, struct cfq_io_context *))
2447 {
2448         struct cfq_io_context *cic;
2449         struct hlist_node *n;
2450
2451         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2452                 func(ioc, cic);
2453 }
2454
2455 /*
2456  * Call func for each cic attached to this ioc.
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         rcu_read_lock();
2463         __call_for_each_cic(ioc, func);
2464         rcu_read_unlock();
2465 }
2466
2467 static void cfq_cic_free_rcu(struct rcu_head *head)
2468 {
2469         struct cfq_io_context *cic;
2470
2471         cic = container_of(head, struct cfq_io_context, rcu_head);
2472
2473         kmem_cache_free(cfq_ioc_pool, cic);
2474         elv_ioc_count_dec(cfq_ioc_count);
2475
2476         if (ioc_gone) {
2477                 /*
2478                  * CFQ scheduler is exiting, grab exit lock and check
2479                  * the pending io context count. If it hits zero,
2480                  * complete ioc_gone and set it back to NULL
2481                  */
2482                 spin_lock(&ioc_gone_lock);
2483                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2484                         complete(ioc_gone);
2485                         ioc_gone = NULL;
2486                 }
2487                 spin_unlock(&ioc_gone_lock);
2488         }
2489 }
2490
2491 static void cfq_cic_free(struct cfq_io_context *cic)
2492 {
2493         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2494 }
2495
2496 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2497 {
2498         unsigned long flags;
2499
2500         BUG_ON(!cic->dead_key);
2501
2502         spin_lock_irqsave(&ioc->lock, flags);
2503         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2504         hlist_del_rcu(&cic->cic_list);
2505         spin_unlock_irqrestore(&ioc->lock, flags);
2506
2507         cfq_cic_free(cic);
2508 }
2509
2510 /*
2511  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2512  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2513  * and ->trim() which is called with the task lock held
2514  */
2515 static void cfq_free_io_context(struct io_context *ioc)
2516 {
2517         /*
2518          * ioc->refcount is zero here, or we are called from elv_unregister(),
2519          * so no more cic's are allowed to be linked into this ioc.  So it
2520          * should be ok to iterate over the known list, we will see all cic's
2521          * since no new ones are added.
2522          */
2523         __call_for_each_cic(ioc, cic_free_func);
2524 }
2525
2526 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2527 {
2528         struct cfq_queue *__cfqq, *next;
2529
2530         if (unlikely(cfqq == cfqd->active_queue)) {
2531                 __cfq_slice_expired(cfqd, cfqq, 0, false);
2532                 cfq_schedule_dispatch(cfqd);
2533         }
2534
2535         /*
2536          * If this queue was scheduled to merge with another queue, be
2537          * sure to drop the reference taken on that queue (and others in
2538          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2539          */
2540         __cfqq = cfqq->new_cfqq;
2541         while (__cfqq) {
2542                 if (__cfqq == cfqq) {
2543                         WARN(1, "cfqq->new_cfqq loop detected\n");
2544                         break;
2545                 }
2546                 next = __cfqq->new_cfqq;
2547                 cfq_put_queue(__cfqq);
2548                 __cfqq = next;
2549         }
2550
2551         cfq_put_queue(cfqq);
2552 }
2553
2554 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2555                                          struct cfq_io_context *cic)
2556 {
2557         struct io_context *ioc = cic->ioc;
2558
2559         list_del_init(&cic->queue_list);
2560
2561         /*
2562          * Make sure key == NULL is seen for dead queues
2563          */
2564         smp_wmb();
2565         cic->dead_key = (unsigned long) cic->key;
2566         cic->key = NULL;
2567
2568         if (ioc->ioc_data == cic)
2569                 rcu_assign_pointer(ioc->ioc_data, NULL);
2570
2571         if (cic->cfqq[BLK_RW_ASYNC]) {
2572                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2573                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2574         }
2575
2576         if (cic->cfqq[BLK_RW_SYNC]) {
2577                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2578                 cic->cfqq[BLK_RW_SYNC] = NULL;
2579         }
2580 }
2581
2582 static void cfq_exit_single_io_context(struct io_context *ioc,
2583                                        struct cfq_io_context *cic)
2584 {
2585         struct cfq_data *cfqd = cic->key;
2586
2587         if (cfqd) {
2588                 struct request_queue *q = cfqd->queue;
2589                 unsigned long flags;
2590
2591                 spin_lock_irqsave(q->queue_lock, flags);
2592
2593                 /*
2594                  * Ensure we get a fresh copy of the ->key to prevent
2595                  * race between exiting task and queue
2596                  */
2597                 smp_read_barrier_depends();
2598                 if (cic->key)
2599                         __cfq_exit_single_io_context(cfqd, cic);
2600
2601                 spin_unlock_irqrestore(q->queue_lock, flags);
2602         }
2603 }
2604
2605 /*
2606  * The process that ioc belongs to has exited, we need to clean up
2607  * and put the internal structures we have that belongs to that process.
