7104ac816fb61966deb68aaa152faa54da5a04fc
[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         /* Sectors dispatched in current dispatch round */
146         unsigned long nr_sectors;
147 };
148
149 /*
150  * First index in the service_trees.
151  * IDLE is handled separately, so it has negative index
152  */
153 enum wl_prio_t {
154         BE_WORKLOAD = 0,
155         RT_WORKLOAD = 1,
156         IDLE_WORKLOAD = 2,
157 };
158
159 /*
160  * Second index in the service_trees.
161  */
162 enum wl_type_t {
163         ASYNC_WORKLOAD = 0,
164         SYNC_NOIDLE_WORKLOAD = 1,
165         SYNC_WORKLOAD = 2
166 };
167
168 /* This is per cgroup per device grouping structure */
169 struct cfq_group {
170         /* group service_tree member */
171         struct rb_node rb_node;
172
173         /* group service_tree key */
174         u64 vdisktime;
175         unsigned int weight;
176         bool on_st;
177
178         /* number of cfqq currently on this group */
179         int nr_cfqq;
180
181         /* Per group busy queus average. Useful for workload slice calc. */
182         unsigned int busy_queues_avg[2];
183         /*
184          * rr lists of queues with requests, onle rr for each priority class.
185          * Counts are embedded in the cfq_rb_root
186          */
187         struct cfq_rb_root service_trees[2][3];
188         struct cfq_rb_root service_tree_idle;
189
190         unsigned long saved_workload_slice;
191         enum wl_type_t saved_workload;
192         enum wl_prio_t saved_serving_prio;
193         struct blkio_group blkg;
194 #ifdef CONFIG_CFQ_GROUP_IOSCHED
195         struct hlist_node cfqd_node;
196         atomic_t ref;
197 #endif
198 };
199
200 /*
201  * Per block device queue structure
202  */
203 struct cfq_data {
204         struct request_queue *queue;
205         /* Root service tree for cfq_groups */
206         struct cfq_rb_root grp_service_tree;
207         struct cfq_group root_group;
208
209         /*
210          * The priority currently being served
211          */
212         enum wl_prio_t serving_prio;
213         enum wl_type_t serving_type;
214         unsigned long workload_expires;
215         struct cfq_group *serving_group;
216         bool noidle_tree_requires_idle;
217
218         /*
219          * Each priority tree is sorted by next_request position.  These
220          * trees are used when determining if two or more queues are
221          * interleaving requests (see cfq_close_cooperator).
222          */
223         struct rb_root prio_trees[CFQ_PRIO_LISTS];
224
225         unsigned int busy_queues;
226
227         int rq_in_driver;
228         int rq_in_flight[2];
229
230         /*
231          * queue-depth detection
232          */
233         int rq_queued;
234         int hw_tag;
235         /*
236          * hw_tag can be
237          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
238          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
239          *  0 => no NCQ
240          */
241         int hw_tag_est_depth;
242         unsigned int hw_tag_samples;
243
244         /*
245          * idle window management
246          */
247         struct timer_list idle_slice_timer;
248         struct work_struct unplug_work;
249
250         struct cfq_queue *active_queue;
251         struct cfq_io_context *active_cic;
252
253         /*
254          * async queue for each priority case
255          */
256         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
257         struct cfq_queue *async_idle_cfqq;
258
259         sector_t last_position;
260
261         /*
262          * tunables, see top of file
263          */
264         unsigned int cfq_quantum;
265         unsigned int cfq_fifo_expire[2];
266         unsigned int cfq_back_penalty;
267         unsigned int cfq_back_max;
268         unsigned int cfq_slice[2];
269         unsigned int cfq_slice_async_rq;
270         unsigned int cfq_slice_idle;
271         unsigned int cfq_latency;
272         unsigned int cfq_group_isolation;
273
274         struct list_head cic_list;
275
276         /*
277          * Fallback dummy cfqq for extreme OOM conditions
278          */
279         struct cfq_queue oom_cfqq;
280
281         unsigned long last_delayed_sync;
282
283         /* List of cfq groups being managed on this device*/
284         struct hlist_head cfqg_list;
285         struct rcu_head rcu;
286 };
287
288 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
289
290 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
291                                             enum wl_prio_t prio,
292                                             enum wl_type_t type)
293 {
294         if (!cfqg)
295                 return NULL;
296
297         if (prio == IDLE_WORKLOAD)
298                 return &cfqg->service_tree_idle;
299
300         return &cfqg->service_trees[prio][type];
301 }
302
303 enum cfqq_state_flags {
304         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
305         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
306         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
307         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
308         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
309         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
310         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
311         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
312         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
313         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
314         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
315         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
316         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
317 };
318
319 #define CFQ_CFQQ_FNS(name)                                              \
320 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
321 {                                                                       \
322         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
323 }                                                                       \
324 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
325 {                                                                       \
326         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
327 }                                                                       \
328 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
329 {                                                                       \
330         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
331 }
332
333 CFQ_CFQQ_FNS(on_rr);
334 CFQ_CFQQ_FNS(wait_request);
335 CFQ_CFQQ_FNS(must_dispatch);
336 CFQ_CFQQ_FNS(must_alloc_slice);
337 CFQ_CFQQ_FNS(fifo_expire);
338 CFQ_CFQQ_FNS(idle_window);
339 CFQ_CFQQ_FNS(prio_changed);
340 CFQ_CFQQ_FNS(slice_new);
341 CFQ_CFQQ_FNS(sync);
342 CFQ_CFQQ_FNS(coop);
343 CFQ_CFQQ_FNS(split_coop);
344 CFQ_CFQQ_FNS(deep);
345 CFQ_CFQQ_FNS(wait_busy);
346 #undef CFQ_CFQQ_FNS
347
348 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
349 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
350         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
351                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
352                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
353
354 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
355         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
356                                 blkg_path(&(cfqg)->blkg), ##args);      \
357
358 #else
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
360         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
361 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
362 #endif
363 #define cfq_log(cfqd, fmt, args...)     \
364         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
365
366 /* Traverses through cfq group service trees */
367 #define for_each_cfqg_st(cfqg, i, j, st) \
368         for (i = 0; i <= IDLE_WORKLOAD; i++) \
369                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
370                         : &cfqg->service_tree_idle; \
371                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
372                         (i == IDLE_WORKLOAD && j == 0); \
373                         j++, st = i < IDLE_WORKLOAD ? \
374                         &cfqg->service_trees[i][j]: NULL) \
375
376
377 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
378 {
379         if (cfq_class_idle(cfqq))
380                 return IDLE_WORKLOAD;
381         if (cfq_class_rt(cfqq))
382                 return RT_WORKLOAD;
383         return BE_WORKLOAD;
384 }
385
386
387 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
388 {
389         if (!cfq_cfqq_sync(cfqq))
390                 return ASYNC_WORKLOAD;
391         if (!cfq_cfqq_idle_window(cfqq))
392                 return SYNC_NOIDLE_WORKLOAD;
393         return SYNC_WORKLOAD;
394 }
395
396 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
397                                         struct cfq_data *cfqd,
398                                         struct cfq_group *cfqg)
399 {
400         if (wl == IDLE_WORKLOAD)
401                 return cfqg->service_tree_idle.count;
402
403         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
404                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
405                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
406 }
407
408 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
409                                         struct cfq_group *cfqg)
410 {
411         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
412                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
413 }
414
415 static void cfq_dispatch_insert(struct request_queue *, struct request *);
416 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
417                                        struct io_context *, gfp_t);
418 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
419                                                 struct io_context *);
420
421 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
422                                             bool is_sync)
423 {
424         return cic->cfqq[is_sync];
425 }
426
427 static inline void cic_set_cfqq(struct cfq_io_context *cic,
428                                 struct cfq_queue *cfqq, bool is_sync)
429 {
430         cic->cfqq[is_sync] = cfqq;
431 }
432
433 /*
434  * We regard a request as SYNC, if it's either a read or has the SYNC bit
435  * set (in which case it could also be direct WRITE).
436  */
437 static inline bool cfq_bio_sync(struct bio *bio)
438 {
439         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
440 }
441
442 /*
443  * scheduler run of queue, if there are requests pending and no one in the
444  * driver that will restart queueing
445  */
446 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
447 {
448         if (cfqd->busy_queues) {
449                 cfq_log(cfqd, "schedule dispatch");
450                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
451         }
452 }
453
454 static int cfq_queue_empty(struct request_queue *q)
455 {
456         struct cfq_data *cfqd = q->elevator->elevator_data;
457
458         return !cfqd->rq_queued;
459 }
460
461 /*
462  * Scale schedule slice based on io priority. Use the sync time slice only
463  * if a queue is marked sync and has sync io queued. A sync queue with async
464  * io only, should not get full sync slice length.
465  */
466 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
467                                  unsigned short prio)
468 {
469         const int base_slice = cfqd->cfq_slice[sync];
470
471         WARN_ON(prio >= IOPRIO_BE_NR);
472
473         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
474 }
475
476 static inline int
477 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
478 {
479         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
480 }
481
482 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
483 {
484         u64 d = delta << CFQ_SERVICE_SHIFT;
485
486         d = d * BLKIO_WEIGHT_DEFAULT;
487         do_div(d, cfqg->weight);
488         return d;
489 }
490
491 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
492 {
493         s64 delta = (s64)(vdisktime - min_vdisktime);
494         if (delta > 0)
495                 min_vdisktime = vdisktime;
496
497         return min_vdisktime;
498 }
499
500 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
501 {
502         s64 delta = (s64)(vdisktime - min_vdisktime);
503         if (delta < 0)
504                 min_vdisktime = vdisktime;
505
506         return min_vdisktime;
507 }
508
509 static void update_min_vdisktime(struct cfq_rb_root *st)
510 {
511         u64 vdisktime = st->min_vdisktime;
512         struct cfq_group *cfqg;
513
514         if (st->active) {
515                 cfqg = rb_entry_cfqg(st->active);
516                 vdisktime = cfqg->vdisktime;
517         }
518
519         if (st->left) {
520                 cfqg = rb_entry_cfqg(st->left);
521                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
522         }
523
524         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
525 }
526
527 /*
528  * get averaged number of queues of RT/BE priority.
529  * average is updated, with a formula that gives more weight to higher numbers,
530  * to quickly follows sudden increases and decrease slowly
531  */
532
533 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
534                                         struct cfq_group *cfqg, bool rt)
535 {
536         unsigned min_q, max_q;
537         unsigned mult  = cfq_hist_divisor - 1;
538         unsigned round = cfq_hist_divisor / 2;
539         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
540
541         min_q = min(cfqg->busy_queues_avg[rt], busy);
542         max_q = max(cfqg->busy_queues_avg[rt], busy);
543         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
544                 cfq_hist_divisor;
545         return cfqg->busy_queues_avg[rt];
546 }
547
548 static inline unsigned
549 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
550 {
551         struct cfq_rb_root *st = &cfqd->grp_service_tree;
552
553         return cfq_target_latency * cfqg->weight / st->total_weight;
554 }
555
556 static inline void
557 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
558 {
559         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
560         if (cfqd->cfq_latency) {
561                 /*
562                  * interested queues (we consider only the ones with the same
563                  * priority class in the cfq group)
564                  */
565                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
566                                                 cfq_class_rt(cfqq));
567                 unsigned sync_slice = cfqd->cfq_slice[1];
568                 unsigned expect_latency = sync_slice * iq;
569                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
570
571                 if (expect_latency > group_slice) {
572                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
573                         /* scale low_slice according to IO priority
574                          * and sync vs async */
575                         unsigned low_slice =
576                                 min(slice, base_low_slice * slice / sync_slice);
577                         /* the adapted slice value is scaled to fit all iqs
578                          * into the target latency */
579                         slice = max(slice * group_slice / expect_latency,
580                                     low_slice);
581                 }
582         }
583         cfqq->slice_start = jiffies;
584         cfqq->slice_end = jiffies + slice;
585         cfqq->allocated_slice = slice;
586         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
587 }
588
589 /*
590  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
591  * isn't valid until the first request from the dispatch is activated
592  * and the slice time set.
593  */
594 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
595 {
596         if (cfq_cfqq_slice_new(cfqq))
597                 return 0;
598         if (time_before(jiffies, cfqq->slice_end))
599                 return 0;
600
601         return 1;
602 }
603
604 /*
605  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
606  * We choose the request that is closest to the head right now. Distance
607  * behind the head is penalized and only allowed to a certain extent.
608  */
609 static struct request *
610 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
611 {
612         sector_t s1, s2, d1 = 0, d2 = 0;
613         unsigned long back_max;
614 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
615 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
616         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
617
618         if (rq1 == NULL || rq1 == rq2)
619                 return rq2;
620         if (rq2 == NULL)
621                 return rq1;
622
623         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
624                 return rq1;
625         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
626                 return rq2;
627         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
628                 return rq1;
629         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
630                 return rq2;
631
632         s1 = blk_rq_pos(rq1);
633         s2 = blk_rq_pos(rq2);
634
635         /*
636          * by definition, 1KiB is 2 sectors
637          */
638         back_max = cfqd->cfq_back_max * 2;
639
640         /*
641          * Strict one way elevator _except_ in the case where we allow
642          * short backward seeks which are biased as twice the cost of a
643          * similar forward seek.
