blk-cgroup: Prepare the base for supporting more than one IO control policies
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
35
36 /*
37  * offset from end of service tree
38  */
39 #define CFQ_IDLE_DELAY          (HZ / 5)
40
41 /*
42  * below this threshold, we consider thinktime immediate
43  */
44 #define CFQ_MIN_TT              (2)
45
46 #define CFQ_SLICE_SCALE         (5)
47 #define CFQ_HW_QUEUE_MIN        (5)
48 #define CFQ_SERVICE_SHIFT       12
49
50 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
54
55 #define RQ_CIC(rq)              \
56         ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
59
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
62
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
66
67 static DEFINE_SPINLOCK(cic_index_lock);
68 static DEFINE_IDA(cic_index_ida);
69
70 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
71 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
72 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73
74 #define sample_valid(samples)   ((samples) > 80)
75 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
76
77 /*
78  * Most of our rbtree usage is for sorting with min extraction, so
79  * if we cache the leftmost node we don't have to walk down the tree
80  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81  * move this into the elevator for the rq sorting as well.
82  */
83 struct cfq_rb_root {
84         struct rb_root rb;
85         struct rb_node *left;
86         unsigned count;
87         unsigned total_weight;
88         u64 min_vdisktime;
89         struct rb_node *active;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         atomic_t ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         struct cfq_group *orig_cfqg;
150 };
151
152 /*
153  * First index in the service_trees.
154  * IDLE is handled separately, so it has negative index
155  */
156 enum wl_prio_t {
157         BE_WORKLOAD = 0,
158         RT_WORKLOAD = 1,
159         IDLE_WORKLOAD = 2,
160 };
161
162 /*
163  * Second index in the service_trees.
164  */
165 enum wl_type_t {
166         ASYNC_WORKLOAD = 0,
167         SYNC_NOIDLE_WORKLOAD = 1,
168         SYNC_WORKLOAD = 2
169 };
170
171 /* This is per cgroup per device grouping structure */
172 struct cfq_group {
173         /* group service_tree member */
174         struct rb_node rb_node;
175
176         /* group service_tree key */
177         u64 vdisktime;
178         unsigned int weight;
179         bool on_st;
180
181         /* number of cfqq currently on this group */
182         int nr_cfqq;
183
184         /* Per group busy queus average. Useful for workload slice calc. */
185         unsigned int busy_queues_avg[2];
186         /*
187          * rr lists of queues with requests, onle rr for each priority class.
188          * Counts are embedded in the cfq_rb_root
189          */
190         struct cfq_rb_root service_trees[2][3];
191         struct cfq_rb_root service_tree_idle;
192
193         unsigned long saved_workload_slice;
194         enum wl_type_t saved_workload;
195         enum wl_prio_t saved_serving_prio;
196         struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198         struct hlist_node cfqd_node;
199         atomic_t ref;
200 #endif
201 };
202
203 /*
204  * Per block device queue structure
205  */
206 struct cfq_data {
207         struct request_queue *queue;
208         /* Root service tree for cfq_groups */
209         struct cfq_rb_root grp_service_tree;
210         struct cfq_group root_group;
211
212         /*
213          * The priority currently being served
214          */
215         enum wl_prio_t serving_prio;
216         enum wl_type_t serving_type;
217         unsigned long workload_expires;
218         struct cfq_group *serving_group;
219         bool noidle_tree_requires_idle;
220
221         /*
222          * Each priority tree is sorted by next_request position.  These
223          * trees are used when determining if two or more queues are
224          * interleaving requests (see cfq_close_cooperator).
