8e0b86a9111a07f38b3284affc73a593d8cbb054
[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)
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 }
920
921 #ifdef CONFIG_CFQ_GROUP_IOSCHED
922 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
923 {
924         if (blkg)
925                 return container_of(blkg, struct cfq_group, blkg);
926         return NULL;
927 }
928
929 void
930 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
931 {
932         cfqg_of_blkg(blkg)->weight = weight;
933 }
934
935 static struct cfq_group *
936 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
937 {
938         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
939         struct cfq_group *cfqg = NULL;
940         void *key = cfqd;
941         int i, j;
942         struct cfq_rb_root *st;
943         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
944         unsigned int major, minor;
945
946         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
947         if (cfqg || !create)
948                 goto done;
949
950         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
951         if (!cfqg)
952                 goto done;
953
954         cfqg->weight = blkcg->weight;
955         for_each_cfqg_st(cfqg, i, j, st)
956                 *st = CFQ_RB_ROOT;
957         RB_CLEAR_NODE(&cfqg->rb_node);
958         blkio_group_init(&cfqg->blkg);
959
960         /*
961          * Take the initial reference that will be released on destroy
962          * This can be thought of a joint reference by cgroup and
963          * elevator which will be dropped by either elevator exit
964          * or cgroup deletion path depending on who is exiting first.
965          */
966         atomic_set(&cfqg->ref, 1);
967
968         /* Add group onto cgroup list */
969         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
970         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
971                                         MKDEV(major, minor));
972
973         /* Add group on cfqd list */
974         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
975
976 done:
977         return cfqg;
978 }
979
980 /*
981  * Search for the cfq group current task belongs to. If create = 1, then also
982  * create the cfq group if it does not exist. request_queue lock must be held.
983  */
984 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
985 {
986         struct cgroup *cgroup;
987         struct cfq_group *cfqg = NULL;
988
989         rcu_read_lock();
990         cgroup = task_cgroup(current, blkio_subsys_id);
991         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
992         if (!cfqg && create)
993                 cfqg = &cfqd->root_group;
994         rcu_read_unlock();
995         return cfqg;
996 }
997
998 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
999 {
1000         /* Currently, all async queues are mapped to root group */
1001         if (!cfq_cfqq_sync(cfqq))
1002                 cfqg = &cfqq->cfqd->root_group;
1003
1004         cfqq->cfqg = cfqg;
1005         /* cfqq reference on cfqg */
1006         atomic_inc(&cfqq->cfqg->ref);
1007 }
1008
1009 static void cfq_put_cfqg(struct cfq_group *cfqg)
1010 {
1011         struct cfq_rb_root *st;
1012         int i, j;
1013
1014         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1015         if (!atomic_dec_and_test(&cfqg->ref))
1016                 return;
1017         for_each_cfqg_st(cfqg, i, j, st)
1018                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1019         kfree(cfqg);
1020 }
1021
1022 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1023 {
1024         /* Something wrong if we are trying to remove same group twice */
1025         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1026
1027         hlist_del_init(&cfqg->cfqd_node);
1028
1029         /*
1030          * Put the reference taken at the time of creation so that when all
1031          * queues are gone, group can be destroyed.
1032          */
1033         cfq_put_cfqg(cfqg);
1034 }
1035
1036 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1037 {
1038         struct hlist_node *pos, *n;
1039         struct cfq_group *cfqg;
1040
1041         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1042                 /*
1043                  * If cgroup removal path got to blk_group first and removed
1044                  * it from cgroup list, then it will take care of destroying
1045                  * cfqg also.
1046                  */
1047                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1048                         cfq_destroy_cfqg(cfqd, cfqg);
1049         }
1050 }
1051
1052 /*
1053  * Blk cgroup controller notification saying that blkio_group object is being
1054  * delinked as associated cgroup object is going away. That also means that
1055  * no new IO will come in this group. So get rid of this group as soon as
1056  * any pending IO in the group is finished.
1057  *
1058  * This function is called under rcu_read_lock(). key is the rcu protected
1059  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1060  * read lock.
1061  *
1062  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1063  * it should not be NULL as even if elevator was exiting, cgroup deltion
1064  * path got to it first.
1065  */
1066 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1067 {
1068         unsigned long  flags;
1069         struct cfq_data *cfqd = key;
1070
1071         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1072         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1073         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1074 }
1075
1076 #else /* GROUP_IOSCHED */
1077 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1078 {
1079         return &cfqd->root_group;
1080 }
1081 static inline void
1082 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1083         cfqq->cfqg = cfqg;
1084 }
1085
1086 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1087 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1088
1089 #endif /* GROUP_IOSCHED */
1090
1091 /*
1092  * The cfqd->service_trees holds all pending cfq_queue's that have
1093  * requests waiting to be processed. It is sorted in the order that
1094  * we will service the queues.
1095  */
1096 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1097                                  bool add_front)
1098 {
1099         struct rb_node **p, *parent;
1100         struct cfq_queue *__cfqq;
1101         unsigned long rb_key;
1102         struct cfq_rb_root *service_tree;
1103         int left;
1104         int new_cfqq = 1;
1105         int group_changed = 0;
1106
1107 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1108         if (!cfqd->cfq_group_isolation
1109             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1110             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1111                 /* Move this cfq to root group */
1112                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1113                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1114                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1115                 cfqq->orig_cfqg = cfqq->cfqg;
1116                 cfqq->cfqg = &cfqd->root_group;
1117                 atomic_inc(&cfqd->root_group.ref);
1118                 group_changed = 1;
1119         } else if (!cfqd->cfq_group_isolation
1120                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1121                 /* cfqq is sequential now needs to go to its original group */
1122                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1123                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1124                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1125                 cfq_put_cfqg(cfqq->cfqg);
1126                 cfqq->cfqg = cfqq->orig_cfqg;
1127                 cfqq->orig_cfqg = NULL;
1128                 group_changed = 1;
1129                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1130         }
1131 #endif
1132
1133         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1134                                                 cfqq_type(cfqq));
1135         if (cfq_class_idle(cfqq)) {
1136                 rb_key = CFQ_IDLE_DELAY;
1137                 parent = rb_last(&service_tree->rb);
1138                 if (parent && parent != &cfqq->rb_node) {
1139                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1140                         rb_key += __cfqq->rb_key;
1141                 } else
1142                         rb_key += jiffies;
1143         } else if (!add_front) {
1144                 /*
1145                  * Get our rb key offset. Subtract any residual slice
1146                  * value carried from last service. A negative resid
1147                  * count indicates slice overrun, and this should position
1148                  * the next service time further away in the tree.
1149                  */
1150                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1151                 rb_key -= cfqq->slice_resid;
1152                 cfqq->slice_resid = 0;
1153         } else {
1154                 rb_key = -HZ;
1155                 __cfqq = cfq_rb_first(service_tree);
1156                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1157         }
1158
1159         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1160                 new_cfqq = 0;
1161                 /*
1162                  * same position, nothing more to do
1163                  */
1164                 if (rb_key == cfqq->rb_key &&
1165                     cfqq->service_tree == service_tree)
1166                         return;
1167
1168                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1169                 cfqq->service_tree = NULL;
1170         }
1171
1172         left = 1;
1173         parent = NULL;
1174         cfqq->service_tree = service_tree;
1175         p = &service_tree->rb.rb_node;
1176         while (*p) {
1177                 struct rb_node **n;
1178
1179                 parent = *p;
1180                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181
1182                 /*
1183                  * sort by key, that represents service time.
1184                  */
1185                 if (time_before(rb_key, __cfqq->rb_key))
1186                         n = &(*p)->rb_left;
1187                 else {
1188                         n = &(*p)->rb_right;
1189                         left = 0;
1190                 }
1191
1192                 p = n;
1193         }
1194
1195         if (left)
1196                 service_tree->left = &cfqq->rb_node;
1197
1198         cfqq->rb_key = rb_key;
1199         rb_link_node(&cfqq->rb_node, parent, p);
1200         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1201         service_tree->count++;
1202         if ((add_front || !new_cfqq) && !group_changed)
1203                 return;
1204         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1205 }
1206
1207 static struct cfq_queue *
1208 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1209                      sector_t sector, struct rb_node **ret_parent,
1210                      struct rb_node ***rb_link)
1211 {
1212         struct rb_node **p, *parent;
1213         struct cfq_queue *cfqq = NULL;
1214
1215         parent = NULL;
1216         p = &root->rb_node;
1217         while (*p) {
1218                 struct rb_node **n;
1219
1220                 parent = *p;
1221                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1222
1223                 /*
1224                  * Sort strictly based on sector.  Smallest to the left,
1225                  * largest to the right.
1226                  */
1227                 if (sector > blk_rq_pos(cfqq->next_rq))
1228                         n = &(*p)->rb_right;
1229                 else if (sector < blk_rq_pos(cfqq->next_rq))
1230                         n = &(*p)->rb_left;
1231                 else
1232                         break;
1233                 p = n;
1234                 cfqq = NULL;
1235         }
1236
1237         *ret_parent = parent;
1238         if (rb_link)
1239                 *rb_link = p;
1240         return cfqq;
1241 }
1242
1243 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1244 {
1245         struct rb_node **p, *parent;
1246         struct cfq_queue *__cfqq;
1247
1248         if (cfqq->p_root) {
1249                 rb_erase(&cfqq->p_node, cfqq->p_root);
1250                 cfqq->p_root = NULL;
1251         }
1252
1253         if (cfq_class_idle(cfqq))
1254                 return;
1255         if (!cfqq->next_rq)
1256                 return;
1257
1258         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1259         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1260                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1261         if (!__cfqq) {
1262                 rb_link_node(&cfqq->p_node, parent, p);
1263                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1264         } else
1265                 cfqq->p_root = NULL;
1266 }
1267
1268 /*
1269  * Update cfqq's position in the service tree.
1270  */
1271 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1272 {
1273         /*
1274          * Resorting requires the cfqq to be on the RR list already.
1275          */
1276         if (cfq_cfqq_on_rr(cfqq)) {
1277                 cfq_service_tree_add(cfqd, cfqq, 0);
1278                 cfq_prio_tree_add(cfqd, cfqq);
1279         }
1280 }
1281
1282 /*
1283  * add to busy list of queues for service, trying to be fair in ordering
1284  * the pending list according to last request service
1285  */
1286 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1287 {
1288         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1289         BUG_ON(cfq_cfqq_on_rr(cfqq));
1290         cfq_mark_cfqq_on_rr(cfqq);
1291         cfqd->busy_queues++;
1292
1293         cfq_resort_rr_list(cfqd, cfqq);
1294 }
1295
1296 /*
1297  * Called when the cfqq no longer has requests pending, remove it from
1298  * the service tree.