644          */
645         if (s1 >= last)
646                 d1 = s1 - last;
647         else if (s1 + back_max >= last)
648                 d1 = (last - s1) * cfqd->cfq_back_penalty;
649         else
650                 wrap |= CFQ_RQ1_WRAP;
651
652         if (s2 >= last)
653                 d2 = s2 - last;
654         else if (s2 + back_max >= last)
655                 d2 = (last - s2) * cfqd->cfq_back_penalty;
656         else
657                 wrap |= CFQ_RQ2_WRAP;
658
659         /* Found required data */
660
661         /*
662          * By doing switch() on the bit mask "wrap" we avoid having to
663          * check two variables for all permutations: --> faster!
664          */
665         switch (wrap) {
666         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
667                 if (d1 < d2)
668                         return rq1;
669                 else if (d2 < d1)
670                         return rq2;
671                 else {
672                         if (s1 >= s2)
673                                 return rq1;
674                         else
675                                 return rq2;
676                 }
677
678         case CFQ_RQ2_WRAP:
679                 return rq1;
680         case CFQ_RQ1_WRAP:
681                 return rq2;
682         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
683         default:
684                 /*
685                  * Since both rqs are wrapped,
686                  * start with the one that's further behind head
687                  * (--> only *one* back seek required),
688                  * since back seek takes more time than forward.
689                  */
690                 if (s1 <= s2)
691                         return rq1;
692                 else
693                         return rq2;
694         }
695 }
696
697 /*
698  * The below is leftmost cache rbtree addon
699  */
700 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
701 {
702         /* Service tree is empty */
703         if (!root->count)
704                 return NULL;
705
706         if (!root->left)
707                 root->left = rb_first(&root->rb);
708
709         if (root->left)
710                 return rb_entry(root->left, struct cfq_queue, rb_node);
711
712         return NULL;
713 }
714
715 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
716 {
717         if (!root->left)
718                 root->left = rb_first(&root->rb);
719
720         if (root->left)
721                 return rb_entry_cfqg(root->left);
722
723         return NULL;
724 }
725
726 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
727 {
728         rb_erase(n, root);
729         RB_CLEAR_NODE(n);
730 }
731
732 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
733 {
734         if (root->left == n)
735                 root->left = NULL;
736         rb_erase_init(n, &root->rb);
737         --root->count;
738 }
739
740 /*
741  * would be nice to take fifo expire time into account as well
742  */
743 static struct request *
744 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
745                   struct request *last)
746 {
747         struct rb_node *rbnext = rb_next(&last->rb_node);
748         struct rb_node *rbprev = rb_prev(&last->rb_node);
749         struct request *next = NULL, *prev = NULL;
750
751         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
752
753         if (rbprev)
754                 prev = rb_entry_rq(rbprev);
755
756         if (rbnext)
757                 next = rb_entry_rq(rbnext);
758         else {
759                 rbnext = rb_first(&cfqq->sort_list);
760                 if (rbnext && rbnext != &last->rb_node)
761                         next = rb_entry_rq(rbnext);
762         }
763
764         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
765 }
766
767 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
768                                       struct cfq_queue *cfqq)
769 {
770         /*
771          * just an approximation, should be ok.
772          */
773         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
774                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
775 }
776
777 static inline s64
778 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
779 {
780         return cfqg->vdisktime - st->min_vdisktime;
781 }
782
783 static void
784 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
785 {
786         struct rb_node **node = &st->rb.rb_node;
787         struct rb_node *parent = NULL;
788         struct cfq_group *__cfqg;
789         s64 key = cfqg_key(st, cfqg);
790         int left = 1;
791
792         while (*node != NULL) {
793                 parent = *node;
794                 __cfqg = rb_entry_cfqg(parent);
795
796                 if (key < cfqg_key(st, __cfqg))
797                         node = &parent->rb_left;
798                 else {
799                         node = &parent->rb_right;
800                         left = 0;
801                 }
802         }
803
804         if (left)
805                 st->left = &cfqg->rb_node;
806
807         rb_link_node(&cfqg->rb_node, parent, node);
808         rb_insert_color(&cfqg->rb_node, &st->rb);
809 }
810
811 static void
812 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
813 {
814         struct cfq_rb_root *st = &cfqd->grp_service_tree;
815         struct cfq_group *__cfqg;
816         struct rb_node *n;
817
818         cfqg->nr_cfqq++;
819         if (cfqg->on_st)
820                 return;
821
822         /*
823          * Currently put the group at the end. Later implement something
824          * so that groups get lesser vtime based on their weights, so that
825          * if group does not loose all if it was not continously backlogged.
826          */
827         n = rb_last(&st->rb);
828         if (n) {
829                 __cfqg = rb_entry_cfqg(n);
830                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
831         } else
832                 cfqg->vdisktime = st->min_vdisktime;
833
834         __cfq_group_service_tree_add(st, cfqg);
835         cfqg->on_st = true;
836         st->total_weight += cfqg->weight;
837 }
838
839 static void
840 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
841 {
842         struct cfq_rb_root *st = &cfqd->grp_service_tree;
843
844         if (st->active == &cfqg->rb_node)
845                 st->active = NULL;
846
847         BUG_ON(cfqg->nr_cfqq < 1);
848         cfqg->nr_cfqq--;
849
850         /* If there are other cfq queues under this group, don't delete it */
851         if (cfqg->nr_cfqq)
852                 return;
853
854         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
855         cfqg->on_st = false;
856         st->total_weight -= cfqg->weight;
857         if (!RB_EMPTY_NODE(&cfqg->rb_node))
858                 cfq_rb_erase(&cfqg->rb_node, st);
859         cfqg->saved_workload_slice = 0;
860         blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
861 }
862
863 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
864 {
865         unsigned int slice_used;
866
867         /*
868          * Queue got expired before even a single request completed or
869          * got expired immediately after first request completion.
870          */
871         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
872                 /*
873                  * Also charge the seek time incurred to the group, otherwise
874                  * if there are mutiple queues in the group, each can dispatch
875                  * a single request on seeky media and cause lots of seek time
876                  * and group will never know it.
877                  */
878                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
879                                         1);
880         } else {
881                 slice_used = jiffies - cfqq->slice_start;
882                 if (slice_used > cfqq->allocated_slice)
883                         slice_used = cfqq->allocated_slice;
884         }
885
886         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
887                                 cfqq->nr_sectors);
888         return slice_used;
889 }
890
891 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
892                                 struct cfq_queue *cfqq)
893 {
894         struct cfq_rb_root *st = &cfqd->grp_service_tree;
895         unsigned int used_sl, charge_sl;
896         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
897                         - cfqg->service_tree_idle.count;
898
899         BUG_ON(nr_sync < 0);
900         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
901
902         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
903                 charge_sl = cfqq->allocated_slice;
904
905         /* Can't update vdisktime while group is on service tree */
906         cfq_rb_erase(&cfqg->rb_node, st);
907         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
908         __cfq_group_service_tree_add(st, cfqg);
909
910         /* This group is being expired. Save the context */
911         if (time_after(cfqd->workload_expires, jiffies)) {
912                 cfqg->saved_workload_slice = cfqd->workload_expires
913                                                 - jiffies;
914                 cfqg->saved_workload = cfqd->serving_type;
915                 cfqg->saved_serving_prio = cfqd->serving_prio;
916         } else
917                 cfqg->saved_workload_slice = 0;
918
919         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
920                                         st->min_vdisktime);
921         blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
922                                                 cfqq->nr_sectors);
923 }
924
925 #ifdef CONFIG_CFQ_GROUP_IOSCHED
926 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
927 {
928         if (blkg)
929                 return container_of(blkg, struct cfq_group, blkg);
930         return NULL;
931 }
932
933 void
934 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
935 {
936         cfqg_of_blkg(blkg)->weight = weight;
937 }
938
939 static struct cfq_group *
940 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
941 {
942         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
943         struct cfq_group *cfqg = NULL;
944         void *key = cfqd;
945         int i, j;
946         struct cfq_rb_root *st;
947         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
948         unsigned int major, minor;
949
950         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
951         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
952                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
953                 cfqg->blkg.dev = MKDEV(major, minor);
954                 goto done;
955         }
956         if (cfqg || !create)
957                 goto done;
958
959         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
960         if (!cfqg)
961                 goto done;
962
963         cfqg->weight = blkcg->weight;
964         for_each_cfqg_st(cfqg, i, j, st)
965                 *st = CFQ_RB_ROOT;
966         RB_CLEAR_NODE(&cfqg->rb_node);
967
968         /*
969          * Take the initial reference that will be released on destroy
970          * This can be thought of a joint reference by cgroup and
971          * elevator which will be dropped by either elevator exit
972          * or cgroup deletion path depending on who is exiting first.
973          */
974         atomic_set(&cfqg->ref, 1);
975
976         /* Add group onto cgroup list */
977         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
978         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
979                                         MKDEV(major, minor));
980
981         /* Add group on cfqd list */
982         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
983
984 done:
985         return cfqg;
986 }
987
988 /*
989  * Search for the cfq group current task belongs to. If create = 1, then also
990  * create the cfq group if it does not exist. request_queue lock must be held.
991  */
992 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
993 {
994         struct cgroup *cgroup;
995         struct cfq_group *cfqg = NULL;
996
997         rcu_read_lock();
998         cgroup = task_cgroup(current, blkio_subsys_id);
999         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1000         if (!cfqg && create)
1001                 cfqg = &cfqd->root_group;
1002         rcu_read_unlock();
1003         return cfqg;
1004 }
1005
1006 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1007 {
1008         /* Currently, all async queues are mapped to root group */
1009         if (!cfq_cfqq_sync(cfqq))
1010                 cfqg = &cfqq->cfqd->root_group;
1011
1012         cfqq->cfqg = cfqg;
1013         /* cfqq reference on cfqg */
1014         atomic_inc(&cfqq->cfqg->ref);
1015 }
1016
1017 static void cfq_put_cfqg(struct cfq_group *cfqg)
1018 {
1019         struct cfq_rb_root *st;
1020         int i, j;
1021
1022         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1023         if (!atomic_dec_and_test(&cfqg->ref))
1024                 return;
1025         for_each_cfqg_st(cfqg, i, j, st)
1026                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1027         kfree(cfqg);
1028 }
1029
1030 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1031 {
1032         /* Something wrong if we are trying to remove same group twice */
1033         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1034
1035         hlist_del_init(&cfqg->cfqd_node);
1036
1037         /*
1038          * Put the reference taken at the time of creation so that when all
1039          * queues are gone, group can be destroyed.
1040          */
1041         cfq_put_cfqg(cfqg);
1042 }
1043
1044 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1045 {
1046         struct hlist_node *pos, *n;
1047         struct cfq_group *cfqg;
1048
1049         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1050                 /*
1051                  * If cgroup removal path got to blk_group first and removed
1052                  * it from cgroup list, then it will take care of destroying
1053                  * cfqg also.
1054                  */
1055                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1056                         cfq_destroy_cfqg(cfqd, cfqg);
1057         }
1058 }
1059
1060 /*
1061  * Blk cgroup controller notification saying that blkio_group object is being
1062  * delinked as associated cgroup object is going away. That also means that
1063  * no new IO will come in this group. So get rid of this group as soon as
1064  * any pending IO in the group is finished.
1065  *
1066  * This function is called under rcu_read_lock(). key is the rcu protected
1067  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1068  * read lock.
1069  *
1070  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1071  * it should not be NULL as even if elevator was exiting, cgroup deltion
1072  * path got to it first.
1073  */
1074 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1075 {
1076         unsigned long  flags;
1077         struct cfq_data *cfqd = key;
1078
1079         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1080         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1081         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1082 }
1083
1084 #else /* GROUP_IOSCHED */
1085 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1086 {
1087         return &cfqd->root_group;
1088 }
1089 static inline void
1090 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1091         cfqq->cfqg = cfqg;
1092 }
1093
1094 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1095 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1096
1097 #endif /* GROUP_IOSCHED */
1098
1099 /*
1100  * The cfqd->service_trees holds all pending cfq_queue's that have
1101  * requests waiting to be processed. It is sorted in the order that
1102  * we will service the queues.
1103  */
1104 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1105                                  bool add_front)
1106 {
1107         struct rb_node **p, *parent;
1108         struct cfq_queue *__cfqq;
1109         unsigned long rb_key;
1110         struct cfq_rb_root *service_tree;
1111         int left;
1112         int new_cfqq = 1;
1113         int group_changed = 0;
1114
1115 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1116         if (!cfqd->cfq_group_isolation
1117             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1118             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1119                 /* Move this cfq to root group */
1120                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1121                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1122                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1123                 cfqq->orig_cfqg = cfqq->cfqg;
1124                 cfqq->cfqg = &cfqd->root_group;
1125                 atomic_inc(&cfqd->root_group.ref);
1126                 group_changed = 1;
1127         } else if (!cfqd->cfq_group_isolation
1128                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1129                 /* cfqq is sequential now needs to go to its original group */
1130                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1131                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1132                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1133                 cfq_put_cfqg(cfqq->cfqg);
1134                 cfqq->cfqg = cfqq->orig_cfqg;
1135                 cfqq->orig_cfqg = NULL;
1136                 group_changed = 1;
1137                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1138         }
1139 #endif
1140
1141         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1142                                                 cfqq_type(cfqq));
1143         if (cfq_class_idle(cfqq)) {
1144                 rb_key = CFQ_IDLE_DELAY;
1145                 parent = rb_last(&service_tree->rb);
1146                 if (parent && parent != &cfqq->rb_node) {
1147                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1148                         rb_key += __cfqq->rb_key;
1149                 } else
1150                         rb_key += jiffies;
1151         } else if (!add_front) {
1152                 /*
1153                  * Get our rb key offset. Subtract any residual slice
1154                  * value carried from last service. A negative resid
1155                  * count indicates slice overrun, and this should position
1156                  * the next service time further away in the tree.