225          */
226         struct rb_root prio_trees[CFQ_PRIO_LISTS];
227
228         unsigned int busy_queues;
229
230         int rq_in_driver;
231         int rq_in_flight[2];
232
233         /*
234          * queue-depth detection
235          */
236         int rq_queued;
237         int hw_tag;
238         /*
239          * hw_tag can be
240          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
241          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
242          *  0 => no NCQ
243          */
244         int hw_tag_est_depth;
245         unsigned int hw_tag_samples;
246
247         /*
248          * idle window management
249          */
250         struct timer_list idle_slice_timer;
251         struct work_struct unplug_work;
252
253         struct cfq_queue *active_queue;
254         struct cfq_io_context *active_cic;
255
256         /*
257          * async queue for each priority case
258          */
259         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
260         struct cfq_queue *async_idle_cfqq;
261
262         sector_t last_position;
263
264         /*
265          * tunables, see top of file
266          */
267         unsigned int cfq_quantum;
268         unsigned int cfq_fifo_expire[2];
269         unsigned int cfq_back_penalty;
270         unsigned int cfq_back_max;
271         unsigned int cfq_slice[2];
272         unsigned int cfq_slice_async_rq;
273         unsigned int cfq_slice_idle;
274         unsigned int cfq_latency;
275         unsigned int cfq_group_isolation;
276
277         unsigned int cic_index;
278         struct list_head cic_list;
279
280         /*
281          * Fallback dummy cfqq for extreme OOM conditions
282          */
283         struct cfq_queue oom_cfqq;
284
285         unsigned long last_delayed_sync;
286
287         /* List of cfq groups being managed on this device*/
288         struct hlist_head cfqg_list;
289         struct rcu_head rcu;
290 };
291
292 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
293
294 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295                                             enum wl_prio_t prio,
296                                             enum wl_type_t type)
297 {
298         if (!cfqg)
299                 return NULL;
300
301         if (prio == IDLE_WORKLOAD)
302                 return &cfqg->service_tree_idle;
303
304         return &cfqg->service_trees[prio][type];
305 }
306
307 enum cfqq_state_flags {
308         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
309         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
310         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
311         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
312         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
313         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
314         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
315         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
316         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
317         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
318         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
319         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
320         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
321 };
322
323 #define CFQ_CFQQ_FNS(name)                                              \
324 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
325 {                                                                       \
326         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
327 }                                                                       \
328 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
329 {                                                                       \
330         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
331 }                                                                       \
332 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
333 {                                                                       \
334         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
335 }
336
337 CFQ_CFQQ_FNS(on_rr);
338 CFQ_CFQQ_FNS(wait_request);
339 CFQ_CFQQ_FNS(must_dispatch);
340 CFQ_CFQQ_FNS(must_alloc_slice);
341 CFQ_CFQQ_FNS(fifo_expire);
342 CFQ_CFQQ_FNS(idle_window);
343 CFQ_CFQQ_FNS(prio_changed);
344 CFQ_CFQQ_FNS(slice_new);
345 CFQ_CFQQ_FNS(sync);
346 CFQ_CFQQ_FNS(coop);
347 CFQ_CFQQ_FNS(split_coop);
348 CFQ_CFQQ_FNS(deep);
349 CFQ_CFQQ_FNS(wait_busy);
350 #undef CFQ_CFQQ_FNS
351
352 #ifdef CONFIG_CFQ_GROUP_IOSCHED
353 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
354         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
355                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
356                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
357
358 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
359         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
360                                 blkg_path(&(cfqg)->blkg), ##args);      \
361
362 #else
363 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
364         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
365 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
366 #endif
367 #define cfq_log(cfqd, fmt, args...)     \
368         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369
370 /* Traverses through cfq group service trees */
371 #define for_each_cfqg_st(cfqg, i, j, st) \
372         for (i = 0; i <= IDLE_WORKLOAD; i++) \
373                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
374                         : &cfqg->service_tree_idle; \
375                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
376                         (i == IDLE_WORKLOAD && j == 0); \
377                         j++, st = i < IDLE_WORKLOAD ? \
378                         &cfqg->service_trees[i][j]: NULL) \
379
380
381 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382 {
383         if (cfq_class_idle(cfqq))
384                 return IDLE_WORKLOAD;
385         if (cfq_class_rt(cfqq))
386                 return RT_WORKLOAD;
387         return BE_WORKLOAD;
388 }
389
390
391 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392 {
393         if (!cfq_cfqq_sync(cfqq))
394                 return ASYNC_WORKLOAD;
395         if (!cfq_cfqq_idle_window(cfqq))
396                 return SYNC_NOIDLE_WORKLOAD;
397         return SYNC_WORKLOAD;
398 }
399
400 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
401                                         struct cfq_data *cfqd,
402                                         struct cfq_group *cfqg)
403 {
404         if (wl == IDLE_WORKLOAD)
405                 return cfqg->service_tree_idle.count;
406
407         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
408                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
409                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
410 }
411
412 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
413                                         struct cfq_group *cfqg)
414 {
415         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
416                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
417 }
418
419 static void cfq_dispatch_insert(struct request_queue *, struct request *);
420 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
421                                        struct io_context *, gfp_t);
422 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
423                                                 struct io_context *);
424
425 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
426                                             bool is_sync)
427 {
428         return cic->cfqq[is_sync];
429 }
430
431 static inline void cic_set_cfqq(struct cfq_io_context *cic,
432                                 struct cfq_queue *cfqq, bool is_sync)
433 {
434         cic->cfqq[is_sync] = cfqq;
435 }
436
437 #define CIC_DEAD_KEY    1ul
438 #define CIC_DEAD_INDEX_SHIFT    1
439
440 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
441 {
442         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
443 }
444
445 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
446 {
447         struct cfq_data *cfqd = cic->key;
448
449         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
450                 return NULL;
451
452         return cfqd;
453 }
454
455 /*
456  * We regard a request as SYNC, if it's either a read or has the SYNC bit
457  * set (in which case it could also be direct WRITE).