1299  */
1300 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1301 {
1302         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1303         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1304         cfq_clear_cfqq_on_rr(cfqq);
1305
1306         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1307                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1308                 cfqq->service_tree = NULL;
1309         }
1310         if (cfqq->p_root) {
1311                 rb_erase(&cfqq->p_node, cfqq->p_root);
1312                 cfqq->p_root = NULL;
1313         }
1314
1315         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1316         BUG_ON(!cfqd->busy_queues);
1317         cfqd->busy_queues--;
1318 }
1319
1320 /*
1321  * rb tree support functions
1322  */
1323 static void cfq_del_rq_rb(struct request *rq)
1324 {
1325         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1326         const int sync = rq_is_sync(rq);
1327
1328         BUG_ON(!cfqq->queued[sync]);
1329         cfqq->queued[sync]--;
1330
1331         elv_rb_del(&cfqq->sort_list, rq);
1332
1333         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1334                 /*
1335                  * Queue will be deleted from service tree when we actually
1336                  * expire it later. Right now just remove it from prio tree
1337                  * as it is empty.
1338                  */
1339                 if (cfqq->p_root) {
1340                         rb_erase(&cfqq->p_node, cfqq->p_root);
1341                         cfqq->p_root = NULL;
1342                 }
1343         }
1344 }
1345
1346 static void cfq_add_rq_rb(struct request *rq)
1347 {
1348         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1349         struct cfq_data *cfqd = cfqq->cfqd;
1350         struct request *__alias, *prev;
1351
1352         cfqq->queued[rq_is_sync(rq)]++;
1353
1354         /*
1355          * looks a little odd, but the first insert might return an alias.
1356          * if that happens, put the alias on the dispatch list
1357          */
1358         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1359                 cfq_dispatch_insert(cfqd->queue, __alias);
1360
1361         if (!cfq_cfqq_on_rr(cfqq))
1362                 cfq_add_cfqq_rr(cfqd, cfqq);
1363
1364         /*
1365          * check if this request is a better next-serve candidate
1366          */
1367         prev = cfqq->next_rq;
1368         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1369
1370         /*
1371          * adjust priority tree position, if ->next_rq changes
1372          */
1373         if (prev != cfqq->next_rq)
1374                 cfq_prio_tree_add(cfqd, cfqq);
1375
1376         BUG_ON(!cfqq->next_rq);
1377 }
1378
1379 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1380 {
1381         elv_rb_del(&cfqq->sort_list, rq);
1382         cfqq->queued[rq_is_sync(rq)]--;
1383         blkiocg_update_request_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1384                                                 rq_is_sync(rq));
1385         cfq_add_rq_rb(rq);
1386         blkiocg_update_request_add_stats(
1387                         &cfqq->cfqg->blkg, &cfqq->cfqd->serving_group->blkg,
1388                         rq_data_dir(rq), rq_is_sync(rq));
1389 }
1390
1391 static struct request *
1392 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1393 {
1394         struct task_struct *tsk = current;
1395         struct cfq_io_context *cic;
1396         struct cfq_queue *cfqq;
1397
1398         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1399         if (!cic)
1400                 return NULL;
1401
1402         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1403         if (cfqq) {
1404                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1405
1406                 return elv_rb_find(&cfqq->sort_list, sector);
1407         }
1408
1409         return NULL;
1410 }
1411
1412 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1413 {
1414         struct cfq_data *cfqd = q->elevator->elevator_data;
1415
1416         cfqd->rq_in_driver++;
1417         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1418                                                 cfqd->rq_in_driver);
1419
1420         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1421 }
1422
1423 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1424 {
1425         struct cfq_data *cfqd = q->elevator->elevator_data;
1426
1427         WARN_ON(!cfqd->rq_in_driver);
1428         cfqd->rq_in_driver--;
1429         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1430                                                 cfqd->rq_in_driver);
1431 }
1432
1433 static void cfq_remove_request(struct request *rq)
1434 {
1435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1436
1437         if (cfqq->next_rq == rq)
1438                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1439
1440         list_del_init(&rq->queuelist);
1441         cfq_del_rq_rb(rq);
1442
1443         cfqq->cfqd->rq_queued--;
1444         blkiocg_update_request_remove_stats(&cfqq->cfqg->blkg, rq_data_dir(rq),
1445                                                 rq_is_sync(rq));
1446         if (rq_is_meta(rq)) {
1447                 WARN_ON(!cfqq->meta_pending);
1448                 cfqq->meta_pending--;
1449         }
1450 }
1451
1452 static int cfq_merge(struct request_queue *q, struct request **req,
1453                      struct bio *bio)
1454 {
1455         struct cfq_data *cfqd = q->elevator->elevator_data;
1456         struct request *__rq;
1457
1458         __rq = cfq_find_rq_fmerge(cfqd, bio);
1459         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1460                 *req = __rq;
1461                 return ELEVATOR_FRONT_MERGE;
1462         }
1463
1464         return ELEVATOR_NO_MERGE;
1465 }
1466
1467 static void cfq_merged_request(struct request_queue *q, struct request *req,
1468                                int type)
1469 {
1470         if (type == ELEVATOR_FRONT_MERGE) {
1471                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1472
1473                 cfq_reposition_rq_rb(cfqq, req);
1474         }
1475 }
1476
1477 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1478                                 struct bio *bio)
1479 {
1480         struct cfq_queue *cfqq = RQ_CFQQ(req);
1481         blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, bio_data_dir(bio),
1482                                         cfq_bio_sync(bio));
1483 }
1484
1485 static void
1486 cfq_merged_requests(struct request_queue *q, struct request *rq,
1487                     struct request *next)
1488 {
1489         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1490         /*
1491          * reposition in fifo if next is older than rq
1492          */
1493         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1494             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1495                 list_move(&rq->queuelist, &next->queuelist);
1496                 rq_set_fifo_time(rq, rq_fifo_time(next));
1497         }
1498
1499         if (cfqq->next_rq == next)
1500                 cfqq->next_rq = rq;
1501         cfq_remove_request(next);
1502         blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, rq_data_dir(next),
1503                                         rq_is_sync(next));
1504 }
1505
1506 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1507                            struct bio *bio)
1508 {
1509         struct cfq_data *cfqd = q->elevator->elevator_data;
1510         struct cfq_io_context *cic;
1511         struct cfq_queue *cfqq;
1512
1513         /*
1514          * Disallow merge of a sync bio into an async request.
1515          */
1516         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1517                 return false;
1518
1519         /*
1520          * Lookup the cfqq that this bio will be queued with. Allow
1521          * merge only if rq is queued there.
1522          */
1523         cic = cfq_cic_lookup(cfqd, current->io_context);
1524         if (!cic)
1525                 return false;
1526
1527         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1528         return cfqq == RQ_CFQQ(rq);
1529 }
1530
1531 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1532                                    struct cfq_queue *cfqq)
1533 {
1534         if (cfqq) {
1535                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1536                                 cfqd->serving_prio, cfqd->serving_type);
1537                 blkiocg_update_set_active_queue_stats(&cfqq->cfqg->blkg);
1538                 cfqq->slice_start = 0;
1539                 cfqq->dispatch_start = jiffies;
1540                 cfqq->allocated_slice = 0;
1541                 cfqq->slice_end = 0;
1542                 cfqq->slice_dispatch = 0;
1543
1544                 cfq_clear_cfqq_wait_request(cfqq);
1545                 cfq_clear_cfqq_must_dispatch(cfqq);
1546                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1547                 cfq_clear_cfqq_fifo_expire(cfqq);
1548                 cfq_mark_cfqq_slice_new(cfqq);
1549
1550                 del_timer(&cfqd->idle_slice_timer);
1551         }
1552
1553         cfqd->active_queue = cfqq;
1554 }
1555
1556 /*
1557  * current cfqq expired its slice (or was too idle), select new one
1558  */
1559 static void
1560 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1561                     bool timed_out)
1562 {
1563         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1564
1565         if (cfq_cfqq_wait_request(cfqq))
1566                 del_timer(&cfqd->idle_slice_timer);
1567
1568         cfq_clear_cfqq_wait_request(cfqq);
1569         cfq_clear_cfqq_wait_busy(cfqq);
1570
1571         /*
1572          * If this cfqq is shared between multiple processes, check to
1573          * make sure that those processes are still issuing I/Os within
1574          * the mean seek distance.  If not, it may be time to break the
1575          * queues apart again.
1576          */
1577         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1578                 cfq_mark_cfqq_split_coop(cfqq);
1579
1580         /*
1581          * store what was left of this slice, if the queue idled/timed out
1582          */
1583         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1584                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1585                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1586         }
1587
1588         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1589
1590         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1591                 cfq_del_cfqq_rr(cfqd, cfqq);
1592
1593         cfq_resort_rr_list(cfqd, cfqq);
1594
1595         if (cfqq == cfqd->active_queue)
1596                 cfqd->active_queue = NULL;
1597
1598         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1599                 cfqd->grp_service_tree.active = NULL;
1600
1601         if (cfqd->active_cic) {
1602                 put_io_context(cfqd->active_cic->ioc);
1603                 cfqd->active_cic = NULL;
1604         }
1605 }
1606
1607 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1608 {
1609         struct cfq_queue *cfqq = cfqd->active_queue;
1610
1611         if (cfqq)
1612                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1613 }
1614
1615 /*
1616  * Get next queue for service. Unless we have a queue preemption,
1617  * we'll simply select the first cfqq in the service tree.
1618  */
1619 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1620 {
1621         struct cfq_rb_root *service_tree =
1622                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1623                                         cfqd->serving_type);
1624
1625         if (!cfqd->rq_queued)
1626                 return NULL;
1627
1628         /* There is nothing to dispatch */
1629         if (!service_tree)
1630                 return NULL;
1631         if (RB_EMPTY_ROOT(&service_tree->rb))
1632                 return NULL;
1633         return cfq_rb_first(service_tree);
1634 }
1635
1636 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1637 {
1638         struct cfq_group *cfqg;
1639         struct cfq_queue *cfqq;
1640         int i, j;
1641         struct cfq_rb_root *st;
1642
1643         if (!cfqd->rq_queued)
1644                 return NULL;
1645
1646         cfqg = cfq_get_next_cfqg(cfqd);
1647         if (!cfqg)
1648                 return NULL;
1649
1650         for_each_cfqg_st(cfqg, i, j, st)
1651                 if ((cfqq = cfq_rb_first(st)) != NULL)
1652                         return cfqq;
1653         return NULL;
1654 }
1655
1656 /*
1657  * Get and set a new active queue for service.
1658  */
1659 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1660                                               struct cfq_queue *cfqq)
1661 {
1662         if (!cfqq)
1663                 cfqq = cfq_get_next_queue(cfqd);
1664
1665         __cfq_set_active_queue(cfqd, cfqq);
1666         return cfqq;
1667 }
1668
1669 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1670                                           struct request *rq)
1671 {
1672         if (blk_rq_pos(rq) >= cfqd->last_position)
1673                 return blk_rq_pos(rq) - cfqd->last_position;
1674         else
1675                 return cfqd->last_position - blk_rq_pos(rq);
1676 }
1677
1678 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1679                                struct request *rq)
1680 {
1681         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1682 }
1683
1684 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1685                                     struct cfq_queue *cur_cfqq)
1686 {
1687         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1688         struct rb_node *parent, *node;
1689         struct cfq_queue *__cfqq;
1690         sector_t sector = cfqd->last_position;
1691
1692         if (RB_EMPTY_ROOT(root))
1693                 return NULL;
1694
1695         /*
1696          * First, if we find a request starting at the end of the last
1697          * request, choose it.