2608  */
2609 static void cfq_exit_io_context(struct io_context *ioc)
2610 {
2611         call_for_each_cic(ioc, cfq_exit_single_io_context);
2612 }
2613
2614 static struct cfq_io_context *
2615 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2616 {
2617         struct cfq_io_context *cic;
2618
2619         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2620                                                         cfqd->queue->node);
2621         if (cic) {
2622                 cic->last_end_request = jiffies;
2623                 INIT_LIST_HEAD(&cic->queue_list);
2624                 INIT_HLIST_NODE(&cic->cic_list);
2625                 cic->dtor = cfq_free_io_context;
2626                 cic->exit = cfq_exit_io_context;
2627                 elv_ioc_count_inc(cfq_ioc_count);
2628         }
2629
2630         return cic;
2631 }
2632
2633 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2634 {
2635         struct task_struct *tsk = current;
2636         int ioprio_class;
2637
2638         if (!cfq_cfqq_prio_changed(cfqq))
2639                 return;
2640
2641         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2642         switch (ioprio_class) {
2643         default:
2644                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2645         case IOPRIO_CLASS_NONE:
2646                 /*
2647                  * no prio set, inherit CPU scheduling settings
2648                  */
2649                 cfqq->ioprio = task_nice_ioprio(tsk);
2650                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2651                 break;
2652         case IOPRIO_CLASS_RT:
2653                 cfqq->ioprio = task_ioprio(ioc);
2654                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2655                 break;
2656         case IOPRIO_CLASS_BE:
2657                 cfqq->ioprio = task_ioprio(ioc);
2658                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2659                 break;
2660         case IOPRIO_CLASS_IDLE:
2661                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2662                 cfqq->ioprio = 7;
2663                 cfq_clear_cfqq_idle_window(cfqq);
2664                 break;
2665         }
2666
2667         /*
2668          * keep track of original prio settings in case we have to temporarily
2669          * elevate the priority of this queue
2670          */
2671         cfqq->org_ioprio = cfqq->ioprio;
2672         cfqq->org_ioprio_class = cfqq->ioprio_class;
2673         cfq_clear_cfqq_prio_changed(cfqq);
2674 }
2675
2676 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2677 {
2678         struct cfq_data *cfqd = cic->key;
2679         struct cfq_queue *cfqq;
2680         unsigned long flags;
2681
2682         if (unlikely(!cfqd))
2683                 return;
2684
2685         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2686
2687         cfqq = cic->cfqq[BLK_RW_ASYNC];
2688         if (cfqq) {
2689                 struct cfq_queue *new_cfqq;
2690                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2691                                                 GFP_ATOMIC);
2692                 if (new_cfqq) {
2693                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2694                         cfq_put_queue(cfqq);
2695                 }
2696         }
2697
2698         cfqq = cic->cfqq[BLK_RW_SYNC];
2699         if (cfqq)
2700                 cfq_mark_cfqq_prio_changed(cfqq);
2701
2702         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2703 }
2704
2705 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2706 {
2707         call_for_each_cic(ioc, changed_ioprio);
2708         ioc->ioprio_changed = 0;
2709 }
2710
2711 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2712                           pid_t pid, bool is_sync)
2713 {
2714         RB_CLEAR_NODE(&cfqq->rb_node);
2715         RB_CLEAR_NODE(&cfqq->p_node);
2716         INIT_LIST_HEAD(&cfqq->fifo);
2717
2718         atomic_set(&cfqq->ref, 0);
2719         cfqq->cfqd = cfqd;
2720
2721         cfq_mark_cfqq_prio_changed(cfqq);
2722
2723         if (is_sync) {
2724                 if (!cfq_class_idle(cfqq))
2725                         cfq_mark_cfqq_idle_window(cfqq);
2726                 cfq_mark_cfqq_sync(cfqq);
2727         }
2728         cfqq->pid = pid;
2729 }
2730
2731 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2732 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2733 {
2734         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2735         struct cfq_data *cfqd = cic->key;
2736         unsigned long flags;
2737         struct request_queue *q;
2738
2739         if (unlikely(!cfqd))
2740                 return;
2741
2742         q = cfqd->queue;
2743
2744         spin_lock_irqsave(q->queue_lock, flags);
2745
2746         if (sync_cfqq) {
2747                 /*
2748                  * Drop reference to sync queue. A new sync queue will be
2749                  * assigned in new group upon arrival of a fresh request.
2750                  */
2751                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2752                 cic_set_cfqq(cic, NULL, 1);
2753                 cfq_put_queue(sync_cfqq);
2754         }
2755
2756         spin_unlock_irqrestore(q->queue_lock, flags);
2757 }
2758
2759 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2760 {
2761         call_for_each_cic(ioc, changed_cgroup);
2762         ioc->cgroup_changed = 0;
2763 }
2764 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2765
2766 static struct cfq_queue *
2767 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2768                      struct io_context *ioc, gfp_t gfp_mask)
2769 {
2770         struct cfq_queue *cfqq, *new_cfqq = NULL;
2771         struct cfq_io_context *cic;
2772         struct cfq_group *cfqg;
2773
2774 retry:
2775         cfqg = cfq_get_cfqg(cfqd, 1);
2776         cic = cfq_cic_lookup(cfqd, ioc);
2777         /* cic always exists here */
2778         cfqq = cic_to_cfqq(cic, is_sync);
2779
2780         /*
2781          * Always try a new alloc if we fell back to the OOM cfqq
2782          * originally, since it should just be a temporary situation.