1157                  */
1158                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1159                 rb_key -= cfqq->slice_resid;
1160                 cfqq->slice_resid = 0;
1161         } else {
1162                 rb_key = -HZ;
1163                 __cfqq = cfq_rb_first(service_tree);
1164                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1165         }
1166
1167         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1168                 new_cfqq = 0;
1169                 /*
1170                  * same position, nothing more to do
1171                  */
1172                 if (rb_key == cfqq->rb_key &&
1173                     cfqq->service_tree == service_tree)
1174                         return;
1175
1176                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1177                 cfqq->service_tree = NULL;
1178         }
1179
1180         left = 1;
1181         parent = NULL;
1182         cfqq->service_tree = service_tree;
1183         p = &service_tree->rb.rb_node;
1184         while (*p) {
1185                 struct rb_node **n;
1186
1187                 parent = *p;
1188                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1189
1190                 /*
1191                  * sort by key, that represents service time.
1192                  */
1193                 if (time_before(rb_key, __cfqq->rb_key))
1194                         n = &(*p)->rb_left;
1195                 else {
1196                         n = &(*p)->rb_right;
1197                         left = 0;
1198                 }
1199
1200                 p = n;
1201         }
1202
1203         if (left)
1204                 service_tree->left = &cfqq->rb_node;
1205
1206         cfqq->rb_key = rb_key;
1207         rb_link_node(&cfqq->rb_node, parent, p);
1208         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1209         service_tree->count++;
1210         if ((add_front || !new_cfqq) && !group_changed)
1211                 return;
1212         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1213 }
1214
1215 static struct cfq_queue *
1216 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1217                      sector_t sector, struct rb_node **ret_parent,
1218                      struct rb_node ***rb_link)
1219 {
1220         struct rb_node **p, *parent;
1221         struct cfq_queue *cfqq = NULL;
1222
1223         parent = NULL;
1224         p = &root->rb_node;
1225         while (*p) {
1226                 struct rb_node **n;
1227
1228                 parent = *p;
1229                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1230
1231                 /*
1232                  * Sort strictly based on sector.  Smallest to the left,
1233                  * largest to the right.
1234                  */
1235                 if (sector > blk_rq_pos(cfqq->next_rq))
1236                         n = &(*p)->rb_right;
1237                 else if (sector < blk_rq_pos(cfqq->next_rq))
1238                         n = &(*p)->rb_left;
1239                 else
1240                         break;
1241                 p = n;
1242                 cfqq = NULL;
1243         }
1244
1245         *ret_parent = parent;
1246         if (rb_link)
1247                 *rb_link = p;
1248         return cfqq;
1249 }
1250
1251 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1252 {
1253         struct rb_node **p, *parent;
1254         struct cfq_queue *__cfqq;
1255
1256         if (cfqq->p_root) {
1257                 rb_erase(&cfqq->p_node, cfqq->p_root);
1258                 cfqq->p_root = NULL;
1259         }
1260
1261         if (cfq_class_idle(cfqq))
1262                 return;
1263         if (!cfqq->next_rq)
1264                 return;
1265
1266         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1267         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1268                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1269         if (!__cfqq) {
1270                 rb_link_node(&cfqq->p_node, parent, p);
1271                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1272         } else
1273                 cfqq->p_root = NULL;
1274 }
1275
1276 /*
1277  * Update cfqq's position in the service tree.
1278  */
1279 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1280 {
1281         /*
1282          * Resorting requires the cfqq to be on the RR list already.
1283          */
1284         if (cfq_cfqq_on_rr(cfqq)) {
1285                 cfq_service_tree_add(cfqd, cfqq, 0);
1286                 cfq_prio_tree_add(cfqd, cfqq);
1287         }
1288 }
1289
1290 /*
1291  * add to busy list of queues for service, trying to be fair in ordering
1292  * the pending list according to last request service
1293  */
1294 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1295 {
1296         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1297         BUG_ON(cfq_cfqq_on_rr(cfqq));
1298         cfq_mark_cfqq_on_rr(cfqq);
1299         cfqd->busy_queues++;
1300
1301         cfq_resort_rr_list(cfqd, cfqq);
1302 }
1303
1304 /*
1305  * Called when the cfqq no longer has requests pending, remove it from
1306  * the service tree.
1307  */
1308 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1309 {
1310         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1311         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1312         cfq_clear_cfqq_on_rr(cfqq);
1313
1314         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1315                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1316                 cfqq->service_tree = NULL;
1317         }
1318         if (cfqq->p_root) {
1319                 rb_erase(&cfqq->p_node, cfqq->p_root);
1320                 cfqq->p_root = NULL;
1321         }
1322
1323         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1324         BUG_ON(!cfqd->busy_queues);
1325         cfqd->busy_queues--;
1326 }
1327
1328 /*
1329  * rb tree support functions
1330  */
1331 static void cfq_del_rq_rb(struct request *rq)
1332 {
1333         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1334         const int sync = rq_is_sync(rq);
1335
1336         BUG_ON(!cfqq->queued[sync]);
1337         cfqq->queued[sync]--;
1338
1339         elv_rb_del(&cfqq->sort_list, rq);
1340
1341         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1342                 /*
1343                  * Queue will be deleted from service tree when we actually
1344                  * expire it later. Right now just remove it from prio tree
1345                  * as it is empty.
1346                  */
1347                 if (cfqq->p_root) {
1348                         rb_erase(&cfqq->p_node, cfqq->p_root);
1349                         cfqq->p_root = NULL;
1350                 }
1351         }
1352 }
1353
1354 static void cfq_add_rq_rb(struct request *rq)
1355 {
1356         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1357         struct cfq_data *cfqd = cfqq->cfqd;
1358         struct request *__alias, *prev;
1359
1360         cfqq->queued[rq_is_sync(rq)]++;
1361
1362         /*
1363          * looks a little odd, but the first insert might return an alias.
1364          * if that happens, put the alias on the dispatch list
1365          */
1366         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1367                 cfq_dispatch_insert(cfqd->queue, __alias);
1368
1369         if (!cfq_cfqq_on_rr(cfqq))
1370                 cfq_add_cfqq_rr(cfqd, cfqq);
1371
1372         /*
1373          * check if this request is a better next-serve candidate
1374          */
1375         prev = cfqq->next_rq;
1376         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1377
1378         /*
1379          * adjust priority tree position, if ->next_rq changes
1380          */
1381         if (prev != cfqq->next_rq)
1382                 cfq_prio_tree_add(cfqd, cfqq);
1383
1384         BUG_ON(!cfqq->next_rq);
1385 }
1386
1387 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1388 {
1389         elv_rb_del(&cfqq->sort_list, rq);
1390         cfqq->queued[rq_is_sync(rq)]--;
1391         cfq_add_rq_rb(rq);
1392 }
1393
1394 static struct request *
1395 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1396 {
1397         struct task_struct *tsk = current;
1398         struct cfq_io_context *cic;
1399         struct cfq_queue *cfqq;
1400
1401         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1402         if (!cic)
1403                 return NULL;
1404
1405         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1406         if (cfqq) {
1407                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1408
1409                 return elv_rb_find(&cfqq->sort_list, sector);
1410         }
1411
1412         return NULL;
1413 }
1414
1415 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1416 {
1417         struct cfq_data *cfqd = q->elevator->elevator_data;
1418
1419         cfqd->rq_in_driver++;
1420         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1421                                                 cfqd->rq_in_driver);
1422
1423         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1424 }
1425
1426 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1427 {
1428         struct cfq_data *cfqd = q->elevator->elevator_data;
1429
1430         WARN_ON(!cfqd->rq_in_driver);
1431         cfqd->rq_in_driver--;
1432         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1433                                                 cfqd->rq_in_driver);
1434 }
1435
1436 static void cfq_remove_request(struct request *rq)
1437 {
1438         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1439
1440         if (cfqq->next_rq == rq)
1441                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1442
1443         list_del_init(&rq->queuelist);
1444         cfq_del_rq_rb(rq);
1445
1446         cfqq->cfqd->rq_queued--;
1447         if (rq_is_meta(rq)) {
1448                 WARN_ON(!cfqq->meta_pending);
1449                 cfqq->meta_pending--;
1450         }
1451 }
1452
1453 static int cfq_merge(struct request_queue *q, struct request **req,
1454                      struct bio *bio)
1455 {
1456         struct cfq_data *cfqd = q->elevator->elevator_data;
1457         struct request *__rq;
1458
1459         __rq = cfq_find_rq_fmerge(cfqd, bio);
1460         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1461                 *req = __rq;
1462                 return ELEVATOR_FRONT_MERGE;
1463         }
1464
1465         return ELEVATOR_NO_MERGE;
1466 }
1467
1468 static void cfq_merged_request(struct request_queue *q, struct request *req,
1469                                int type)
1470 {
1471         if (type == ELEVATOR_FRONT_MERGE) {
1472                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1473
1474                 cfq_reposition_rq_rb(cfqq, req);
1475         }
1476 }
1477
1478 static void
1479 cfq_merged_requests(struct request_queue *q, struct request *rq,
1480                     struct request *next)
1481 {
1482         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1483         /*
1484          * reposition in fifo if next is older than rq
1485          */
1486         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1487             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1488                 list_move(&rq->queuelist, &next->queuelist);
1489                 rq_set_fifo_time(rq, rq_fifo_time(next));
1490         }
1491
1492         if (cfqq->next_rq == next)
1493                 cfqq->next_rq = rq;
1494         cfq_remove_request(next);
1495 }
1496
1497 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1498                            struct bio *bio)
1499 {
1500         struct cfq_data *cfqd = q->elevator->elevator_data;
1501         struct cfq_io_context *cic;
1502         struct cfq_queue *cfqq;
1503
1504         /*
1505          * Disallow merge of a sync bio into an async request.
1506          */
1507         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1508                 return false;
1509
1510         /*
1511          * Lookup the cfqq that this bio will be queued with. Allow
1512          * merge only if rq is queued there.
1513          */
1514         cic = cfq_cic_lookup(cfqd, current->io_context);
1515         if (!cic)
1516                 return false;
1517
1518         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1519         return cfqq == RQ_CFQQ(rq);
1520 }
1521
1522 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1523                                    struct cfq_queue *cfqq)
1524 {
1525         if (cfqq) {
1526                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1527                                 cfqd->serving_prio, cfqd->serving_type);
1528                 cfqq->slice_start = 0;
1529                 cfqq->dispatch_start = jiffies;
1530                 cfqq->allocated_slice = 0;
1531                 cfqq->slice_end = 0;
1532                 cfqq->slice_dispatch = 0;
1533                 cfqq->nr_sectors = 0;
1534
1535                 cfq_clear_cfqq_wait_request(cfqq);
1536                 cfq_clear_cfqq_must_dispatch(cfqq);
1537                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1538                 cfq_clear_cfqq_fifo_expire(cfqq);
1539                 cfq_mark_cfqq_slice_new(cfqq);
1540
1541                 del_timer(&cfqd->idle_slice_timer);
1542         }
1543
1544         cfqd->active_queue = cfqq;
1545 }
1546
1547 /*
1548  * current cfqq expired its slice (or was too idle), select new one
1549  */
1550 static void
1551 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1552                     bool timed_out)
1553 {
1554         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1555
1556         if (cfq_cfqq_wait_request(cfqq))
1557                 del_timer(&cfqd->idle_slice_timer);
1558
1559         cfq_clear_cfqq_wait_request(cfqq);
1560         cfq_clear_cfqq_wait_busy(cfqq);
1561
1562         /*
1563          * If this cfqq is shared between multiple processes, check to
1564          * make sure that those processes are still issuing I/Os within
1565          * the mean seek distance.  If not, it may be time to break the
1566          * queues apart again.
1567          */
1568         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1569                 cfq_mark_cfqq_split_coop(cfqq);
1570
1571         /*
1572          * store what was left of this slice, if the queue idled/timed out
1573          */
1574         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1575                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1576                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1577         }
1578
1579         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1580
1581         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1582                 cfq_del_cfqq_rr(cfqd, cfqq);
1583
1584         cfq_resort_rr_list(cfqd, cfqq);
1585
1586         if (cfqq == cfqd->active_queue)
1587                 cfqd->active_queue = NULL;
1588
1589         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1590                 cfqd->grp_service_tree.active = NULL;
1591
1592         if (cfqd->active_cic) {
1593                 put_io_context(cfqd->active_cic->ioc);
1594                 cfqd->active_cic = NULL;
1595         }
1596 }
1597
1598 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1599 {
1600         struct cfq_queue *cfqq = cfqd->active_queue;
1601
1602         if (cfqq)
1603                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1604 }
1605
1606 /*
1607  * Get next queue for service. Unless we have a queue preemption,
1608  * we'll simply select the first cfqq in the service tree.