1698          */
1699         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1700         if (__cfqq)
1701                 return __cfqq;
1702
1703         /*
1704          * If the exact sector wasn't found, the parent of the NULL leaf
1705          * will contain the closest sector.
1706          */
1707         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1708         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1709                 return __cfqq;
1710
1711         if (blk_rq_pos(__cfqq->next_rq) < sector)
1712                 node = rb_next(&__cfqq->p_node);
1713         else
1714                 node = rb_prev(&__cfqq->p_node);
1715         if (!node)
1716                 return NULL;
1717
1718         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1719         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1720                 return __cfqq;
1721
1722         return NULL;
1723 }
1724
1725 /*
1726  * cfqd - obvious
1727  * cur_cfqq - passed in so that we don't decide that the current queue is
1728  *            closely cooperating with itself.
1729  *
1730  * So, basically we're assuming that that cur_cfqq has dispatched at least
1731  * one request, and that cfqd->last_position reflects a position on the disk
1732  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1733  * assumption.
1734  */
1735 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1736                                               struct cfq_queue *cur_cfqq)
1737 {
1738         struct cfq_queue *cfqq;
1739
1740         if (cfq_class_idle(cur_cfqq))
1741                 return NULL;
1742         if (!cfq_cfqq_sync(cur_cfqq))
1743                 return NULL;
1744         if (CFQQ_SEEKY(cur_cfqq))
1745                 return NULL;
1746
1747         /*
1748          * Don't search priority tree if it's the only queue in the group.
1749          */
1750         if (cur_cfqq->cfqg->nr_cfqq == 1)
1751                 return NULL;
1752
1753         /*
1754          * We should notice if some of the queues are cooperating, eg
1755          * working closely on the same area of the disk. In that case,
1756          * we can group them together and don't waste time idling.
1757          */
1758         cfqq = cfqq_close(cfqd, cur_cfqq);
1759         if (!cfqq)
1760                 return NULL;
1761
1762         /* If new queue belongs to different cfq_group, don't choose it */
1763         if (cur_cfqq->cfqg != cfqq->cfqg)
1764                 return NULL;
1765
1766         /*
1767          * It only makes sense to merge sync queues.
1768          */
1769         if (!cfq_cfqq_sync(cfqq))
1770                 return NULL;
1771         if (CFQQ_SEEKY(cfqq))
1772                 return NULL;
1773
1774         /*
1775          * Do not merge queues of different priority classes
1776          */
1777         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1778                 return NULL;
1779
1780         return cfqq;
1781 }
1782
1783 /*
1784  * Determine whether we should enforce idle window for this queue.
1785  */
1786
1787 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1788 {
1789         enum wl_prio_t prio = cfqq_prio(cfqq);
1790         struct cfq_rb_root *service_tree = cfqq->service_tree;
1791
1792         BUG_ON(!service_tree);
1793         BUG_ON(!service_tree->count);
1794
1795         /* We never do for idle class queues. */
1796         if (prio == IDLE_WORKLOAD)
1797                 return false;
1798
1799         /* We do for queues that were marked with idle window flag. */
1800         if (cfq_cfqq_idle_window(cfqq) &&
1801            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1802                 return true;
1803
1804         /*
1805          * Otherwise, we do only if they are the last ones
1806          * in their service tree.
1807          */
1808         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1809                 return 1;
1810         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1811                         service_tree->count);
1812         return 0;
1813 }
1814
1815 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1816 {
1817         struct cfq_queue *cfqq = cfqd->active_queue;
1818         struct cfq_io_context *cic;
1819         unsigned long sl;
1820
1821         /*
1822          * SSD device without seek penalty, disable idling. But only do so
1823          * for devices that support queuing, otherwise we still have a problem
1824          * with sync vs async workloads.
1825          */
1826         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1827                 return;
1828
1829         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1830         WARN_ON(cfq_cfqq_slice_new(cfqq));
1831
1832         /*
1833          * idle is disabled, either manually or by past process history
1834          */
1835         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1836                 return;
1837
1838         /*
1839          * still active requests from this queue, don't idle
1840          */
1841         if (cfqq->dispatched)
1842                 return;
1843
1844         /*
1845          * task has exited, don't wait
1846          */
1847         cic = cfqd->active_cic;
1848         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1849                 return;
1850
1851         /*
1852          * If our average think time is larger than the remaining time
1853          * slice, then don't idle. This avoids overrunning the allotted
1854          * time slice.
1855          */
1856         if (sample_valid(cic->ttime_samples) &&
1857             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1858                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1859                                 cic->ttime_mean);
1860                 return;
1861         }
1862
1863         cfq_mark_cfqq_wait_request(cfqq);
1864
1865         sl = cfqd->cfq_slice_idle;
1866
1867         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1868         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1869 }
1870
1871 /*
1872  * Move request from internal lists to the request queue dispatch list.
1873  */
1874 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1875 {
1876         struct cfq_data *cfqd = q->elevator->elevator_data;
1877         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1878
1879         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1880
1881         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1882         cfq_remove_request(rq);
1883         cfqq->dispatched++;
1884         elv_dispatch_sort(q, rq);
1885
1886         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1887         blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1888                                         rq_data_dir(rq), rq_is_sync(rq));
1889 }
1890
1891 /*
1892  * return expired entry, or NULL to just start from scratch in rbtree
1893  */
1894 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1895 {
1896         struct request *rq = NULL;
1897
1898         if (cfq_cfqq_fifo_expire(cfqq))
1899                 return NULL;
1900
1901         cfq_mark_cfqq_fifo_expire(cfqq);
1902
1903         if (list_empty(&cfqq->fifo))
1904                 return NULL;
1905
1906         rq = rq_entry_fifo(cfqq->fifo.next);
1907         if (time_before(jiffies, rq_fifo_time(rq)))
1908                 rq = NULL;
1909
1910         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1911         return rq;
1912 }
1913
1914 static inline int
1915 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1916 {
1917         const int base_rq = cfqd->cfq_slice_async_rq;
1918
1919         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1920
1921         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1922 }
1923
1924 /*
1925  * Must be called with the queue_lock held.
1926  */
1927 static int cfqq_process_refs(struct cfq_queue *cfqq)
1928 {
1929         int process_refs, io_refs;
1930
1931         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1932         process_refs = atomic_read(&cfqq->ref) - io_refs;
1933         BUG_ON(process_refs < 0);
1934         return process_refs;
1935 }
1936
1937 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1938 {
1939         int process_refs, new_process_refs;
1940         struct cfq_queue *__cfqq;
1941
1942         /* Avoid a circular list and skip interim queue merges */
1943         while ((__cfqq = new_cfqq->new_cfqq)) {
1944                 if (__cfqq == cfqq)
1945                         return;
1946                 new_cfqq = __cfqq;
1947         }
1948
1949         process_refs = cfqq_process_refs(cfqq);
1950         /*
1951          * If the process for the cfqq has gone away, there is no
1952          * sense in merging the queues.
1953          */
1954         if (process_refs == 0)
1955                 return;
1956
1957         /*
1958          * Merge in the direction of the lesser amount of work.
1959          */
1960         new_process_refs = cfqq_process_refs(new_cfqq);
1961         if (new_process_refs >= process_refs) {
1962                 cfqq->new_cfqq = new_cfqq;
1963                 atomic_add(process_refs, &new_cfqq->ref);
1964         } else {
1965                 new_cfqq->new_cfqq = cfqq;
1966                 atomic_add(new_process_refs, &cfqq->ref);
1967         }
1968 }
1969
1970 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1971                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1972 {
1973         struct cfq_queue *queue;
1974         int i;
1975         bool key_valid = false;
1976         unsigned long lowest_key = 0;
1977         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1978
1979         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1980                 /* select the one with lowest rb_key */
1981                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1982                 if (queue &&
1983                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1984                         lowest_key = queue->rb_key;
1985                         cur_best = i;
1986                         key_valid = true;
1987                 }
1988         }
1989
1990         return cur_best;
1991 }
1992
1993 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1994 {
1995         unsigned slice;
1996         unsigned count;
1997         struct cfq_rb_root *st;
1998         unsigned group_slice;
1999
2000         if (!cfqg) {
2001                 cfqd->serving_prio = IDLE_WORKLOAD;
2002                 cfqd->workload_expires = jiffies + 1;
2003                 return;
2004         }
2005
2006         /* Choose next priority. RT > BE > IDLE */
2007         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2008                 cfqd->serving_prio = RT_WORKLOAD;
2009         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2010                 cfqd->serving_prio = BE_WORKLOAD;
2011         else {
2012                 cfqd->serving_prio = IDLE_WORKLOAD;
2013                 cfqd->workload_expires = jiffies + 1;
2014                 return;
2015         }
2016
2017         /*
2018          * For RT and BE, we have to choose also the type
2019          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2020          * expiration time
2021          */
2022         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2023         count = st->count;
2024
2025         /*
2026          * check workload expiration, and that we still have other queues ready
2027          */
2028         if (count && !time_after(jiffies, cfqd->workload_expires))
2029                 return;
2030
2031         /* otherwise select new workload type */
2032         cfqd->serving_type =
2033                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2034         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2035         count = st->count;
2036
2037         /*
2038          * the workload slice is computed as a fraction of target latency
2039          * proportional to the number of queues in that workload, over
2040          * all the queues in the same priority class
2041          */
2042         group_slice = cfq_group_slice(cfqd, cfqg);
2043
2044         slice = group_slice * count /
2045                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2046                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2047
2048         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2049                 unsigned int tmp;
2050
2051                 /*
2052                  * Async queues are currently system wide. Just taking
2053                  * proportion of queues with-in same group will lead to higher
2054                  * async ratio system wide as generally root group is going
2055                  * to have higher weight. A more accurate thing would be to
2056                  * calculate system wide asnc/sync ratio.
2057                  */
2058                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2059                 tmp = tmp/cfqd->busy_queues;
2060                 slice = min_t(unsigned, slice, tmp);
2061
2062                 /* async workload slice is scaled down according to
2063                  * the sync/async slice ratio. */
2064                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2065         } else
2066                 /* sync workload slice is at least 2 * cfq_slice_idle */
2067                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2068
2069         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2070         cfq_log(cfqd, "workload slice:%d", slice);
2071         cfqd->workload_expires = jiffies + slice;
2072         cfqd->noidle_tree_requires_idle = false;
2073 }
2074
2075 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2076 {
2077         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2078         struct cfq_group *cfqg;
2079
2080         if (RB_EMPTY_ROOT(&st->rb))
2081                 return NULL;
2082         cfqg = cfq_rb_first_group(st);
2083         st->active = &cfqg->rb_node;
2084         update_min_vdisktime(st);
2085         return cfqg;
2086 }
2087
2088 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2089 {
2090         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2091
2092         cfqd->serving_group = cfqg;
2093
2094         /* Restore the workload type data */
2095         if (cfqg->saved_workload_slice) {
2096                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2097                 cfqd->serving_type = cfqg->saved_workload;
2098                 cfqd->serving_prio = cfqg->saved_serving_prio;
2099         } else
2100                 cfqd->workload_expires = jiffies - 1;
2101
2102         choose_service_tree(cfqd, cfqg);
2103 }
2104
2105 /*
2106  * Select a queue for service. If we have a current active queue,
2107  * check whether to continue servicing it, or retrieve and set a new one.