2783          */
2784         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2785                 cfqq = NULL;
2786                 if (new_cfqq) {
2787                         cfqq = new_cfqq;
2788                         new_cfqq = NULL;
2789                 } else if (gfp_mask & __GFP_WAIT) {
2790                         spin_unlock_irq(cfqd->queue->queue_lock);
2791                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2792                                         gfp_mask | __GFP_ZERO,
2793                                         cfqd->queue->node);
2794                         spin_lock_irq(cfqd->queue->queue_lock);
2795                         if (new_cfqq)
2796                                 goto retry;
2797                 } else {
2798                         cfqq = kmem_cache_alloc_node(cfq_pool,
2799                                         gfp_mask | __GFP_ZERO,
2800                                         cfqd->queue->node);
2801                 }
2802
2803                 if (cfqq) {
2804                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2805                         cfq_init_prio_data(cfqq, ioc);
2806                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2807                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2808                 } else
2809                         cfqq = &cfqd->oom_cfqq;
2810         }
2811
2812         if (new_cfqq)
2813                 kmem_cache_free(cfq_pool, new_cfqq);
2814
2815         return cfqq;
2816 }
2817
2818 static struct cfq_queue **
2819 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2820 {
2821         switch (ioprio_class) {
2822         case IOPRIO_CLASS_RT:
2823                 return &cfqd->async_cfqq[0][ioprio];
2824         case IOPRIO_CLASS_BE:
2825                 return &cfqd->async_cfqq[1][ioprio];
2826         case IOPRIO_CLASS_IDLE:
2827                 return &cfqd->async_idle_cfqq;
2828         default:
2829                 BUG();
2830         }
2831 }
2832
2833 static struct cfq_queue *
2834 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2835               gfp_t gfp_mask)
2836 {
2837         const int ioprio = task_ioprio(ioc);
2838         const int ioprio_class = task_ioprio_class(ioc);
2839         struct cfq_queue **async_cfqq = NULL;
2840         struct cfq_queue *cfqq = NULL;
2841
2842         if (!is_sync) {
2843                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2844                 cfqq = *async_cfqq;
2845         }
2846
2847         if (!cfqq)
2848                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2849
2850         /*
2851          * pin the queue now that it's allocated, scheduler exit will prune it
2852          */
2853         if (!is_sync && !(*async_cfqq)) {
2854                 atomic_inc(&cfqq->ref);
2855                 *async_cfqq = cfqq;
2856         }
2857
2858         atomic_inc(&cfqq->ref);
2859         return cfqq;
2860 }
2861
2862 /*
2863  * We drop cfq io contexts lazily, so we may find a dead one.
2864  */
2865 static void
2866 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2867                   struct cfq_io_context *cic)
2868 {
2869         unsigned long flags;
2870
2871         WARN_ON(!list_empty(&cic->queue_list));
2872
2873         spin_lock_irqsave(&ioc->lock, flags);
2874
2875         BUG_ON(ioc->ioc_data == cic);
2876
2877         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2878         hlist_del_rcu(&cic->cic_list);
2879         spin_unlock_irqrestore(&ioc->lock, flags);
2880
2881         cfq_cic_free(cic);
2882 }
2883
2884 static struct cfq_io_context *
2885 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2886 {
2887         struct cfq_io_context *cic;
2888         unsigned long flags;
2889         void *k;
2890
2891         if (unlikely(!ioc))
2892                 return NULL;
2893
2894         rcu_read_lock();
2895
2896         /*
2897          * we maintain a last-hit cache, to avoid browsing over the tree
2898          */
2899         cic = rcu_dereference(ioc->ioc_data);
2900         if (cic && cic->key == cfqd) {
2901                 rcu_read_unlock();
2902                 return cic;
2903         }
2904
2905         do {
2906                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2907                 rcu_read_unlock();
2908                 if (!cic)
2909                         break;
2910                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2911                 k = cic->key;
2912                 if (unlikely(!k)) {
2913                         cfq_drop_dead_cic(cfqd, ioc, cic);
2914                         rcu_read_lock();
2915                         continue;
2916                 }
2917
2918                 spin_lock_irqsave(&ioc->lock, flags);
2919                 rcu_assign_pointer(ioc->ioc_data, cic);
2920                 spin_unlock_irqrestore(&ioc->lock, flags);
2921                 break;
2922         } while (1);
2923
2924         return cic;
2925 }
2926
2927 /*
2928  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2929  * the process specific cfq io context when entered from the block layer.
2930  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2931  */
2932 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2933                         struct cfq_io_context *cic, gfp_t gfp_mask)
2934 {
2935         unsigned long flags;
2936         int ret;
2937
2938         ret = radix_tree_preload(gfp_mask);
2939         if (!ret) {
2940                 cic->ioc = ioc;
2941                 cic->key = cfqd;
2942
2943                 spin_lock_irqsave(&ioc->lock, flags);
2944                 ret = radix_tree_insert(&ioc->radix_root,
2945                                                 (unsigned long) cfqd, cic);
2946                 if (!ret)
2947                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2948                 spin_unlock_irqrestore(&ioc->lock, flags);
2949
2950                 radix_tree_preload_end();
2951
2952                 if (!ret) {
2953                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2954                         list_add(&cic->queue_list, &cfqd->cic_list);
2955                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2956                 }
2957         }
2958
2959         if (ret)
2960                 printk(KERN_ERR "cfq: cic link failed!\n");
2961
2962         return ret;
2963 }
2964
2965 /*
2966  * Setup general io context and cfq io context. There can be several cfq
2967  * io contexts per general io context, if this process is doing io to more
2968  * than one device managed by cfq.