1609  */
1610 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1611 {
1612         struct cfq_rb_root *service_tree =
1613                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1614                                         cfqd->serving_type);
1615
1616         if (!cfqd->rq_queued)
1617                 return NULL;
1618
1619         /* There is nothing to dispatch */
1620         if (!service_tree)
1621                 return NULL;
1622         if (RB_EMPTY_ROOT(&service_tree->rb))
1623                 return NULL;
1624         return cfq_rb_first(service_tree);
1625 }
1626
1627 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1628 {
1629         struct cfq_group *cfqg;
1630         struct cfq_queue *cfqq;
1631         int i, j;
1632         struct cfq_rb_root *st;
1633
1634         if (!cfqd->rq_queued)
1635                 return NULL;
1636
1637         cfqg = cfq_get_next_cfqg(cfqd);
1638         if (!cfqg)
1639                 return NULL;
1640
1641         for_each_cfqg_st(cfqg, i, j, st)
1642                 if ((cfqq = cfq_rb_first(st)) != NULL)
1643                         return cfqq;
1644         return NULL;
1645 }
1646
1647 /*
1648  * Get and set a new active queue for service.
1649  */
1650 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1651                                               struct cfq_queue *cfqq)
1652 {
1653         if (!cfqq)
1654                 cfqq = cfq_get_next_queue(cfqd);
1655
1656         __cfq_set_active_queue(cfqd, cfqq);
1657         return cfqq;
1658 }
1659
1660 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1661                                           struct request *rq)
1662 {
1663         if (blk_rq_pos(rq) >= cfqd->last_position)
1664                 return blk_rq_pos(rq) - cfqd->last_position;
1665         else
1666                 return cfqd->last_position - blk_rq_pos(rq);
1667 }
1668
1669 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1670                                struct request *rq)
1671 {
1672         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1673 }
1674
1675 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1676                                     struct cfq_queue *cur_cfqq)
1677 {
1678         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1679         struct rb_node *parent, *node;
1680         struct cfq_queue *__cfqq;
1681         sector_t sector = cfqd->last_position;
1682
1683         if (RB_EMPTY_ROOT(root))
1684                 return NULL;
1685
1686         /*
1687          * First, if we find a request starting at the end of the last
1688          * request, choose it.
1689          */
1690         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1691         if (__cfqq)
1692                 return __cfqq;
1693
1694         /*
1695          * If the exact sector wasn't found, the parent of the NULL leaf
1696          * will contain the closest sector.
1697          */
1698         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1699         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1700                 return __cfqq;
1701
1702         if (blk_rq_pos(__cfqq->next_rq) < sector)
1703                 node = rb_next(&__cfqq->p_node);
1704         else
1705                 node = rb_prev(&__cfqq->p_node);
1706         if (!node)
1707                 return NULL;
1708
1709         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1710         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1711                 return __cfqq;
1712
1713         return NULL;
1714 }
1715
1716 /*
1717  * cfqd - obvious
1718  * cur_cfqq - passed in so that we don't decide that the current queue is
1719  *            closely cooperating with itself.
1720  *
1721  * So, basically we're assuming that that cur_cfqq has dispatched at least
1722  * one request, and that cfqd->last_position reflects a position on the disk
1723  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1724  * assumption.
1725  */
1726 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1727                                               struct cfq_queue *cur_cfqq)
1728 {
1729         struct cfq_queue *cfqq;
1730
1731         if (cfq_class_idle(cur_cfqq))
1732                 return NULL;
1733         if (!cfq_cfqq_sync(cur_cfqq))
1734                 return NULL;
1735         if (CFQQ_SEEKY(cur_cfqq))
1736                 return NULL;
1737
1738         /*
1739          * Don't search priority tree if it's the only queue in the group.
1740          */
1741         if (cur_cfqq->cfqg->nr_cfqq == 1)
1742                 return NULL;
1743
1744         /*
1745          * We should notice if some of the queues are cooperating, eg
1746          * working closely on the same area of the disk. In that case,
1747          * we can group them together and don't waste time idling.
1748          */
1749         cfqq = cfqq_close(cfqd, cur_cfqq);
1750         if (!cfqq)
1751                 return NULL;
1752
1753         /* If new queue belongs to different cfq_group, don't choose it */
1754         if (cur_cfqq->cfqg != cfqq->cfqg)
1755                 return NULL;
1756
1757         /*
1758          * It only makes sense to merge sync queues.
1759          */
1760         if (!cfq_cfqq_sync(cfqq))
1761                 return NULL;
1762         if (CFQQ_SEEKY(cfqq))
1763                 return NULL;
1764
1765         /*
1766          * Do not merge queues of different priority classes
1767          */
1768         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1769                 return NULL;
1770
1771         return cfqq;
1772 }
1773
1774 /*
1775  * Determine whether we should enforce idle window for this queue.
1776  */
1777
1778 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1779 {
1780         enum wl_prio_t prio = cfqq_prio(cfqq);
1781         struct cfq_rb_root *service_tree = cfqq->service_tree;
1782
1783         BUG_ON(!service_tree);
1784         BUG_ON(!service_tree->count);
1785
1786         /* We never do for idle class queues. */
1787         if (prio == IDLE_WORKLOAD)
1788                 return false;
1789
1790         /* We do for queues that were marked with idle window flag. */
1791         if (cfq_cfqq_idle_window(cfqq) &&
1792            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1793                 return true;
1794
1795         /*
1796          * Otherwise, we do only if they are the last ones
1797          * in their service tree.
1798          */
1799         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1800                 return 1;
1801         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1802                         service_tree->count);
1803         return 0;
1804 }
1805
1806 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1807 {
1808         struct cfq_queue *cfqq = cfqd->active_queue;
1809         struct cfq_io_context *cic;
1810         unsigned long sl;
1811
1812         /*
1813          * SSD device without seek penalty, disable idling. But only do so
1814          * for devices that support queuing, otherwise we still have a problem
1815          * with sync vs async workloads.
1816          */
1817         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1818                 return;
1819
1820         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1821         WARN_ON(cfq_cfqq_slice_new(cfqq));
1822
1823         /*
1824          * idle is disabled, either manually or by past process history
1825          */
1826         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1827                 return;
1828
1829         /*
1830          * still active requests from this queue, don't idle
1831          */
1832         if (cfqq->dispatched)
1833                 return;
1834
1835         /*
1836          * task has exited, don't wait
1837          */
1838         cic = cfqd->active_cic;
1839         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1840                 return;
1841
1842         /*
1843          * If our average think time is larger than the remaining time
1844          * slice, then don't idle. This avoids overrunning the allotted
1845          * time slice.
1846          */
1847         if (sample_valid(cic->ttime_samples) &&
1848             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1849                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1850                                 cic->ttime_mean);
1851                 return;
1852         }
1853
1854         cfq_mark_cfqq_wait_request(cfqq);
1855
1856         sl = cfqd->cfq_slice_idle;
1857
1858         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1859         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1860 }
1861
1862 /*
1863  * Move request from internal lists to the request queue dispatch list.
1864  */
1865 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1866 {
1867         struct cfq_data *cfqd = q->elevator->elevator_data;
1868         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1869
1870         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1871
1872         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1873         cfq_remove_request(rq);
1874         cfqq->dispatched++;
1875         elv_dispatch_sort(q, rq);
1876
1877         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1878         cfqq->nr_sectors += blk_rq_sectors(rq);
1879 }
1880
1881 /*
1882  * return expired entry, or NULL to just start from scratch in rbtree
1883  */
1884 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1885 {
1886         struct request *rq = NULL;
1887
1888         if (cfq_cfqq_fifo_expire(cfqq))
1889                 return NULL;
1890
1891         cfq_mark_cfqq_fifo_expire(cfqq);
1892
1893         if (list_empty(&cfqq->fifo))
1894                 return NULL;
1895
1896         rq = rq_entry_fifo(cfqq->fifo.next);
1897         if (time_before(jiffies, rq_fifo_time(rq)))
1898                 rq = NULL;
1899
1900         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1901         return rq;
1902 }
1903
1904 static inline int
1905 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1906 {
1907         const int base_rq = cfqd->cfq_slice_async_rq;
1908
1909         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1910
1911         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1912 }
1913
1914 /*
1915  * Must be called with the queue_lock held.
1916  */
1917 static int cfqq_process_refs(struct cfq_queue *cfqq)
1918 {
1919         int process_refs, io_refs;
1920
1921         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1922         process_refs = atomic_read(&cfqq->ref) - io_refs;
1923         BUG_ON(process_refs < 0);
1924         return process_refs;
1925 }
1926
1927 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1928 {
1929         int process_refs, new_process_refs;
1930         struct cfq_queue *__cfqq;
1931
1932         /* Avoid a circular list and skip interim queue merges */
1933         while ((__cfqq = new_cfqq->new_cfqq)) {
1934                 if (__cfqq == cfqq)
1935                         return;
1936                 new_cfqq = __cfqq;
1937         }
1938
1939         process_refs = cfqq_process_refs(cfqq);
1940         /*
1941          * If the process for the cfqq has gone away, there is no
1942          * sense in merging the queues.
1943          */
1944         if (process_refs == 0)
1945                 return;
1946
1947         /*
1948          * Merge in the direction of the lesser amount of work.
1949          */
1950         new_process_refs = cfqq_process_refs(new_cfqq);
1951         if (new_process_refs >= process_refs) {
1952                 cfqq->new_cfqq = new_cfqq;
1953                 atomic_add(process_refs, &new_cfqq->ref);
1954         } else {
1955                 new_cfqq->new_cfqq = cfqq;
1956                 atomic_add(new_process_refs, &cfqq->ref);
1957         }
1958 }
1959
1960 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1961                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1962 {
1963         struct cfq_queue *queue;
1964         int i;
1965         bool key_valid = false;
1966         unsigned long lowest_key = 0;
1967         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1968
1969         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1970                 /* select the one with lowest rb_key */
1971                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1972                 if (queue &&
1973                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1974                         lowest_key = queue->rb_key;
1975                         cur_best = i;
1976                         key_valid = true;
1977                 }
1978         }
1979
1980         return cur_best;
1981 }
1982
1983 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1984 {
1985         unsigned slice;
1986         unsigned count;
1987         struct cfq_rb_root *st;
1988         unsigned group_slice;
1989
1990         if (!cfqg) {
1991                 cfqd->serving_prio = IDLE_WORKLOAD;
1992                 cfqd->workload_expires = jiffies + 1;
1993                 return;
1994         }
1995
1996         /* Choose next priority. RT > BE > IDLE */
1997         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1998                 cfqd->serving_prio = RT_WORKLOAD;
1999         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2000                 cfqd->serving_prio = BE_WORKLOAD;
2001         else {
2002                 cfqd->serving_prio = IDLE_WORKLOAD;
2003                 cfqd->workload_expires = jiffies + 1;
2004                 return;
2005         }
2006
2007         /*
2008          * For RT and BE, we have to choose also the type
2009          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2010          * expiration time
2011          */
2012         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2013         count = st->count;
2014
2015         /*
2016          * check workload expiration, and that we still have other queues ready
2017          */
2018         if (count && !time_after(jiffies, cfqd->workload_expires))
2019                 return;
2020
2021         /* otherwise select new workload type */
2022         cfqd->serving_type =
2023                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2024         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2025         count = st->count;
2026
2027         /*
2028          * the workload slice is computed as a fraction of target latency
2029          * proportional to the number of queues in that workload, over
2030          * all the queues in the same priority class
2031          */
2032         group_slice = cfq_group_slice(cfqd, cfqg);
2033
2034         slice = group_slice * count /
2035                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2036                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2037
2038         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2039                 unsigned int tmp;
2040
2041                 /*
2042                  * Async queues are currently system wide. Just taking
2043                  * proportion of queues with-in same group will lead to higher
2044                  * async ratio system wide as generally root group is going
2045                  * to have higher weight. A more accurate thing would be to
2046                  * calculate system wide asnc/sync ratio.
2047                  */
2048                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2049                 tmp = tmp/cfqd->busy_queues;
2050                 slice = min_t(unsigned, slice, tmp);
2051
2052                 /* async workload slice is scaled down according to
2053                  * the sync/async slice ratio. */
2054                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2055         } else
2056                 /* sync workload slice is at least 2 * cfq_slice_idle */
2057                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2058
2059         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2060         cfq_log(cfqd, "workload slice:%d", slice);
2061         cfqd->workload_expires = jiffies + slice;
2062         cfqd->noidle_tree_requires_idle = false;
2063 }
2064
2065 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2066 {
2067         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2068         struct cfq_group *cfqg;
2069
2070         if (RB_EMPTY_ROOT(&st->rb))
2071                 return NULL;
2072         cfqg = cfq_rb_first_group(st);
2073         st->active = &cfqg->rb_node;
2074         update_min_vdisktime(st);
2075         return cfqg;
2076 }
2077
2078 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2079 {
2080         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2081
2082         cfqd->serving_group = cfqg;
2083
2084         /* Restore the workload type data */
2085         if (cfqg->saved_workload_slice) {
2086                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2087                 cfqd->serving_type = cfqg->saved_workload;
2088                 cfqd->serving_prio = cfqg->saved_serving_prio;
2089         } else
2090                 cfqd->workload_expires = jiffies - 1;
2091
2092         choose_service_tree(cfqd, cfqg);
2093 }
2094
2095 /*
2096  * Select a queue for service. If we have a current active queue,
2097  * check whether to continue servicing it, or retrieve and set a new one.
2098  */
2099 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2100 {
2101         struct cfq_queue *cfqq, *new_cfqq = NULL;
2102
2103         cfqq = cfqd->active_queue;
2104         if (!cfqq)
2105                 goto new_queue;
2106
2107         if (!cfqd->rq_queued)
2108                 return NULL;
2109
2110         /*
2111          * We were waiting for group to get backlogged. Expire the queue
2112          */
2113         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2114                 goto expire;
2115
2116         /*
2117          * The active queue has run out of time, expire it and select new.