2108  */
2109 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2110 {
2111         struct cfq_queue *cfqq, *new_cfqq = NULL;
2112
2113         cfqq = cfqd->active_queue;
2114         if (!cfqq)
2115                 goto new_queue;
2116
2117         if (!cfqd->rq_queued)
2118                 return NULL;
2119
2120         /*
2121          * We were waiting for group to get backlogged. Expire the queue
2122          */
2123         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2124                 goto expire;
2125
2126         /*
2127          * The active queue has run out of time, expire it and select new.
2128          */
2129         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2130                 /*
2131                  * If slice had not expired at the completion of last request
2132                  * we might not have turned on wait_busy flag. Don't expire
2133                  * the queue yet. Allow the group to get backlogged.
2134                  *
2135                  * The very fact that we have used the slice, that means we
2136                  * have been idling all along on this queue and it should be
2137                  * ok to wait for this request to complete.
2138                  */
2139                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2140                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2141                         cfqq = NULL;
2142                         goto keep_queue;
2143                 } else
2144                         goto expire;
2145         }
2146
2147         /*
2148          * The active queue has requests and isn't expired, allow it to
2149          * dispatch.
2150          */
2151         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2152                 goto keep_queue;
2153
2154         /*
2155          * If another queue has a request waiting within our mean seek
2156          * distance, let it run.  The expire code will check for close
2157          * cooperators and put the close queue at the front of the service
2158          * tree.  If possible, merge the expiring queue with the new cfqq.
2159          */
2160         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2161         if (new_cfqq) {
2162                 if (!cfqq->new_cfqq)
2163                         cfq_setup_merge(cfqq, new_cfqq);
2164                 goto expire;
2165         }
2166
2167         /*
2168          * No requests pending. If the active queue still has requests in
2169          * flight or is idling for a new request, allow either of these
2170          * conditions to happen (or time out) before selecting a new queue.
2171          */
2172         if (timer_pending(&cfqd->idle_slice_timer) ||
2173             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2174                 cfqq = NULL;
2175                 goto keep_queue;
2176         }
2177
2178 expire:
2179         cfq_slice_expired(cfqd, 0);
2180 new_queue:
2181         /*
2182          * Current queue expired. Check if we have to switch to a new
2183          * service tree
2184          */
2185         if (!new_cfqq)
2186                 cfq_choose_cfqg(cfqd);
2187
2188         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2189 keep_queue:
2190         return cfqq;
2191 }
2192
2193 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2194 {
2195         int dispatched = 0;
2196
2197         while (cfqq->next_rq) {
2198                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2199                 dispatched++;
2200         }
2201
2202         BUG_ON(!list_empty(&cfqq->fifo));
2203
2204         /* By default cfqq is not expired if it is empty. Do it explicitly */
2205         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2206         return dispatched;
2207 }
2208
2209 /*
2210  * Drain our current requests. Used for barriers and when switching
2211  * io schedulers on-the-fly.
2212  */
2213 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2214 {
2215         struct cfq_queue *cfqq;
2216         int dispatched = 0;
2217
2218         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2219                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2220
2221         cfq_slice_expired(cfqd, 0);
2222         BUG_ON(cfqd->busy_queues);
2223
2224         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2225         return dispatched;
2226 }
2227
2228 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2229         struct cfq_queue *cfqq)
2230 {
2231         /* the queue hasn't finished any request, can't estimate */
2232         if (cfq_cfqq_slice_new(cfqq))
2233                 return 1;
2234         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2235                 cfqq->slice_end))
2236                 return 1;
2237
2238         return 0;
2239 }
2240
2241 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2242 {
2243         unsigned int max_dispatch;
2244
2245         /*
2246          * Drain async requests before we start sync IO
2247          */
2248         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2249                 return false;
2250
2251         /*
2252          * If this is an async queue and we have sync IO in flight, let it wait
2253          */
2254         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2255                 return false;
2256
2257         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2258         if (cfq_class_idle(cfqq))
2259                 max_dispatch = 1;
2260
2261         /*
2262          * Does this cfqq already have too much IO in flight?
2263          */
2264         if (cfqq->dispatched >= max_dispatch) {
2265                 /*
2266                  * idle queue must always only have a single IO in flight
2267                  */
2268                 if (cfq_class_idle(cfqq))
2269                         return false;
2270
2271                 /*
2272                  * We have other queues, don't allow more IO from this one
2273                  */
2274                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2275                         return false;
2276
2277                 /*
2278                  * Sole queue user, no limit
2279                  */
2280                 if (cfqd->busy_queues == 1)
2281                         max_dispatch = -1;
2282                 else
2283                         /*
2284                          * Normally we start throttling cfqq when cfq_quantum/2
2285                          * requests have been dispatched. But we can drive
2286                          * deeper queue depths at the beginning of slice
2287                          * subjected to upper limit of cfq_quantum.
2288                          * */
2289                         max_dispatch = cfqd->cfq_quantum;
2290         }
2291
2292         /*
2293          * Async queues must wait a bit before being allowed dispatch.
2294          * We also ramp up the dispatch depth gradually for async IO,
2295          * based on the last sync IO we serviced
2296          */
2297         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2298                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2299                 unsigned int depth;
2300
2301                 depth = last_sync / cfqd->cfq_slice[1];
2302                 if (!depth && !cfqq->dispatched)
2303                         depth = 1;
2304                 if (depth < max_dispatch)
2305                         max_dispatch = depth;
2306         }
2307
2308         /*
2309          * If we're below the current max, allow a dispatch
2310          */
2311         return cfqq->dispatched < max_dispatch;
2312 }
2313
2314 /*
2315  * Dispatch a request from cfqq, moving them to the request queue
2316  * dispatch list.
2317  */
2318 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2319 {
2320         struct request *rq;
2321
2322         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2323
2324         if (!cfq_may_dispatch(cfqd, cfqq))
2325                 return false;
2326
2327         /*
2328          * follow expired path, else get first next available
2329          */
2330         rq = cfq_check_fifo(cfqq);
2331         if (!rq)
2332                 rq = cfqq->next_rq;
2333
2334         /*
2335          * insert request into driver dispatch list
2336          */
2337         cfq_dispatch_insert(cfqd->queue, rq);
2338
2339         if (!cfqd->active_cic) {
2340                 struct cfq_io_context *cic = RQ_CIC(rq);
2341
2342                 atomic_long_inc(&cic->ioc->refcount);
2343                 cfqd->active_cic = cic;
2344         }
2345
2346         return true;
2347 }
2348
2349 /*
2350  * Find the cfqq that we need to service and move a request from that to the
2351  * dispatch list
2352  */
2353 static int cfq_dispatch_requests(struct request_queue *q, int force)
2354 {
2355         struct cfq_data *cfqd = q->elevator->elevator_data;
2356         struct cfq_queue *cfqq;
2357
2358         if (!cfqd->busy_queues)
2359                 return 0;
2360
2361         if (unlikely(force))
2362                 return cfq_forced_dispatch(cfqd);
2363
2364         cfqq = cfq_select_queue(cfqd);
2365         if (!cfqq)
2366                 return 0;
2367
2368         /*
2369          * Dispatch a request from this cfqq, if it is allowed
2370          */
2371         if (!cfq_dispatch_request(cfqd, cfqq))
2372                 return 0;
2373
2374         cfqq->slice_dispatch++;
2375         cfq_clear_cfqq_must_dispatch(cfqq);
2376
2377         /*
2378          * expire an async queue immediately if it has used up its slice. idle
2379          * queue always expire after 1 dispatch round.
2380          */
2381         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2382             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2383             cfq_class_idle(cfqq))) {
2384                 cfqq->slice_end = jiffies + 1;
2385                 cfq_slice_expired(cfqd, 0);
2386         }
2387
2388         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2389         return 1;
2390 }
2391
2392 /*
2393  * task holds one reference to the queue, dropped when task exits. each rq
2394  * in-flight on this queue also holds a reference, dropped when rq is freed.
2395  *
2396  * Each cfq queue took a reference on the parent group. Drop it now.
2397  * queue lock must be held here.
2398  */
2399 static void cfq_put_queue(struct cfq_queue *cfqq)
2400 {
2401         struct cfq_data *cfqd = cfqq->cfqd;
2402         struct cfq_group *cfqg, *orig_cfqg;
2403
2404         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2405
2406         if (!atomic_dec_and_test(&cfqq->ref))
2407                 return;
2408
2409         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2410         BUG_ON(rb_first(&cfqq->sort_list));
2411         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2412         cfqg = cfqq->cfqg;
2413         orig_cfqg = cfqq->orig_cfqg;
2414
2415         if (unlikely(cfqd->active_queue == cfqq)) {
2416                 __cfq_slice_expired(cfqd, cfqq, 0);
2417                 cfq_schedule_dispatch(cfqd);
2418         }
2419
2420         BUG_ON(cfq_cfqq_on_rr(cfqq));
2421         kmem_cache_free(cfq_pool, cfqq);
2422         cfq_put_cfqg(cfqg);
2423         if (orig_cfqg)
2424                 cfq_put_cfqg(orig_cfqg);
2425 }
2426
2427 /*
2428  * Must always be called with the rcu_read_lock() held
2429  */
2430 static void
2431 __call_for_each_cic(struct io_context *ioc,
2432                     void (*func)(struct io_context *, struct cfq_io_context *))
2433 {
2434         struct cfq_io_context *cic;
2435         struct hlist_node *n;
2436
2437         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2438                 func(ioc, cic);
2439 }
2440
2441 /*
2442  * Call func for each cic attached to this ioc.
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         rcu_read_lock();
2449         __call_for_each_cic(ioc, func);
2450         rcu_read_unlock();
2451 }
2452
2453 static void cfq_cic_free_rcu(struct rcu_head *head)
2454 {
2455         struct cfq_io_context *cic;
2456
2457         cic = container_of(head, struct cfq_io_context, rcu_head);
2458
2459         kmem_cache_free(cfq_ioc_pool, cic);
2460         elv_ioc_count_dec(cfq_ioc_count);
2461
2462         if (ioc_gone) {
2463                 /*
2464                  * CFQ scheduler is exiting, grab exit lock and check
2465                  * the pending io context count. If it hits zero,
2466                  * complete ioc_gone and set it back to NULL
2467                  */
2468                 spin_lock(&ioc_gone_lock);
2469                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2470                         complete(ioc_gone);
2471                         ioc_gone = NULL;
2472                 }
2473                 spin_unlock(&ioc_gone_lock);
2474         }
2475 }
2476
2477 static void cfq_cic_free(struct cfq_io_context *cic)
2478 {
2479         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2480 }
2481
2482 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2483 {
2484         unsigned long flags;
2485
2486         BUG_ON(!cic->dead_key);
2487
2488         spin_lock_irqsave(&ioc->lock, flags);
2489         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2490         hlist_del_rcu(&cic->cic_list);
2491         spin_unlock_irqrestore(&ioc->lock, flags);
2492
2493         cfq_cic_free(cic);
2494 }
2495
2496 /*
2497  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2498  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2499  * and ->trim() which is called with the task lock held
2500  */
2501 static void cfq_free_io_context(struct io_context *ioc)
2502 {
2503         /*
2504          * ioc->refcount is zero here, or we are called from elv_unregister(),
2505          * so no more cic's are allowed to be linked into this ioc.  So it
2506          * should be ok to iterate over the known list, we will see all cic's
2507          * since no new ones are added.