2969  */
2970 static struct cfq_io_context *
2971 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2972 {
2973         struct io_context *ioc = NULL;
2974         struct cfq_io_context *cic;
2975
2976         might_sleep_if(gfp_mask & __GFP_WAIT);
2977
2978         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2979         if (!ioc)
2980                 return NULL;
2981
2982         cic = cfq_cic_lookup(cfqd, ioc);
2983         if (cic)
2984                 goto out;
2985
2986         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2987         if (cic == NULL)
2988                 goto err;
2989
2990         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2991                 goto err_free;
2992
2993 out:
2994         smp_read_barrier_depends();
2995         if (unlikely(ioc->ioprio_changed))
2996                 cfq_ioc_set_ioprio(ioc);
2997
2998 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2999         if (unlikely(ioc->cgroup_changed))
3000                 cfq_ioc_set_cgroup(ioc);
3001 #endif
3002         return cic;
3003 err_free:
3004         cfq_cic_free(cic);
3005 err:
3006         put_io_context(ioc);
3007         return NULL;
3008 }
3009
3010 static void
3011 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3012 {
3013         unsigned long elapsed = jiffies - cic->last_end_request;
3014         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3015
3016         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3017         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3018         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3019 }
3020
3021 static void
3022 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3023                        struct request *rq)
3024 {
3025         sector_t sdist = 0;
3026         sector_t n_sec = blk_rq_sectors(rq);
3027         if (cfqq->last_request_pos) {
3028                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3029                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3030                 else
3031                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3032         }
3033
3034         cfqq->seek_history <<= 1;
3035         if (blk_queue_nonrot(cfqd->queue))
3036                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3037         else
3038                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3039 }
3040
3041 /*
3042  * Disable idle window if the process thinks too long or seeks so much that
3043  * it doesn't matter
3044  */
3045 static void
3046 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3047                        struct cfq_io_context *cic)
3048 {
3049         int old_idle, enable_idle;
3050
3051         /*
3052          * Don't idle for async or idle io prio class
3053          */
3054         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3055                 return;
3056
3057         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3058
3059         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3060                 cfq_mark_cfqq_deep(cfqq);
3061
3062         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3063             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3064                 enable_idle = 0;
3065         else if (sample_valid(cic->ttime_samples)) {
3066                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3067                         enable_idle = 0;
3068                 else
3069                         enable_idle = 1;
3070         }
3071
3072         if (old_idle != enable_idle) {
3073                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3074                 if (enable_idle)
3075                         cfq_mark_cfqq_idle_window(cfqq);
3076                 else
3077                         cfq_clear_cfqq_idle_window(cfqq);
3078         }
3079 }
3080
3081 /*
3082  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3083  * no or if we aren't sure, a 1 will cause a preempt.
3084  */
3085 static bool
3086 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3087                    struct request *rq)
3088 {
3089         struct cfq_queue *cfqq;
3090
3091         cfqq = cfqd->active_queue;
3092         if (!cfqq)
3093                 return false;
3094
3095         if (cfq_class_idle(new_cfqq))
3096                 return false;
3097
3098         if (cfq_class_idle(cfqq))
3099                 return true;
3100
3101         /*
3102          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3103          */
3104         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3105                 return false;
3106
3107         /*
3108          * if the new request is sync, but the currently running queue is
3109          * not, let the sync request have priority.
3110          */
3111         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3112                 return true;
3113
3114         if (new_cfqq->cfqg != cfqq->cfqg)
3115                 return false;
3116
3117         if (cfq_slice_used(cfqq))
3118                 return true;
3119
3120         /* Allow preemption only if we are idling on sync-noidle tree */
3121         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3122             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3123             new_cfqq->service_tree->count == 2 &&
3124             RB_EMPTY_ROOT(&cfqq->sort_list))
3125                 return true;
3126
3127         /*
3128          * So both queues are sync. Let the new request get disk time if
3129          * it's a metadata request and the current queue is doing regular IO.
3130          */
3131         if (rq_is_meta(rq) && !cfqq->meta_pending)
3132                 return true;
3133
3134         /*
3135          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3136          */
3137         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3138                 return true;
3139
3140         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3141                 return false;
3142
3143         /*
3144          * if this request is as-good as one we would expect from the
3145          * current cfqq, let it preempt
3146          */
3147         if (cfq_rq_close(cfqd, cfqq, rq))
3148                 return true;
3149
3150         return false;
3151 }
3152
3153 /*
3154  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3155  * let it have half of its nominal slice.
3156  */
3157 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3158 {
3159         cfq_log_cfqq(cfqd, cfqq, "preempt");
3160         cfq_slice_expired(cfqd, 1, false);
3161
3162         /*
3163          * Put the new queue at the front of the of the current list,
3164          * so we know that it will be selected next.
3165          */
3166         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3167
3168         cfq_service_tree_add(cfqd, cfqq, 1);
3169
3170         cfqq->slice_end = 0;
3171         cfq_mark_cfqq_slice_new(cfqq);
3172 }
3173
3174 /*
3175  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3176  * something we should do about it
3177  */
3178 static void
3179 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3180                 struct request *rq)
3181 {
3182         struct cfq_io_context *cic = RQ_CIC(rq);
3183
3184         cfqd->rq_queued++;
3185         if (rq_is_meta(rq))
3186                 cfqq->meta_pending++;
3187
3188         cfq_update_io_thinktime(cfqd, cic);
3189         cfq_update_io_seektime(cfqd, cfqq, rq);
3190         cfq_update_idle_window(cfqd, cfqq, cic);
3191
3192         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3193
3194         if (cfqq == cfqd->active_queue) {
3195                 /*
3196                  * Remember that we saw a request from this process, but
3197                  * don't start queuing just yet. Otherwise we risk seeing lots
3198                  * of tiny requests, because we disrupt the normal plugging
3199                  * and merging. If the request is already larger than a single
3200                  * page, let it rip immediately. For that case we assume that
3201                  * merging is already done. Ditto for a busy system that
3202                  * has other work pending, don't risk delaying until the
3203                  * idle timer unplug to continue working.
3204                  */
3205                 if (cfq_cfqq_wait_request(cfqq)) {
3206                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3207                             cfqd->busy_queues > 1) {
3208                                 cfq_del_timer(cfqd, cfqq);
3209                                 cfq_clear_cfqq_wait_request(cfqq);
3210                                 __blk_run_queue(cfqd->queue);
3211                         } else
3212                                 cfq_mark_cfqq_must_dispatch(cfqq);
3213                 }
3214         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3215                 /*
3216                  * not the active queue - expire current slice if it is
3217                  * idle and has expired it's mean thinktime or this new queue
3218                  * has some old slice time left and is of higher priority or
3219                  * this new queue is RT and the current one is BE
3220                  */
3221                 cfq_preempt_queue(cfqd, cfqq);
3222                 __blk_run_queue(cfqd->queue);
3223         }
3224 }
3225
3226 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3227 {
3228         struct cfq_data *cfqd = q->elevator->elevator_data;
3229         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3230
3231         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3232         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3233
3234         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3235         list_add_tail(&rq->queuelist, &cfqq->fifo);
3236         cfq_add_rq_rb(rq);
3237
3238         blkiocg_update_request_add_stats(&cfqq->cfqg->blkg,
3239                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3240                         rq_is_sync(rq));
3241         cfq_rq_enqueued(cfqd, cfqq, rq);
3242 }
3243
3244 /*
3245  * Update hw_tag based on peak queue depth over 50 samples under
3246  * sufficient load.