458  */
459 static inline bool cfq_bio_sync(struct bio *bio)
460 {
461         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
462 }
463
464 /*
465  * scheduler run of queue, if there are requests pending and no one in the
466  * driver that will restart queueing
467  */
468 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
469 {
470         if (cfqd->busy_queues) {
471                 cfq_log(cfqd, "schedule dispatch");
472                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
473         }
474 }
475
476 static int cfq_queue_empty(struct request_queue *q)
477 {
478         struct cfq_data *cfqd = q->elevator->elevator_data;
479
480         return !cfqd->rq_queued;
481 }
482
483 /*
484  * Scale schedule slice based on io priority. Use the sync time slice only
485  * if a queue is marked sync and has sync io queued. A sync queue with async
486  * io only, should not get full sync slice length.
487  */
488 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
489                                  unsigned short prio)
490 {
491         const int base_slice = cfqd->cfq_slice[sync];
492
493         WARN_ON(prio >= IOPRIO_BE_NR);
494
495         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
496 }
497
498 static inline int
499 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
500 {
501         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
502 }
503
504 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
505 {
506         u64 d = delta << CFQ_SERVICE_SHIFT;
507
508         d = d * BLKIO_WEIGHT_DEFAULT;
509         do_div(d, cfqg->weight);
510         return d;
511 }
512
513 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
514 {
515         s64 delta = (s64)(vdisktime - min_vdisktime);
516         if (delta > 0)
517                 min_vdisktime = vdisktime;
518
519         return min_vdisktime;
520 }
521
522 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
523 {
524         s64 delta = (s64)(vdisktime - min_vdisktime);
525         if (delta < 0)
526                 min_vdisktime = vdisktime;
527
528         return min_vdisktime;
529 }
530
531 static void update_min_vdisktime(struct cfq_rb_root *st)
532 {
533         u64 vdisktime = st->min_vdisktime;
534         struct cfq_group *cfqg;
535
536         if (st->active) {
537                 cfqg = rb_entry_cfqg(st->active);
538                 vdisktime = cfqg->vdisktime;
539         }
540
541         if (st->left) {
542                 cfqg = rb_entry_cfqg(st->left);
543                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
544         }
545
546         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
547 }
548
549 /*
550  * get averaged number of queues of RT/BE priority.
551  * average is updated, with a formula that gives more weight to higher numbers,
552  * to quickly follows sudden increases and decrease slowly
553  */
554
555 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
556                                         struct cfq_group *cfqg, bool rt)
557 {
558         unsigned min_q, max_q;
559         unsigned mult  = cfq_hist_divisor - 1;
560         unsigned round = cfq_hist_divisor / 2;
561         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
562
563         min_q = min(cfqg->busy_queues_avg[rt], busy);
564         max_q = max(cfqg->busy_queues_avg[rt], busy);
565         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
566                 cfq_hist_divisor;
567         return cfqg->busy_queues_avg[rt];
568 }
569
570 static inline unsigned
571 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
572 {
573         struct cfq_rb_root *st = &cfqd->grp_service_tree;
574
575         return cfq_target_latency * cfqg->weight / st->total_weight;
576 }
577
578 static inline void
579 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
580 {
581         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
582         if (cfqd->cfq_latency) {
583                 /*
584                  * interested queues (we consider only the ones with the same
585                  * priority class in the cfq group)
586                  */
587                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
588                                                 cfq_class_rt(cfqq));
589                 unsigned sync_slice = cfqd->cfq_slice[1];
590                 unsigned expect_latency = sync_slice * iq;
591                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
592
593                 if (expect_latency > group_slice) {
594                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
595                         /* scale low_slice according to IO priority
596                          * and sync vs async */
597                         unsigned low_slice =
598                                 min(slice, base_low_slice * slice / sync_slice);
599                         /* the adapted slice value is scaled to fit all iqs
600                          * into the target latency */
601                         slice = max(slice * group_slice / expect_latency,
602                                     low_slice);
603                 }
604         }
605         cfqq->slice_start = jiffies;
606         cfqq->slice_end = jiffies + slice;
607         cfqq->allocated_slice = slice;
608         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
609 }
610
611 /*
612  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
613  * isn't valid until the first request from the dispatch is activated
614  * and the slice time set.