2508          */
2509         __call_for_each_cic(ioc, cic_free_func);
2510 }
2511
2512 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2513 {
2514         struct cfq_queue *__cfqq, *next;
2515
2516         if (unlikely(cfqq == cfqd->active_queue)) {
2517                 __cfq_slice_expired(cfqd, cfqq, 0);
2518                 cfq_schedule_dispatch(cfqd);
2519         }
2520
2521         /*
2522          * If this queue was scheduled to merge with another queue, be
2523          * sure to drop the reference taken on that queue (and others in
2524          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2525          */
2526         __cfqq = cfqq->new_cfqq;
2527         while (__cfqq) {
2528                 if (__cfqq == cfqq) {
2529                         WARN(1, "cfqq->new_cfqq loop detected\n");
2530                         break;
2531                 }
2532                 next = __cfqq->new_cfqq;
2533                 cfq_put_queue(__cfqq);
2534                 __cfqq = next;
2535         }
2536
2537         cfq_put_queue(cfqq);
2538 }
2539
2540 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2541                                          struct cfq_io_context *cic)
2542 {
2543         struct io_context *ioc = cic->ioc;
2544
2545         list_del_init(&cic->queue_list);
2546
2547         /*
2548          * Make sure key == NULL is seen for dead queues
2549          */
2550         smp_wmb();
2551         cic->dead_key = (unsigned long) cic->key;
2552         cic->key = NULL;
2553
2554         if (ioc->ioc_data == cic)
2555                 rcu_assign_pointer(ioc->ioc_data, NULL);
2556
2557         if (cic->cfqq[BLK_RW_ASYNC]) {
2558                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2559                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2560         }
2561
2562         if (cic->cfqq[BLK_RW_SYNC]) {
2563                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2564                 cic->cfqq[BLK_RW_SYNC] = NULL;
2565         }
2566 }
2567
2568 static void cfq_exit_single_io_context(struct io_context *ioc,
2569                                        struct cfq_io_context *cic)
2570 {
2571         struct cfq_data *cfqd = cic->key;
2572
2573         if (cfqd) {
2574                 struct request_queue *q = cfqd->queue;
2575                 unsigned long flags;
2576
2577                 spin_lock_irqsave(q->queue_lock, flags);
2578
2579                 /*
2580                  * Ensure we get a fresh copy of the ->key to prevent
2581                  * race between exiting task and queue
2582                  */
2583                 smp_read_barrier_depends();
2584                 if (cic->key)
2585                         __cfq_exit_single_io_context(cfqd, cic);
2586
2587                 spin_unlock_irqrestore(q->queue_lock, flags);
2588         }
2589 }
2590
2591 /*
2592  * The process that ioc belongs to has exited, we need to clean up
2593  * and put the internal structures we have that belongs to that process.
2594  */
2595 static void cfq_exit_io_context(struct io_context *ioc)
2596 {
2597         call_for_each_cic(ioc, cfq_exit_single_io_context);
2598 }
2599
2600 static struct cfq_io_context *
2601 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2602 {
2603         struct cfq_io_context *cic;
2604
2605         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2606                                                         cfqd->queue->node);
2607         if (cic) {
2608                 cic->last_end_request = jiffies;
2609                 INIT_LIST_HEAD(&cic->queue_list);
2610                 INIT_HLIST_NODE(&cic->cic_list);
2611                 cic->dtor = cfq_free_io_context;
2612                 cic->exit = cfq_exit_io_context;
2613                 elv_ioc_count_inc(cfq_ioc_count);
2614         }
2615
2616         return cic;
2617 }
2618
2619 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2620 {
2621         struct task_struct *tsk = current;
2622         int ioprio_class;
2623
2624         if (!cfq_cfqq_prio_changed(cfqq))
2625                 return;
2626
2627         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2628         switch (ioprio_class) {
2629         default:
2630                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2631         case IOPRIO_CLASS_NONE:
2632                 /*
2633                  * no prio set, inherit CPU scheduling settings
2634                  */
2635                 cfqq->ioprio = task_nice_ioprio(tsk);
2636                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2637                 break;
2638         case IOPRIO_CLASS_RT:
2639                 cfqq->ioprio = task_ioprio(ioc);
2640                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2641                 break;
2642         case IOPRIO_CLASS_BE:
2643                 cfqq->ioprio = task_ioprio(ioc);
2644                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2645                 break;
2646         case IOPRIO_CLASS_IDLE:
2647                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2648                 cfqq->ioprio = 7;
2649                 cfq_clear_cfqq_idle_window(cfqq);
2650                 break;
2651         }
2652
2653         /*
2654          * keep track of original prio settings in case we have to temporarily
2655          * elevate the priority of this queue
2656          */
2657         cfqq->org_ioprio = cfqq->ioprio;
2658         cfqq->org_ioprio_class = cfqq->ioprio_class;
2659         cfq_clear_cfqq_prio_changed(cfqq);
2660 }
2661
2662 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2663 {
2664         struct cfq_data *cfqd = cic->key;
2665         struct cfq_queue *cfqq;
2666         unsigned long flags;
2667
2668         if (unlikely(!cfqd))
2669                 return;
2670
2671         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2672
2673         cfqq = cic->cfqq[BLK_RW_ASYNC];
2674         if (cfqq) {
2675                 struct cfq_queue *new_cfqq;
2676                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2677                                                 GFP_ATOMIC);
2678                 if (new_cfqq) {
2679                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2680                         cfq_put_queue(cfqq);
2681                 }
2682         }
2683
2684         cfqq = cic->cfqq[BLK_RW_SYNC];
2685         if (cfqq)
2686                 cfq_mark_cfqq_prio_changed(cfqq);
2687
2688         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2689 }
2690
2691 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2692 {
2693         call_for_each_cic(ioc, changed_ioprio);
2694         ioc->ioprio_changed = 0;
2695 }
2696
2697 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2698                           pid_t pid, bool is_sync)
2699 {
2700         RB_CLEAR_NODE(&cfqq->rb_node);
2701         RB_CLEAR_NODE(&cfqq->p_node);
2702         INIT_LIST_HEAD(&cfqq->fifo);
2703
2704         atomic_set(&cfqq->ref, 0);
2705         cfqq->cfqd = cfqd;
2706
2707         cfq_mark_cfqq_prio_changed(cfqq);
2708
2709         if (is_sync) {
2710                 if (!cfq_class_idle(cfqq))
2711                         cfq_mark_cfqq_idle_window(cfqq);
2712                 cfq_mark_cfqq_sync(cfqq);
2713         }
2714         cfqq->pid = pid;
2715 }
2716
2717 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2718 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2719 {
2720         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2721         struct cfq_data *cfqd = cic->key;
2722         unsigned long flags;
2723         struct request_queue *q;
2724
2725         if (unlikely(!cfqd))
2726                 return;
2727
2728         q = cfqd->queue;
2729
2730         spin_lock_irqsave(q->queue_lock, flags);
2731
2732         if (sync_cfqq) {
2733                 /*
2734                  * Drop reference to sync queue. A new sync queue will be
2735                  * assigned in new group upon arrival of a fresh request.
2736                  */
2737                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2738                 cic_set_cfqq(cic, NULL, 1);
2739                 cfq_put_queue(sync_cfqq);
2740         }
2741
2742         spin_unlock_irqrestore(q->queue_lock, flags);
2743 }
2744
2745 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2746 {
2747         call_for_each_cic(ioc, changed_cgroup);
2748         ioc->cgroup_changed = 0;
2749 }
2750 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2751
2752 static struct cfq_queue *
2753 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2754                      struct io_context *ioc, gfp_t gfp_mask)
2755 {
2756         struct cfq_queue *cfqq, *new_cfqq = NULL;
2757         struct cfq_io_context *cic;
2758         struct cfq_group *cfqg;
2759
2760 retry:
2761         cfqg = cfq_get_cfqg(cfqd, 1);
2762         cic = cfq_cic_lookup(cfqd, ioc);
2763         /* cic always exists here */
2764         cfqq = cic_to_cfqq(cic, is_sync);
2765
2766         /*
2767          * Always try a new alloc if we fell back to the OOM cfqq
2768          * originally, since it should just be a temporary situation.
2769          */
2770         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2771                 cfqq = NULL;
2772                 if (new_cfqq) {
2773                         cfqq = new_cfqq;
2774                         new_cfqq = NULL;
2775                 } else if (gfp_mask & __GFP_WAIT) {
2776                         spin_unlock_irq(cfqd->queue->queue_lock);
2777                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2778                                         gfp_mask | __GFP_ZERO,
2779                                         cfqd->queue->node);
2780                         spin_lock_irq(cfqd->queue->queue_lock);
2781                         if (new_cfqq)
2782                                 goto retry;
2783                 } else {
2784                         cfqq = kmem_cache_alloc_node(cfq_pool,
2785                                         gfp_mask | __GFP_ZERO,
2786                                         cfqd->queue->node);
2787                 }
2788
2789                 if (cfqq) {
2790                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2791                         cfq_init_prio_data(cfqq, ioc);
2792                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2793                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2794                 } else
2795                         cfqq = &cfqd->oom_cfqq;
2796         }
2797
2798         if (new_cfqq)
2799                 kmem_cache_free(cfq_pool, new_cfqq);
2800
2801         return cfqq;
2802 }
2803
2804 static struct cfq_queue **
2805 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2806 {
2807         switch (ioprio_class) {
2808         case IOPRIO_CLASS_RT:
2809                 return &cfqd->async_cfqq[0][ioprio];
2810         case IOPRIO_CLASS_BE:
2811                 return &cfqd->async_cfqq[1][ioprio];
2812         case IOPRIO_CLASS_IDLE:
2813                 return &cfqd->async_idle_cfqq;
2814         default:
2815                 BUG();
2816         }
2817 }
2818
2819 static struct cfq_queue *
2820 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2821               gfp_t gfp_mask)
2822 {
2823         const int ioprio = task_ioprio(ioc);
2824         const int ioprio_class = task_ioprio_class(ioc);
2825         struct cfq_queue **async_cfqq = NULL;
2826         struct cfq_queue *cfqq = NULL;
2827
2828         if (!is_sync) {
2829                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2830                 cfqq = *async_cfqq;
2831         }
2832
2833         if (!cfqq)
2834                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2835
2836         /*
2837          * pin the queue now that it's allocated, scheduler exit will prune it
2838          */
2839         if (!is_sync && !(*async_cfqq)) {
2840                 atomic_inc(&cfqq->ref);
2841                 *async_cfqq = cfqq;
2842         }
2843
2844         atomic_inc(&cfqq->ref);
2845         return cfqq;
2846 }
2847
2848 /*
2849  * We drop cfq io contexts lazily, so we may find a dead one.
