cfq-iosched: rethink seeky detection for SSDs
[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 = 4;
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_SECT_THR_NONROT    (sector_t)(2 * 32)
51 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
52
53 #define RQ_CIC(rq)              \
54         ((struct cfq_io_context *) (rq)->elevator_private)
55 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
56
57 static struct kmem_cache *cfq_pool;
58 static struct kmem_cache *cfq_ioc_pool;
59
60 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
61 static struct completion *ioc_gone;
62 static DEFINE_SPINLOCK(ioc_gone_lock);
63
64 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67
68 #define sample_valid(samples)   ((samples) > 80)
69 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
70
71 /*
72  * Most of our rbtree usage is for sorting with min extraction, so
73  * if we cache the leftmost node we don't have to walk down the tree
74  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
75  * move this into the elevator for the rq sorting as well.
76  */
77 struct cfq_rb_root {
78         struct rb_root rb;
79         struct rb_node *left;
80         unsigned count;
81         u64 min_vdisktime;
82         struct rb_node *active;
83         unsigned total_weight;
84 };
85 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
86
87 /*
88  * Per process-grouping structure
89  */
90 struct cfq_queue {
91         /* reference count */
92         atomic_t ref;
93         /* various state flags, see below */
94         unsigned int flags;
95         /* parent cfq_data */
96         struct cfq_data *cfqd;
97         /* service_tree member */
98         struct rb_node rb_node;
99         /* service_tree key */
100         unsigned long rb_key;
101         /* prio tree member */
102         struct rb_node p_node;
103         /* prio tree root we belong to, if any */
104         struct rb_root *p_root;
105         /* sorted list of pending requests */
106         struct rb_root sort_list;
107         /* if fifo isn't expired, next request to serve */
108         struct request *next_rq;
109         /* requests queued in sort_list */
110         int queued[2];
111         /* currently allocated requests */
112         int allocated[2];
113         /* fifo list of requests in sort_list */
114         struct list_head fifo;
115
116         /* time when queue got scheduled in to dispatch first request. */
117         unsigned long dispatch_start;
118         unsigned int allocated_slice;
119         unsigned int slice_dispatch;
120         /* time when first request from queue completed and slice started. */
121         unsigned long slice_start;
122         unsigned long slice_end;
123         long slice_resid;
124
125         /* pending metadata requests */
126         int meta_pending;
127         /* number of requests that are on the dispatch list or inside driver */
128         int dispatched;
129
130         /* io prio of this group */
131         unsigned short ioprio, org_ioprio;
132         unsigned short ioprio_class, org_ioprio_class;
133
134         pid_t pid;
135
136         u32 seek_history;
137         sector_t last_request_pos;
138
139         struct cfq_rb_root *service_tree;
140         struct cfq_queue *new_cfqq;
141         struct cfq_group *cfqg;
142         struct cfq_group *orig_cfqg;
143         /* Sectors dispatched in current dispatch round */
144         unsigned long nr_sectors;
145 };
146
147 /*
148  * First index in the service_trees.
149  * IDLE is handled separately, so it has negative index
150  */
151 enum wl_prio_t {
152         BE_WORKLOAD = 0,
153         RT_WORKLOAD = 1,
154         IDLE_WORKLOAD = 2,
155 };
156
157 /*
158  * Second index in the service_trees.
159  */
160 enum wl_type_t {
161         ASYNC_WORKLOAD = 0,
162         SYNC_NOIDLE_WORKLOAD = 1,
163         SYNC_WORKLOAD = 2
164 };
165
166 /* This is per cgroup per device grouping structure */
167 struct cfq_group {
168         /* group service_tree member */
169         struct rb_node rb_node;
170
171         /* group service_tree key */
172         u64 vdisktime;
173         unsigned int weight;
174         bool on_st;
175
176         /* number of cfqq currently on this group */
177         int nr_cfqq;
178
179         /* Per group busy queus average. Useful for workload slice calc. */
180         unsigned int busy_queues_avg[2];
181         /*
182          * rr lists of queues with requests, onle rr for each priority class.
183          * Counts are embedded in the cfq_rb_root
184          */
185         struct cfq_rb_root service_trees[2][3];
186         struct cfq_rb_root service_tree_idle;
187
188         unsigned long saved_workload_slice;
189         enum wl_type_t saved_workload;
190         enum wl_prio_t saved_serving_prio;
191         struct blkio_group blkg;
192 #ifdef CONFIG_CFQ_GROUP_IOSCHED
193         struct hlist_node cfqd_node;
194         atomic_t ref;
195 #endif
196 };
197
198 /*
199  * Per block device queue structure
200  */
201 struct cfq_data {
202         struct request_queue *queue;
203         /* Root service tree for cfq_groups */
204         struct cfq_rb_root grp_service_tree;
205         struct cfq_group root_group;
206
207         /*
208          * The priority currently being served
209          */
210         enum wl_prio_t serving_prio;
211         enum wl_type_t serving_type;
212         unsigned long workload_expires;
213         struct cfq_group *serving_group;
214         bool noidle_tree_requires_idle;
215
216         /*
217          * Each priority tree is sorted by next_request position.  These
218          * trees are used when determining if two or more queues are
219          * interleaving requests (see cfq_close_cooperator).
220          */
221         struct rb_root prio_trees[CFQ_PRIO_LISTS];
222
223         unsigned int busy_queues;
224
225         int rq_in_driver[2];
226         int sync_flight;
227
228         /*
229          * queue-depth detection
230          */
231         int rq_queued;
232         int hw_tag;
233         /*
234          * hw_tag can be
235          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
236          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
237          *  0 => no NCQ
238          */
239         int hw_tag_est_depth;
240         unsigned int hw_tag_samples;
241
242         /*
243          * idle window management
244          */
245         struct timer_list idle_slice_timer;
246         struct work_struct unplug_work;
247
248         struct cfq_queue *active_queue;
249         struct cfq_io_context *active_cic;
250
251         /*
252          * async queue for each priority case
253          */
254         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
255         struct cfq_queue *async_idle_cfqq;
256
257         sector_t last_position;
258
259         /*
260          * tunables, see top of file
261          */
262         unsigned int cfq_quantum;
263         unsigned int cfq_fifo_expire[2];
264         unsigned int cfq_back_penalty;
265         unsigned int cfq_back_max;
266         unsigned int cfq_slice[2];
267         unsigned int cfq_slice_async_rq;
268         unsigned int cfq_slice_idle;
269         unsigned int cfq_latency;
270         unsigned int cfq_group_isolation;
271
272         struct list_head cic_list;
273
274         /*
275          * Fallback dummy cfqq for extreme OOM conditions
276          */
277         struct cfq_queue oom_cfqq;
278
279         unsigned long last_delayed_sync;
280
281         /* List of cfq groups being managed on this device*/
282         struct hlist_head cfqg_list;
283         struct rcu_head rcu;
284 };
285
286 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
287
288 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
289                                             enum wl_prio_t prio,
290                                             enum wl_type_t type)
291 {
292         if (!cfqg)
293                 return NULL;
294
295         if (prio == IDLE_WORKLOAD)
296                 return &cfqg->service_tree_idle;
297
298         return &cfqg->service_trees[prio][type];
299 }
300
301 enum cfqq_state_flags {
302         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
303         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
304         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
305         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
306         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
307         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
308         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
309         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
310         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
311         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
312         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
313         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
314         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
315 };
316
317 #define CFQ_CFQQ_FNS(name)                                              \
318 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
319 {                                                                       \
320         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
321 }                                                                       \
322 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
323 {                                                                       \
324         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
325 }                                                                       \
326 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
327 {                                                                       \
328         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
329 }
330
331 CFQ_CFQQ_FNS(on_rr);
332 CFQ_CFQQ_FNS(wait_request);
333 CFQ_CFQQ_FNS(must_dispatch);
334 CFQ_CFQQ_FNS(must_alloc_slice);
335 CFQ_CFQQ_FNS(fifo_expire);
336 CFQ_CFQQ_FNS(idle_window);
337 CFQ_CFQQ_FNS(prio_changed);
338 CFQ_CFQQ_FNS(slice_new);
339 CFQ_CFQQ_FNS(sync);
340 CFQ_CFQQ_FNS(coop);
341 CFQ_CFQQ_FNS(split_coop);
342 CFQ_CFQQ_FNS(deep);
343 CFQ_CFQQ_FNS(wait_busy);
344 #undef CFQ_CFQQ_FNS
345
346 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
347 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
348         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
349                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
350                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
351
352 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
353         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
354                                 blkg_path(&(cfqg)->blkg), ##args);      \
355
356 #else
357 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
358         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
359 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
360 #endif
361 #define cfq_log(cfqd, fmt, args...)     \
362         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
363
364 /* Traverses through cfq group service trees */
365 #define for_each_cfqg_st(cfqg, i, j, st) \
366         for (i = 0; i <= IDLE_WORKLOAD; i++) \
367                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
368                         : &cfqg->service_tree_idle; \
369                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
370                         (i == IDLE_WORKLOAD && j == 0); \
371                         j++, st = i < IDLE_WORKLOAD ? \
372                         &cfqg->service_trees[i][j]: NULL) \
373
374
375 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
376 {
377         if (cfq_class_idle(cfqq))
378                 return IDLE_WORKLOAD;
379         if (cfq_class_rt(cfqq))
380                 return RT_WORKLOAD;
381         return BE_WORKLOAD;
382 }
383
384
385 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
386 {
387         if (!cfq_cfqq_sync(cfqq))
388                 return ASYNC_WORKLOAD;
389         if (!cfq_cfqq_idle_window(cfqq))
390                 return SYNC_NOIDLE_WORKLOAD;
391         return SYNC_WORKLOAD;
392 }
393
394 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
395                                         struct cfq_data *cfqd,
396                                         struct cfq_group *cfqg)
397 {
398         if (wl == IDLE_WORKLOAD)
399                 return cfqg->service_tree_idle.count;
400
401         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
402                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
403                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
404 }
405
406 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
407                                         struct cfq_group *cfqg)
408 {
409         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
410                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
411 }
412
413 static void cfq_dispatch_insert(struct request_queue *, struct request *);
414 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
415                                        struct io_context *, gfp_t);
416 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
417                                                 struct io_context *);
418
419 static inline int rq_in_driver(struct cfq_data *cfqd)
420 {
421         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
422 }
423
424 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
425                                             bool is_sync)
426 {
427         return cic->cfqq[is_sync];
428 }
429
430 static inline void cic_set_cfqq(struct cfq_io_context *cic,
431                                 struct cfq_queue *cfqq, bool is_sync)
432 {
433         cic->cfqq[is_sync] = cfqq;
434 }
435
436 /*
437  * We regard a request as SYNC, if it's either a read or has the SYNC bit
438  * set (in which case it could also be direct WRITE).
439  */
440 static inline bool cfq_bio_sync(struct bio *bio)
441 {
442         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
443 }
444
445 /*
446  * scheduler run of queue, if there are requests pending and no one in the
447  * driver that will restart queueing
448  */
449 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
450 {
451         if (cfqd->busy_queues) {
452                 cfq_log(cfqd, "schedule dispatch");
453                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
454         }
455 }
456
457 static int cfq_queue_empty(struct request_queue *q)
458 {
459         struct cfq_data *cfqd = q->elevator->elevator_data;
460
461         return !cfqd->rq_queued;
462 }
463
464 /*
465  * Scale schedule slice based on io priority. Use the sync time slice only
466  * if a queue is marked sync and has sync io queued. A sync queue with async
467  * io only, should not get full sync slice length.
468  */
469 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
470                                  unsigned short prio)
471 {
472         const int base_slice = cfqd->cfq_slice[sync];
473
474         WARN_ON(prio >= IOPRIO_BE_NR);
475
476         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
477 }
478
479 static inline int
480 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
481 {
482         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
483 }
484
485 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
486 {
487         u64 d = delta << CFQ_SERVICE_SHIFT;
488
489         d = d * BLKIO_WEIGHT_DEFAULT;
490         do_div(d, cfqg->weight);
491         return d;
492 }
493
494 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
495 {
496         s64 delta = (s64)(vdisktime - min_vdisktime);
497         if (delta > 0)
498                 min_vdisktime = vdisktime;
499
500         return min_vdisktime;
501 }
502
503 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
504 {
505         s64 delta = (s64)(vdisktime - min_vdisktime);
506         if (delta < 0)
507                 min_vdisktime = vdisktime;
508
509         return min_vdisktime;
510 }
511
512 static void update_min_vdisktime(struct cfq_rb_root *st)
513 {
514         u64 vdisktime = st->min_vdisktime;
515         struct cfq_group *cfqg;
516
517         if (st->active) {
518                 cfqg = rb_entry_cfqg(st->active);
519                 vdisktime = cfqg->vdisktime;
520         }
521
522         if (st->left) {
523                 cfqg = rb_entry_cfqg(st->left);
524                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
525         }
526
527         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
528 }
529
530 /*
531  * get averaged number of queues of RT/BE priority.