615  */
616 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
617 {
618         if (cfq_cfqq_slice_new(cfqq))
619                 return 0;
620         if (time_before(jiffies, cfqq->slice_end))
621                 return 0;
622
623         return 1;
624 }
625
626 /*
627  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
628  * We choose the request that is closest to the head right now. Distance
629  * behind the head is penalized and only allowed to a certain extent.
630  */
631 static struct request *
632 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
633 {
634         sector_t s1, s2, d1 = 0, d2 = 0;
635         unsigned long back_max;
636 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
637 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
638         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
639
640         if (rq1 == NULL || rq1 == rq2)
641                 return rq2;
642         if (rq2 == NULL)
643                 return rq1;
644
645         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
646                 return rq1;
647         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
648                 return rq2;
649         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
650                 return rq1;
651         else if ((rq2->cmd_flags & REQ_META) &&
652                  !(rq1->cmd_flags & REQ_META))
653                 return rq2;
654
655         s1 = blk_rq_pos(rq1);
656         s2 = blk_rq_pos(rq2);
657
658         /*
659          * by definition, 1KiB is 2 sectors
660          */
661         back_max = cfqd->cfq_back_max * 2;
662
663         /*
664          * Strict one way elevator _except_ in the case where we allow
665          * short backward seeks which are biased as twice the cost of a
666          * similar forward seek.
667          */
668         if (s1 >= last)
669                 d1 = s1 - last;
670         else if (s1 + back_max >= last)
671                 d1 = (last - s1) * cfqd->cfq_back_penalty;
672         else
673                 wrap |= CFQ_RQ1_WRAP;
674
675         if (s2 >= last)
676                 d2 = s2 - last;
677         else if (s2 + back_max >= last)
678                 d2 = (last - s2) * cfqd->cfq_back_penalty;
679         else
680                 wrap |= CFQ_RQ2_WRAP;
681
682         /* Found required data */
683
684         /*
685          * By doing switch() on the bit mask "wrap" we avoid having to
686          * check two variables for all permutations: --> faster!
687          */
688         switch (wrap) {
689         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
690                 if (d1 < d2)
691                         return rq1;
692                 else if (d2 < d1)
693                         return rq2;
694                 else {
695                         if (s1 >= s2)
696                                 return rq1;
697                         else
698                                 return rq2;
699                 }
700
701         case CFQ_RQ2_WRAP:
702                 return rq1;
703         case CFQ_RQ1_WRAP:
704                 return rq2;
705         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
706         default:
707                 /*
708                  * Since both rqs are wrapped,
709                  * start with the one that's further behind head
710                  * (--> only *one* back seek required),
711                  * since back seek takes more time than forward.
712                  */
713                 if (s1 <= s2)
714                         return rq1;
715                 else
716                         return rq2;
717         }
718 }
719
720 /*
721  * The below is leftmost cache rbtree addon
722  */
723 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
724 {
725         /* Service tree is empty */
726         if (!root->count)
727                 return NULL;
728
729         if (!root->left)
730                 root->left = rb_first(&root->rb);
731
732         if (root->left)
733                 return rb_entry(root->left, struct cfq_queue, rb_node);
734
735         return NULL;
736 }
737
738 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
739 {
740         if (!root->left)
741                 root->left = rb_first(&root->rb);
742
743         if (root->left)
744                 return rb_entry_cfqg(root->left);
745
746         return NULL;
747 }
748
749 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
750 {
751         rb_erase(n, root);
752         RB_CLEAR_NODE(n);
753 }
754
755 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
756 {
757         if (root->left == n)
758                 root->left = NULL;
759         rb_erase_init(n, &root->rb);
760         --root->count;
761 }
762
763 /*
764  * would be nice to take fifo expire time into account as well
765  */
766 static struct request *
767 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
768                   struct request *last)
769 {
770         struct rb_node *rbnext = rb_next(&last->rb_node);
771         struct rb_node *rbprev = rb_prev(&last->rb_node);
772         struct request *next = NULL, *prev = NULL;
773
774         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
775
776         if (rbprev)
777                 prev = rb_entry_rq(rbprev);
778
779         if (rbnext)
780                 next = rb_entry_rq(rbnext);
781         else {
782                 rbnext = rb_first(&cfqq->sort_list);
783                 if (rbnext && rbnext != &last->rb_node)
784                         next = rb_entry_rq(rbnext);
785         }
786
787         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
788 }
789
790 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
791                                       struct cfq_queue *cfqq)
792 {
793         /*
794          * just an approximation, should be ok.