2850  */
2851 static void
2852 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2853                   struct cfq_io_context *cic)
2854 {
2855         unsigned long flags;
2856
2857         WARN_ON(!list_empty(&cic->queue_list));
2858
2859         spin_lock_irqsave(&ioc->lock, flags);
2860
2861         BUG_ON(ioc->ioc_data == cic);
2862
2863         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2864         hlist_del_rcu(&cic->cic_list);
2865         spin_unlock_irqrestore(&ioc->lock, flags);
2866
2867         cfq_cic_free(cic);
2868 }
2869
2870 static struct cfq_io_context *
2871 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2872 {
2873         struct cfq_io_context *cic;
2874         unsigned long flags;
2875         void *k;
2876
2877         if (unlikely(!ioc))
2878                 return NULL;
2879
2880         rcu_read_lock();
2881
2882         /*
2883          * we maintain a last-hit cache, to avoid browsing over the tree
2884          */
2885         cic = rcu_dereference(ioc->ioc_data);
2886         if (cic && cic->key == cfqd) {
2887                 rcu_read_unlock();
2888                 return cic;
2889         }
2890
2891         do {
2892                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2893                 rcu_read_unlock();
2894                 if (!cic)
2895                         break;
2896                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2897                 k = cic->key;
2898                 if (unlikely(!k)) {
2899                         cfq_drop_dead_cic(cfqd, ioc, cic);
2900                         rcu_read_lock();
2901                         continue;
2902                 }
2903
2904                 spin_lock_irqsave(&ioc->lock, flags);
2905                 rcu_assign_pointer(ioc->ioc_data, cic);
2906                 spin_unlock_irqrestore(&ioc->lock, flags);
2907                 break;
2908         } while (1);
2909
2910         return cic;
2911 }
2912
2913 /*
2914  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2915  * the process specific cfq io context when entered from the block layer.
2916  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2917  */
2918 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2919                         struct cfq_io_context *cic, gfp_t gfp_mask)
2920 {
2921         unsigned long flags;
2922         int ret;
2923
2924         ret = radix_tree_preload(gfp_mask);
2925         if (!ret) {
2926                 cic->ioc = ioc;
2927                 cic->key = cfqd;
2928
2929                 spin_lock_irqsave(&ioc->lock, flags);
2930                 ret = radix_tree_insert(&ioc->radix_root,
2931                                                 (unsigned long) cfqd, cic);
2932                 if (!ret)
2933                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2934                 spin_unlock_irqrestore(&ioc->lock, flags);
2935
2936                 radix_tree_preload_end();
2937
2938                 if (!ret) {
2939                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2940                         list_add(&cic->queue_list, &cfqd->cic_list);
2941                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2942                 }
2943         }
2944
2945         if (ret)
2946                 printk(KERN_ERR "cfq: cic link failed!\n");
2947
2948         return ret;
2949 }
2950
2951 /*
2952  * Setup general io context and cfq io context. There can be several cfq
2953  * io contexts per general io context, if this process is doing io to more
2954  * than one device managed by cfq.
2955  */
2956 static struct cfq_io_context *
2957 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2958 {
2959         struct io_context *ioc = NULL;
2960         struct cfq_io_context *cic;
2961
2962         might_sleep_if(gfp_mask & __GFP_WAIT);
2963
2964         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2965         if (!ioc)
2966                 return NULL;
2967
2968         cic = cfq_cic_lookup(cfqd, ioc);
2969         if (cic)
2970                 goto out;
2971
2972         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2973         if (cic == NULL)
2974                 goto err;
2975
2976         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2977                 goto err_free;
2978
2979 out:
2980         smp_read_barrier_depends();
2981         if (unlikely(ioc->ioprio_changed))
2982                 cfq_ioc_set_ioprio(ioc);
2983
2984 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2985         if (unlikely(ioc->cgroup_changed))
2986                 cfq_ioc_set_cgroup(ioc);
2987 #endif
2988         return cic;
2989 err_free:
2990         cfq_cic_free(cic);
2991 err:
2992         put_io_context(ioc);
2993         return NULL;
2994 }
2995
2996 static void
2997 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2998 {
2999         unsigned long elapsed = jiffies - cic->last_end_request;
3000         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3001
3002         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3003         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3004         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3005 }
3006
3007 static void
3008 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3009                        struct request *rq)
3010 {
3011         sector_t sdist = 0;
3012         sector_t n_sec = blk_rq_sectors(rq);
3013         if (cfqq->last_request_pos) {
3014                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3015                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3016                 else
3017                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3018         }
3019
3020         cfqq->seek_history <<= 1;
3021         if (blk_queue_nonrot(cfqd->queue))
3022                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3023         else
3024                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3025 }
3026
3027 /*
3028  * Disable idle window if the process thinks too long or seeks so much that
3029  * it doesn't matter
3030  */
3031 static void
3032 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3033                        struct cfq_io_context *cic)
3034 {
3035         int old_idle, enable_idle;
3036
3037         /*
3038          * Don't idle for async or idle io prio class
3039          */
3040         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3041                 return;
3042
3043         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3044
3045         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3046                 cfq_mark_cfqq_deep(cfqq);
3047
3048         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3049             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3050                 enable_idle = 0;
3051         else if (sample_valid(cic->ttime_samples)) {
3052                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3053                         enable_idle = 0;
3054                 else
3055                         enable_idle = 1;
3056         }
3057
3058         if (old_idle != enable_idle) {
3059                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3060                 if (enable_idle)
3061                         cfq_mark_cfqq_idle_window(cfqq);
3062                 else
3063                         cfq_clear_cfqq_idle_window(cfqq);
3064         }
3065 }
3066
3067 /*
3068  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3069  * no or if we aren't sure, a 1 will cause a preempt.
3070  */
3071 static bool
3072 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3073                    struct request *rq)
3074 {
3075         struct cfq_queue *cfqq;
3076
3077         cfqq = cfqd->active_queue;
3078         if (!cfqq)
3079                 return false;
3080
3081         if (cfq_class_idle(new_cfqq))
3082                 return false;
3083
3084         if (cfq_class_idle(cfqq))
3085                 return true;
3086
3087         /*
3088          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3089          */
3090         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3091                 return false;
3092
3093         /*
3094          * if the new request is sync, but the currently running queue is
3095          * not, let the sync request have priority.
3096          */
3097         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3098                 return true;
3099
3100         if (new_cfqq->cfqg != cfqq->cfqg)
3101                 return false;
3102
3103         if (cfq_slice_used(cfqq))
3104                 return true;
3105
3106         /* Allow preemption only if we are idling on sync-noidle tree */
3107         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3108             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3109             new_cfqq->service_tree->count == 2 &&
3110             RB_EMPTY_ROOT(&cfqq->sort_list))
3111                 return true;
3112
3113         /*
3114          * So both queues are sync. Let the new request get disk time if
3115          * it's a metadata request and the current queue is doing regular IO.
3116          */
3117         if (rq_is_meta(rq) && !cfqq->meta_pending)
3118                 return true;
3119
3120         /*
3121          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3122          */
3123         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3124                 return true;
3125
3126         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3127                 return false;
3128
3129         /*
3130          * if this request is as-good as one we would expect from the
3131          * current cfqq, let it preempt
3132          */
3133         if (cfq_rq_close(cfqd, cfqq, rq))
3134                 return true;
3135
3136         return false;
3137 }
3138
3139 /*
3140  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3141  * let it have half of its nominal slice.
3142  */
3143 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3144 {
3145         cfq_log_cfqq(cfqd, cfqq, "preempt");
3146         cfq_slice_expired(cfqd, 1);
3147
3148         /*
3149          * Put the new queue at the front of the of the current list,
3150          * so we know that it will be selected next.
3151          */
3152         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3153
3154         cfq_service_tree_add(cfqd, cfqq, 1);
3155
3156         cfqq->slice_end = 0;
3157         cfq_mark_cfqq_slice_new(cfqq);
3158 }
3159
3160 /*
3161  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3162  * something we should do about it
3163  */
3164 static void
3165 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3166                 struct request *rq)
3167 {
3168         struct cfq_io_context *cic = RQ_CIC(rq);
3169
3170         cfqd->rq_queued++;
3171         if (rq_is_meta(rq))
3172                 cfqq->meta_pending++;
3173
3174         cfq_update_io_thinktime(cfqd, cic);
3175         cfq_update_io_seektime(cfqd, cfqq, rq);
3176         cfq_update_idle_window(cfqd, cfqq, cic);
3177
3178         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3179
3180         if (cfqq == cfqd->active_queue) {
3181                 /*
3182                  * Remember that we saw a request from this process, but
3183                  * don't start queuing just yet. Otherwise we risk seeing lots
3184                  * of tiny requests, because we disrupt the normal plugging
3185                  * and merging. If the request is already larger than a single
3186                  * page, let it rip immediately. For that case we assume that
3187                  * merging is already done. Ditto for a busy system that
3188                  * has other work pending, don't risk delaying until the
3189                  * idle timer unplug to continue working.
3190                  */
3191                 if (cfq_cfqq_wait_request(cfqq)) {
3192                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3193                             cfqd->busy_queues > 1) {
3194                                 del_timer(&cfqd->idle_slice_timer);
3195                                 cfq_clear_cfqq_wait_request(cfqq);
3196                                 __blk_run_queue(cfqd->queue);
3197                         } else
3198                                 cfq_mark_cfqq_must_dispatch(cfqq);
3199                 }
3200         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3201                 /*
3202                  * not the active queue - expire current slice if it is
3203                  * idle and has expired it's mean thinktime or this new queue
3204                  * has some old slice time left and is of higher priority or
3205                  * this new queue is RT and the current one is BE
3206                  */
3207                 cfq_preempt_queue(cfqd, cfqq);
3208                 __blk_run_queue(cfqd->queue);
3209         }
3210 }
3211
3212 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3213 {
3214         struct cfq_data *cfqd = q->elevator->elevator_data;
3215         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3216
3217         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3218         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3219
3220         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3221         list_add_tail(&rq->queuelist, &cfqq->fifo);
3222         cfq_add_rq_rb(rq);
3223
3224         blkiocg_update_request_add_stats(&cfqq->cfqg->blkg,
3225                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3226                         rq_is_sync(rq));
3227         cfq_rq_enqueued(cfqd, cfqq, rq);
3228 }
3229
3230 /*
3231  * Update hw_tag based on peak queue depth over 50 samples under
3232  * sufficient load.