532  * average is updated, with a formula that gives more weight to higher numbers,
533  * to quickly follows sudden increases and decrease slowly
534  */
535
536 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
537                                         struct cfq_group *cfqg, bool rt)
538 {
539         unsigned min_q, max_q;
540         unsigned mult  = cfq_hist_divisor - 1;
541         unsigned round = cfq_hist_divisor / 2;
542         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
543
544         min_q = min(cfqg->busy_queues_avg[rt], busy);
545         max_q = max(cfqg->busy_queues_avg[rt], busy);
546         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
547                 cfq_hist_divisor;
548         return cfqg->busy_queues_avg[rt];
549 }
550
551 static inline unsigned
552 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
553 {
554         struct cfq_rb_root *st = &cfqd->grp_service_tree;
555
556         return cfq_target_latency * cfqg->weight / st->total_weight;
557 }
558
559 static inline void
560 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
561 {
562         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
563         if (cfqd->cfq_latency) {
564                 /*
565                  * interested queues (we consider only the ones with the same
566                  * priority class in the cfq group)
567                  */
568                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
569                                                 cfq_class_rt(cfqq));
570                 unsigned sync_slice = cfqd->cfq_slice[1];
571                 unsigned expect_latency = sync_slice * iq;
572                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
573
574                 if (expect_latency > group_slice) {
575                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
576                         /* scale low_slice according to IO priority
577                          * and sync vs async */
578                         unsigned low_slice =
579                                 min(slice, base_low_slice * slice / sync_slice);
580                         /* the adapted slice value is scaled to fit all iqs
581                          * into the target latency */
582                         slice = max(slice * group_slice / expect_latency,
583                                     low_slice);
584                 }
585         }
586         cfqq->slice_start = jiffies;
587         cfqq->slice_end = jiffies + slice;
588         cfqq->allocated_slice = slice;
589         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
590 }
591
592 /*
593  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
594  * isn't valid until the first request from the dispatch is activated
595  * and the slice time set.
596  */
597 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
598 {
599         if (cfq_cfqq_slice_new(cfqq))
600                 return 0;
601         if (time_before(jiffies, cfqq->slice_end))
602                 return 0;
603
604         return 1;
605 }
606
607 /*
608  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
609  * We choose the request that is closest to the head right now. Distance
610  * behind the head is penalized and only allowed to a certain extent.
611  */
612 static struct request *
613 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
614 {
615         sector_t s1, s2, d1 = 0, d2 = 0;
616         unsigned long back_max;
617 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
618 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
619         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
620
621         if (rq1 == NULL || rq1 == rq2)
622                 return rq2;
623         if (rq2 == NULL)
624                 return rq1;
625
626         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
627                 return rq1;
628         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
629                 return rq2;
630         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
631                 return rq1;
632         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
633                 return rq2;
634
635         s1 = blk_rq_pos(rq1);
636         s2 = blk_rq_pos(rq2);
637
638         /*
639          * by definition, 1KiB is 2 sectors
640          */
641         back_max = cfqd->cfq_back_max * 2;
642
643         /*
644          * Strict one way elevator _except_ in the case where we allow
645          * short backward seeks which are biased as twice the cost of a
646          * similar forward seek.
647          */
648         if (s1 >= last)
649                 d1 = s1 - last;
650         else if (s1 + back_max >= last)
651                 d1 = (last - s1) * cfqd->cfq_back_penalty;
652         else
653                 wrap |= CFQ_RQ1_WRAP;
654
655         if (s2 >= last)
656                 d2 = s2 - last;
657         else if (s2 + back_max >= last)
658                 d2 = (last - s2) * cfqd->cfq_back_penalty;
659         else
660                 wrap |= CFQ_RQ2_WRAP;
661
662         /* Found required data */
663
664         /*
665          * By doing switch() on the bit mask "wrap" we avoid having to
666          * check two variables for all permutations: --> faster!
667          */
668         switch (wrap) {
669         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
670                 if (d1 < d2)
671                         return rq1;
672                 else if (d2 < d1)
673                         return rq2;
674                 else {
675                         if (s1 >= s2)
676                                 return rq1;
677                         else
678                                 return rq2;
679                 }
680
681         case CFQ_RQ2_WRAP:
682                 return rq1;
683         case CFQ_RQ1_WRAP:
684                 return rq2;
685         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
686         default:
687                 /*
688                  * Since both rqs are wrapped,
689                  * start with the one that's further behind head
690                  * (--> only *one* back seek required),
691                  * since back seek takes more time than forward.
692                  */
693                 if (s1 <= s2)
694                         return rq1;
695                 else
696                         return rq2;
697         }
698 }
699
700 /*
701  * The below is leftmost cache rbtree addon
702  */
703 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
704 {
705         /* Service tree is empty */
706         if (!root->count)
707                 return NULL;
708
709         if (!root->left)
710                 root->left = rb_first(&root->rb);
711
712         if (root->left)
713                 return rb_entry(root->left, struct cfq_queue, rb_node);
714
715         return NULL;
716 }
717
718 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
719 {
720         if (!root->left)
721                 root->left = rb_first(&root->rb);
722
723         if (root->left)
724                 return rb_entry_cfqg(root->left);
725
726         return NULL;
727 }
728
729 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
730 {
731         rb_erase(n, root);
732         RB_CLEAR_NODE(n);
733 }
734
735 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
736 {
737         if (root->left == n)
738                 root->left = NULL;
739         rb_erase_init(n, &root->rb);
740         --root->count;
741 }
742
743 /*
744  * would be nice to take fifo expire time into account as well
745  */
746 static struct request *
747 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
748                   struct request *last)
749 {
750         struct rb_node *rbnext = rb_next(&last->rb_node);
751         struct rb_node *rbprev = rb_prev(&last->rb_node);
752         struct request *next = NULL, *prev = NULL;
753
754         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
755
756         if (rbprev)
757                 prev = rb_entry_rq(rbprev);
758
759         if (rbnext)
760                 next = rb_entry_rq(rbnext);
761         else {
762                 rbnext = rb_first(&cfqq->sort_list);
763                 if (rbnext && rbnext != &last->rb_node)
764                         next = rb_entry_rq(rbnext);
765         }
766
767         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
768 }
769
770 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
771                                       struct cfq_queue *cfqq)
772 {
773         /*
774          * just an approximation, should be ok.
775          */
776         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
777                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
778 }
779
780 static inline s64
781 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
782 {
783         return cfqg->vdisktime - st->min_vdisktime;
784 }
785
786 static void
787 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
788 {
789         struct rb_node **node = &st->rb.rb_node;
790         struct rb_node *parent = NULL;
791         struct cfq_group *__cfqg;
792         s64 key = cfqg_key(st, cfqg);
793         int left = 1;
794
795         while (*node != NULL) {
796                 parent = *node;
797                 __cfqg = rb_entry_cfqg(parent);
798
799                 if (key < cfqg_key(st, __cfqg))
800                         node = &parent->rb_left;
801                 else {
802                         node = &parent->rb_right;
803                         left = 0;
804                 }
805         }
806
807         if (left)
808                 st->left = &cfqg->rb_node;
809
810         rb_link_node(&cfqg->rb_node, parent, node);
811         rb_insert_color(&cfqg->rb_node, &st->rb);
812 }
813
814 static void
815 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
816 {
817         struct cfq_rb_root *st = &cfqd->grp_service_tree;
818         struct cfq_group *__cfqg;
819         struct rb_node *n;
820
821         cfqg->nr_cfqq++;
822         if (cfqg->on_st)
823                 return;
824
825         /*
826          * Currently put the group at the end. Later implement something
827          * so that groups get lesser vtime based on their weights, so that
828          * if group does not loose all if it was not continously backlogged.
829          */
830         n = rb_last(&st->rb);
831         if (n) {
832                 __cfqg = rb_entry_cfqg(n);
833                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
834         } else
835                 cfqg->vdisktime = st->min_vdisktime;
836
837         __cfq_group_service_tree_add(st, cfqg);
838         cfqg->on_st = true;
839         st->total_weight += cfqg->weight;
840 }
841
842 static void
843 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
844 {
845         struct cfq_rb_root *st = &cfqd->grp_service_tree;
846
847         if (st->active == &cfqg->rb_node)
848                 st->active = NULL;
849
850         BUG_ON(cfqg->nr_cfqq < 1);
851         cfqg->nr_cfqq--;
852
853         /* If there are other cfq queues under this group, don't delete it */
854         if (cfqg->nr_cfqq)
855                 return;
856
857         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
858         cfqg->on_st = false;
859         st->total_weight -= cfqg->weight;
860         if (!RB_EMPTY_NODE(&cfqg->rb_node))
861                 cfq_rb_erase(&cfqg->rb_node, st);
862         cfqg->saved_workload_slice = 0;
863         blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
864 }
865
866 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
867 {
868         unsigned int slice_used;
869
870         /*
871          * Queue got expired before even a single request completed or
872          * got expired immediately after first request completion.
873          */
874         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
875                 /*
876                  * Also charge the seek time incurred to the group, otherwise
877                  * if there are mutiple queues in the group, each can dispatch
878                  * a single request on seeky media and cause lots of seek time
879                  * and group will never know it.
880                  */
881                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
882                                         1);
883         } else {
884                 slice_used = jiffies - cfqq->slice_start;
885                 if (slice_used > cfqq->allocated_slice)
886                         slice_used = cfqq->allocated_slice;
887         }
888
889         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
890                                 cfqq->nr_sectors);
891         return slice_used;
892 }
893
894 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
895                                 struct cfq_queue *cfqq)
896 {
897         struct cfq_rb_root *st = &cfqd->grp_service_tree;
898         unsigned int used_sl, charge_sl;
899         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
900                         - cfqg->service_tree_idle.count;
901
902         BUG_ON(nr_sync < 0);
903         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
904
905         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
906                 charge_sl = cfqq->allocated_slice;
907
908         /* Can't update vdisktime while group is on service tree */
909         cfq_rb_erase(&cfqg->rb_node, st);
910         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
911         __cfq_group_service_tree_add(st, cfqg);
912
913         /* This group is being expired. Save the context */
914         if (time_after(cfqd->workload_expires, jiffies)) {
915                 cfqg->saved_workload_slice = cfqd->workload_expires
916                                                 - jiffies;
917                 cfqg->saved_workload = cfqd->serving_type;
918                 cfqg->saved_serving_prio = cfqd->serving_prio;
919         } else
920                 cfqg->saved_workload_slice = 0;
921
922         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
923                                         st->min_vdisktime);
924         blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
925                                                 cfqq->nr_sectors);
926 }
927
928 #ifdef CONFIG_CFQ_GROUP_IOSCHED
929 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
930 {
931         if (blkg)
932                 return container_of(blkg, struct cfq_group, blkg);
933         return NULL;
934 }
935
936 void
937 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
938 {
939         cfqg_of_blkg(blkg)->weight = weight;
940 }
941
942 static struct cfq_group *
943 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
944 {
945         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
946         struct cfq_group *cfqg = NULL;
947         void *key = cfqd;
948         int i, j;
949         struct cfq_rb_root *st;
950         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
951         unsigned int major, minor;
952
953         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
954         if (cfqg || !create)
955                 goto done;
956
957         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
958         if (!cfqg)
959                 goto done;
960
961         cfqg->weight = blkcg->weight;
962         for_each_cfqg_st(cfqg, i, j, st)
963                 *st = CFQ_RB_ROOT;
964         RB_CLEAR_NODE(&cfqg->rb_node);
965
966         /*
967          * Take the initial reference that will be released on destroy
968          * This can be thought of a joint reference by cgroup and
969          * elevator which will be dropped by either elevator exit
970          * or cgroup deletion path depending on who is exiting first.
971          */
972         atomic_set(&cfqg->ref, 1);
973
974         /* Add group onto cgroup list */
975         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
976         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
977                                         MKDEV(major, minor));
978
979         /* Add group on cfqd list */
980         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
981
982 done:
983         return cfqg;
984 }
985
986 /*
987  * Search for the cfq group current task belongs to. If create = 1, then also
988  * create the cfq group if it does not exist. request_queue lock must be held.
989  */
990 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
991 {
992         struct cgroup *cgroup;
993         struct cfq_group *cfqg = NULL;
994
995         rcu_read_lock();
996         cgroup = task_cgroup(current, blkio_subsys_id);
997         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
998         if (!cfqg && create)
999                 cfqg = &cfqd->root_group;
1000         rcu_read_unlock();
1001         return cfqg;
1002 }
1003
1004 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1005 {
1006         /* Currently, all async queues are mapped to root group */
1007         if (!cfq_cfqq_sync(cfqq))
1008                 cfqg = &cfqq->cfqd->root_group;
1009
1010         cfqq->cfqg = cfqg;
1011         /* cfqq reference on cfqg */
1012         atomic_inc(&cfqq->cfqg->ref);
1013 }
1014
1015 static void cfq_put_cfqg(struct cfq_group *cfqg)
1016 {
1017         struct cfq_rb_root *st;
1018         int i, j;
1019
1020         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1021         if (!atomic_dec_and_test(&cfqg->ref))
1022                 return;
1023         for_each_cfqg_st(cfqg, i, j, st)
1024                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1025         kfree(cfqg);
1026 }
1027
1028 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1029 {
1030         /* Something wrong if we are trying to remove same group twice */
1031         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1032
1033         hlist_del_init(&cfqg->cfqd_node);
1034
1035         /*
1036          * Put the reference taken at the time of creation so that when all
1037          * queues are gone, group can be destroyed.
1038          */
1039         cfq_put_cfqg(cfqg);
1040 }
1041
1042 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1043 {
1044         struct hlist_node *pos, *n;
1045         struct cfq_group *cfqg;
1046
1047         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1048                 /*
1049                  * If cgroup removal path got to blk_group first and removed
1050                  * it from cgroup list, then it will take care of destroying
1051                  * cfqg also.
1052                  */
1053                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1054                         cfq_destroy_cfqg(cfqd, cfqg);
1055         }
1056 }
1057
1058 /*
1059  * Blk cgroup controller notification saying that blkio_group object is being
1060  * delinked as associated cgroup object is going away. That also means that
1061  * no new IO will come in this group. So get rid of this group as soon as
1062  * any pending IO in the group is finished.
1063  *
1064  * This function is called under rcu_read_lock(). key is the rcu protected
1065  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1066  * read lock.
1067  *
1068  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1069  * it should not be NULL as even if elevator was exiting, cgroup deltion
1070  * path got to it first.