2118          */
2119         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2120                 /*
2121                  * If slice had not expired at the completion of last request
2122                  * we might not have turned on wait_busy flag. Don't expire
2123                  * the queue yet. Allow the group to get backlogged.
2124                  *
2125                  * The very fact that we have used the slice, that means we
2126                  * have been idling all along on this queue and it should be
2127                  * ok to wait for this request to complete.
2128                  */
2129                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2130                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2131                         cfqq = NULL;
2132                         goto keep_queue;
2133                 } else
2134                         goto expire;
2135         }
2136
2137         /*
2138          * The active queue has requests and isn't expired, allow it to
2139          * dispatch.
2140          */
2141         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2142                 goto keep_queue;
2143
2144         /*
2145          * If another queue has a request waiting within our mean seek
2146          * distance, let it run.  The expire code will check for close
2147          * cooperators and put the close queue at the front of the service
2148          * tree.  If possible, merge the expiring queue with the new cfqq.
2149          */
2150         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2151         if (new_cfqq) {
2152                 if (!cfqq->new_cfqq)
2153                         cfq_setup_merge(cfqq, new_cfqq);
2154                 goto expire;
2155         }
2156
2157         /*
2158          * No requests pending. If the active queue still has requests in
2159          * flight or is idling for a new request, allow either of these
2160          * conditions to happen (or time out) before selecting a new queue.
2161          */
2162         if (timer_pending(&cfqd->idle_slice_timer) ||
2163             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2164                 cfqq = NULL;
2165                 goto keep_queue;
2166         }
2167
2168 expire:
2169         cfq_slice_expired(cfqd, 0);
2170 new_queue:
2171         /*
2172          * Current queue expired. Check if we have to switch to a new
2173          * service tree
2174          */
2175         if (!new_cfqq)
2176                 cfq_choose_cfqg(cfqd);
2177
2178         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2179 keep_queue:
2180         return cfqq;
2181 }
2182
2183 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2184 {
2185         int dispatched = 0;
2186
2187         while (cfqq->next_rq) {
2188                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2189                 dispatched++;
2190         }
2191
2192         BUG_ON(!list_empty(&cfqq->fifo));
2193
2194         /* By default cfqq is not expired if it is empty. Do it explicitly */
2195         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2196         return dispatched;
2197 }
2198
2199 /*
2200  * Drain our current requests. Used for barriers and when switching
2201  * io schedulers on-the-fly.
2202  */
2203 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2204 {
2205         struct cfq_queue *cfqq;
2206         int dispatched = 0;
2207
2208         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2209                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2210
2211         cfq_slice_expired(cfqd, 0);
2212         BUG_ON(cfqd->busy_queues);
2213
2214         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2215         return dispatched;
2216 }
2217
2218 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2219         struct cfq_queue *cfqq)
2220 {
2221         /* the queue hasn't finished any request, can't estimate */
2222         if (cfq_cfqq_slice_new(cfqq))
2223                 return 1;
2224         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2225                 cfqq->slice_end))
2226                 return 1;
2227
2228         return 0;
2229 }
2230
2231 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2232 {
2233         unsigned int max_dispatch;
2234
2235         /*
2236          * Drain async requests before we start sync IO
2237          */
2238         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2239                 return false;
2240
2241         /*
2242          * If this is an async queue and we have sync IO in flight, let it wait
2243          */
2244         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2245                 return false;
2246
2247         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2248         if (cfq_class_idle(cfqq))
2249                 max_dispatch = 1;
2250
2251         /*
2252          * Does this cfqq already have too much IO in flight?
2253          */
2254         if (cfqq->dispatched >= max_dispatch) {
2255                 /*
2256                  * idle queue must always only have a single IO in flight
2257                  */
2258                 if (cfq_class_idle(cfqq))
2259                         return false;
2260
2261                 /*
2262                  * We have other queues, don't allow more IO from this one
2263                  */
2264                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2265                         return false;
2266
2267                 /*
2268                  * Sole queue user, no limit
2269                  */
2270                 if (cfqd->busy_queues == 1)
2271                         max_dispatch = -1;
2272                 else
2273                         /*
2274                          * Normally we start throttling cfqq when cfq_quantum/2
2275                          * requests have been dispatched. But we can drive
2276                          * deeper queue depths at the beginning of slice
2277                          * subjected to upper limit of cfq_quantum.
2278                          * */
2279                         max_dispatch = cfqd->cfq_quantum;
2280         }
2281
2282         /*
2283          * Async queues must wait a bit before being allowed dispatch.
2284          * We also ramp up the dispatch depth gradually for async IO,
2285          * based on the last sync IO we serviced
2286          */
2287         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2288                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2289                 unsigned int depth;
2290
2291                 depth = last_sync / cfqd->cfq_slice[1];
2292                 if (!depth && !cfqq->dispatched)
2293                         depth = 1;
2294                 if (depth < max_dispatch)
2295                         max_dispatch = depth;
2296         }
2297
2298         /*
2299          * If we're below the current max, allow a dispatch
2300          */
2301         return cfqq->dispatched < max_dispatch;
2302 }
2303
2304 /*
2305  * Dispatch a request from cfqq, moving them to the request queue
2306  * dispatch list.
2307  */
2308 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2309 {
2310         struct request *rq;
2311
2312         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2313
2314         if (!cfq_may_dispatch(cfqd, cfqq))
2315                 return false;
2316
2317         /*
2318          * follow expired path, else get first next available
2319          */
2320         rq = cfq_check_fifo(cfqq);
2321         if (!rq)
2322                 rq = cfqq->next_rq;
2323
2324         /*
2325          * insert request into driver dispatch list
2326          */
2327         cfq_dispatch_insert(cfqd->queue, rq);
2328
2329         if (!cfqd->active_cic) {
2330                 struct cfq_io_context *cic = RQ_CIC(rq);
2331
2332                 atomic_long_inc(&cic->ioc->refcount);
2333                 cfqd->active_cic = cic;
2334         }
2335
2336         return true;
2337 }
2338
2339 /*
2340  * Find the cfqq that we need to service and move a request from that to the
2341  * dispatch list
2342  */
2343 static int cfq_dispatch_requests(struct request_queue *q, int force)
2344 {
2345         struct cfq_data *cfqd = q->elevator->elevator_data;
2346         struct cfq_queue *cfqq;
2347
2348         if (!cfqd->busy_queues)
2349                 return 0;
2350
2351         if (unlikely(force))
2352                 return cfq_forced_dispatch(cfqd);
2353
2354         cfqq = cfq_select_queue(cfqd);
2355         if (!cfqq)
2356                 return 0;
2357
2358         /*
2359          * Dispatch a request from this cfqq, if it is allowed
2360          */
2361         if (!cfq_dispatch_request(cfqd, cfqq))
2362                 return 0;
2363
2364         cfqq->slice_dispatch++;
2365         cfq_clear_cfqq_must_dispatch(cfqq);
2366
2367         /*
2368          * expire an async queue immediately if it has used up its slice. idle
2369          * queue always expire after 1 dispatch round.
2370          */
2371         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2372             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2373             cfq_class_idle(cfqq))) {
2374                 cfqq->slice_end = jiffies + 1;
2375                 cfq_slice_expired(cfqd, 0);
2376         }
2377
2378         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2379         return 1;
2380 }
2381
2382 /*
2383  * task holds one reference to the queue, dropped when task exits. each rq
2384  * in-flight on this queue also holds a reference, dropped when rq is freed.
2385  *
2386  * Each cfq queue took a reference on the parent group. Drop it now.
2387  * queue lock must be held here.
2388  */
2389 static void cfq_put_queue(struct cfq_queue *cfqq)
2390 {
2391         struct cfq_data *cfqd = cfqq->cfqd;
2392         struct cfq_group *cfqg, *orig_cfqg;
2393
2394         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2395
2396         if (!atomic_dec_and_test(&cfqq->ref))
2397                 return;
2398
2399         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2400         BUG_ON(rb_first(&cfqq->sort_list));
2401         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2402         cfqg = cfqq->cfqg;
2403         orig_cfqg = cfqq->orig_cfqg;
2404
2405         if (unlikely(cfqd->active_queue == cfqq)) {
2406                 __cfq_slice_expired(cfqd, cfqq, 0);
2407                 cfq_schedule_dispatch(cfqd);
2408         }
2409
2410         BUG_ON(cfq_cfqq_on_rr(cfqq));
2411         kmem_cache_free(cfq_pool, cfqq);
2412         cfq_put_cfqg(cfqg);
2413         if (orig_cfqg)
2414                 cfq_put_cfqg(orig_cfqg);
2415 }
2416
2417 /*
2418  * Must always be called with the rcu_read_lock() held
2419  */
2420 static void
2421 __call_for_each_cic(struct io_context *ioc,
2422                     void (*func)(struct io_context *, struct cfq_io_context *))
2423 {
2424         struct cfq_io_context *cic;
2425         struct hlist_node *n;
2426
2427         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2428                 func(ioc, cic);
2429 }
2430
2431 /*
2432  * Call func for each cic attached to this ioc.
2433  */
2434 static void
2435 call_for_each_cic(struct io_context *ioc,
2436                   void (*func)(struct io_context *, struct cfq_io_context *))
2437 {
2438         rcu_read_lock();
2439         __call_for_each_cic(ioc, func);
2440         rcu_read_unlock();
2441 }
2442
2443 static void cfq_cic_free_rcu(struct rcu_head *head)
2444 {
2445         struct cfq_io_context *cic;
2446
2447         cic = container_of(head, struct cfq_io_context, rcu_head);
2448
2449         kmem_cache_free(cfq_ioc_pool, cic);
2450         elv_ioc_count_dec(cfq_ioc_count);
2451
2452         if (ioc_gone) {
2453                 /*
2454                  * CFQ scheduler is exiting, grab exit lock and check
2455                  * the pending io context count. If it hits zero,
2456                  * complete ioc_gone and set it back to NULL
2457                  */
2458                 spin_lock(&ioc_gone_lock);
2459                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2460                         complete(ioc_gone);
2461                         ioc_gone = NULL;
2462                 }
2463                 spin_unlock(&ioc_gone_lock);
2464         }
2465 }
2466
2467 static void cfq_cic_free(struct cfq_io_context *cic)
2468 {
2469         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2470 }
2471
2472 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2473 {
2474         unsigned long flags;
2475
2476         BUG_ON(!cic->dead_key);
2477
2478         spin_lock_irqsave(&ioc->lock, flags);
2479         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2480         hlist_del_rcu(&cic->cic_list);
2481         spin_unlock_irqrestore(&ioc->lock, flags);
2482
2483         cfq_cic_free(cic);
2484 }
2485
2486 /*
2487  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2488  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2489  * and ->trim() which is called with the task lock held
2490  */
2491 static void cfq_free_io_context(struct io_context *ioc)
2492 {
2493         /*
2494          * ioc->refcount is zero here, or we are called from elv_unregister(),
2495          * so no more cic's are allowed to be linked into this ioc.  So it
2496          * should be ok to iterate over the known list, we will see all cic's
2497          * since no new ones are added.
2498          */
2499         __call_for_each_cic(ioc, cic_free_func);
2500 }
2501
2502 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2503 {
2504         struct cfq_queue *__cfqq, *next;
2505
2506         if (unlikely(cfqq == cfqd->active_queue)) {
2507                 __cfq_slice_expired(cfqd, cfqq, 0);
2508                 cfq_schedule_dispatch(cfqd);
2509         }
2510
2511         /*
2512          * If this queue was scheduled to merge with another queue, be
2513          * sure to drop the reference taken on that queue (and others in
2514          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2515          */
2516         __cfqq = cfqq->new_cfqq;
2517         while (__cfqq) {
2518                 if (__cfqq == cfqq) {
2519                         WARN(1, "cfqq->new_cfqq loop detected\n");
2520                         break;
2521                 }
2522                 next = __cfqq->new_cfqq;
2523                 cfq_put_queue(__cfqq);
2524                 __cfqq = next;
2525         }
2526
2527         cfq_put_queue(cfqq);
2528 }
2529
2530 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2531                                          struct cfq_io_context *cic)
2532 {
2533         struct io_context *ioc = cic->ioc;
2534
2535         list_del_init(&cic->queue_list);
2536
2537         /*
2538          * Make sure key == NULL is seen for dead queues
2539          */
2540         smp_wmb();
2541         cic->dead_key = (unsigned long) cic->key;
2542         cic->key = NULL;
2543
2544         if (ioc->ioc_data == cic)
2545                 rcu_assign_pointer(ioc->ioc_data, NULL);
2546
2547         if (cic->cfqq[BLK_RW_ASYNC]) {
2548                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2549                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2550         }
2551
2552         if (cic->cfqq[BLK_RW_SYNC]) {
2553                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2554                 cic->cfqq[BLK_RW_SYNC] = NULL;
2555         }
2556 }
2557
2558 static void cfq_exit_single_io_context(struct io_context *ioc,
2559                                        struct cfq_io_context *cic)
2560 {
2561         struct cfq_data *cfqd = cic->key;
2562
2563         if (cfqd) {
2564                 struct request_queue *q = cfqd->queue;
2565                 unsigned long flags;
2566
2567                 spin_lock_irqsave(q->queue_lock, flags);
2568
2569                 /*
2570                  * Ensure we get a fresh copy of the ->key to prevent
2571                  * race between exiting task and queue
2572                  */
2573                 smp_read_barrier_depends();
2574                 if (cic->key)
2575                         __cfq_exit_single_io_context(cfqd, cic);
2576
2577                 spin_unlock_irqrestore(q->queue_lock, flags);
2578         }
2579 }
2580
2581 /*
2582  * The process that ioc belongs to has exited, we need to clean up
2583  * and put the internal structures we have that belongs to that process.