3233  */
3234 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3235 {
3236         struct cfq_queue *cfqq = cfqd->active_queue;
3237
3238         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3239                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3240
3241         if (cfqd->hw_tag == 1)
3242                 return;
3243
3244         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3245             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3246                 return;
3247
3248         /*
3249          * If active queue hasn't enough requests and can idle, cfq might not
3250          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3251          * case
3252          */
3253         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3254             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3255             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3256                 return;
3257
3258         if (cfqd->hw_tag_samples++ < 50)
3259                 return;
3260
3261         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3262                 cfqd->hw_tag = 1;
3263         else
3264                 cfqd->hw_tag = 0;
3265 }
3266
3267 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3268 {
3269         struct cfq_io_context *cic = cfqd->active_cic;
3270
3271         /* If there are other queues in the group, don't wait */
3272         if (cfqq->cfqg->nr_cfqq > 1)
3273                 return false;
3274
3275         if (cfq_slice_used(cfqq))
3276                 return true;
3277
3278         /* if slice left is less than think time, wait busy */
3279         if (cic && sample_valid(cic->ttime_samples)
3280             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3281                 return true;
3282
3283         /*
3284          * If think times is less than a jiffy than ttime_mean=0 and above
3285          * will not be true. It might happen that slice has not expired yet
3286          * but will expire soon (4-5 ns) during select_queue(). To cover the
3287          * case where think time is less than a jiffy, mark the queue wait
3288          * busy if only 1 jiffy is left in the slice.
3289          */
3290         if (cfqq->slice_end - jiffies == 1)
3291                 return true;
3292
3293         return false;
3294 }
3295
3296 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3297 {
3298         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3299         struct cfq_data *cfqd = cfqq->cfqd;
3300         const int sync = rq_is_sync(rq);
3301         unsigned long now;
3302
3303         now = jiffies;
3304         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3305
3306         cfq_update_hw_tag(cfqd);
3307
3308         WARN_ON(!cfqd->rq_in_driver);
3309         WARN_ON(!cfqq->dispatched);
3310         cfqd->rq_in_driver--;
3311         cfqq->dispatched--;
3312         blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3313                         rq_io_start_time_ns(rq), rq_data_dir(rq),
3314                         rq_is_sync(rq));
3315
3316         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3317
3318         if (sync) {
3319                 RQ_CIC(rq)->last_end_request = now;
3320                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3321                         cfqd->last_delayed_sync = now;
3322         }
3323
3324         /*
3325          * If this is the active queue, check if it needs to be expired,
3326          * or if we want to idle in case it has no pending requests.
3327          */
3328         if (cfqd->active_queue == cfqq) {
3329                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3330
3331                 if (cfq_cfqq_slice_new(cfqq)) {
3332                         cfq_set_prio_slice(cfqd, cfqq);
3333                         cfq_clear_cfqq_slice_new(cfqq);
3334                 }
3335
3336                 /*
3337                  * Should we wait for next request to come in before we expire
3338                  * the queue.
3339                  */
3340                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3341                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3342                         cfq_mark_cfqq_wait_busy(cfqq);
3343                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3344                 }
3345
3346                 /*
3347                  * Idling is not enabled on:
3348                  * - expired queues
3349                  * - idle-priority queues
3350                  * - async queues
3351                  * - queues with still some requests queued
3352                  * - when there is a close cooperator
3353                  */
3354                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3355                         cfq_slice_expired(cfqd, 1);
3356                 else if (sync && cfqq_empty &&
3357                          !cfq_close_cooperator(cfqd, cfqq)) {
3358                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3359                         /*
3360                          * Idling is enabled for SYNC_WORKLOAD.
3361                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3362                          * only if we processed at least one !rq_noidle request
3363                          */
3364                         if (cfqd->serving_type == SYNC_WORKLOAD
3365                             || cfqd->noidle_tree_requires_idle
3366                             || cfqq->cfqg->nr_cfqq == 1)
3367                                 cfq_arm_slice_timer(cfqd);
3368                 }
3369         }
3370
3371         if (!cfqd->rq_in_driver)
3372                 cfq_schedule_dispatch(cfqd);
3373 }
3374
3375 /*
3376  * we temporarily boost lower priority queues if they are holding fs exclusive
3377  * resources. they are boosted to normal prio (CLASS_BE/4)
3378  */
3379 static void cfq_prio_boost(struct cfq_queue *cfqq)
3380 {
3381         if (has_fs_excl()) {
3382                 /*
3383                  * boost idle prio on transactions that would lock out other
3384                  * users of the filesystem
3385                  */
3386                 if (cfq_class_idle(cfqq))
3387                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3388                 if (cfqq->ioprio > IOPRIO_NORM)
3389                         cfqq->ioprio = IOPRIO_NORM;
3390         } else {
3391                 /*
3392                  * unboost the queue (if needed)
3393                  */
3394                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3395                 cfqq->ioprio = cfqq->org_ioprio;
3396         }
3397 }
3398
3399 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3400 {
3401         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3402                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3403                 return ELV_MQUEUE_MUST;
3404         }
3405
3406         return ELV_MQUEUE_MAY;
3407 }
3408
3409 static int cfq_may_queue(struct request_queue *q, int rw)
3410 {
3411         struct cfq_data *cfqd = q->elevator->elevator_data;
3412         struct task_struct *tsk = current;
3413         struct cfq_io_context *cic;
3414         struct cfq_queue *cfqq;
3415
3416         /*
3417          * don't force setup of a queue from here, as a call to may_queue
3418          * does not necessarily imply that a request actually will be queued.
3419          * so just lookup a possibly existing queue, or return 'may queue'
3420          * if that fails
3421          */
3422         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3423         if (!cic)
3424                 return ELV_MQUEUE_MAY;
3425
3426         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3427         if (cfqq) {
3428                 cfq_init_prio_data(cfqq, cic->ioc);
3429                 cfq_prio_boost(cfqq);
3430
3431                 return __cfq_may_queue(cfqq);
3432         }
3433
3434         return ELV_MQUEUE_MAY;
3435 }
3436
3437 /*
3438  * queue lock held here
3439  */
3440 static void cfq_put_request(struct request *rq)
3441 {
3442         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3443
3444         if (cfqq) {
3445                 const int rw = rq_data_dir(rq);
3446
3447                 BUG_ON(!cfqq->allocated[rw]);
3448                 cfqq->allocated[rw]--;
3449
3450                 put_io_context(RQ_CIC(rq)->ioc);
3451
3452                 rq->elevator_private = NULL;
3453                 rq->elevator_private2 = NULL;
3454
3455                 cfq_put_queue(cfqq);
3456         }
3457 }
3458
3459 static struct cfq_queue *
3460 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3461                 struct cfq_queue *cfqq)
3462 {
3463         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3464         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3465         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3466         cfq_put_queue(cfqq);
3467         return cic_to_cfqq(cic, 1);
3468 }
3469
3470 /*
3471  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3472  * was the last process referring to said cfqq.
3473  */
3474 static struct cfq_queue *
3475 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3476 {
3477         if (cfqq_process_refs(cfqq) == 1) {
3478                 cfqq->pid = current->pid;
3479                 cfq_clear_cfqq_coop(cfqq);
3480                 cfq_clear_cfqq_split_coop(cfqq);
3481                 return cfqq;
3482         }
3483
3484         cic_set_cfqq(cic, NULL, 1);
3485         cfq_put_queue(cfqq);
3486         return NULL;
3487 }
3488 /*
3489  * Allocate cfq data structures associated with this request.
3490  */
3491 static int
3492 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3493 {
3494         struct cfq_data *cfqd = q->elevator->elevator_data;
3495         struct cfq_io_context *cic;
3496         const int rw = rq_data_dir(rq);
3497         const bool is_sync = rq_is_sync(rq);
3498         struct cfq_queue *cfqq;
3499         unsigned long flags;
3500
3501         might_sleep_if(gfp_mask & __GFP_WAIT);
3502
3503         cic = cfq_get_io_context(cfqd, gfp_mask);
3504
3505         spin_lock_irqsave(q->queue_lock, flags);
3506
3507         if (!cic)
3508                 goto queue_fail;
3509
3510 new_queue:
3511         cfqq = cic_to_cfqq(cic, is_sync);
3512         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3513                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3514                 cic_set_cfqq(cic, cfqq, is_sync);
3515         } else {
3516                 /*
3517                  * If the queue was seeky for too long, break it apart.
3518                  */
3519                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3520                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3521                         cfqq = split_cfqq(cic, cfqq);
3522                         if (!cfqq)
3523                                 goto new_queue;
3524                 }
3525
3526                 /*
3527                  * Check to see if this queue is scheduled to merge with
3528                  * another, closely cooperating queue.  The merging of
3529                  * queues happens here as it must be done in process context.
3530                  * The reference on new_cfqq was taken in merge_cfqqs.