1071  */
1072 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1073 {
1074         unsigned long  flags;
1075         struct cfq_data *cfqd = key;
1076
1077         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1078         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1079         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1080 }
1081
1082 #else /* GROUP_IOSCHED */
1083 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1084 {
1085         return &cfqd->root_group;
1086 }
1087 static inline void
1088 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1089         cfqq->cfqg = cfqg;
1090 }
1091
1092 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1093 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1094
1095 #endif /* GROUP_IOSCHED */
1096
1097 /*
1098  * The cfqd->service_trees holds all pending cfq_queue's that have
1099  * requests waiting to be processed. It is sorted in the order that
1100  * we will service the queues.
1101  */
1102 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1103                                  bool add_front)
1104 {
1105         struct rb_node **p, *parent;
1106         struct cfq_queue *__cfqq;
1107         unsigned long rb_key;
1108         struct cfq_rb_root *service_tree;
1109         int left;
1110         int new_cfqq = 1;
1111         int group_changed = 0;
1112
1113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1114         if (!cfqd->cfq_group_isolation
1115             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1116             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1117                 /* Move this cfq to root group */
1118                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1119                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1120                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1121                 cfqq->orig_cfqg = cfqq->cfqg;
1122                 cfqq->cfqg = &cfqd->root_group;
1123                 atomic_inc(&cfqd->root_group.ref);
1124                 group_changed = 1;
1125         } else if (!cfqd->cfq_group_isolation
1126                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1127                 /* cfqq is sequential now needs to go to its original group */
1128                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1129                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1130                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1131                 cfq_put_cfqg(cfqq->cfqg);
1132                 cfqq->cfqg = cfqq->orig_cfqg;
1133                 cfqq->orig_cfqg = NULL;
1134                 group_changed = 1;
1135                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1136         }
1137 #endif
1138
1139         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1140                                                 cfqq_type(cfqq));
1141         if (cfq_class_idle(cfqq)) {
1142                 rb_key = CFQ_IDLE_DELAY;
1143                 parent = rb_last(&service_tree->rb);
1144                 if (parent && parent != &cfqq->rb_node) {
1145                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1146                         rb_key += __cfqq->rb_key;
1147                 } else
1148                         rb_key += jiffies;
1149         } else if (!add_front) {
1150                 /*
1151                  * Get our rb key offset. Subtract any residual slice
1152                  * value carried from last service. A negative resid
1153                  * count indicates slice overrun, and this should position
1154                  * the next service time further away in the tree.
1155                  */
1156                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1157                 rb_key -= cfqq->slice_resid;
1158                 cfqq->slice_resid = 0;
1159         } else {
1160                 rb_key = -HZ;
1161                 __cfqq = cfq_rb_first(service_tree);
1162                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1163         }
1164
1165         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1166                 new_cfqq = 0;
1167                 /*
1168                  * same position, nothing more to do
1169                  */
1170                 if (rb_key == cfqq->rb_key &&
1171                     cfqq->service_tree == service_tree)
1172                         return;
1173
1174                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1175                 cfqq->service_tree = NULL;
1176         }
1177
1178         left = 1;
1179         parent = NULL;
1180         cfqq->service_tree = service_tree;
1181         p = &service_tree->rb.rb_node;
1182         while (*p) {
1183                 struct rb_node **n;
1184
1185                 parent = *p;
1186                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1187
1188                 /*
1189                  * sort by key, that represents service time.
1190                  */
1191                 if (time_before(rb_key, __cfqq->rb_key))
1192                         n = &(*p)->rb_left;
1193                 else {
1194                         n = &(*p)->rb_right;
1195                         left = 0;
1196                 }
1197
1198                 p = n;
1199         }
1200
1201         if (left)
1202                 service_tree->left = &cfqq->rb_node;
1203
1204         cfqq->rb_key = rb_key;
1205         rb_link_node(&cfqq->rb_node, parent, p);
1206         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1207         service_tree->count++;
1208         if ((add_front || !new_cfqq) && !group_changed)
1209                 return;
1210         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1211 }
1212
1213 static struct cfq_queue *
1214 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1215                      sector_t sector, struct rb_node **ret_parent,
1216                      struct rb_node ***rb_link)
1217 {
1218         struct rb_node **p, *parent;
1219         struct cfq_queue *cfqq = NULL;
1220
1221         parent = NULL;
1222         p = &root->rb_node;
1223         while (*p) {
1224                 struct rb_node **n;
1225
1226                 parent = *p;
1227                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1228
1229                 /*
1230                  * Sort strictly based on sector.  Smallest to the left,
1231                  * largest to the right.
1232                  */
1233                 if (sector > blk_rq_pos(cfqq->next_rq))
1234                         n = &(*p)->rb_right;
1235                 else if (sector < blk_rq_pos(cfqq->next_rq))
1236                         n = &(*p)->rb_left;
1237                 else
1238                         break;
1239                 p = n;
1240                 cfqq = NULL;
1241         }
1242
1243         *ret_parent = parent;
1244         if (rb_link)
1245                 *rb_link = p;
1246         return cfqq;
1247 }
1248
1249 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1250 {
1251         struct rb_node **p, *parent;
1252         struct cfq_queue *__cfqq;
1253
1254         if (cfqq->p_root) {
1255                 rb_erase(&cfqq->p_node, cfqq->p_root);
1256                 cfqq->p_root = NULL;
1257         }
1258
1259         if (cfq_class_idle(cfqq))
1260                 return;
1261         if (!cfqq->next_rq)
1262                 return;
1263
1264         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1265         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1266                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1267         if (!__cfqq) {
1268                 rb_link_node(&cfqq->p_node, parent, p);
1269                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1270         } else
1271                 cfqq->p_root = NULL;
1272 }
1273
1274 /*
1275  * Update cfqq's position in the service tree.
1276  */
1277 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1278 {
1279         /*
1280          * Resorting requires the cfqq to be on the RR list already.
1281          */
1282         if (cfq_cfqq_on_rr(cfqq)) {
1283                 cfq_service_tree_add(cfqd, cfqq, 0);
1284                 cfq_prio_tree_add(cfqd, cfqq);
1285         }
1286 }
1287
1288 /*
1289  * add to busy list of queues for service, trying to be fair in ordering
1290  * the pending list according to last request service
1291  */
1292 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1293 {
1294         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1295         BUG_ON(cfq_cfqq_on_rr(cfqq));
1296         cfq_mark_cfqq_on_rr(cfqq);
1297         cfqd->busy_queues++;
1298
1299         cfq_resort_rr_list(cfqd, cfqq);
1300 }
1301
1302 /*
1303  * Called when the cfqq no longer has requests pending, remove it from
1304  * the service tree.
1305  */
1306 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1307 {
1308         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1309         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1310         cfq_clear_cfqq_on_rr(cfqq);
1311
1312         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1313                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1314                 cfqq->service_tree = NULL;
1315         }
1316         if (cfqq->p_root) {
1317                 rb_erase(&cfqq->p_node, cfqq->p_root);
1318                 cfqq->p_root = NULL;
1319         }
1320
1321         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1322         BUG_ON(!cfqd->busy_queues);
1323         cfqd->busy_queues--;
1324 }
1325
1326 /*
1327  * rb tree support functions
1328  */
1329 static void cfq_del_rq_rb(struct request *rq)
1330 {
1331         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1332         const int sync = rq_is_sync(rq);
1333
1334         BUG_ON(!cfqq->queued[sync]);
1335         cfqq->queued[sync]--;
1336
1337         elv_rb_del(&cfqq->sort_list, rq);
1338
1339         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1340                 /*
1341                  * Queue will be deleted from service tree when we actually
1342                  * expire it later. Right now just remove it from prio tree
1343                  * as it is empty.
1344                  */
1345                 if (cfqq->p_root) {
1346                         rb_erase(&cfqq->p_node, cfqq->p_root);
1347                         cfqq->p_root = NULL;
1348                 }
1349         }
1350 }
1351
1352 static void cfq_add_rq_rb(struct request *rq)
1353 {
1354         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1355         struct cfq_data *cfqd = cfqq->cfqd;
1356         struct request *__alias, *prev;
1357
1358         cfqq->queued[rq_is_sync(rq)]++;
1359
1360         /*
1361          * looks a little odd, but the first insert might return an alias.
1362          * if that happens, put the alias on the dispatch list
1363          */
1364         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1365                 cfq_dispatch_insert(cfqd->queue, __alias);
1366
1367         if (!cfq_cfqq_on_rr(cfqq))
1368                 cfq_add_cfqq_rr(cfqd, cfqq);
1369
1370         /*
1371          * check if this request is a better next-serve candidate
1372          */
1373         prev = cfqq->next_rq;
1374         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1375
1376         /*
1377          * adjust priority tree position, if ->next_rq changes
1378          */
1379         if (prev != cfqq->next_rq)
1380                 cfq_prio_tree_add(cfqd, cfqq);
1381
1382         BUG_ON(!cfqq->next_rq);
1383 }
1384
1385 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1386 {
1387         elv_rb_del(&cfqq->sort_list, rq);
1388         cfqq->queued[rq_is_sync(rq)]--;
1389         cfq_add_rq_rb(rq);
1390 }
1391
1392 static struct request *
1393 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1394 {
1395         struct task_struct *tsk = current;
1396         struct cfq_io_context *cic;
1397         struct cfq_queue *cfqq;
1398
1399         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1400         if (!cic)
1401                 return NULL;
1402
1403         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1404         if (cfqq) {
1405                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1406
1407                 return elv_rb_find(&cfqq->sort_list, sector);
1408         }
1409
1410         return NULL;
1411 }
1412
1413 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1414 {
1415         struct cfq_data *cfqd = q->elevator->elevator_data;
1416
1417         cfqd->rq_in_driver[rq_is_sync(rq)]++;
1418         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1419                                                 rq_in_driver(cfqd));
1420
1421         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1422 }
1423
1424 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1425 {
1426         struct cfq_data *cfqd = q->elevator->elevator_data;
1427         const int sync = rq_is_sync(rq);
1428
1429         WARN_ON(!cfqd->rq_in_driver[sync]);
1430         cfqd->rq_in_driver[sync]--;
1431         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1432                                                 rq_in_driver(cfqd));
1433 }
1434
1435 static void cfq_remove_request(struct request *rq)
1436 {
1437         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1438
1439         if (cfqq->next_rq == rq)
1440                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1441
1442         list_del_init(&rq->queuelist);
1443         cfq_del_rq_rb(rq);
1444
1445         cfqq->cfqd->rq_queued--;
1446         if (rq_is_meta(rq)) {
1447                 WARN_ON(!cfqq->meta_pending);
1448                 cfqq->meta_pending--;
1449         }
1450 }
1451
1452 static int cfq_merge(struct request_queue *q, struct request **req,
1453                      struct bio *bio)
1454 {
1455         struct cfq_data *cfqd = q->elevator->elevator_data;
1456         struct request *__rq;
1457
1458         __rq = cfq_find_rq_fmerge(cfqd, bio);
1459         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1460                 *req = __rq;
1461                 return ELEVATOR_FRONT_MERGE;
1462         }
1463
1464         return ELEVATOR_NO_MERGE;
1465 }
1466
1467 static void cfq_merged_request(struct request_queue *q, struct request *req,
1468                                int type)
1469 {
1470         if (type == ELEVATOR_FRONT_MERGE) {
1471                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1472
1473                 cfq_reposition_rq_rb(cfqq, req);
1474         }
1475 }
1476
1477 static void
1478 cfq_merged_requests(struct request_queue *q, struct request *rq,
1479                     struct request *next)
1480 {
1481         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1482         /*
1483          * reposition in fifo if next is older than rq
1484          */
1485         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1486             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1487                 list_move(&rq->queuelist, &next->queuelist);
1488                 rq_set_fifo_time(rq, rq_fifo_time(next));
1489         }
1490
1491         if (cfqq->next_rq == next)
1492                 cfqq->next_rq = rq;
1493         cfq_remove_request(next);
1494 }
1495
1496 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1497                            struct bio *bio)
1498 {
1499         struct cfq_data *cfqd = q->elevator->elevator_data;
1500         struct cfq_io_context *cic;
1501         struct cfq_queue *cfqq;
1502
1503         /*
1504          * Disallow merge of a sync bio into an async request.
1505          */
1506         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1507                 return false;
1508
1509         /*
1510          * Lookup the cfqq that this bio will be queued with. Allow
1511          * merge only if rq is queued there.
1512          */
1513         cic = cfq_cic_lookup(cfqd, current->io_context);
1514         if (!cic)
1515                 return false;
1516
1517         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1518         return cfqq == RQ_CFQQ(rq);
1519 }
1520
1521 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1522                                    struct cfq_queue *cfqq)
1523 {
1524         if (cfqq) {
1525                 cfq_log_cfqq(cfqd, cfqq, "set_active");
1526                 cfqq->slice_start = 0;
1527                 cfqq->dispatch_start = jiffies;
1528                 cfqq->allocated_slice = 0;
1529                 cfqq->slice_end = 0;
1530                 cfqq->slice_dispatch = 0;
1531                 cfqq->nr_sectors = 0;
1532
1533                 cfq_clear_cfqq_wait_request(cfqq);
1534                 cfq_clear_cfqq_must_dispatch(cfqq);
1535                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1536                 cfq_clear_cfqq_fifo_expire(cfqq);
1537                 cfq_mark_cfqq_slice_new(cfqq);
1538
1539                 del_timer(&cfqd->idle_slice_timer);
1540         }
1541
1542         cfqd->active_queue = cfqq;
1543 }
1544
1545 /*
1546  * current cfqq expired its slice (or was too idle), select new one
1547  */
1548 static void
1549 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1550                     bool timed_out)
1551 {
1552         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1553
1554         if (cfq_cfqq_wait_request(cfqq))
1555                 del_timer(&cfqd->idle_slice_timer);
1556
1557         cfq_clear_cfqq_wait_request(cfqq);
1558         cfq_clear_cfqq_wait_busy(cfqq);
1559
1560         /*
1561          * If this cfqq is shared between multiple processes, check to
1562          * make sure that those processes are still issuing I/Os within
1563          * the mean seek distance.  If not, it may be time to break the
1564          * queues apart again.