3247  */
3248 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3249 {
3250         struct cfq_queue *cfqq = cfqd->active_queue;
3251
3252         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3253                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3254
3255         if (cfqd->hw_tag == 1)
3256                 return;
3257
3258         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3259             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3260                 return;
3261
3262         /*
3263          * If active queue hasn't enough requests and can idle, cfq might not
3264          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3265          * case
3266          */
3267         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3268             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3269             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3270                 return;
3271
3272         if (cfqd->hw_tag_samples++ < 50)
3273                 return;
3274
3275         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3276                 cfqd->hw_tag = 1;
3277         else
3278                 cfqd->hw_tag = 0;
3279 }
3280
3281 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3282 {
3283         struct cfq_io_context *cic = cfqd->active_cic;
3284
3285         /* If there are other queues in the group, don't wait */
3286         if (cfqq->cfqg->nr_cfqq > 1)
3287                 return false;
3288
3289         if (cfq_slice_used(cfqq))
3290                 return true;
3291
3292         /* if slice left is less than think time, wait busy */
3293         if (cic && sample_valid(cic->ttime_samples)
3294             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3295                 return true;
3296
3297         /*
3298          * If think times is less than a jiffy than ttime_mean=0 and above
3299          * will not be true. It might happen that slice has not expired yet
3300          * but will expire soon (4-5 ns) during select_queue(). To cover the
3301          * case where think time is less than a jiffy, mark the queue wait
3302          * busy if only 1 jiffy is left in the slice.
3303          */
3304         if (cfqq->slice_end - jiffies == 1)
3305                 return true;
3306
3307         return false;
3308 }
3309
3310 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3311 {
3312         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3313         struct cfq_data *cfqd = cfqq->cfqd;
3314         const int sync = rq_is_sync(rq);
3315         unsigned long now;
3316
3317         now = jiffies;
3318         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3319
3320         cfq_update_hw_tag(cfqd);
3321
3322         WARN_ON(!cfqd->rq_in_driver);
3323         WARN_ON(!cfqq->dispatched);
3324         cfqd->rq_in_driver--;
3325         cfqq->dispatched--;
3326         blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3327                         rq_io_start_time_ns(rq), rq_data_dir(rq),
3328                         rq_is_sync(rq));
3329
3330         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3331
3332         if (sync) {
3333                 RQ_CIC(rq)->last_end_request = now;
3334                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3335                         cfqd->last_delayed_sync = now;
3336         }
3337
3338         /*
3339          * If this is the active queue, check if it needs to be expired,
3340          * or if we want to idle in case it has no pending requests.
3341          */
3342         if (cfqd->active_queue == cfqq) {
3343                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3344
3345                 if (cfq_cfqq_slice_new(cfqq)) {
3346                         cfq_set_prio_slice(cfqd, cfqq);
3347                         cfq_clear_cfqq_slice_new(cfqq);
3348                 }
3349
3350                 /*
3351                  * Should we wait for next request to come in before we expire
3352                  * the queue.
3353                  */
3354                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3355                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3356                         cfq_mark_cfqq_wait_busy(cfqq);
3357                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3358                 }
3359
3360                 /*
3361                  * Idling is not enabled on:
3362                  * - expired queues
3363                  * - idle-priority queues
3364                  * - async queues
3365                  * - queues with still some requests queued
3366                  * - when there is a close cooperator
3367                  */
3368                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3369                         cfq_slice_expired(cfqd, 1, false);
3370                 else if (sync && cfqq_empty &&
3371                          !cfq_close_cooperator(cfqd, cfqq)) {
3372                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3373                         /*
3374                          * Idling is enabled for SYNC_WORKLOAD.
3375                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3376                          * only if we processed at least one !rq_noidle request
3377                          */
3378                         if (cfqd->serving_type == SYNC_WORKLOAD
3379                             || cfqd->noidle_tree_requires_idle
3380                             || cfqq->cfqg->nr_cfqq == 1)
3381                                 cfq_arm_slice_timer(cfqd);
3382                 }
3383         }
3384
3385         if (!cfqd->rq_in_driver)
3386                 cfq_schedule_dispatch(cfqd);
3387 }
3388
3389 /*
3390  * we temporarily boost lower priority queues if they are holding fs exclusive
3391  * resources. they are boosted to normal prio (CLASS_BE/4)
3392  */
3393 static void cfq_prio_boost(struct cfq_queue *cfqq)
3394 {
3395         if (has_fs_excl()) {
3396                 /*
3397                  * boost idle prio on transactions that would lock out other
3398                  * users of the filesystem
3399                  */
3400                 if (cfq_class_idle(cfqq))
3401                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3402                 if (cfqq->ioprio > IOPRIO_NORM)
3403                         cfqq->ioprio = IOPRIO_NORM;
3404         } else {
3405                 /*
3406                  * unboost the queue (if needed)
3407                  */
3408                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3409                 cfqq->ioprio = cfqq->org_ioprio;
3410         }
3411 }
3412
3413 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3414 {
3415         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3416                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3417                 return ELV_MQUEUE_MUST;
3418         }
3419
3420         return ELV_MQUEUE_MAY;
3421 }
3422
3423 static int cfq_may_queue(struct request_queue *q, int rw)
3424 {
3425         struct cfq_data *cfqd = q->elevator->elevator_data;
3426         struct task_struct *tsk = current;
3427         struct cfq_io_context *cic;
3428         struct cfq_queue *cfqq;
3429
3430         /*
3431          * don't force setup of a queue from here, as a call to may_queue
3432          * does not necessarily imply that a request actually will be queued.