2584  */
2585 static void cfq_exit_io_context(struct io_context *ioc)
2586 {
2587         call_for_each_cic(ioc, cfq_exit_single_io_context);
2588 }
2589
2590 static struct cfq_io_context *
2591 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2592 {
2593         struct cfq_io_context *cic;
2594
2595         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2596                                                         cfqd->queue->node);
2597         if (cic) {
2598                 cic->last_end_request = jiffies;
2599                 INIT_LIST_HEAD(&cic->queue_list);
2600                 INIT_HLIST_NODE(&cic->cic_list);
2601                 cic->dtor = cfq_free_io_context;
2602                 cic->exit = cfq_exit_io_context;
2603                 elv_ioc_count_inc(cfq_ioc_count);
2604         }
2605
2606         return cic;
2607 }
2608
2609 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2610 {
2611         struct task_struct *tsk = current;
2612         int ioprio_class;
2613
2614         if (!cfq_cfqq_prio_changed(cfqq))
2615                 return;
2616
2617         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2618         switch (ioprio_class) {
2619         default:
2620                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2621         case IOPRIO_CLASS_NONE:
2622                 /*
2623                  * no prio set, inherit CPU scheduling settings
2624                  */
2625                 cfqq->ioprio = task_nice_ioprio(tsk);
2626                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2627                 break;
2628         case IOPRIO_CLASS_RT:
2629                 cfqq->ioprio = task_ioprio(ioc);
2630                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2631                 break;
2632         case IOPRIO_CLASS_BE:
2633                 cfqq->ioprio = task_ioprio(ioc);
2634                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2635                 break;
2636         case IOPRIO_CLASS_IDLE:
2637                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2638                 cfqq->ioprio = 7;
2639                 cfq_clear_cfqq_idle_window(cfqq);
2640                 break;
2641         }
2642
2643         /*
2644          * keep track of original prio settings in case we have to temporarily
2645          * elevate the priority of this queue
2646          */
2647         cfqq->org_ioprio = cfqq->ioprio;
2648         cfqq->org_ioprio_class = cfqq->ioprio_class;
2649         cfq_clear_cfqq_prio_changed(cfqq);
2650 }
2651
2652 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2653 {
2654         struct cfq_data *cfqd = cic->key;
2655         struct cfq_queue *cfqq;
2656         unsigned long flags;
2657
2658         if (unlikely(!cfqd))
2659                 return;
2660
2661         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2662
2663         cfqq = cic->cfqq[BLK_RW_ASYNC];
2664         if (cfqq) {
2665                 struct cfq_queue *new_cfqq;
2666                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2667                                                 GFP_ATOMIC);
2668                 if (new_cfqq) {
2669                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2670                         cfq_put_queue(cfqq);
2671                 }
2672         }
2673
2674         cfqq = cic->cfqq[BLK_RW_SYNC];
2675         if (cfqq)
2676                 cfq_mark_cfqq_prio_changed(cfqq);
2677
2678         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2679 }
2680
2681 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2682 {
2683         call_for_each_cic(ioc, changed_ioprio);
2684         ioc->ioprio_changed = 0;
2685 }
2686
2687 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2688                           pid_t pid, bool is_sync)
2689 {
2690         RB_CLEAR_NODE(&cfqq->rb_node);
2691         RB_CLEAR_NODE(&cfqq->p_node);
2692         INIT_LIST_HEAD(&cfqq->fifo);
2693
2694         atomic_set(&cfqq->ref, 0);
2695         cfqq->cfqd = cfqd;
2696
2697         cfq_mark_cfqq_prio_changed(cfqq);
2698
2699         if (is_sync) {
2700                 if (!cfq_class_idle(cfqq))
2701                         cfq_mark_cfqq_idle_window(cfqq);
2702                 cfq_mark_cfqq_sync(cfqq);
2703         }
2704         cfqq->pid = pid;
2705 }
2706
2707 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2708 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2709 {
2710         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2711         struct cfq_data *cfqd = cic->key;
2712         unsigned long flags;
2713         struct request_queue *q;
2714
2715         if (unlikely(!cfqd))
2716                 return;
2717
2718         q = cfqd->queue;
2719
2720         spin_lock_irqsave(q->queue_lock, flags);
2721
2722         if (sync_cfqq) {
2723                 /*
2724                  * Drop reference to sync queue. A new sync queue will be
2725                  * assigned in new group upon arrival of a fresh request.
2726                  */
2727                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2728                 cic_set_cfqq(cic, NULL, 1);
2729                 cfq_put_queue(sync_cfqq);
2730         }
2731
2732         spin_unlock_irqrestore(q->queue_lock, flags);
2733 }
2734
2735 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2736 {
2737         call_for_each_cic(ioc, changed_cgroup);
2738         ioc->cgroup_changed = 0;
2739 }
2740 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2741
2742 static struct cfq_queue *
2743 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2744                      struct io_context *ioc, gfp_t gfp_mask)
2745 {
2746         struct cfq_queue *cfqq, *new_cfqq = NULL;
2747         struct cfq_io_context *cic;
2748         struct cfq_group *cfqg;
2749
2750 retry:
2751         cfqg = cfq_get_cfqg(cfqd, 1);
2752         cic = cfq_cic_lookup(cfqd, ioc);
2753         /* cic always exists here */
2754         cfqq = cic_to_cfqq(cic, is_sync);
2755
2756         /*
2757          * Always try a new alloc if we fell back to the OOM cfqq
2758          * originally, since it should just be a temporary situation.
2759          */
2760         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2761                 cfqq = NULL;
2762                 if (new_cfqq) {
2763                         cfqq = new_cfqq;
2764                         new_cfqq = NULL;
2765                 } else if (gfp_mask & __GFP_WAIT) {
2766                         spin_unlock_irq(cfqd->queue->queue_lock);
2767                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2768                                         gfp_mask | __GFP_ZERO,
2769                                         cfqd->queue->node);
2770                         spin_lock_irq(cfqd->queue->queue_lock);
2771                         if (new_cfqq)
2772                                 goto retry;
2773                 } else {
2774                         cfqq = kmem_cache_alloc_node(cfq_pool,
2775                                         gfp_mask | __GFP_ZERO,
2776                                         cfqd->queue->node);
2777                 }
2778
2779                 if (cfqq) {
2780                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2781                         cfq_init_prio_data(cfqq, ioc);
2782                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2783                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2784                 } else
2785                         cfqq = &cfqd->oom_cfqq;
2786         }
2787
2788         if (new_cfqq)
2789                 kmem_cache_free(cfq_pool, new_cfqq);
2790
2791         return cfqq;
2792 }
2793
2794 static struct cfq_queue **
2795 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2796 {
2797         switch (ioprio_class) {
2798         case IOPRIO_CLASS_RT:
2799                 return &cfqd->async_cfqq[0][ioprio];
2800         case IOPRIO_CLASS_BE:
2801                 return &cfqd->async_cfqq[1][ioprio];
2802         case IOPRIO_CLASS_IDLE:
2803                 return &cfqd->async_idle_cfqq;
2804         default:
2805                 BUG();
2806         }
2807 }
2808
2809 static struct cfq_queue *
2810 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2811               gfp_t gfp_mask)
2812 {
2813         const int ioprio = task_ioprio(ioc);
2814         const int ioprio_class = task_ioprio_class(ioc);
2815         struct cfq_queue **async_cfqq = NULL;
2816         struct cfq_queue *cfqq = NULL;
2817
2818         if (!is_sync) {
2819                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2820                 cfqq = *async_cfqq;
2821         }
2822
2823         if (!cfqq)
2824                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2825
2826         /*
2827          * pin the queue now that it's allocated, scheduler exit will prune it
2828          */
2829         if (!is_sync && !(*async_cfqq)) {
2830                 atomic_inc(&cfqq->ref);
2831                 *async_cfqq = cfqq;
2832         }
2833
2834         atomic_inc(&cfqq->ref);
2835         return cfqq;
2836 }
2837
2838 /*
2839  * We drop cfq io contexts lazily, so we may find a dead one.
2840  */
2841 static void
2842 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2843                   struct cfq_io_context *cic)
2844 {
2845         unsigned long flags;
2846
2847         WARN_ON(!list_empty(&cic->queue_list));
2848
2849         spin_lock_irqsave(&ioc->lock, flags);
2850
2851         BUG_ON(ioc->ioc_data == cic);
2852
2853         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2854         hlist_del_rcu(&cic->cic_list);
2855         spin_unlock_irqrestore(&ioc->lock, flags);
2856
2857         cfq_cic_free(cic);
2858 }
2859
2860 static struct cfq_io_context *
2861 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2862 {
2863         struct cfq_io_context *cic;
2864         unsigned long flags;
2865         void *k;
2866
2867         if (unlikely(!ioc))
2868                 return NULL;
2869
2870         rcu_read_lock();
2871
2872         /*
2873          * we maintain a last-hit cache, to avoid browsing over the tree
2874          */
2875         cic = rcu_dereference(ioc->ioc_data);
2876         if (cic && cic->key == cfqd) {
2877                 rcu_read_unlock();
2878                 return cic;
2879         }
2880
2881         do {
2882                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2883                 rcu_read_unlock();
2884                 if (!cic)
2885                         break;
2886                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2887                 k = cic->key;
2888                 if (unlikely(!k)) {
2889                         cfq_drop_dead_cic(cfqd, ioc, cic);
2890                         rcu_read_lock();
2891                         continue;
2892                 }
2893
2894                 spin_lock_irqsave(&ioc->lock, flags);
2895                 rcu_assign_pointer(ioc->ioc_data, cic);
2896                 spin_unlock_irqrestore(&ioc->lock, flags);
2897                 break;
2898         } while (1);
2899
2900         return cic;
2901 }
2902
2903 /*
2904  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2905  * the process specific cfq io context when entered from the block layer.
2906  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2907  */
2908 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2909                         struct cfq_io_context *cic, gfp_t gfp_mask)
2910 {
2911         unsigned long flags;
2912         int ret;
2913
2914         ret = radix_tree_preload(gfp_mask);
2915         if (!ret) {
2916                 cic->ioc = ioc;
2917                 cic->key = cfqd;
2918
2919                 spin_lock_irqsave(&ioc->lock, flags);
2920                 ret = radix_tree_insert(&ioc->radix_root,
2921                                                 (unsigned long) cfqd, cic);
2922                 if (!ret)
2923                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2924                 spin_unlock_irqrestore(&ioc->lock, flags);
2925
2926                 radix_tree_preload_end();
2927
2928                 if (!ret) {
2929                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2930                         list_add(&cic->queue_list, &cfqd->cic_list);
2931                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2932                 }
2933         }
2934
2935         if (ret)
2936                 printk(KERN_ERR "cfq: cic link failed!\n");
2937
2938         return ret;
2939 }
2940
2941 /*
2942  * Setup general io context and cfq io context. There can be several cfq
2943  * io contexts per general io context, if this process is doing io to more
2944  * than one device managed by cfq.
2945  */
2946 static struct cfq_io_context *
2947 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2948 {
2949         struct io_context *ioc = NULL;
2950         struct cfq_io_context *cic;
2951
2952         might_sleep_if(gfp_mask & __GFP_WAIT);
2953
2954         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2955         if (!ioc)
2956                 return NULL;
2957
2958         cic = cfq_cic_lookup(cfqd, ioc);
2959         if (cic)
2960                 goto out;
2961
2962         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2963         if (cic == NULL)
2964                 goto err;
2965
2966         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2967                 goto err_free;
2968
2969 out:
2970         smp_read_barrier_depends();
2971         if (unlikely(ioc->ioprio_changed))
2972                 cfq_ioc_set_ioprio(ioc);
2973
2974 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2975         if (unlikely(ioc->cgroup_changed))
2976                 cfq_ioc_set_cgroup(ioc);
2977 #endif
2978         return cic;
2979 err_free:
2980         cfq_cic_free(cic);
2981 err:
2982         put_io_context(ioc);
2983         return NULL;
2984 }
2985
2986 static void
2987 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2988 {
2989         unsigned long elapsed = jiffies - cic->last_end_request;
2990         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2991
2992         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2993         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2994         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2995 }
2996
2997 static void
2998 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2999                        struct request *rq)
3000 {
3001         sector_t sdist = 0;
3002         sector_t n_sec = blk_rq_sectors(rq);
3003         if (cfqq->last_request_pos) {
3004                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3005                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3006                 else
3007                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3008         }
3009
3010         cfqq->seek_history <<= 1;
3011         if (blk_queue_nonrot(cfqd->queue))
3012                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3013         else
3014                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3015 }
3016
3017 /*
3018  * Disable idle window if the process thinks too long or seeks so much that
3019  * it doesn't matter
3020  */
3021 static void
3022 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3023                        struct cfq_io_context *cic)
3024 {
3025         int old_idle, enable_idle;
3026
3027         /*
3028          * Don't idle for async or idle io prio class
3029          */
3030         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3031                 return;
3032
3033         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3034
3035         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3036                 cfq_mark_cfqq_deep(cfqq);
3037
3038         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3039             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3040                 enable_idle = 0;
3041         else if (sample_valid(cic->ttime_samples)) {
3042                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3043                         enable_idle = 0;
3044                 else
3045                         enable_idle = 1;
3046         }
3047
3048         if (old_idle != enable_idle) {
3049                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3050                 if (enable_idle)
3051                         cfq_mark_cfqq_idle_window(cfqq);
3052                 else
3053                         cfq_clear_cfqq_idle_window(cfqq);
3054         }
3055 }
3056
3057 /*
3058  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3059  * no or if we aren't sure, a 1 will cause a preempt.