1565          */
1566         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1567                 cfq_mark_cfqq_split_coop(cfqq);
1568
1569         /*
1570          * store what was left of this slice, if the queue idled/timed out
1571          */
1572         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1573                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1574                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1575         }
1576
1577         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1578
1579         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1580                 cfq_del_cfqq_rr(cfqd, cfqq);
1581
1582         cfq_resort_rr_list(cfqd, cfqq);
1583
1584         if (cfqq == cfqd->active_queue)
1585                 cfqd->active_queue = NULL;
1586
1587         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1588                 cfqd->grp_service_tree.active = NULL;
1589
1590         if (cfqd->active_cic) {
1591                 put_io_context(cfqd->active_cic->ioc);
1592                 cfqd->active_cic = NULL;
1593         }
1594 }
1595
1596 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1597 {
1598         struct cfq_queue *cfqq = cfqd->active_queue;
1599
1600         if (cfqq)
1601                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1602 }
1603
1604 /*
1605  * Get next queue for service. Unless we have a queue preemption,
1606  * we'll simply select the first cfqq in the service tree.
1607  */
1608 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1609 {
1610         struct cfq_rb_root *service_tree =
1611                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1612                                         cfqd->serving_type);
1613
1614         if (!cfqd->rq_queued)
1615                 return NULL;
1616
1617         /* There is nothing to dispatch */
1618         if (!service_tree)
1619                 return NULL;
1620         if (RB_EMPTY_ROOT(&service_tree->rb))
1621                 return NULL;
1622         return cfq_rb_first(service_tree);
1623 }
1624
1625 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1626 {
1627         struct cfq_group *cfqg;
1628         struct cfq_queue *cfqq;
1629         int i, j;
1630         struct cfq_rb_root *st;
1631
1632         if (!cfqd->rq_queued)
1633                 return NULL;
1634
1635         cfqg = cfq_get_next_cfqg(cfqd);
1636         if (!cfqg)
1637                 return NULL;
1638
1639         for_each_cfqg_st(cfqg, i, j, st)
1640                 if ((cfqq = cfq_rb_first(st)) != NULL)
1641                         return cfqq;
1642         return NULL;
1643 }
1644
1645 /*
1646  * Get and set a new active queue for service.
1647  */
1648 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1649                                               struct cfq_queue *cfqq)
1650 {
1651         if (!cfqq)
1652                 cfqq = cfq_get_next_queue(cfqd);
1653
1654         __cfq_set_active_queue(cfqd, cfqq);
1655         return cfqq;
1656 }
1657
1658 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1659                                           struct request *rq)
1660 {
1661         if (blk_rq_pos(rq) >= cfqd->last_position)
1662                 return blk_rq_pos(rq) - cfqd->last_position;
1663         else
1664                 return cfqd->last_position - blk_rq_pos(rq);
1665 }
1666
1667 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1668                                struct request *rq, bool for_preempt)
1669 {
1670         return cfq_dist_from_last(cfqd, rq) <= CFQQ_SEEK_THR;
1671 }
1672
1673 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1674                                     struct cfq_queue *cur_cfqq)
1675 {
1676         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1677         struct rb_node *parent, *node;
1678         struct cfq_queue *__cfqq;
1679         sector_t sector = cfqd->last_position;
1680
1681         if (RB_EMPTY_ROOT(root))
1682                 return NULL;
1683
1684         /*
1685          * First, if we find a request starting at the end of the last
1686          * request, choose it.
1687          */
1688         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1689         if (__cfqq)
1690                 return __cfqq;
1691
1692         /*
1693          * If the exact sector wasn't found, the parent of the NULL leaf
1694          * will contain the closest sector.
1695          */
1696         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1697         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1698                 return __cfqq;
1699
1700         if (blk_rq_pos(__cfqq->next_rq) < sector)
1701                 node = rb_next(&__cfqq->p_node);
1702         else
1703                 node = rb_prev(&__cfqq->p_node);
1704         if (!node)
1705                 return NULL;
1706
1707         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1708         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1709                 return __cfqq;
1710
1711         return NULL;
1712 }
1713
1714 /*
1715  * cfqd - obvious
1716  * cur_cfqq - passed in so that we don't decide that the current queue is
1717  *            closely cooperating with itself.
1718  *
1719  * So, basically we're assuming that that cur_cfqq has dispatched at least
1720  * one request, and that cfqd->last_position reflects a position on the disk
1721  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1722  * assumption.
1723  */
1724 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1725                                               struct cfq_queue *cur_cfqq)
1726 {
1727         struct cfq_queue *cfqq;
1728
1729         if (!cfq_cfqq_sync(cur_cfqq))
1730                 return NULL;
1731         if (CFQQ_SEEKY(cur_cfqq))
1732                 return NULL;
1733
1734         /*
1735          * Don't search priority tree if it's the only queue in the group.
1736          */
1737         if (cur_cfqq->cfqg->nr_cfqq == 1)
1738                 return NULL;
1739
1740         /*
1741          * We should notice if some of the queues are cooperating, eg
1742          * working closely on the same area of the disk. In that case,
1743          * we can group them together and don't waste time idling.
1744          */
1745         cfqq = cfqq_close(cfqd, cur_cfqq);
1746         if (!cfqq)
1747                 return NULL;
1748
1749         /* If new queue belongs to different cfq_group, don't choose it */
1750         if (cur_cfqq->cfqg != cfqq->cfqg)
1751                 return NULL;
1752
1753         /*
1754          * It only makes sense to merge sync queues.
1755          */
1756         if (!cfq_cfqq_sync(cfqq))
1757                 return NULL;
1758         if (CFQQ_SEEKY(cfqq))
1759                 return NULL;
1760
1761         /*
1762          * Do not merge queues of different priority classes
1763          */
1764         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1765                 return NULL;
1766
1767         return cfqq;
1768 }
1769
1770 /*
1771  * Determine whether we should enforce idle window for this queue.
1772  */
1773
1774 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1775 {
1776         enum wl_prio_t prio = cfqq_prio(cfqq);
1777         struct cfq_rb_root *service_tree = cfqq->service_tree;
1778
1779         BUG_ON(!service_tree);
1780         BUG_ON(!service_tree->count);
1781
1782         /* We never do for idle class queues. */
1783         if (prio == IDLE_WORKLOAD)
1784                 return false;
1785
1786         /* We do for queues that were marked with idle window flag. */
1787         if (cfq_cfqq_idle_window(cfqq) &&
1788            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1789                 return true;
1790
1791         /*
1792          * Otherwise, we do only if they are the last ones
1793          * in their service tree.
1794          */
1795         return service_tree->count == 1 && cfq_cfqq_sync(cfqq);
1796 }
1797
1798 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1799 {
1800         struct cfq_queue *cfqq = cfqd->active_queue;
1801         struct cfq_io_context *cic;
1802         unsigned long sl;
1803
1804         /*
1805          * SSD device without seek penalty, disable idling. But only do so
1806          * for devices that support queuing, otherwise we still have a problem
1807          * with sync vs async workloads.
1808          */
1809         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1810                 return;
1811
1812         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1813         WARN_ON(cfq_cfqq_slice_new(cfqq));
1814
1815         /*
1816          * idle is disabled, either manually or by past process history
1817          */
1818         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1819                 return;
1820
1821         /*
1822          * still active requests from this queue, don't idle
1823          */
1824         if (cfqq->dispatched)
1825                 return;
1826
1827         /*
1828          * task has exited, don't wait
1829          */
1830         cic = cfqd->active_cic;
1831         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1832                 return;
1833
1834         /*
1835          * If our average think time is larger than the remaining time
1836          * slice, then don't idle. This avoids overrunning the allotted
1837          * time slice.
1838          */
1839         if (sample_valid(cic->ttime_samples) &&
1840             (cfqq->slice_end - jiffies < cic->ttime_mean))
1841                 return;
1842
1843         cfq_mark_cfqq_wait_request(cfqq);
1844
1845         sl = cfqd->cfq_slice_idle;
1846
1847         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1848         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1849 }
1850
1851 /*
1852  * Move request from internal lists to the request queue dispatch list.
1853  */
1854 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1855 {
1856         struct cfq_data *cfqd = q->elevator->elevator_data;
1857         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1858
1859         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1860
1861         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1862         cfq_remove_request(rq);
1863         cfqq->dispatched++;
1864         elv_dispatch_sort(q, rq);
1865
1866         if (cfq_cfqq_sync(cfqq))
1867                 cfqd->sync_flight++;
1868         cfqq->nr_sectors += blk_rq_sectors(rq);
1869 }
1870
1871 /*
1872  * return expired entry, or NULL to just start from scratch in rbtree
1873  */
1874 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1875 {
1876         struct request *rq = NULL;
1877
1878         if (cfq_cfqq_fifo_expire(cfqq))
1879                 return NULL;
1880
1881         cfq_mark_cfqq_fifo_expire(cfqq);
1882
1883         if (list_empty(&cfqq->fifo))
1884                 return NULL;
1885
1886         rq = rq_entry_fifo(cfqq->fifo.next);
1887         if (time_before(jiffies, rq_fifo_time(rq)))
1888                 rq = NULL;
1889
1890         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1891         return rq;
1892 }
1893
1894 static inline int
1895 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1896 {
1897         const int base_rq = cfqd->cfq_slice_async_rq;
1898
1899         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1900
1901         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1902 }
1903
1904 /*
1905  * Must be called with the queue_lock held.
1906  */
1907 static int cfqq_process_refs(struct cfq_queue *cfqq)
1908 {
1909         int process_refs, io_refs;
1910
1911         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1912         process_refs = atomic_read(&cfqq->ref) - io_refs;
1913         BUG_ON(process_refs < 0);
1914         return process_refs;
1915 }
1916
1917 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1918 {
1919         int process_refs, new_process_refs;
1920         struct cfq_queue *__cfqq;
1921
1922         /* Avoid a circular list and skip interim queue merges */
1923         while ((__cfqq = new_cfqq->new_cfqq)) {
1924                 if (__cfqq == cfqq)
1925                         return;
1926                 new_cfqq = __cfqq;
1927         }
1928
1929         process_refs = cfqq_process_refs(cfqq);
1930         /*
1931          * If the process for the cfqq has gone away, there is no
1932          * sense in merging the queues.
1933          */
1934         if (process_refs == 0)
1935                 return;
1936
1937         /*
1938          * Merge in the direction of the lesser amount of work.
1939          */
1940         new_process_refs = cfqq_process_refs(new_cfqq);
1941         if (new_process_refs >= process_refs) {
1942                 cfqq->new_cfqq = new_cfqq;
1943                 atomic_add(process_refs, &new_cfqq->ref);
1944         } else {
1945                 new_cfqq->new_cfqq = cfqq;
1946                 atomic_add(new_process_refs, &cfqq->ref);
1947         }
1948 }
1949
1950 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1951                                 struct cfq_group *cfqg, enum wl_prio_t prio)
1952 {
1953         struct cfq_queue *queue;
1954         int i;
1955         bool key_valid = false;
1956         unsigned long lowest_key = 0;
1957         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1958
1959         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1960                 /* select the one with lowest rb_key */
1961                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1962                 if (queue &&
1963                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1964                         lowest_key = queue->rb_key;
1965                         cur_best = i;
1966                         key_valid = true;
1967                 }
1968         }
1969
1970         return cur_best;
1971 }
1972
1973 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1974 {
1975         unsigned slice;
1976         unsigned count;
1977         struct cfq_rb_root *st;
1978         unsigned group_slice;
1979
1980         if (!cfqg) {
1981                 cfqd->serving_prio = IDLE_WORKLOAD;
1982                 cfqd->workload_expires = jiffies + 1;
1983                 return;
1984         }
1985
1986         /* Choose next priority. RT > BE > IDLE */
1987         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1988                 cfqd->serving_prio = RT_WORKLOAD;
1989         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1990                 cfqd->serving_prio = BE_WORKLOAD;
1991         else {
1992                 cfqd->serving_prio = IDLE_WORKLOAD;
1993                 cfqd->workload_expires = jiffies + 1;
1994                 return;
1995         }
1996
1997         /*
1998          * For RT and BE, we have to choose also the type
1999          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2000          * expiration time
2001          */
2002         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2003         count = st->count;
2004
2005         /*
2006          * check workload expiration, and that we still have other queues ready
2007          */
2008         if (count && !time_after(jiffies, cfqd->workload_expires))
2009                 return;
2010
2011         /* otherwise select new workload type */
2012         cfqd->serving_type =
2013                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2014         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2015         count = st->count;
2016
2017         /*
2018          * the workload slice is computed as a fraction of target latency
2019          * proportional to the number of queues in that workload, over
2020          * all the queues in the same priority class
2021          */
2022         group_slice = cfq_group_slice(cfqd, cfqg);
2023
2024         slice = group_slice * count /
2025                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2026                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2027
2028         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2029                 unsigned int tmp;
2030
2031                 /*
2032                  * Async queues are currently system wide. Just taking
2033                  * proportion of queues with-in same group will lead to higher
2034                  * async ratio system wide as generally root group is going
2035                  * to have higher weight. A more accurate thing would be to
2036                  * calculate system wide asnc/sync ratio.