3433          * so just lookup a possibly existing queue, or return 'may queue'
3434          * if that fails
3435          */
3436         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3437         if (!cic)
3438                 return ELV_MQUEUE_MAY;
3439
3440         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3441         if (cfqq) {
3442                 cfq_init_prio_data(cfqq, cic->ioc);
3443                 cfq_prio_boost(cfqq);
3444
3445                 return __cfq_may_queue(cfqq);
3446         }
3447
3448         return ELV_MQUEUE_MAY;
3449 }
3450
3451 /*
3452  * queue lock held here
3453  */
3454 static void cfq_put_request(struct request *rq)
3455 {
3456         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3457
3458         if (cfqq) {
3459                 const int rw = rq_data_dir(rq);
3460
3461                 BUG_ON(!cfqq->allocated[rw]);
3462                 cfqq->allocated[rw]--;
3463
3464                 put_io_context(RQ_CIC(rq)->ioc);
3465
3466                 rq->elevator_private = NULL;
3467                 rq->elevator_private2 = NULL;
3468
3469                 cfq_put_queue(cfqq);
3470         }
3471 }
3472
3473 static struct cfq_queue *
3474 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3475                 struct cfq_queue *cfqq)
3476 {
3477         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3478         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3479         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3480         cfq_put_queue(cfqq);
3481         return cic_to_cfqq(cic, 1);
3482 }
3483
3484 /*
3485  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3486  * was the last process referring to said cfqq.
3487  */
3488 static struct cfq_queue *
3489 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3490 {
3491         if (cfqq_process_refs(cfqq) == 1) {
3492                 cfqq->pid = current->pid;
3493                 cfq_clear_cfqq_coop(cfqq);
3494                 cfq_clear_cfqq_split_coop(cfqq);
3495                 return cfqq;
3496         }
3497
3498         cic_set_cfqq(cic, NULL, 1);
3499         cfq_put_queue(cfqq);
3500         return NULL;
3501 }
3502 /*
3503  * Allocate cfq data structures associated with this request.
3504  */
3505 static int
3506 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3507 {
3508         struct cfq_data *cfqd = q->elevator->elevator_data;
3509         struct cfq_io_context *cic;
3510         const int rw = rq_data_dir(rq);
3511         const bool is_sync = rq_is_sync(rq);
3512         struct cfq_queue *cfqq;
3513         unsigned long flags;
3514
3515         might_sleep_if(gfp_mask & __GFP_WAIT);
3516
3517         cic = cfq_get_io_context(cfqd, gfp_mask);
3518
3519         spin_lock_irqsave(q->queue_lock, flags);
3520
3521         if (!cic)
3522                 goto queue_fail;
3523
3524 new_queue:
3525         cfqq = cic_to_cfqq(cic, is_sync);
3526         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3527                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3528                 cic_set_cfqq(cic, cfqq, is_sync);
3529         } else {
3530                 /*
3531                  * If the queue was seeky for too long, break it apart.
3532                  */
3533                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3534                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3535                         cfqq = split_cfqq(cic, cfqq);
3536                         if (!cfqq)
3537                                 goto new_queue;
3538                 }
3539
3540                 /*
3541                  * Check to see if this queue is scheduled to merge with
3542                  * another, closely cooperating queue.  The merging of
3543                  * queues happens here as it must be done in process context.
3544                  * The reference on new_cfqq was taken in merge_cfqqs.
3545                  */
3546                 if (cfqq->new_cfqq)
3547                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3548         }
3549
3550         cfqq->allocated[rw]++;
3551         atomic_inc(&cfqq->ref);
3552
3553         spin_unlock_irqrestore(q->queue_lock, flags);
3554
3555         rq->elevator_private = cic;
3556         rq->elevator_private2 = cfqq;
3557         return 0;
3558
3559 queue_fail:
3560         if (cic)
3561                 put_io_context(cic->ioc);
3562
3563         cfq_schedule_dispatch(cfqd);
3564         spin_unlock_irqrestore(q->queue_lock, flags);
3565         cfq_log(cfqd, "set_request fail");
3566         return 1;
3567 }
3568
3569 static void cfq_kick_queue(struct work_struct *work)
3570 {
3571         struct cfq_data *cfqd =
3572                 container_of(work, struct cfq_data, unplug_work);
3573         struct request_queue *q = cfqd->queue;
3574
3575         spin_lock_irq(q->queue_lock);
3576         __blk_run_queue(cfqd->queue);
3577         spin_unlock_irq(q->queue_lock);
3578 }
3579
3580 /*
3581  * Timer running if the active_queue is currently idling inside its time slice
3582  */
3583 static void cfq_idle_slice_timer(unsigned long data)
3584 {
3585         struct cfq_data *cfqd = (struct cfq_data *) data;
3586         struct cfq_queue *cfqq;
3587         unsigned long flags;
3588         int timed_out = 1;
3589
3590         cfq_log(cfqd, "idle timer fired");
3591
3592         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3593
3594         cfqq = cfqd->active_queue;
3595         if (cfqq) {
3596                 timed_out = 0;
3597
3598                 /*
3599                  * We saw a request before the queue expired, let it through
3600                  */
3601                 if (cfq_cfqq_must_dispatch(cfqq))
3602                         goto out_kick;
3603
3604                 /*
3605                  * expired
3606                  */
3607                 if (cfq_slice_used(cfqq))
3608                         goto expire;
3609
3610                 /*
3611                  * only expire and reinvoke request handler, if there are
3612                  * other queues with pending requests
3613                  */
3614                 if (!cfqd->busy_queues)