3060  */
3061 static bool
3062 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3063                    struct request *rq)
3064 {
3065         struct cfq_queue *cfqq;
3066
3067         cfqq = cfqd->active_queue;
3068         if (!cfqq)
3069                 return false;
3070
3071         if (cfq_class_idle(new_cfqq))
3072                 return false;
3073
3074         if (cfq_class_idle(cfqq))
3075                 return true;
3076
3077         /*
3078          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3079          */
3080         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3081                 return false;
3082
3083         /*
3084          * if the new request is sync, but the currently running queue is
3085          * not, let the sync request have priority.
3086          */
3087         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3088                 return true;
3089
3090         if (new_cfqq->cfqg != cfqq->cfqg)
3091                 return false;
3092
3093         if (cfq_slice_used(cfqq))
3094                 return true;
3095
3096         /* Allow preemption only if we are idling on sync-noidle tree */
3097         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3098             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3099             new_cfqq->service_tree->count == 2 &&
3100             RB_EMPTY_ROOT(&cfqq->sort_list))
3101                 return true;
3102
3103         /*
3104          * So both queues are sync. Let the new request get disk time if
3105          * it's a metadata request and the current queue is doing regular IO.
3106          */
3107         if (rq_is_meta(rq) && !cfqq->meta_pending)
3108                 return true;
3109
3110         /*
3111          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3112          */
3113         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3114                 return true;
3115
3116         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3117                 return false;
3118
3119         /*
3120          * if this request is as-good as one we would expect from the
3121          * current cfqq, let it preempt
3122          */
3123         if (cfq_rq_close(cfqd, cfqq, rq))
3124                 return true;
3125
3126         return false;
3127 }
3128
3129 /*
3130  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3131  * let it have half of its nominal slice.
3132  */
3133 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3134 {
3135         cfq_log_cfqq(cfqd, cfqq, "preempt");
3136         cfq_slice_expired(cfqd, 1);
3137
3138         /*
3139          * Put the new queue at the front of the of the current list,
3140          * so we know that it will be selected next.
3141          */
3142         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3143
3144         cfq_service_tree_add(cfqd, cfqq, 1);
3145
3146         cfqq->slice_end = 0;
3147         cfq_mark_cfqq_slice_new(cfqq);
3148 }
3149
3150 /*
3151  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3152  * something we should do about it
3153  */
3154 static void
3155 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3156                 struct request *rq)
3157 {
3158         struct cfq_io_context *cic = RQ_CIC(rq);
3159
3160         cfqd->rq_queued++;
3161         if (rq_is_meta(rq))
3162                 cfqq->meta_pending++;
3163
3164         cfq_update_io_thinktime(cfqd, cic);
3165         cfq_update_io_seektime(cfqd, cfqq, rq);
3166         cfq_update_idle_window(cfqd, cfqq, cic);
3167
3168         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3169
3170         if (cfqq == cfqd->active_queue) {
3171                 /*
3172                  * Remember that we saw a request from this process, but
3173                  * don't start queuing just yet. Otherwise we risk seeing lots
3174                  * of tiny requests, because we disrupt the normal plugging
3175                  * and merging. If the request is already larger than a single
3176                  * page, let it rip immediately. For that case we assume that
3177                  * merging is already done. Ditto for a busy system that
3178                  * has other work pending, don't risk delaying until the
3179                  * idle timer unplug to continue working.
3180                  */
3181                 if (cfq_cfqq_wait_request(cfqq)) {
3182                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3183                             cfqd->busy_queues > 1) {
3184                                 del_timer(&cfqd->idle_slice_timer);
3185                                 cfq_clear_cfqq_wait_request(cfqq);
3186                                 __blk_run_queue(cfqd->queue);
3187                         } else
3188                                 cfq_mark_cfqq_must_dispatch(cfqq);
3189                 }
3190         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3191                 /*
3192                  * not the active queue - expire current slice if it is
3193                  * idle and has expired it's mean thinktime or this new queue
3194                  * has some old slice time left and is of higher priority or
3195                  * this new queue is RT and the current one is BE
3196                  */
3197                 cfq_preempt_queue(cfqd, cfqq);
3198                 __blk_run_queue(cfqd->queue);
3199         }
3200 }
3201
3202 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3203 {
3204         struct cfq_data *cfqd = q->elevator->elevator_data;
3205         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3206
3207         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3208         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3209
3210         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3211         list_add_tail(&rq->queuelist, &cfqq->fifo);
3212         cfq_add_rq_rb(rq);
3213
3214         cfq_rq_enqueued(cfqd, cfqq, rq);
3215 }
3216
3217 /*
3218  * Update hw_tag based on peak queue depth over 50 samples under
3219  * sufficient load.
3220  */
3221 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3222 {
3223         struct cfq_queue *cfqq = cfqd->active_queue;
3224
3225         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3226                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3227
3228         if (cfqd->hw_tag == 1)
3229                 return;
3230
3231         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3232             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3233                 return;
3234
3235         /*
3236          * If active queue hasn't enough requests and can idle, cfq might not
3237          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3238          * case
3239          */
3240         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3241             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3242             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3243                 return;
3244
3245         if (cfqd->hw_tag_samples++ < 50)
3246                 return;
3247
3248         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3249                 cfqd->hw_tag = 1;
3250         else
3251                 cfqd->hw_tag = 0;
3252 }
3253
3254 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3255 {
3256         struct cfq_io_context *cic = cfqd->active_cic;
3257
3258         /* If there are other queues in the group, don't wait */
3259         if (cfqq->cfqg->nr_cfqq > 1)
3260                 return false;
3261
3262         if (cfq_slice_used(cfqq))
3263                 return true;
3264
3265         /* if slice left is less than think time, wait busy */
3266         if (cic && sample_valid(cic->ttime_samples)
3267             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3268                 return true;
3269
3270         /*
3271          * If think times is less than a jiffy than ttime_mean=0 and above
3272          * will not be true. It might happen that slice has not expired yet
3273          * but will expire soon (4-5 ns) during select_queue(). To cover the
3274          * case where think time is less than a jiffy, mark the queue wait
3275          * busy if only 1 jiffy is left in the slice.
3276          */
3277         if (cfqq->slice_end - jiffies == 1)
3278                 return true;
3279
3280         return false;
3281 }
3282
3283 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3284 {
3285         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3286         struct cfq_data *cfqd = cfqq->cfqd;
3287         const int sync = rq_is_sync(rq);
3288         unsigned long now;
3289
3290         now = jiffies;
3291         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3292
3293         cfq_update_hw_tag(cfqd);
3294
3295         WARN_ON(!cfqd->rq_in_driver);
3296         WARN_ON(!cfqq->dispatched);
3297         cfqd->rq_in_driver--;
3298         cfqq->dispatched--;
3299
3300         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3301
3302         if (sync) {
3303                 RQ_CIC(rq)->last_end_request = now;
3304                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3305                         cfqd->last_delayed_sync = now;
3306         }
3307
3308         /*
3309          * If this is the active queue, check if it needs to be expired,
3310          * or if we want to idle in case it has no pending requests.
3311          */
3312         if (cfqd->active_queue == cfqq) {
3313                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3314
3315                 if (cfq_cfqq_slice_new(cfqq)) {
3316                         cfq_set_prio_slice(cfqd, cfqq);
3317                         cfq_clear_cfqq_slice_new(cfqq);
3318                 }
3319
3320                 /*
3321                  * Should we wait for next request to come in before we expire
3322                  * the queue.
3323                  */
3324                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3325                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3326                         cfq_mark_cfqq_wait_busy(cfqq);
3327                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3328                 }
3329
3330                 /*
3331                  * Idling is not enabled on:
3332                  * - expired queues
3333                  * - idle-priority queues
3334                  * - async queues
3335                  * - queues with still some requests queued
3336                  * - when there is a close cooperator
3337                  */
3338                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3339                         cfq_slice_expired(cfqd, 1);
3340                 else if (sync && cfqq_empty &&
3341                          !cfq_close_cooperator(cfqd, cfqq)) {
3342                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3343                         /*
3344                          * Idling is enabled for SYNC_WORKLOAD.
3345                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3346                          * only if we processed at least one !rq_noidle request
3347                          */
3348                         if (cfqd->serving_type == SYNC_WORKLOAD
3349                             || cfqd->noidle_tree_requires_idle
3350                             || cfqq->cfqg->nr_cfqq == 1)
3351                                 cfq_arm_slice_timer(cfqd);
3352                 }
3353         }
3354
3355         if (!cfqd->rq_in_driver)
3356                 cfq_schedule_dispatch(cfqd);
3357 }
3358
3359 /*
3360  * we temporarily boost lower priority queues if they are holding fs exclusive
3361  * resources. they are boosted to normal prio (CLASS_BE/4)
3362  */
3363 static void cfq_prio_boost(struct cfq_queue *cfqq)
3364 {
3365         if (has_fs_excl()) {
3366                 /*
3367                  * boost idle prio on transactions that would lock out other
3368                  * users of the filesystem
3369                  */
3370                 if (cfq_class_idle(cfqq))
3371                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3372                 if (cfqq->ioprio > IOPRIO_NORM)
3373                         cfqq->ioprio = IOPRIO_NORM;
3374         } else {
3375                 /*
3376                  * unboost the queue (if needed)
3377                  */
3378                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3379                 cfqq->ioprio = cfqq->org_ioprio;
3380         }
3381 }
3382
3383 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3384 {
3385         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3386                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3387                 return ELV_MQUEUE_MUST;
3388         }
3389
3390         return ELV_MQUEUE_MAY;
3391 }
3392
3393 static int cfq_may_queue(struct request_queue *q, int rw)
3394 {
3395         struct cfq_data *cfqd = q->elevator->elevator_data;
3396         struct task_struct *tsk = current;
3397         struct cfq_io_context *cic;
3398         struct cfq_queue *cfqq;
3399
3400         /*
3401          * don't force setup of a queue from here, as a call to may_queue
3402          * does not necessarily imply that a request actually will be queued.
3403          * so just lookup a possibly existing queue, or return 'may queue'
3404          * if that fails
3405          */
3406         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3407         if (!cic)
3408                 return ELV_MQUEUE_MAY;
3409
3410         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3411         if (cfqq) {
3412                 cfq_init_prio_data(cfqq, cic->ioc);
3413                 cfq_prio_boost(cfqq);
3414
3415                 return __cfq_may_queue(cfqq);
3416         }
3417
3418         return ELV_MQUEUE_MAY;
3419 }
3420
3421 /*
3422  * queue lock held here
3423  */
3424 static void cfq_put_request(struct request *rq)
3425 {
3426         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3427
3428         if (cfqq) {
3429                 const int rw = rq_data_dir(rq);
3430
3431                 BUG_ON(!cfqq->allocated[rw]);
3432                 cfqq->allocated[rw]--;
3433
3434                 put_io_context(RQ_CIC(rq)->ioc);
3435
3436                 rq->elevator_private = NULL;
3437                 rq->elevator_private2 = NULL;
3438
3439                 cfq_put_queue(cfqq);
3440         }
3441 }
3442
3443 static struct cfq_queue *
3444 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3445                 struct cfq_queue *cfqq)
3446 {
3447         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3448         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3449         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3450         cfq_put_queue(cfqq);
3451         return cic_to_cfqq(cic, 1);
3452 }
3453
3454 /*
3455  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3456  * was the last process referring to said cfqq.
3457  */
3458 static struct cfq_queue *
3459 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3460 {
3461         if (cfqq_process_refs(cfqq) == 1) {
3462                 cfqq->pid = current->pid;
3463                 cfq_clear_cfqq_coop(cfqq);
3464                 cfq_clear_cfqq_split_coop(cfqq);
3465                 return cfqq;
3466         }
3467
3468         cic_set_cfqq(cic, NULL, 1);
3469         cfq_put_queue(cfqq);
3470         return NULL;
3471 }
3472 /*
3473  * Allocate cfq data structures associated with this request.
3474  */
3475 static int
3476 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3477 {
3478         struct cfq_data *cfqd = q->elevator->elevator_data;
3479         struct cfq_io_context *cic;
3480         const int rw = rq_data_dir(rq);
3481         const bool is_sync = rq_is_sync(rq);
3482         struct cfq_queue *cfqq;
3483         unsigned long flags;
3484
3485         might_sleep_if(gfp_mask & __GFP_WAIT);
3486
3487         cic = cfq_get_io_context(cfqd, gfp_mask);
3488
3489         spin_lock_irqsave(q->queue_lock, flags);
3490
3491         if (!cic)
3492                 goto queue_fail;
3493
3494 new_queue:
3495         cfqq = cic_to_cfqq(cic, is_sync);
3496         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3497                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3498                 cic_set_cfqq(cic, cfqq, is_sync);
3499         } else {
3500                 /*
3501                  * If the queue was seeky for too long, break it apart.
3502                  */
3503                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3504                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3505                         cfqq = split_cfqq(cic, cfqq);
3506                         if (!cfqq)
3507                                 goto new_queue;
3508                 }
3509
3510                 /*
3511                  * Check to see if this queue is scheduled to merge with
3512                  * another, closely cooperating queue.  The merging of
3513                  * queues happens here as it must be done in process context.