2037                  */
2038                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2039                 tmp = tmp/cfqd->busy_queues;
2040                 slice = min_t(unsigned, slice, tmp);
2041
2042                 /* async workload slice is scaled down according to
2043                  * the sync/async slice ratio. */
2044                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2045         } else
2046                 /* sync workload slice is at least 2 * cfq_slice_idle */
2047                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2048
2049         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2050         cfqd->workload_expires = jiffies + slice;
2051         cfqd->noidle_tree_requires_idle = false;
2052 }
2053
2054 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2055 {
2056         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2057         struct cfq_group *cfqg;
2058
2059         if (RB_EMPTY_ROOT(&st->rb))
2060                 return NULL;
2061         cfqg = cfq_rb_first_group(st);
2062         st->active = &cfqg->rb_node;
2063         update_min_vdisktime(st);
2064         return cfqg;
2065 }
2066
2067 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2068 {
2069         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2070
2071         cfqd->serving_group = cfqg;
2072
2073         /* Restore the workload type data */
2074         if (cfqg->saved_workload_slice) {
2075                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2076                 cfqd->serving_type = cfqg->saved_workload;
2077                 cfqd->serving_prio = cfqg->saved_serving_prio;
2078         } else
2079                 cfqd->workload_expires = jiffies - 1;
2080
2081         choose_service_tree(cfqd, cfqg);
2082 }
2083
2084 /*
2085  * Select a queue for service. If we have a current active queue,
2086  * check whether to continue servicing it, or retrieve and set a new one.
2087  */
2088 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2089 {
2090         struct cfq_queue *cfqq, *new_cfqq = NULL;
2091
2092         cfqq = cfqd->active_queue;
2093         if (!cfqq)
2094                 goto new_queue;
2095
2096         if (!cfqd->rq_queued)
2097                 return NULL;
2098
2099         /*
2100          * We were waiting for group to get backlogged. Expire the queue
2101          */
2102         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2103                 goto expire;
2104
2105         /*
2106          * The active queue has run out of time, expire it and select new.
2107          */
2108         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2109                 /*
2110                  * If slice had not expired at the completion of last request
2111                  * we might not have turned on wait_busy flag. Don't expire
2112                  * the queue yet. Allow the group to get backlogged.
2113                  *
2114                  * The very fact that we have used the slice, that means we
2115                  * have been idling all along on this queue and it should be
2116                  * ok to wait for this request to complete.
2117                  */
2118                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2119                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2120                         cfqq = NULL;
2121                         goto keep_queue;
2122                 } else
2123                         goto expire;
2124         }
2125
2126         /*
2127          * The active queue has requests and isn't expired, allow it to
2128          * dispatch.
2129          */
2130         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2131                 goto keep_queue;
2132
2133         /*
2134          * If another queue has a request waiting within our mean seek
2135          * distance, let it run.  The expire code will check for close
2136          * cooperators and put the close queue at the front of the service
2137          * tree.  If possible, merge the expiring queue with the new cfqq.
2138          */
2139         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2140         if (new_cfqq) {
2141                 if (!cfqq->new_cfqq)
2142                         cfq_setup_merge(cfqq, new_cfqq);
2143                 goto expire;
2144         }
2145
2146         /*
2147          * No requests pending. If the active queue still has requests in
2148          * flight or is idling for a new request, allow either of these
2149          * conditions to happen (or time out) before selecting a new queue.
2150          */
2151         if (timer_pending(&cfqd->idle_slice_timer) ||
2152             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2153                 cfqq = NULL;
2154                 goto keep_queue;
2155         }
2156
2157 expire:
2158         cfq_slice_expired(cfqd, 0);
2159 new_queue:
2160         /*
2161          * Current queue expired. Check if we have to switch to a new
2162          * service tree
2163          */
2164         if (!new_cfqq)
2165                 cfq_choose_cfqg(cfqd);
2166
2167         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2168 keep_queue:
2169         return cfqq;
2170 }
2171
2172 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2173 {
2174         int dispatched = 0;
2175
2176         while (cfqq->next_rq) {
2177                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2178                 dispatched++;
2179         }
2180
2181         BUG_ON(!list_empty(&cfqq->fifo));
2182
2183         /* By default cfqq is not expired if it is empty. Do it explicitly */
2184         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2185         return dispatched;
2186 }
2187
2188 /*
2189  * Drain our current requests. Used for barriers and when switching
2190  * io schedulers on-the-fly.
2191  */
2192 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2193 {
2194         struct cfq_queue *cfqq;
2195         int dispatched = 0;
2196
2197         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2198                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2199
2200         cfq_slice_expired(cfqd, 0);
2201         BUG_ON(cfqd->busy_queues);
2202
2203         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2204         return dispatched;
2205 }
2206
2207 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2208 {
2209         unsigned int max_dispatch;
2210
2211         /*
2212          * Drain async requests before we start sync IO
2213          */
2214         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2215                 return false;
2216
2217         /*
2218          * If this is an async queue and we have sync IO in flight, let it wait
2219          */
2220         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2221                 return false;
2222
2223         max_dispatch = cfqd->cfq_quantum;
2224         if (cfq_class_idle(cfqq))
2225                 max_dispatch = 1;
2226
2227         /*
2228          * Does this cfqq already have too much IO in flight?
2229          */
2230         if (cfqq->dispatched >= max_dispatch) {
2231                 /*
2232                  * idle queue must always only have a single IO in flight
2233                  */
2234                 if (cfq_class_idle(cfqq))
2235                         return false;
2236
2237                 /*
2238                  * We have other queues, don't allow more IO from this one
2239                  */
2240                 if (cfqd->busy_queues > 1)
2241                         return false;
2242
2243                 /*
2244                  * Sole queue user, no limit
2245                  */
2246                 max_dispatch = -1;
2247         }
2248
2249         /*
2250          * Async queues must wait a bit before being allowed dispatch.
2251          * We also ramp up the dispatch depth gradually for async IO,
2252          * based on the last sync IO we serviced
2253          */
2254         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2255                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2256                 unsigned int depth;
2257
2258                 depth = last_sync / cfqd->cfq_slice[1];
2259                 if (!depth && !cfqq->dispatched)
2260                         depth = 1;
2261                 if (depth < max_dispatch)
2262                         max_dispatch = depth;
2263         }
2264
2265         /*
2266          * If we're below the current max, allow a dispatch
2267          */
2268         return cfqq->dispatched < max_dispatch;
2269 }
2270
2271 /*
2272  * Dispatch a request from cfqq, moving them to the request queue
2273  * dispatch list.
2274  */
2275 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2276 {
2277         struct request *rq;
2278
2279         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2280
2281         if (!cfq_may_dispatch(cfqd, cfqq))
2282                 return false;
2283
2284         /*
2285          * follow expired path, else get first next available
2286          */
2287         rq = cfq_check_fifo(cfqq);
2288         if (!rq)
2289                 rq = cfqq->next_rq;
2290
2291         /*
2292          * insert request into driver dispatch list
2293          */
2294         cfq_dispatch_insert(cfqd->queue, rq);
2295
2296         if (!cfqd->active_cic) {
2297                 struct cfq_io_context *cic = RQ_CIC(rq);
2298
2299                 atomic_long_inc(&cic->ioc->refcount);
2300                 cfqd->active_cic = cic;
2301         }
2302
2303         return true;
2304 }
2305
2306 /*
2307  * Find the cfqq that we need to service and move a request from that to the
2308  * dispatch list
2309  */
2310 static int cfq_dispatch_requests(struct request_queue *q, int force)
2311 {
2312         struct cfq_data *cfqd = q->elevator->elevator_data;
2313         struct cfq_queue *cfqq;
2314
2315         if (!cfqd->busy_queues)
2316                 return 0;
2317
2318         if (unlikely(force))
2319                 return cfq_forced_dispatch(cfqd);
2320
2321         cfqq = cfq_select_queue(cfqd);
2322         if (!cfqq)
2323                 return 0;
2324
2325         /*
2326          * Dispatch a request from this cfqq, if it is allowed
2327          */
2328         if (!cfq_dispatch_request(cfqd, cfqq))
2329                 return 0;
2330
2331         cfqq->slice_dispatch++;
2332         cfq_clear_cfqq_must_dispatch(cfqq);
2333
2334         /*
2335          * expire an async queue immediately if it has used up its slice. idle
2336          * queue always expire after 1 dispatch round.
2337          */
2338         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2339             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2340             cfq_class_idle(cfqq))) {
2341                 cfqq->slice_end = jiffies + 1;
2342                 cfq_slice_expired(cfqd, 0);
2343         }
2344
2345         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2346         return 1;
2347 }
2348
2349 /*
2350  * task holds one reference to the queue, dropped when task exits. each rq
2351  * in-flight on this queue also holds a reference, dropped when rq is freed.
2352  *
2353  * Each cfq queue took a reference on the parent group. Drop it now.
2354  * queue lock must be held here.
2355  */
2356 static void cfq_put_queue(struct cfq_queue *cfqq)
2357 {
2358         struct cfq_data *cfqd = cfqq->cfqd;
2359         struct cfq_group *cfqg, *orig_cfqg;
2360
2361         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2362
2363         if (!atomic_dec_and_test(&cfqq->ref))
2364                 return;
2365
2366         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2367         BUG_ON(rb_first(&cfqq->sort_list));
2368         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2369         cfqg = cfqq->cfqg;
2370         orig_cfqg = cfqq->orig_cfqg;
2371
2372         if (unlikely(cfqd->active_queue == cfqq)) {
2373                 __cfq_slice_expired(cfqd, cfqq, 0);
2374                 cfq_schedule_dispatch(cfqd);
2375         }
2376
2377         BUG_ON(cfq_cfqq_on_rr(cfqq));
2378         kmem_cache_free(cfq_pool, cfqq);
2379         cfq_put_cfqg(cfqg);
2380         if (orig_cfqg)
2381                 cfq_put_cfqg(orig_cfqg);
2382 }
2383
2384 /*
2385  * Must always be called with the rcu_read_lock() held
2386  */
2387 static void
2388 __call_for_each_cic(struct io_context *ioc,
2389                     void (*func)(struct io_context *, struct cfq_io_context *))
2390 {
2391         struct cfq_io_context *cic;
2392         struct hlist_node *n;
2393
2394         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2395                 func(ioc, cic);
2396 }
2397
2398 /*
2399  * Call func for each cic attached to this ioc.
2400  */
2401 static void
2402 call_for_each_cic(struct io_context *ioc,
2403                   void (*func)(struct io_context *, struct cfq_io_context *))
2404 {
2405         rcu_read_lock();
2406         __call_for_each_cic(ioc, func);
2407         rcu_read_unlock();
2408 }
2409
2410 static void cfq_cic_free_rcu(struct rcu_head *head)
2411 {
2412         struct cfq_io_context *cic;
2413
2414         cic = container_of(head, struct cfq_io_context, rcu_head);
2415
2416         kmem_cache_free(cfq_ioc_pool, cic);
2417         elv_ioc_count_dec(cfq_ioc_count);
2418
2419         if (ioc_gone) {
2420                 /*
2421                  * CFQ scheduler is exiting, grab exit lock and check
2422                  * the pending io context count. If it hits zero,
2423                  * complete ioc_gone and set it back to NULL
2424                  */
2425                 spin_lock(&ioc_gone_lock);
2426                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2427                         complete(ioc_gone);
2428                         ioc_gone = NULL;
2429                 }
2430                 spin_unlock(&ioc_gone_lock);
2431         }
2432 }
2433
2434 static void cfq_cic_free(struct cfq_io_context *cic)
2435 {
2436         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2437 }
2438
2439 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2440 {
2441         unsigned long flags;
2442
2443         BUG_ON(!cic->dead_key);
2444
2445         spin_lock_irqsave(&ioc->lock, flags);
2446         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2447         hlist_del_rcu(&cic->cic_list);
2448         spin_unlock_irqrestore(&ioc->lock, flags);
2449
2450         cfq_cic_free(cic);
2451 }
2452
2453 /*
2454  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2455  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2456  * and ->trim() which is called with the task lock held
2457  */
2458 static void cfq_free_io_context(struct io_context *ioc)
2459 {
2460         /*
2461          * ioc->refcount is zero here, or we are called from elv_unregister(),
2462          * so no more cic's are allowed to be linked into this ioc.  So it
2463          * should be ok to iterate over the known list, we will see all cic's
2464          * since no new ones are added.
2465          */
2466         __call_for_each_cic(ioc, cic_free_func);
2467 }
2468
2469 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2470 {
2471         struct cfq_queue *__cfqq, *next;
2472
2473         if (unlikely(cfqq == cfqd->active_queue)) {
2474                 __cfq_slice_expired(cfqd, cfqq, 0);
2475                 cfq_schedule_dispatch(cfqd);
2476         }
2477
2478         /*
2479          * If this queue was scheduled to merge with another queue, be
2480          * sure to drop the reference taken on that queue (and others in
2481          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2482          */
2483         __cfqq = cfqq->new_cfqq;
2484         while (__cfqq) {
2485                 if (__cfqq == cfqq) {
2486                         WARN(1, "cfqq->new_cfqq loop detected\n");
2487                         break;
2488                 }
2489                 next = __cfqq->new_cfqq;
2490                 cfq_put_queue(__cfqq);
2491                 __cfqq = next;
2492         }
2493
2494         cfq_put_queue(cfqq);
2495 }
2496
2497 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2498                                          struct cfq_io_context *cic)
2499 {
2500         struct io_context *ioc = cic->ioc;
2501
2502         list_del_init(&cic->queue_list);
2503
2504         /*
2505          * Make sure key == NULL is seen for dead queues
2506          */
2507         smp_wmb();
2508         cic->dead_key = (unsigned long) cic->key;
2509         cic->key = NULL;
2510
2511         if (ioc->ioc_data == cic)
2512                 rcu_assign_pointer(ioc->ioc_data, NULL);
2513
2514         if (cic->cfqq[BLK_RW_ASYNC]) {
2515                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2516                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2517         }
2518
2519         if (cic->cfqq[BLK_RW_SYNC]) {
2520                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2521                 cic->cfqq[BLK_RW_SYNC] = NULL;
2522         }
2523 }
2524
2525 static void cfq_exit_single_io_context(struct io_context *ioc,
2526                                        struct cfq_io_context *cic)
2527 {
2528         struct cfq_data *cfqd = cic->key;
2529
2530         if (cfqd) {
2531                 struct request_queue *q = cfqd->queue;
2532                 unsigned long flags;
2533
2534                 spin_lock_irqsave(q->queue_lock, flags);
2535
2536                 /*
2537                  * Ensure we get a fresh copy of the ->key to prevent
2538                  * race between exiting task and queue
2539                  */
2540                 smp_read_barrier_depends();
2541                 if (cic->key)
2542                         __cfq_exit_single_io_context(cfqd, cic);
2543
2544                 spin_unlock_irqrestore(q->queue_lock, flags);
2545         }
2546 }
2547
2548 /*
2549  * The process that ioc belongs to has exited, we need to clean up
2550  * and put the internal structures we have that belongs to that process.