3615                         goto out_cont;
3616
3617                 /*
3618                  * not expired and it has a request pending, let it dispatch
3619                  */
3620                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3621                         goto out_kick;
3622
3623                 /*
3624                  * Queue depth flag is reset only when the idle didn't succeed
3625                  */
3626                 cfq_clear_cfqq_deep(cfqq);
3627         }
3628 expire:
3629         cfq_slice_expired(cfqd, timed_out, false);
3630 out_kick:
3631         cfq_schedule_dispatch(cfqd);
3632 out_cont:
3633         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3634 }
3635
3636 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3637 {
3638         del_timer_sync(&cfqd->idle_slice_timer);
3639         cancel_work_sync(&cfqd->unplug_work);
3640 }
3641
3642 static void cfq_put_async_queues(struct cfq_data *cfqd)
3643 {
3644         int i;
3645
3646         for (i = 0; i < IOPRIO_BE_NR; i++) {
3647                 if (cfqd->async_cfqq[0][i])
3648                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3649                 if (cfqd->async_cfqq[1][i])
3650                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3651         }
3652
3653         if (cfqd->async_idle_cfqq)
3654                 cfq_put_queue(cfqd->async_idle_cfqq);
3655 }
3656
3657 static void cfq_cfqd_free(struct rcu_head *head)
3658 {
3659         kfree(container_of(head, struct cfq_data, rcu));
3660 }
3661
3662 static void cfq_exit_queue(struct elevator_queue *e)
3663 {
3664         struct cfq_data *cfqd = e->elevator_data;
3665         struct request_queue *q = cfqd->queue;
3666
3667         cfq_shutdown_timer_wq(cfqd);
3668
3669         spin_lock_irq(q->queue_lock);
3670
3671         if (cfqd->active_queue)
3672                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0, false);
3673
3674         while (!list_empty(&cfqd->cic_list)) {
3675                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3676                                                         struct cfq_io_context,
3677                                                         queue_list);
3678
3679                 __cfq_exit_single_io_context(cfqd, cic);
3680         }
3681
3682         cfq_put_async_queues(cfqd);
3683         cfq_release_cfq_groups(cfqd);
3684         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3685
3686         spin_unlock_irq(q->queue_lock);
3687
3688         cfq_shutdown_timer_wq(cfqd);
3689
3690         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3691         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3692 }
3693
3694 static void *cfq_init_queue(struct request_queue *q)
3695 {
3696         struct cfq_data *cfqd;
3697         int i, j;
3698         struct cfq_group *cfqg;
3699         struct cfq_rb_root *st;
3700
3701         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3702         if (!cfqd)
3703                 return NULL;
3704
3705         /* Init root service tree */
3706         cfqd->grp_service_tree = CFQ_RB_ROOT;
3707
3708         /* Init root group */
3709         cfqg = &cfqd->root_group;
3710         for_each_cfqg_st(cfqg, i, j, st)
3711                 *st = CFQ_RB_ROOT;
3712         RB_CLEAR_NODE(&cfqg->rb_node);
3713
3714         /* Give preference to root group over other groups */
3715         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3716
3717 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3718         /*
3719          * Take a reference to root group which we never drop. This is just
3720          * to make sure that cfq_put_cfqg() does not try to kfree root group
3721          */
3722         atomic_set(&cfqg->ref, 1);
3723         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3724                                         0);
3725 #endif
3726         /*
3727          * Not strictly needed (since RB_ROOT just clears the node and we
3728          * zeroed cfqd on alloc), but better be safe in case someone decides
3729          * to add magic to the rb code
3730          */
3731         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3732                 cfqd->prio_trees[i] = RB_ROOT;
3733
3734         /*
3735          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3736          * Grab a permanent reference to it, so that the normal code flow
3737          * will not attempt to free it.
3738          */
3739         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3740         atomic_inc(&cfqd->oom_cfqq.ref);
3741         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3742
3743         INIT_LIST_HEAD(&cfqd->cic_list);
3744
3745         cfqd->queue = q;
3746
3747         init_timer(&cfqd->idle_slice_timer);
3748         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3749         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3750
3751         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3752
3753         cfqd->cfq_quantum = cfq_quantum;
3754         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3755         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3756         cfqd->cfq_back_max = cfq_back_max;
3757         cfqd->cfq_back_penalty = cfq_back_penalty;
3758         cfqd->cfq_slice[0] = cfq_slice_async;
3759         cfqd->cfq_slice[1] = cfq_slice_sync;
3760         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3761         cfqd->cfq_slice_idle = cfq_slice_idle;
3762         cfqd->cfq_latency = 1;
3763         cfqd->cfq_group_isolation = 0;
3764         cfqd->hw_tag = -1;
3765         /*
3766          * we optimistically start assuming sync ops weren't delayed in last
3767          * second, in order to have larger depth for async operations.
3768          */
3769         cfqd->last_delayed_sync = jiffies - HZ;
3770         INIT_RCU_HEAD(&cfqd->rcu);
3771         return cfqd;
3772 }
3773
3774 static void cfq_slab_kill(void)
3775 {
3776         /*
3777          * Caller already ensured that pending RCU callbacks are completed,
3778          * so we should have no busy allocations at this point.