3531                  */
3532                 if (cfqq->new_cfqq)
3533                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3534         }
3535
3536         cfqq->allocated[rw]++;
3537         atomic_inc(&cfqq->ref);
3538
3539         spin_unlock_irqrestore(q->queue_lock, flags);
3540
3541         rq->elevator_private = cic;
3542         rq->elevator_private2 = cfqq;
3543         return 0;
3544
3545 queue_fail:
3546         if (cic)
3547                 put_io_context(cic->ioc);
3548
3549         cfq_schedule_dispatch(cfqd);
3550         spin_unlock_irqrestore(q->queue_lock, flags);
3551         cfq_log(cfqd, "set_request fail");
3552         return 1;
3553 }
3554
3555 static void cfq_kick_queue(struct work_struct *work)
3556 {
3557         struct cfq_data *cfqd =
3558                 container_of(work, struct cfq_data, unplug_work);
3559         struct request_queue *q = cfqd->queue;
3560
3561         spin_lock_irq(q->queue_lock);
3562         __blk_run_queue(cfqd->queue);
3563         spin_unlock_irq(q->queue_lock);
3564 }
3565
3566 /*
3567  * Timer running if the active_queue is currently idling inside its time slice
3568  */
3569 static void cfq_idle_slice_timer(unsigned long data)
3570 {
3571         struct cfq_data *cfqd = (struct cfq_data *) data;
3572         struct cfq_queue *cfqq;
3573         unsigned long flags;
3574         int timed_out = 1;
3575
3576         cfq_log(cfqd, "idle timer fired");
3577
3578         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3579
3580         cfqq = cfqd->active_queue;
3581         if (cfqq) {
3582                 timed_out = 0;
3583
3584                 /*
3585                  * We saw a request before the queue expired, let it through
3586                  */
3587                 if (cfq_cfqq_must_dispatch(cfqq))
3588                         goto out_kick;
3589
3590                 /*
3591                  * expired
3592                  */
3593                 if (cfq_slice_used(cfqq))
3594                         goto expire;
3595
3596                 /*
3597                  * only expire and reinvoke request handler, if there are
3598                  * other queues with pending requests
3599                  */
3600                 if (!cfqd->busy_queues)
3601                         goto out_cont;
3602
3603                 /*
3604                  * not expired and it has a request pending, let it dispatch
3605                  */
3606                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3607                         goto out_kick;
3608
3609                 /*
3610                  * Queue depth flag is reset only when the idle didn't succeed
3611                  */
3612                 cfq_clear_cfqq_deep(cfqq);
3613         }
3614 expire:
3615         cfq_slice_expired(cfqd, timed_out);
3616 out_kick:
3617         cfq_schedule_dispatch(cfqd);
3618 out_cont:
3619         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3620 }
3621
3622 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3623 {
3624         del_timer_sync(&cfqd->idle_slice_timer);
3625         cancel_work_sync(&cfqd->unplug_work);
3626 }
3627
3628 static void cfq_put_async_queues(struct cfq_data *cfqd)
3629 {
3630         int i;
3631
3632         for (i = 0; i < IOPRIO_BE_NR; i++) {
3633                 if (cfqd->async_cfqq[0][i])
3634                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3635                 if (cfqd->async_cfqq[1][i])
3636                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3637         }
3638
3639         if (cfqd->async_idle_cfqq)
3640                 cfq_put_queue(cfqd->async_idle_cfqq);
3641 }
3642
3643 static void cfq_cfqd_free(struct rcu_head *head)
3644 {
3645         kfree(container_of(head, struct cfq_data, rcu));
3646 }
3647
3648 static void cfq_exit_queue(struct elevator_queue *e)
3649 {
3650         struct cfq_data *cfqd = e->elevator_data;
3651         struct request_queue *q = cfqd->queue;
3652
3653         cfq_shutdown_timer_wq(cfqd);
3654
3655         spin_lock_irq(q->queue_lock);
3656
3657         if (cfqd->active_queue)
3658                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3659
3660         while (!list_empty(&cfqd->cic_list)) {
3661                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3662                                                         struct cfq_io_context,
3663                                                         queue_list);
3664
3665                 __cfq_exit_single_io_context(cfqd, cic);
3666         }
3667
3668         cfq_put_async_queues(cfqd);
3669         cfq_release_cfq_groups(cfqd);
3670         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3671
3672         spin_unlock_irq(q->queue_lock);
3673
3674         cfq_shutdown_timer_wq(cfqd);
3675
3676         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3677         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3678 }
3679
3680 static void *cfq_init_queue(struct request_queue *q)
3681 {
3682         struct cfq_data *cfqd;
3683         int i, j;
3684         struct cfq_group *cfqg;
3685         struct cfq_rb_root *st;
3686
3687         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3688         if (!cfqd)
3689                 return NULL;
3690
3691         /* Init root service tree */
3692         cfqd->grp_service_tree = CFQ_RB_ROOT;
3693
3694         /* Init root group */
3695         cfqg = &cfqd->root_group;
3696         for_each_cfqg_st(cfqg, i, j, st)
3697                 *st = CFQ_RB_ROOT;
3698         RB_CLEAR_NODE(&cfqg->rb_node);
3699
3700         /* Give preference to root group over other groups */
3701         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3702
3703 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3704         /*
3705          * Take a reference to root group which we never drop. This is just
3706          * to make sure that cfq_put_cfqg() does not try to kfree root group
3707          */
3708         atomic_set(&cfqg->ref, 1);
3709         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3710                                         0);
3711 #endif
3712         /*
3713          * Not strictly needed (since RB_ROOT just clears the node and we
3714          * zeroed cfqd on alloc), but better be safe in case someone decides
3715          * to add magic to the rb code
3716          */
3717         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3718                 cfqd->prio_trees[i] = RB_ROOT;
3719
3720         /*
3721          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3722          * Grab a permanent reference to it, so that the normal code flow
3723          * will not attempt to free it.
3724          */
3725         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3726         atomic_inc(&cfqd->oom_cfqq.ref);
3727         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3728
3729         INIT_LIST_HEAD(&cfqd->cic_list);
3730
3731         cfqd->queue = q;
3732
3733         init_timer(&cfqd->idle_slice_timer);
3734         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3735         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3736
3737         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3738
3739         cfqd->cfq_quantum = cfq_quantum;
3740         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3741         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3742         cfqd->cfq_back_max = cfq_back_max;
3743         cfqd->cfq_back_penalty = cfq_back_penalty;
3744         cfqd->cfq_slice[0] = cfq_slice_async;
3745         cfqd->cfq_slice[1] = cfq_slice_sync;
3746         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3747         cfqd->cfq_slice_idle = cfq_slice_idle;
3748         cfqd->cfq_latency = 1;
3749         cfqd->cfq_group_isolation = 0;
3750         cfqd->hw_tag = -1;
3751         /*
3752          * we optimistically start assuming sync ops weren't delayed in last
3753          * second, in order to have larger depth for async operations.
3754          */
3755         cfqd->last_delayed_sync = jiffies - HZ;
3756         INIT_RCU_HEAD(&cfqd->rcu);
3757         return cfqd;
3758 }
3759
3760 static void cfq_slab_kill(void)
3761 {
3762         /*
3763          * Caller already ensured that pending RCU callbacks are completed,
3764          * so we should have no busy allocations at this point.
3765          */
3766         if (cfq_pool)
3767                 kmem_cache_destroy(cfq_pool);
3768         if (cfq_ioc_pool)
3769                 kmem_cache_destroy(cfq_ioc_pool);
3770 }
3771
3772 static int __init cfq_slab_setup(void)
3773 {
3774         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3775         if (!cfq_pool)
3776                 goto fail;
3777
3778         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3779         if (!cfq_ioc_pool)
3780                 goto fail;
3781
3782         return 0;
3783 fail:
3784         cfq_slab_kill();
3785         return -ENOMEM;
3786 }
3787
3788 /*
3789  * sysfs parts below -->
3790  */
3791 static ssize_t
3792 cfq_var_show(unsigned int var, char *page)
3793 {
3794         return sprintf(page, "%d\n", var);
3795 }
3796
3797 static ssize_t
3798 cfq_var_store(unsigned int *var, const char *page, size_t count)
3799 {
3800         char *p = (char *) page;
3801
3802         *var = simple_strtoul(p, &p, 10);
3803         return count;
3804 }
3805
3806 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3807 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3808 {                                                                       \
3809         struct cfq_data *cfqd = e->elevator_data;                       \
3810         unsigned int __data = __VAR;                                    \
3811         if (__CONV)                                                     \
3812                 __data = jiffies_to_msecs(__data);                      \
3813         return cfq_var_show(__data, (page));                            \
3814 }
3815 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3816 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3817 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3818 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3819 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3820 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3821 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3822 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3823 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3824 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3825 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3826 #undef SHOW_FUNCTION
3827
3828 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3829 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3830 {                                                                       \
3831         struct cfq_data *cfqd = e->elevator_data;                       \
3832         unsigned int __data;                                            \
3833         int ret = cfq_var_store(&__data, (page), count);                \
3834         if (__data < (MIN))                                             \
3835                 __data = (MIN);                                         \
3836         else if (__data > (MAX))                                        \
3837                 __data = (MAX);                                         \
3838         if (__CONV)                                                     \
3839                 *(__PTR) = msecs_to_jiffies(__data);                    \
3840         else                                                            \
3841                 *(__PTR) = __data;                                      \
3842         return ret;                                                     \
3843 }
3844 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3845 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3846                 UINT_MAX, 1);
3847 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3848                 UINT_MAX, 1);
3849 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3850 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3851                 UINT_MAX, 0);
3852 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3853 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3854 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3855 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3856                 UINT_MAX, 0);
3857 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3858 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3859 #undef STORE_FUNCTION
3860
3861 #define CFQ_ATTR(name) \
3862         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3863
3864 static struct elv_fs_entry cfq_attrs[] = {
3865         CFQ_ATTR(quantum),
3866         CFQ_ATTR(fifo_expire_sync),
3867         CFQ_ATTR(fifo_expire_async),
3868         CFQ_ATTR(back_seek_max),
3869         CFQ_ATTR(back_seek_penalty),
3870         CFQ_ATTR(slice_sync),
3871         CFQ_ATTR(slice_async),
3872         CFQ_ATTR(slice_async_rq),
3873         CFQ_ATTR(slice_idle),
3874         CFQ_ATTR(low_latency),
3875         CFQ_ATTR(group_isolation),
3876         __ATTR_NULL
3877 };
3878
3879 static struct elevator_type iosched_cfq = {
3880         .ops = {
3881                 .elevator_merge_fn =            cfq_merge,
3882                 .elevator_merged_fn =           cfq_merged_request,
3883                 .elevator_merge_req_fn =        cfq_merged_requests,
3884                 .elevator_allow_merge_fn =      cfq_allow_merge,
3885                 .elevator_bio_merged_fn =       cfq_bio_merged,
3886                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3887                 .elevator_add_req_fn =          cfq_insert_request,
3888                 .elevator_activate_req_fn =     cfq_activate_request,
3889                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3890                 .elevator_queue_empty_fn =      cfq_queue_empty,
3891                 .elevator_completed_req_fn =    cfq_completed_request,
3892                 .elevator_former_req_fn =       elv_rb_former_request,
3893                 .elevator_latter_req_fn =       elv_rb_latter_request,
3894                 .elevator_set_req_fn =          cfq_set_request,
3895                 .elevator_put_req_fn =          cfq_put_request,
3896                 .elevator_may_queue_fn =        cfq_may_queue,
3897                 .elevator_init_fn =             cfq_init_queue,
3898                 .elevator_exit_fn =             cfq_exit_queue,
3899                 .trim =                         cfq_free_io_context,
3900         },
3901         .elevator_attrs =       cfq_attrs,
3902         .elevator_name =        "cfq",
3903         .elevator_owner =       THIS_MODULE,
3904 };
3905
3906 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3907 static struct blkio_policy_type blkio_policy_cfq = {
3908         .ops = {
3909                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3910                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3911         },
3912 };
3913 #else
3914 static struct blkio_policy_type blkio_policy_cfq;
3915 #endif
3916
3917 static int __init cfq_init(void)
3918 {
3919         /*
3920          * could be 0 on HZ < 1000 setups
3921          */
3922         if (!cfq_slice_async)
3923                 cfq_slice_async = 1;
3924         if (!cfq_slice_idle)
3925                 cfq_slice_idle = 1;
3926
3927         if (cfq_slab_setup())
3928                 return -ENOMEM;
3929
3930         elv_register(&iosched_cfq);
3931         blkio_policy_register(&blkio_policy_cfq);
3932
3933         return 0;
3934 }
3935
3936 static void __exit cfq_exit(void)
3937 {
3938         DECLARE_COMPLETION_ONSTACK(all_gone);
3939         blkio_policy_unregister(&blkio_policy_cfq);
3940         elv_unregister(&iosched_cfq);
3941         ioc_gone = &all_gone;
3942         /* ioc_gone's update must be visible before reading ioc_count */
3943         smp_wmb();
3944
3945         /*
3946          * this also protects us from entering cfq_slab_kill() with
3947          * pending RCU callbacks
3948          */
3949         if (elv_ioc_count_read(cfq_ioc_count))
3950                 wait_for_completion(&all_gone);
3951         cfq_slab_kill();
3952 }
3953
3954 module_init(cfq_init);
3955 module_exit(cfq_exit);
3956
3957 MODULE_AUTHOR("Jens Axboe");
3958 MODULE_LICENSE("GPL");
3959 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");