2551  */
2552 static void cfq_exit_io_context(struct io_context *ioc)
2553 {
2554         call_for_each_cic(ioc, cfq_exit_single_io_context);
2555 }
2556
2557 static struct cfq_io_context *
2558 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2559 {
2560         struct cfq_io_context *cic;
2561
2562         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2563                                                         cfqd->queue->node);
2564         if (cic) {
2565                 cic->last_end_request = jiffies;
2566                 INIT_LIST_HEAD(&cic->queue_list);
2567                 INIT_HLIST_NODE(&cic->cic_list);
2568                 cic->dtor = cfq_free_io_context;
2569                 cic->exit = cfq_exit_io_context;
2570                 elv_ioc_count_inc(cfq_ioc_count);
2571         }
2572
2573         return cic;
2574 }
2575
2576 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2577 {
2578         struct task_struct *tsk = current;
2579         int ioprio_class;
2580
2581         if (!cfq_cfqq_prio_changed(cfqq))
2582                 return;
2583
2584         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2585         switch (ioprio_class) {
2586         default:
2587                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2588         case IOPRIO_CLASS_NONE:
2589                 /*
2590                  * no prio set, inherit CPU scheduling settings
2591                  */
2592                 cfqq->ioprio = task_nice_ioprio(tsk);
2593                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2594                 break;
2595         case IOPRIO_CLASS_RT:
2596                 cfqq->ioprio = task_ioprio(ioc);
2597                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2598                 break;
2599         case IOPRIO_CLASS_BE:
2600                 cfqq->ioprio = task_ioprio(ioc);
2601                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2602                 break;
2603         case IOPRIO_CLASS_IDLE:
2604                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2605                 cfqq->ioprio = 7;
2606                 cfq_clear_cfqq_idle_window(cfqq);
2607                 break;
2608         }
2609
2610         /*
2611          * keep track of original prio settings in case we have to temporarily
2612          * elevate the priority of this queue
2613          */
2614         cfqq->org_ioprio = cfqq->ioprio;
2615         cfqq->org_ioprio_class = cfqq->ioprio_class;
2616         cfq_clear_cfqq_prio_changed(cfqq);
2617 }
2618
2619 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2620 {
2621         struct cfq_data *cfqd = cic->key;
2622         struct cfq_queue *cfqq;
2623         unsigned long flags;
2624
2625         if (unlikely(!cfqd))
2626                 return;
2627
2628         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2629
2630         cfqq = cic->cfqq[BLK_RW_ASYNC];
2631         if (cfqq) {
2632                 struct cfq_queue *new_cfqq;
2633                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2634                                                 GFP_ATOMIC);
2635                 if (new_cfqq) {
2636                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2637                         cfq_put_queue(cfqq);
2638                 }
2639         }
2640
2641         cfqq = cic->cfqq[BLK_RW_SYNC];
2642         if (cfqq)
2643                 cfq_mark_cfqq_prio_changed(cfqq);
2644
2645         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2646 }
2647
2648 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2649 {
2650         call_for_each_cic(ioc, changed_ioprio);
2651         ioc->ioprio_changed = 0;
2652 }
2653
2654 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2655                           pid_t pid, bool is_sync)
2656 {
2657         RB_CLEAR_NODE(&cfqq->rb_node);
2658         RB_CLEAR_NODE(&cfqq->p_node);
2659         INIT_LIST_HEAD(&cfqq->fifo);
2660
2661         atomic_set(&cfqq->ref, 0);
2662         cfqq->cfqd = cfqd;
2663
2664         cfq_mark_cfqq_prio_changed(cfqq);
2665
2666         if (is_sync) {
2667                 if (!cfq_class_idle(cfqq))
2668                         cfq_mark_cfqq_idle_window(cfqq);
2669                 cfq_mark_cfqq_sync(cfqq);
2670         }
2671         cfqq->pid = pid;
2672 }
2673
2674 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2675 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2676 {
2677         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2678         struct cfq_data *cfqd = cic->key;
2679         unsigned long flags;
2680         struct request_queue *q;
2681
2682         if (unlikely(!cfqd))
2683                 return;
2684
2685         q = cfqd->queue;
2686
2687         spin_lock_irqsave(q->queue_lock, flags);
2688
2689         if (sync_cfqq) {
2690                 /*
2691                  * Drop reference to sync queue. A new sync queue will be
2692                  * assigned in new group upon arrival of a fresh request.
2693                  */
2694                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2695                 cic_set_cfqq(cic, NULL, 1);
2696                 cfq_put_queue(sync_cfqq);
2697         }
2698
2699         spin_unlock_irqrestore(q->queue_lock, flags);
2700 }
2701
2702 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2703 {
2704         call_for_each_cic(ioc, changed_cgroup);
2705         ioc->cgroup_changed = 0;
2706 }
2707 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2708
2709 static struct cfq_queue *
2710 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2711                      struct io_context *ioc, gfp_t gfp_mask)
2712 {
2713         struct cfq_queue *cfqq, *new_cfqq = NULL;
2714         struct cfq_io_context *cic;
2715         struct cfq_group *cfqg;
2716
2717 retry:
2718         cfqg = cfq_get_cfqg(cfqd, 1);
2719         cic = cfq_cic_lookup(cfqd, ioc);
2720         /* cic always exists here */
2721         cfqq = cic_to_cfqq(cic, is_sync);
2722
2723         /*
2724          * Always try a new alloc if we fell back to the OOM cfqq
2725          * originally, since it should just be a temporary situation.
2726          */
2727         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2728                 cfqq = NULL;
2729                 if (new_cfqq) {
2730                         cfqq = new_cfqq;
2731                         new_cfqq = NULL;
2732                 } else if (gfp_mask & __GFP_WAIT) {
2733                         spin_unlock_irq(cfqd->queue->queue_lock);
2734                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2735                                         gfp_mask | __GFP_ZERO,
2736                                         cfqd->queue->node);
2737                         spin_lock_irq(cfqd->queue->queue_lock);
2738                         if (new_cfqq)
2739                                 goto retry;
2740                 } else {
2741                         cfqq = kmem_cache_alloc_node(cfq_pool,
2742                                         gfp_mask | __GFP_ZERO,
2743                                         cfqd->queue->node);
2744                 }
2745
2746                 if (cfqq) {
2747                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2748                         cfq_init_prio_data(cfqq, ioc);
2749                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2750                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2751                 } else
2752                         cfqq = &cfqd->oom_cfqq;
2753         }
2754
2755         if (new_cfqq)
2756                 kmem_cache_free(cfq_pool, new_cfqq);
2757
2758         return cfqq;
2759 }
2760
2761 static struct cfq_queue **
2762 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2763 {
2764         switch (ioprio_class) {
2765         case IOPRIO_CLASS_RT:
2766                 return &cfqd->async_cfqq[0][ioprio];
2767         case IOPRIO_CLASS_BE:
2768                 return &cfqd->async_cfqq[1][ioprio];
2769         case IOPRIO_CLASS_IDLE:
2770                 return &cfqd->async_idle_cfqq;
2771         default:
2772                 BUG();
2773         }
2774 }
2775
2776 static struct cfq_queue *
2777 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2778               gfp_t gfp_mask)
2779 {
2780         const int ioprio = task_ioprio(ioc);
2781         const int ioprio_class = task_ioprio_class(ioc);
2782         struct cfq_queue **async_cfqq = NULL;
2783         struct cfq_queue *cfqq = NULL;
2784
2785         if (!is_sync) {
2786                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2787                 cfqq = *async_cfqq;
2788         }
2789
2790         if (!cfqq)
2791                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2792
2793         /*
2794          * pin the queue now that it's allocated, scheduler exit will prune it
2795          */
2796         if (!is_sync && !(*async_cfqq)) {
2797                 atomic_inc(&cfqq->ref);
2798                 *async_cfqq = cfqq;
2799         }
2800
2801         atomic_inc(&cfqq->ref);
2802         return cfqq;
2803 }
2804
2805 /*
2806  * We drop cfq io contexts lazily, so we may find a dead one.
2807  */
2808 static void
2809 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2810                   struct cfq_io_context *cic)
2811 {
2812         unsigned long flags;
2813
2814         WARN_ON(!list_empty(&cic->queue_list));
2815
2816         spin_lock_irqsave(&ioc->lock, flags);
2817
2818         BUG_ON(ioc->ioc_data == cic);
2819
2820         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2821         hlist_del_rcu(&cic->cic_list);
2822         spin_unlock_irqrestore(&ioc->lock, flags);
2823
2824         cfq_cic_free(cic);
2825 }
2826
2827 static struct cfq_io_context *
2828 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2829 {
2830         struct cfq_io_context *cic;
2831         unsigned long flags;
2832         void *k;
2833
2834         if (unlikely(!ioc))
2835                 return NULL;
2836
2837         rcu_read_lock();
2838
2839         /*
2840          * we maintain a last-hit cache, to avoid browsing over the tree
2841          */
2842         cic = rcu_dereference(ioc->ioc_data);
2843         if (cic && cic->key == cfqd) {
2844                 rcu_read_unlock();
2845                 return cic;
2846         }
2847
2848         do {
2849                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2850                 rcu_read_unlock();
2851                 if (!cic)
2852                         break;
2853                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2854                 k = cic->key;
2855                 if (unlikely(!k)) {
2856                         cfq_drop_dead_cic(cfqd, ioc, cic);
2857                         rcu_read_lock();
2858                         continue;
2859                 }
2860
2861                 spin_lock_irqsave(&ioc->lock, flags);
2862                 rcu_assign_pointer(ioc->ioc_data, cic);
2863                 spin_unlock_irqrestore(&ioc->lock, flags);
2864                 break;
2865         } while (1);
2866
2867         return cic;
2868 }
2869
2870 /*
2871  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2872  * the process specific cfq io context when entered from the block layer.
2873  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2874  */
2875 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2876                         struct cfq_io_context *cic, gfp_t gfp_mask)
2877 {
2878         unsigned long flags;
2879         int ret;
2880
2881         ret = radix_tree_preload(gfp_mask);
2882         if (!ret) {
2883                 cic->ioc = ioc;
2884                 cic->key = cfqd;
2885
2886                 spin_lock_irqsave(&ioc->lock, flags);
2887                 ret = radix_tree_insert(&ioc->radix_root,
2888                                                 (unsigned long) cfqd, cic);
2889                 if (!ret)
2890                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2891                 spin_unlock_irqrestore(&ioc->lock, flags);
2892
2893                 radix_tree_preload_end();
2894
2895                 if (!ret) {
2896                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2897                         list_add(&cic->queue_list, &cfqd->cic_list);
2898                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2899                 }
2900         }
2901
2902         if (ret)
2903                 printk(KERN_ERR "cfq: cic link failed!\n");
2904
2905         return ret;
2906 }
2907
2908 /*
2909  * Setup general io context and cfq io context. There can be several cfq
2910  * io contexts per general io context, if this process is doing io to more
2911  * than one device managed by cfq.
2912  */
2913 static struct cfq_io_context *
2914 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2915 {
2916         struct io_context *ioc = NULL;
2917         struct cfq_io_context *cic;
2918
2919         might_sleep_if(gfp_mask & __GFP_WAIT);
2920
2921         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2922         if (!ioc)
2923                 return NULL;
2924
2925         cic = cfq_cic_lookup(cfqd, ioc);
2926         if (cic)
2927                 goto out;
2928
2929         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2930         if (cic == NULL)
2931                 goto err;
2932
2933         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2934                 goto err_free;
2935
2936 out:
2937         smp_read_barrier_depends();
2938         if (unlikely(ioc->ioprio_changed))
2939                 cfq_ioc_set_ioprio(ioc);
2940
2941 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2942         if (unlikely(ioc->cgroup_changed))
2943                 cfq_ioc_set_cgroup(ioc);
2944 #endif
2945         return cic;
2946 err_free:
2947         cfq_cic_free(cic);
2948 err:
2949         put_io_context(ioc);
2950         return NULL;
2951 }
2952
2953 static void
2954 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2955 {
2956         unsigned long elapsed = jiffies - cic->last_end_request;
2957         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2958
2959         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2960         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2961         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2962 }
2963
2964 static void
2965 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2966                        struct request *rq)
2967 {
2968         sector_t sdist = 0;
2969         sector_t n_sec = blk_rq_sectors(rq);
2970         if (cfqq->last_request_pos) {
2971                 if (cfqq->last_request_pos < blk_rq_pos(rq))
2972                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2973                 else
2974                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2975         }
2976
2977         cfqq->seek_history <<= 1;
2978         if (blk_queue_nonrot(cfqd->queue))
2979                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2980         else
2981                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2982 }
2983
2984 /*
2985  * Disable idle window if the process thinks too long or seeks so much that
2986  * it doesn't matter
2987  */
2988 static void
2989 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2990                        struct cfq_io_context *cic)
2991 {
2992         int old_idle, enable_idle;
2993
2994         /*
2995          * Don't idle for async or idle io prio class
2996          */
2997         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2998                 return;
2999
3000         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3001
3002         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3003                 cfq_mark_cfqq_deep(cfqq);
3004
3005         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3006             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3007                 enable_idle = 0;
3008         else if (sample_valid(cic->ttime_samples)) {
3009                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3010                         enable_idle = 0;
3011                 else
3012                         enable_idle = 1;
3013         }
3014
3015         if (old_idle != enable_idle) {
3016                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3017                 if (enable_idle)
3018                         cfq_mark_cfqq_idle_window(cfqq);
3019                 else
3020                         cfq_clear_cfqq_idle_window(cfqq);
3021         }
3022 }
3023
3024 /*
3025  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3026  * no or if we aren't sure, a 1 will cause a preempt.