3779          */
3780         if (cfq_pool)
3781                 kmem_cache_destroy(cfq_pool);
3782         if (cfq_ioc_pool)
3783                 kmem_cache_destroy(cfq_ioc_pool);
3784 }
3785
3786 static int __init cfq_slab_setup(void)
3787 {
3788         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3789         if (!cfq_pool)
3790                 goto fail;
3791
3792         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3793         if (!cfq_ioc_pool)
3794                 goto fail;
3795
3796         return 0;
3797 fail:
3798         cfq_slab_kill();
3799         return -ENOMEM;
3800 }
3801
3802 /*
3803  * sysfs parts below -->
3804  */
3805 static ssize_t
3806 cfq_var_show(unsigned int var, char *page)
3807 {
3808         return sprintf(page, "%d\n", var);
3809 }
3810
3811 static ssize_t
3812 cfq_var_store(unsigned int *var, const char *page, size_t count)
3813 {
3814         char *p = (char *) page;
3815
3816         *var = simple_strtoul(p, &p, 10);
3817         return count;
3818 }
3819
3820 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3821 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3822 {                                                                       \
3823         struct cfq_data *cfqd = e->elevator_data;                       \
3824         unsigned int __data = __VAR;                                    \
3825         if (__CONV)                                                     \
3826                 __data = jiffies_to_msecs(__data);                      \
3827         return cfq_var_show(__data, (page));                            \
3828 }
3829 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3830 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3831 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3832 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3833 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3834 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3835 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3836 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3837 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3838 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3839 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3840 #undef SHOW_FUNCTION
3841
3842 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3843 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3844 {                                                                       \
3845         struct cfq_data *cfqd = e->elevator_data;                       \
3846         unsigned int __data;                                            \
3847         int ret = cfq_var_store(&__data, (page), count);                \
3848         if (__data < (MIN))                                             \
3849                 __data = (MIN);                                         \
3850         else if (__data > (MAX))                                        \
3851                 __data = (MAX);                                         \
3852         if (__CONV)                                                     \
3853                 *(__PTR) = msecs_to_jiffies(__data);                    \
3854         else                                                            \
3855                 *(__PTR) = __data;                                      \
3856         return ret;                                                     \
3857 }
3858 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3859 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3860                 UINT_MAX, 1);
3861 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3862                 UINT_MAX, 1);
3863 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3864 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3865                 UINT_MAX, 0);
3866 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3867 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3868 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3869 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3870                 UINT_MAX, 0);
3871 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3872 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3873 #undef STORE_FUNCTION
3874
3875 #define CFQ_ATTR(name) \
3876         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3877
3878 static struct elv_fs_entry cfq_attrs[] = {
3879         CFQ_ATTR(quantum),
3880         CFQ_ATTR(fifo_expire_sync),
3881         CFQ_ATTR(fifo_expire_async),
3882         CFQ_ATTR(back_seek_max),
3883         CFQ_ATTR(back_seek_penalty),
3884         CFQ_ATTR(slice_sync),
3885         CFQ_ATTR(slice_async),
3886         CFQ_ATTR(slice_async_rq),
3887         CFQ_ATTR(slice_idle),
3888         CFQ_ATTR(low_latency),
3889         CFQ_ATTR(group_isolation),
3890         __ATTR_NULL
3891 };
3892
3893 static struct elevator_type iosched_cfq = {
3894         .ops = {
3895                 .elevator_merge_fn =            cfq_merge,
3896                 .elevator_merged_fn =           cfq_merged_request,
3897                 .elevator_merge_req_fn =        cfq_merged_requests,
3898                 .elevator_allow_merge_fn =      cfq_allow_merge,
3899                 .elevator_bio_merged_fn =       cfq_bio_merged,
3900                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3901                 .elevator_add_req_fn =          cfq_insert_request,
3902                 .elevator_activate_req_fn =     cfq_activate_request,
3903                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3904                 .elevator_queue_empty_fn =      cfq_queue_empty,
3905                 .elevator_completed_req_fn =    cfq_completed_request,
3906                 .elevator_former_req_fn =       elv_rb_former_request,
3907                 .elevator_latter_req_fn =       elv_rb_latter_request,
3908                 .elevator_set_req_fn =          cfq_set_request,
3909                 .elevator_put_req_fn =          cfq_put_request,
3910                 .elevator_may_queue_fn =        cfq_may_queue,
3911                 .elevator_init_fn =             cfq_init_queue,
3912                 .elevator_exit_fn =             cfq_exit_queue,
3913                 .trim =                         cfq_free_io_context,
3914         },
3915         .elevator_attrs =       cfq_attrs,
3916         .elevator_name =        "cfq",
3917         .elevator_owner =       THIS_MODULE,
3918 };
3919
3920 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3921 static struct blkio_policy_type blkio_policy_cfq = {
3922         .ops = {
3923                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3924                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3925         },
3926 };
3927 #else
3928 static struct blkio_policy_type blkio_policy_cfq;
3929 #endif
3930
3931 static int __init cfq_init(void)
3932 {
3933         /*
3934          * could be 0 on HZ < 1000 setups
3935          */
3936         if (!cfq_slice_async)
3937                 cfq_slice_async = 1;
3938         if (!cfq_slice_idle)
3939                 cfq_slice_idle = 1;
3940
3941         if (cfq_slab_setup())
3942                 return -ENOMEM;
3943
3944         elv_register(&iosched_cfq);
3945         blkio_policy_register(&blkio_policy_cfq);
3946
3947         return 0;
3948 }
3949
3950 static void __exit cfq_exit(void)
3951 {
3952         DECLARE_COMPLETION_ONSTACK(all_gone);
3953         blkio_policy_unregister(&blkio_policy_cfq);
3954         elv_unregister(&iosched_cfq);
3955         ioc_gone = &all_gone;
3956         /* ioc_gone's update must be visible before reading ioc_count */
3957         smp_wmb();
3958
3959         /*
3960          * this also protects us from entering cfq_slab_kill() with
3961          * pending RCU callbacks
3962          */
3963         if (elv_ioc_count_read(cfq_ioc_count))
3964                 wait_for_completion(&all_gone);
3965         cfq_slab_kill();
3966 }
3967
3968 module_init(cfq_init);
3969 module_exit(cfq_exit);
3970
3971 MODULE_AUTHOR("Jens Axboe");
3972 MODULE_LICENSE("GPL");
3973 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");