3514                  * The reference on new_cfqq was taken in merge_cfqqs.
3515                  */
3516                 if (cfqq->new_cfqq)
3517                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3518         }
3519
3520         cfqq->allocated[rw]++;
3521         atomic_inc(&cfqq->ref);
3522
3523         spin_unlock_irqrestore(q->queue_lock, flags);
3524
3525         rq->elevator_private = cic;
3526         rq->elevator_private2 = cfqq;
3527         return 0;
3528
3529 queue_fail:
3530         if (cic)
3531                 put_io_context(cic->ioc);
3532
3533         cfq_schedule_dispatch(cfqd);
3534         spin_unlock_irqrestore(q->queue_lock, flags);
3535         cfq_log(cfqd, "set_request fail");
3536         return 1;
3537 }
3538
3539 static void cfq_kick_queue(struct work_struct *work)
3540 {
3541         struct cfq_data *cfqd =
3542                 container_of(work, struct cfq_data, unplug_work);
3543         struct request_queue *q = cfqd->queue;
3544
3545         spin_lock_irq(q->queue_lock);
3546         __blk_run_queue(cfqd->queue);
3547         spin_unlock_irq(q->queue_lock);
3548 }
3549
3550 /*
3551  * Timer running if the active_queue is currently idling inside its time slice
3552  */
3553 static void cfq_idle_slice_timer(unsigned long data)
3554 {
3555         struct cfq_data *cfqd = (struct cfq_data *) data;
3556         struct cfq_queue *cfqq;
3557         unsigned long flags;
3558         int timed_out = 1;
3559
3560         cfq_log(cfqd, "idle timer fired");
3561
3562         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3563
3564         cfqq = cfqd->active_queue;
3565         if (cfqq) {
3566                 timed_out = 0;
3567
3568                 /*
3569                  * We saw a request before the queue expired, let it through
3570                  */
3571                 if (cfq_cfqq_must_dispatch(cfqq))
3572                         goto out_kick;
3573
3574                 /*
3575                  * expired
3576                  */
3577                 if (cfq_slice_used(cfqq))
3578                         goto expire;
3579
3580                 /*
3581                  * only expire and reinvoke request handler, if there are
3582                  * other queues with pending requests
3583                  */
3584                 if (!cfqd->busy_queues)
3585                         goto out_cont;
3586
3587                 /*
3588                  * not expired and it has a request pending, let it dispatch
3589                  */
3590                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3591                         goto out_kick;
3592
3593                 /*
3594                  * Queue depth flag is reset only when the idle didn't succeed
3595                  */
3596                 cfq_clear_cfqq_deep(cfqq);
3597         }
3598 expire:
3599         cfq_slice_expired(cfqd, timed_out);
3600 out_kick:
3601         cfq_schedule_dispatch(cfqd);
3602 out_cont:
3603         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3604 }
3605
3606 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3607 {
3608         del_timer_sync(&cfqd->idle_slice_timer);
3609         cancel_work_sync(&cfqd->unplug_work);
3610 }
3611
3612 static void cfq_put_async_queues(struct cfq_data *cfqd)
3613 {
3614         int i;
3615
3616         for (i = 0; i < IOPRIO_BE_NR; i++) {
3617                 if (cfqd->async_cfqq[0][i])
3618                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3619                 if (cfqd->async_cfqq[1][i])
3620                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3621         }
3622
3623         if (cfqd->async_idle_cfqq)
3624                 cfq_put_queue(cfqd->async_idle_cfqq);
3625 }
3626
3627 static void cfq_cfqd_free(struct rcu_head *head)
3628 {
3629         kfree(container_of(head, struct cfq_data, rcu));
3630 }
3631
3632 static void cfq_exit_queue(struct elevator_queue *e)
3633 {
3634         struct cfq_data *cfqd = e->elevator_data;
3635         struct request_queue *q = cfqd->queue;
3636
3637         cfq_shutdown_timer_wq(cfqd);
3638
3639         spin_lock_irq(q->queue_lock);
3640
3641         if (cfqd->active_queue)
3642                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3643
3644         while (!list_empty(&cfqd->cic_list)) {
3645                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3646                                                         struct cfq_io_context,
3647                                                         queue_list);
3648
3649                 __cfq_exit_single_io_context(cfqd, cic);
3650         }
3651
3652         cfq_put_async_queues(cfqd);
3653         cfq_release_cfq_groups(cfqd);
3654         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3655
3656         spin_unlock_irq(q->queue_lock);
3657
3658         cfq_shutdown_timer_wq(cfqd);
3659
3660         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3661         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3662 }
3663
3664 static void *cfq_init_queue(struct request_queue *q)
3665 {
3666         struct cfq_data *cfqd;
3667         int i, j;
3668         struct cfq_group *cfqg;
3669         struct cfq_rb_root *st;
3670
3671         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3672         if (!cfqd)
3673                 return NULL;
3674
3675         /* Init root service tree */
3676         cfqd->grp_service_tree = CFQ_RB_ROOT;
3677
3678         /* Init root group */
3679         cfqg = &cfqd->root_group;
3680         for_each_cfqg_st(cfqg, i, j, st)
3681                 *st = CFQ_RB_ROOT;
3682         RB_CLEAR_NODE(&cfqg->rb_node);
3683
3684         /* Give preference to root group over other groups */
3685         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3686
3687 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3688         /*
3689          * Take a reference to root group which we never drop. This is just
3690          * to make sure that cfq_put_cfqg() does not try to kfree root group
3691          */
3692         atomic_set(&cfqg->ref, 1);
3693         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3694                                         0);
3695 #endif
3696         /*
3697          * Not strictly needed (since RB_ROOT just clears the node and we
3698          * zeroed cfqd on alloc), but better be safe in case someone decides
3699          * to add magic to the rb code
3700          */
3701         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3702                 cfqd->prio_trees[i] = RB_ROOT;
3703
3704         /*
3705          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3706          * Grab a permanent reference to it, so that the normal code flow
3707          * will not attempt to free it.
3708          */
3709         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3710         atomic_inc(&cfqd->oom_cfqq.ref);
3711         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3712
3713         INIT_LIST_HEAD(&cfqd->cic_list);
3714
3715         cfqd->queue = q;
3716
3717         init_timer(&cfqd->idle_slice_timer);
3718         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3719         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3720
3721         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3722
3723         cfqd->cfq_quantum = cfq_quantum;
3724         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3725         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3726         cfqd->cfq_back_max = cfq_back_max;
3727         cfqd->cfq_back_penalty = cfq_back_penalty;
3728         cfqd->cfq_slice[0] = cfq_slice_async;
3729         cfqd->cfq_slice[1] = cfq_slice_sync;
3730         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3731         cfqd->cfq_slice_idle = cfq_slice_idle;
3732         cfqd->cfq_latency = 1;
3733         cfqd->cfq_group_isolation = 0;
3734         cfqd->hw_tag = -1;
3735         /*
3736          * we optimistically start assuming sync ops weren't delayed in last
3737          * second, in order to have larger depth for async operations.
3738          */
3739         cfqd->last_delayed_sync = jiffies - HZ;
3740         INIT_RCU_HEAD(&cfqd->rcu);
3741         return cfqd;
3742 }
3743
3744 static void cfq_slab_kill(void)
3745 {
3746         /*
3747          * Caller already ensured that pending RCU callbacks are completed,
3748          * so we should have no busy allocations at this point.
3749          */
3750         if (cfq_pool)
3751                 kmem_cache_destroy(cfq_pool);
3752         if (cfq_ioc_pool)
3753                 kmem_cache_destroy(cfq_ioc_pool);
3754 }
3755
3756 static int __init cfq_slab_setup(void)
3757 {
3758         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3759         if (!cfq_pool)
3760                 goto fail;
3761
3762         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3763         if (!cfq_ioc_pool)
3764                 goto fail;
3765
3766         return 0;
3767 fail:
3768         cfq_slab_kill();
3769         return -ENOMEM;
3770 }
3771
3772 /*
3773  * sysfs parts below -->
3774  */
3775 static ssize_t
3776 cfq_var_show(unsigned int var, char *page)
3777 {
3778         return sprintf(page, "%d\n", var);
3779 }
3780
3781 static ssize_t
3782 cfq_var_store(unsigned int *var, const char *page, size_t count)
3783 {
3784         char *p = (char *) page;
3785
3786         *var = simple_strtoul(p, &p, 10);
3787         return count;
3788 }
3789
3790 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3791 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3792 {                                                                       \
3793         struct cfq_data *cfqd = e->elevator_data;                       \
3794         unsigned int __data = __VAR;                                    \
3795         if (__CONV)                                                     \
3796                 __data = jiffies_to_msecs(__data);                      \
3797         return cfq_var_show(__data, (page));                            \
3798 }
3799 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3800 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3801 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3802 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3803 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3804 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3805 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3806 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3807 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3808 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3809 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3810 #undef SHOW_FUNCTION
3811
3812 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3813 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3814 {                                                                       \
3815         struct cfq_data *cfqd = e->elevator_data;                       \
3816         unsigned int __data;                                            \
3817         int ret = cfq_var_store(&__data, (page), count);                \
3818         if (__data < (MIN))                                             \
3819                 __data = (MIN);                                         \
3820         else if (__data > (MAX))                                        \
3821                 __data = (MAX);                                         \
3822         if (__CONV)                                                     \
3823                 *(__PTR) = msecs_to_jiffies(__data);                    \
3824         else                                                            \
3825                 *(__PTR) = __data;                                      \
3826         return ret;                                                     \
3827 }
3828 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3829 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3830                 UINT_MAX, 1);
3831 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3832                 UINT_MAX, 1);
3833 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3834 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3835                 UINT_MAX, 0);
3836 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3837 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3838 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3839 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3840                 UINT_MAX, 0);
3841 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3842 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3843 #undef STORE_FUNCTION
3844
3845 #define CFQ_ATTR(name) \
3846         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3847
3848 static struct elv_fs_entry cfq_attrs[] = {
3849         CFQ_ATTR(quantum),
3850         CFQ_ATTR(fifo_expire_sync),
3851         CFQ_ATTR(fifo_expire_async),
3852         CFQ_ATTR(back_seek_max),
3853         CFQ_ATTR(back_seek_penalty),
3854         CFQ_ATTR(slice_sync),
3855         CFQ_ATTR(slice_async),
3856         CFQ_ATTR(slice_async_rq),
3857         CFQ_ATTR(slice_idle),
3858         CFQ_ATTR(low_latency),
3859         CFQ_ATTR(group_isolation),
3860         __ATTR_NULL
3861 };
3862
3863 static struct elevator_type iosched_cfq = {
3864         .ops = {
3865                 .elevator_merge_fn =            cfq_merge,
3866                 .elevator_merged_fn =           cfq_merged_request,
3867                 .elevator_merge_req_fn =        cfq_merged_requests,
3868                 .elevator_allow_merge_fn =      cfq_allow_merge,
3869                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3870                 .elevator_add_req_fn =          cfq_insert_request,
3871                 .elevator_activate_req_fn =     cfq_activate_request,
3872                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3873                 .elevator_queue_empty_fn =      cfq_queue_empty,
3874                 .elevator_completed_req_fn =    cfq_completed_request,
3875                 .elevator_former_req_fn =       elv_rb_former_request,
3876                 .elevator_latter_req_fn =       elv_rb_latter_request,
3877                 .elevator_set_req_fn =          cfq_set_request,
3878                 .elevator_put_req_fn =          cfq_put_request,
3879                 .elevator_may_queue_fn =        cfq_may_queue,
3880                 .elevator_init_fn =             cfq_init_queue,
3881                 .elevator_exit_fn =             cfq_exit_queue,
3882                 .trim =                         cfq_free_io_context,
3883         },
3884         .elevator_attrs =       cfq_attrs,
3885         .elevator_name =        "cfq",
3886         .elevator_owner =       THIS_MODULE,
3887 };
3888
3889 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3890 static struct blkio_policy_type blkio_policy_cfq = {
3891         .ops = {
3892                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3893                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3894         },
3895 };
3896 #else
3897 static struct blkio_policy_type blkio_policy_cfq;
3898 #endif
3899
3900 static int __init cfq_init(void)
3901 {
3902         /*
3903          * could be 0 on HZ < 1000 setups
3904          */
3905         if (!cfq_slice_async)
3906                 cfq_slice_async = 1;
3907         if (!cfq_slice_idle)
3908                 cfq_slice_idle = 1;
3909
3910         if (cfq_slab_setup())
3911                 return -ENOMEM;
3912
3913         elv_register(&iosched_cfq);
3914         blkio_policy_register(&blkio_policy_cfq);
3915
3916         return 0;
3917 }
3918
3919 static void __exit cfq_exit(void)
3920 {
3921         DECLARE_COMPLETION_ONSTACK(all_gone);
3922         blkio_policy_unregister(&blkio_policy_cfq);
3923         elv_unregister(&iosched_cfq);
3924         ioc_gone = &all_gone;
3925         /* ioc_gone's update must be visible before reading ioc_count */
3926         smp_wmb();
3927
3928         /*
3929          * this also protects us from entering cfq_slab_kill() with
3930          * pending RCU callbacks
3931          */
3932         if (elv_ioc_count_read(cfq_ioc_count))
3933                 wait_for_completion(&all_gone);
3934         cfq_slab_kill();
3935 }
3936
3937 module_init(cfq_init);
3938 module_exit(cfq_exit);
3939
3940 MODULE_AUTHOR("Jens Axboe");
3941 MODULE_LICENSE("GPL");
3942 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");