3027  */
3028 static bool
3029 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3030                    struct request *rq)
3031 {
3032         struct cfq_queue *cfqq;
3033
3034         cfqq = cfqd->active_queue;
3035         if (!cfqq)
3036                 return false;
3037
3038         if (cfq_class_idle(new_cfqq))
3039                 return false;
3040
3041         if (cfq_class_idle(cfqq))
3042                 return true;
3043
3044         /*
3045          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3046          */
3047         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3048                 return false;
3049
3050         /*
3051          * if the new request is sync, but the currently running queue is
3052          * not, let the sync request have priority.
3053          */
3054         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3055                 return true;
3056
3057         if (new_cfqq->cfqg != cfqq->cfqg)
3058                 return false;
3059
3060         if (cfq_slice_used(cfqq))
3061                 return true;
3062
3063         /* Allow preemption only if we are idling on sync-noidle tree */
3064         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3065             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3066             new_cfqq->service_tree->count == 2 &&
3067             RB_EMPTY_ROOT(&cfqq->sort_list))
3068                 return true;
3069
3070         /*
3071          * So both queues are sync. Let the new request get disk time if
3072          * it's a metadata request and the current queue is doing regular IO.
3073          */
3074         if (rq_is_meta(rq) && !cfqq->meta_pending)
3075                 return true;
3076
3077         /*
3078          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3079          */
3080         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3081                 return true;
3082
3083         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3084                 return false;
3085
3086         /*
3087          * if this request is as-good as one we would expect from the
3088          * current cfqq, let it preempt
3089          */
3090         if (cfq_rq_close(cfqd, cfqq, rq, true))
3091                 return true;
3092
3093         return false;
3094 }
3095
3096 /*
3097  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3098  * let it have half of its nominal slice.
3099  */
3100 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3101 {
3102         cfq_log_cfqq(cfqd, cfqq, "preempt");
3103         cfq_slice_expired(cfqd, 1);
3104
3105         /*
3106          * Put the new queue at the front of the of the current list,
3107          * so we know that it will be selected next.
3108          */
3109         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3110
3111         cfq_service_tree_add(cfqd, cfqq, 1);
3112
3113         cfqq->slice_end = 0;
3114         cfq_mark_cfqq_slice_new(cfqq);
3115 }
3116
3117 /*
3118  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3119  * something we should do about it
3120  */
3121 static void
3122 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3123                 struct request *rq)
3124 {
3125         struct cfq_io_context *cic = RQ_CIC(rq);
3126
3127         cfqd->rq_queued++;
3128         if (rq_is_meta(rq))
3129                 cfqq->meta_pending++;
3130
3131         cfq_update_io_thinktime(cfqd, cic);
3132         cfq_update_io_seektime(cfqd, cfqq, rq);
3133         cfq_update_idle_window(cfqd, cfqq, cic);
3134
3135         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3136
3137         if (cfqq == cfqd->active_queue) {
3138                 /*
3139                  * Remember that we saw a request from this process, but
3140                  * don't start queuing just yet. Otherwise we risk seeing lots
3141                  * of tiny requests, because we disrupt the normal plugging
3142                  * and merging. If the request is already larger than a single
3143                  * page, let it rip immediately. For that case we assume that
3144                  * merging is already done. Ditto for a busy system that
3145                  * has other work pending, don't risk delaying until the
3146                  * idle timer unplug to continue working.
3147                  */
3148                 if (cfq_cfqq_wait_request(cfqq)) {
3149                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3150                             cfqd->busy_queues > 1) {
3151                                 del_timer(&cfqd->idle_slice_timer);
3152                                 cfq_clear_cfqq_wait_request(cfqq);
3153                                 __blk_run_queue(cfqd->queue);
3154                         } else
3155                                 cfq_mark_cfqq_must_dispatch(cfqq);
3156                 }
3157         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3158                 /*
3159                  * not the active queue - expire current slice if it is
3160                  * idle and has expired it's mean thinktime or this new queue
3161                  * has some old slice time left and is of higher priority or
3162                  * this new queue is RT and the current one is BE
3163                  */
3164                 cfq_preempt_queue(cfqd, cfqq);
3165                 __blk_run_queue(cfqd->queue);
3166         }
3167 }
3168
3169 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3170 {
3171         struct cfq_data *cfqd = q->elevator->elevator_data;
3172         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3173
3174         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3175         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3176
3177         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3178         list_add_tail(&rq->queuelist, &cfqq->fifo);
3179         cfq_add_rq_rb(rq);
3180
3181         cfq_rq_enqueued(cfqd, cfqq, rq);
3182 }
3183
3184 /*
3185  * Update hw_tag based on peak queue depth over 50 samples under
3186  * sufficient load.
3187  */
3188 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3189 {
3190         struct cfq_queue *cfqq = cfqd->active_queue;
3191
3192         if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3193                 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3194
3195         if (cfqd->hw_tag == 1)
3196                 return;
3197
3198         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3199             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3200                 return;
3201
3202         /*
3203          * If active queue hasn't enough requests and can idle, cfq might not
3204          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3205          * case
3206          */
3207         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3208             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3209             CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3210                 return;
3211
3212         if (cfqd->hw_tag_samples++ < 50)
3213                 return;
3214
3215         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3216                 cfqd->hw_tag = 1;
3217         else
3218                 cfqd->hw_tag = 0;
3219 }
3220
3221 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3222 {
3223         struct cfq_io_context *cic = cfqd->active_cic;
3224
3225         /* If there are other queues in the group, don't wait */
3226         if (cfqq->cfqg->nr_cfqq > 1)
3227                 return false;
3228
3229         if (cfq_slice_used(cfqq))
3230                 return true;
3231
3232         /* if slice left is less than think time, wait busy */
3233         if (cic && sample_valid(cic->ttime_samples)
3234             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3235                 return true;
3236
3237         /*
3238          * If think times is less than a jiffy than ttime_mean=0 and above
3239          * will not be true. It might happen that slice has not expired yet
3240          * but will expire soon (4-5 ns) during select_queue(). To cover the
3241          * case where think time is less than a jiffy, mark the queue wait
3242          * busy if only 1 jiffy is left in the slice.
3243          */
3244         if (cfqq->slice_end - jiffies == 1)
3245                 return true;
3246
3247         return false;
3248 }
3249
3250 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3251 {
3252         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3253         struct cfq_data *cfqd = cfqq->cfqd;
3254         const int sync = rq_is_sync(rq);
3255         unsigned long now;
3256
3257         now = jiffies;
3258         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3259
3260         cfq_update_hw_tag(cfqd);
3261
3262         WARN_ON(!cfqd->rq_in_driver[sync]);
3263         WARN_ON(!cfqq->dispatched);
3264         cfqd->rq_in_driver[sync]--;
3265         cfqq->dispatched--;
3266
3267         if (cfq_cfqq_sync(cfqq))
3268                 cfqd->sync_flight--;
3269
3270         if (sync) {
3271                 RQ_CIC(rq)->last_end_request = now;
3272                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3273                         cfqd->last_delayed_sync = now;
3274         }
3275
3276         /*
3277          * If this is the active queue, check if it needs to be expired,
3278          * or if we want to idle in case it has no pending requests.
3279          */
3280         if (cfqd->active_queue == cfqq) {
3281                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3282
3283                 if (cfq_cfqq_slice_new(cfqq)) {
3284                         cfq_set_prio_slice(cfqd, cfqq);
3285                         cfq_clear_cfqq_slice_new(cfqq);
3286                 }
3287
3288                 /*
3289                  * Should we wait for next request to come in before we expire
3290                  * the queue.
3291                  */
3292                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3293                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3294                         cfq_mark_cfqq_wait_busy(cfqq);
3295                 }
3296
3297                 /*
3298                  * Idling is not enabled on:
3299                  * - expired queues
3300                  * - idle-priority queues
3301                  * - async queues
3302                  * - queues with still some requests queued
3303                  * - when there is a close cooperator
3304                  */
3305                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3306                         cfq_slice_expired(cfqd, 1);
3307                 else if (sync && cfqq_empty &&
3308                          !cfq_close_cooperator(cfqd, cfqq)) {
3309                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3310                         /*
3311                          * Idling is enabled for SYNC_WORKLOAD.
3312                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3313                          * only if we processed at least one !rq_noidle request
3314                          */
3315                         if (cfqd->serving_type == SYNC_WORKLOAD
3316                             || cfqd->noidle_tree_requires_idle
3317                             || cfqq->cfqg->nr_cfqq == 1)
3318                                 cfq_arm_slice_timer(cfqd);
3319                 }
3320         }
3321
3322         if (!rq_in_driver(cfqd))
3323                 cfq_schedule_dispatch(cfqd);
3324 }
3325
3326 /*
3327  * we temporarily boost lower priority queues if they are holding fs exclusive
3328  * resources. they are boosted to normal prio (CLASS_BE/4)
3329  */
3330 static void cfq_prio_boost(struct cfq_queue *cfqq)
3331 {
3332         if (has_fs_excl()) {
3333                 /*
3334                  * boost idle prio on transactions that would lock out other
3335                  * users of the filesystem
3336                  */
3337                 if (cfq_class_idle(cfqq))
3338                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3339                 if (cfqq->ioprio > IOPRIO_NORM)
3340                         cfqq->ioprio = IOPRIO_NORM;
3341         } else {
3342                 /*
3343                  * unboost the queue (if needed)
3344                  */
3345                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3346                 cfqq->ioprio = cfqq->org_ioprio;
3347         }
3348 }
3349
3350 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3351 {
3352         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3353                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3354                 return ELV_MQUEUE_MUST;
3355         }
3356
3357         return ELV_MQUEUE_MAY;
3358 }
3359
3360 static int cfq_may_queue(struct request_queue *q, int rw)
3361 {
3362         struct cfq_data *cfqd = q->elevator->elevator_data;
3363         struct task_struct *tsk = current;
3364         struct cfq_io_context *cic;
3365         struct cfq_queue *cfqq;
3366
3367         /*
3368          * don't force setup of a queue from here, as a call to may_queue
3369          * does not necessarily imply that a request actually will be queued.
3370          * so just lookup a possibly existing queue, or return 'may queue'
3371          * if that fails
3372          */
3373         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3374         if (!cic)
3375                 return ELV_MQUEUE_MAY;
3376
3377         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3378         if (cfqq) {
3379                 cfq_init_prio_data(cfqq, cic->ioc);
3380                 cfq_prio_boost(cfqq);
3381
3382                 return __cfq_may_queue(cfqq);
3383         }
3384
3385         return ELV_MQUEUE_MAY;
3386 }
3387
3388 /*
3389  * queue lock held here
3390  */
3391 static void cfq_put_request(struct request *rq)
3392 {
3393         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3394
3395         if (cfqq) {
3396                 const int rw = rq_data_dir(rq);
3397
3398                 BUG_ON(!cfqq->allocated[rw]);
3399                 cfqq->allocated[rw]--;
3400
3401                 put_io_context(RQ_CIC(rq)->ioc);
3402
3403                 rq->elevator_private = NULL;
3404                 rq->elevator_private2 = NULL;
3405
3406                 cfq_put_queue(cfqq);
3407         }
3408 }
3409
3410 static struct cfq_queue *
3411 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3412                 struct cfq_queue *cfqq)
3413 {
3414         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3415         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3416         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3417         cfq_put_queue(cfqq);
3418         return cic_to_cfqq(cic, 1);
3419 }
3420
3421 /*
3422  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3423  * was the last process referring to said cfqq.
3424  */
3425 static struct cfq_queue *
3426 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3427 {
3428         if (cfqq_process_refs(cfqq) == 1) {
3429                 cfqq->pid = current->pid;
3430                 cfq_clear_cfqq_coop(cfqq);
3431                 cfq_clear_cfqq_split_coop(cfqq);
3432                 return cfqq;
3433         }
3434
3435         cic_set_cfqq(cic, NULL, 1);
3436         cfq_put_queue(cfqq);
3437         return NULL;
3438 }
3439 /*
3440  * Allocate cfq data structures associated with this request.
3441  */
3442 static int
3443 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3444 {
3445         struct cfq_data *cfqd = q->elevator->elevator_data;
3446         struct cfq_io_context *cic;
3447         const int rw = rq_data_dir(rq);
3448         const bool is_sync = rq_is_sync(rq);
3449         struct cfq_queue *cfqq;
3450         unsigned long flags;
3451
3452         might_sleep_if(gfp_mask & __GFP_WAIT);
3453
3454         cic = cfq_get_io_context(cfqd, gfp_mask);
3455
3456         spin_lock_irqsave(q->queue_lock, flags);
3457
3458         if (!cic)
3459                 goto queue_fail;
3460
3461 new_queue:
3462         cfqq = cic_to_cfqq(cic, is_sync);
3463         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3464                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3465                 cic_set_cfqq(cic, cfqq, is_sync);
3466         } else {
3467                 /*
3468                  * If the queue was seeky for too long, break it apart.
3469                  */
3470                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3471                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3472                         cfqq = split_cfqq(cic, cfqq);
3473                         if (!cfqq)
3474                                 goto new_queue;
3475                 }
3476
3477                 /*
3478                  * Check to see if this queue is scheduled to merge with
3479                  * another, closely cooperating queue.  The merging of
3480                  * queues happens here as it must be done in process context.
3481                  * The reference on new_cfqq was taken in merge_cfqqs.
3482                  */
3483                 if (cfqq->new_cfqq)
3484                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3485         }
3486
3487         cfqq->allocated[rw]++;
3488         atomic_inc(&cfqq->ref);
3489
3490         spin_unlock_irqrestore(q->queue_lock, flags);
3491
3492         rq->elevator_private = cic;
3493         rq->elevator_private2 = cfqq;
3494         return 0;
3495
3496 queue_fail:
3497         if (cic)
3498                 put_io_context(cic->ioc);
3499
3500         cfq_schedule_dispatch(cfqd);
3501         spin_unlock_irqrestore(q->queue_lock, flags);
3502         cfq_log(cfqd, "set_request fail");
3503         return 1;
3504 }
3505
3506 static void cfq_kick_queue(struct work_struct *work)
3507 {
3508         struct cfq_data *cfqd =
3509                 container_of(work, struct cfq_data, unplug_work);
3510         struct request_queue *q = cfqd->queue;
3511
3512         spin_lock_irq(q->queue_lock);
3513         __blk_run_queue(cfqd->queue);
3514         spin_unlock_irq(q->queue_lock);
3515 }
3516
3517 /*
3518  * Timer running if the active_queue is currently idling inside its time slice
3519  */
3520 static void cfq_idle_slice_timer(unsigned long data)
3521 {
3522         struct cfq_data *cfqd = (struct cfq_data *) data;
3523         struct cfq_queue *cfqq;
3524         unsigned long flags;
3525         int timed_out = 1;
3526
3527         cfq_log(cfqd, "idle timer fired");
3528
3529         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3530
3531         cfqq = cfqd->active_queue;
3532         if (cfqq) {
3533                 timed_out = 0;
3534
3535                 /*
3536                  * We saw a request before the queue expired, let it through
3537                  */
3538                 if (cfq_cfqq_must_dispatch(cfqq))
3539                         goto out_kick;
3540
3541                 /*
3542                  * expired
3543                  */
3544                 if (cfq_slice_used(cfqq))
3545                         goto expire;
3546
3547                 /*
3548                  * only expire and reinvoke request handler, if there are
3549                  * other queues with pending requests
3550                  */
3551                 if (!cfqd->busy_queues)
3552                         goto out_cont;
3553
3554                 /*
3555                  * not expired and it has a request pending, let it dispatch
3556                  */
3557                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3558                         goto out_kick;
3559
3560                 /*
3561                  * Queue depth flag is reset only when the idle didn't succeed
3562                  */
3563                 cfq_clear_cfqq_deep(cfqq);
3564         }
3565 expire:
3566         cfq_slice_expired(cfqd, timed_out);
3567 out_kick:
3568         cfq_schedule_dispatch(cfqd);
3569 out_cont:
3570         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3571 }
3572
3573 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3574 {
3575         del_timer_sync(&cfqd->idle_slice_timer);
3576         cancel_work_sync(&cfqd->unplug_work);
3577 }
3578
3579 static void cfq_put_async_queues(struct cfq_data *cfqd)
3580 {
3581         int i;
3582
3583         for (i = 0; i < IOPRIO_BE_NR; i++) {
3584                 if (cfqd->async_cfqq[0][i])
3585                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3586                 if (cfqd->async_cfqq[1][i])
3587                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3588         }
3589
3590         if (cfqd->async_idle_cfqq)
3591                 cfq_put_queue(cfqd->async_idle_cfqq);
3592 }
3593
3594 static void cfq_cfqd_free(struct rcu_head *head)
3595 {
3596         kfree(container_of(head, struct cfq_data, rcu));
3597 }
3598
3599 static void cfq_exit_queue(struct elevator_queue *e)
3600 {
3601         struct cfq_data *cfqd = e->elevator_data;
3602         struct request_queue *q = cfqd->queue;
3603
3604         cfq_shutdown_timer_wq(cfqd);
3605
3606         spin_lock_irq(q->queue_lock);
3607
3608         if (cfqd->active_queue)
3609                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3610
3611         while (!list_empty(&cfqd->cic_list)) {
3612                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3613                                                         struct cfq_io_context,
3614                                                         queue_list);
3615
3616                 __cfq_exit_single_io_context(cfqd, cic);
3617         }
3618
3619         cfq_put_async_queues(cfqd);
3620         cfq_release_cfq_groups(cfqd);
3621         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3622
3623         spin_unlock_irq(q->queue_lock);
3624
3625         cfq_shutdown_timer_wq(cfqd);
3626
3627         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3628         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3629 }
3630
3631 static void *cfq_init_queue(struct request_queue *q)
3632 {
3633         struct cfq_data *cfqd;
3634         int i, j;
3635         struct cfq_group *cfqg;
3636         struct cfq_rb_root *st;
3637
3638         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3639         if (!cfqd)
3640                 return NULL;
3641
3642         /* Init root service tree */
3643         cfqd->grp_service_tree = CFQ_RB_ROOT;
3644
3645         /* Init root group */
3646         cfqg = &cfqd->root_group;
3647         for_each_cfqg_st(cfqg, i, j, st)
3648                 *st = CFQ_RB_ROOT;
3649         RB_CLEAR_NODE(&cfqg->rb_node);
3650
3651         /* Give preference to root group over other groups */
3652         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3653
3654 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3655         /*
3656          * Take a reference to root group which we never drop. This is just
3657          * to make sure that cfq_put_cfqg() does not try to kfree root group
3658          */
3659         atomic_set(&cfqg->ref, 1);
3660         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3661                                         0);
3662 #endif
3663         /*
3664          * Not strictly needed (since RB_ROOT just clears the node and we
3665          * zeroed cfqd on alloc), but better be safe in case someone decides
3666          * to add magic to the rb code
3667          */
3668         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3669                 cfqd->prio_trees[i] = RB_ROOT;
3670
3671         /*
3672          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3673          * Grab a permanent reference to it, so that the normal code flow
3674          * will not attempt to free it.
3675          */
3676         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3677         atomic_inc(&cfqd->oom_cfqq.ref);
3678         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3679
3680         INIT_LIST_HEAD(&cfqd->cic_list);
3681
3682         cfqd->queue = q;
3683
3684         init_timer(&cfqd->idle_slice_timer);
3685         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3686         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3687
3688         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3689
3690         cfqd->cfq_quantum = cfq_quantum;
3691         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3692         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3693         cfqd->cfq_back_max = cfq_back_max;
3694         cfqd->cfq_back_penalty = cfq_back_penalty;
3695         cfqd->cfq_slice[0] = cfq_slice_async;
3696         cfqd->cfq_slice[1] = cfq_slice_sync;
3697         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3698         cfqd->cfq_slice_idle = cfq_slice_idle;
3699         cfqd->cfq_latency = 1;
3700         cfqd->cfq_group_isolation = 0;
3701         cfqd->hw_tag = -1;
3702         /*
3703          * we optimistically start assuming sync ops weren't delayed in last
3704          * second, in order to have larger depth for async operations.
3705          */
3706         cfqd->last_delayed_sync = jiffies - HZ;
3707         INIT_RCU_HEAD(&cfqd->rcu);
3708         return cfqd;
3709 }
3710
3711 static void cfq_slab_kill(void)
3712 {
3713         /*
3714          * Caller already ensured that pending RCU callbacks are completed,
3715          * so we should have no busy allocations at this point.
3716          */
3717         if (cfq_pool)
3718                 kmem_cache_destroy(cfq_pool);
3719         if (cfq_ioc_pool)
3720                 kmem_cache_destroy(cfq_ioc_pool);
3721 }
3722
3723 static int __init cfq_slab_setup(void)
3724 {
3725         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3726         if (!cfq_pool)
3727                 goto fail;
3728
3729         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3730         if (!cfq_ioc_pool)
3731                 goto fail;
3732
3733         return 0;
3734 fail:
3735         cfq_slab_kill();
3736         return -ENOMEM;
3737 }
3738
3739 /*
3740  * sysfs parts below -->
3741  */
3742 static ssize_t
3743 cfq_var_show(unsigned int var, char *page)
3744 {
3745         return sprintf(page, "%d\n", var);
3746 }
3747
3748 static ssize_t
3749 cfq_var_store(unsigned int *var, const char *page, size_t count)
3750 {
3751         char *p = (char *) page;
3752
3753         *var = simple_strtoul(p, &p, 10);
3754         return count;
3755 }
3756
3757 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3758 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3759 {                                                                       \
3760         struct cfq_data *cfqd = e->elevator_data;                       \
3761         unsigned int __data = __VAR;                                    \
3762         if (__CONV)                                                     \
3763                 __data = jiffies_to_msecs(__data);                      \
3764         return cfq_var_show(__data, (page));                            \
3765 }
3766 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3767 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3768 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3769 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3770 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3771 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3772 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3773 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3774 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3775 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3776 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3777 #undef SHOW_FUNCTION
3778
3779 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3780 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3781 {                                                                       \
3782         struct cfq_data *cfqd = e->elevator_data;                       \
3783         unsigned int __data;                                            \
3784         int ret = cfq_var_store(&__data, (page), count);                \
3785         if (__data < (MIN))                                             \
3786                 __data = (MIN);                                         \
3787         else if (__data > (MAX))                                        \
3788                 __data = (MAX);                                         \
3789         if (__CONV)                                                     \
3790                 *(__PTR) = msecs_to_jiffies(__data);                    \
3791         else                                                            \
3792                 *(__PTR) = __data;                                      \
3793         return ret;                                                     \
3794 }
3795 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3796 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3797                 UINT_MAX, 1);
3798 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3799                 UINT_MAX, 1);
3800 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3801 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3802                 UINT_MAX, 0);
3803 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3804 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3805 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3806 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3807                 UINT_MAX, 0);
3808 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3809 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3810 #undef STORE_FUNCTION
3811
3812 #define CFQ_ATTR(name) \
3813         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3814
3815 static struct elv_fs_entry cfq_attrs[] = {
3816         CFQ_ATTR(quantum),
3817         CFQ_ATTR(fifo_expire_sync),
3818         CFQ_ATTR(fifo_expire_async),
3819         CFQ_ATTR(back_seek_max),
3820         CFQ_ATTR(back_seek_penalty),
3821         CFQ_ATTR(slice_sync),
3822         CFQ_ATTR(slice_async),
3823         CFQ_ATTR(slice_async_rq),
3824         CFQ_ATTR(slice_idle),
3825         CFQ_ATTR(low_latency),
3826         CFQ_ATTR(group_isolation),
3827         __ATTR_NULL
3828 };
3829
3830 static struct elevator_type iosched_cfq = {
3831         .ops = {
3832                 .elevator_merge_fn =            cfq_merge,
3833                 .elevator_merged_fn =           cfq_merged_request,
3834                 .elevator_merge_req_fn =        cfq_merged_requests,
3835                 .elevator_allow_merge_fn =      cfq_allow_merge,
3836                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3837                 .elevator_add_req_fn =          cfq_insert_request,
3838                 .elevator_activate_req_fn =     cfq_activate_request,
3839                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3840                 .elevator_queue_empty_fn =      cfq_queue_empty,
3841                 .elevator_completed_req_fn =    cfq_completed_request,
3842                 .elevator_former_req_fn =       elv_rb_former_request,
3843                 .elevator_latter_req_fn =       elv_rb_latter_request,
3844                 .elevator_set_req_fn =          cfq_set_request,
3845                 .elevator_put_req_fn =          cfq_put_request,
3846                 .elevator_may_queue_fn =        cfq_may_queue,
3847                 .elevator_init_fn =             cfq_init_queue,
3848                 .elevator_exit_fn =             cfq_exit_queue,
3849                 .trim =                         cfq_free_io_context,
3850         },
3851         .elevator_attrs =       cfq_attrs,
3852         .elevator_name =        "cfq",
3853         .elevator_owner =       THIS_MODULE,
3854 };
3855
3856 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3857 static struct blkio_policy_type blkio_policy_cfq = {
3858         .ops = {
3859                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3860                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3861         },
3862 };
3863 #else
3864 static struct blkio_policy_type blkio_policy_cfq;
3865 #endif
3866
3867 static int __init cfq_init(void)
3868 {
3869         /*
3870          * could be 0 on HZ < 1000 setups
3871          */
3872         if (!cfq_slice_async)
3873                 cfq_slice_async = 1;
3874         if (!cfq_slice_idle)
3875                 cfq_slice_idle = 1;
3876
3877         if (cfq_slab_setup())
3878                 return -ENOMEM;
3879
3880         elv_register(&iosched_cfq);
3881         blkio_policy_register(&blkio_policy_cfq);
3882
3883         return 0;
3884 }
3885
3886 static void __exit cfq_exit(void)
3887 {
3888         DECLARE_COMPLETION_ONSTACK(all_gone);
3889         blkio_policy_unregister(&blkio_policy_cfq);
3890         elv_unregister(&iosched_cfq);
3891         ioc_gone = &all_gone;
3892         /* ioc_gone's update must be visible before reading ioc_count */
3893         smp_wmb();
3894
3895         /*
3896          * this also protects us from entering cfq_slab_kill() with
3897          * pending RCU callbacks
3898          */
3899         if (elv_ioc_count_read(cfq_ioc_count))
3900                 wait_for_completion(&all_gone);
3901         cfq_slab_kill();
3902 }
3903
3904 module_init(cfq_init);
3905 module_exit(cfq_exit);
3906
3907 MODULE_AUTHOR("Jens Axboe");
3908 MODULE_LICENSE("GPL");
3909 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");