795          */
796         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
797                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
798 }
799
800 static inline s64
801 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
802 {
803         return cfqg->vdisktime - st->min_vdisktime;
804 }
805
806 static void
807 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
808 {
809         struct rb_node **node = &st->rb.rb_node;
810         struct rb_node *parent = NULL;
811         struct cfq_group *__cfqg;
812         s64 key = cfqg_key(st, cfqg);
813         int left = 1;
814
815         while (*node != NULL) {
816                 parent = *node;
817                 __cfqg = rb_entry_cfqg(parent);
818
819                 if (key < cfqg_key(st, __cfqg))
820                         node = &parent->rb_left;
821                 else {
822                         node = &parent->rb_right;
823                         left = 0;
824                 }
825         }
826
827         if (left)
828                 st->left = &cfqg->rb_node;
829
830         rb_link_node(&cfqg->rb_node, parent, node);
831         rb_insert_color(&cfqg->rb_node, &st->rb);
832 }
833
834 static void
835 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
836 {
837         struct cfq_rb_root *st = &cfqd->grp_service_tree;
838         struct cfq_group *__cfqg;
839         struct rb_node *n;
840
841         cfqg->nr_cfqq++;
842         if (cfqg->on_st)
843                 return;
844
845         /*
846          * Currently put the group at the end. Later implement something
847          * so that groups get lesser vtime based on their weights, so that
848          * if group does not loose all if it was not continously backlogged.
849          */
850         n = rb_last(&st->rb);
851         if (n) {
852                 __cfqg = rb_entry_cfqg(n);
853                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
854         } else
855                 cfqg->vdisktime = st->min_vdisktime;
856
857         __cfq_group_service_tree_add(st, cfqg);
858         cfqg->on_st = true;
859         st->total_weight += cfqg->weight;
860 }
861
862 static void
863 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
864 {
865         struct cfq_rb_root *st = &cfqd->grp_service_tree;
866
867         if (st->active == &cfqg->rb_node)
868                 st->active = NULL;
869
870         BUG_ON(cfqg->nr_cfqq < 1);
871         cfqg->nr_cfqq--;
872
873         /* If there are other cfq queues under this group, don't delete it */
874         if (cfqg->nr_cfqq)
875                 return;
876
877         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
878         cfqg->on_st = false;
879         st->total_weight -= cfqg->weight;
880         if (!RB_EMPTY_NODE(&cfqg->rb_node))
881                 cfq_rb_erase(&cfqg->rb_node, st);
882         cfqg->saved_workload_slice = 0;
883         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
884 }
885
886 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
887 {
888         unsigned int slice_used;
889
890         /*
891          * Queue got expired before even a single request completed or
892          * got expired immediately after first request completion.
893          */
894         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
895                 /*
896                  * Also charge the seek time incurred to the group, otherwise
897                  * if there are mutiple queues in the group, each can dispatch
898                  * a single request on seeky media and cause lots of seek time
899                  * and group will never know it.
900                  */
901                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
902                                         1);
903         } else {
904                 slice_used = jiffies - cfqq->slice_start;
905                 if (slice_used > cfqq->allocated_slice)
906                         slice_used = cfqq->allocated_slice;
907         }
908
909         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
910         return slice_used;
911 }
912
913 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
914                                 struct cfq_queue *cfqq)
915 {
916         struct cfq_rb_root *st = &cfqd->grp_service_tree;
917         unsigned int used_sl, charge_sl;
918         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
919                         - cfqg->service_tree_idle.count;
920
921         BUG_ON(nr_sync < 0);
922         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
923
924         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
925                 charge_sl = cfqq->allocated_slice;
926
927         /* Can't update vdisktime while group is on service tree */
928         cfq_rb_erase(&cfqg->rb_node, st);
929         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
930         __cfq_group_service_tree_add(st, cfqg);
931
932         /* This group is being expired. Save the context */
933         if (time_after(cfqd->workload_expires, jiffies)) {
934                 cfqg->saved_workload_slice = cfqd->workload_expires
935                                                 - jiffies;
936                 cfqg->saved_workload = cfqd->serving_type;
937                 cfqg->saved_serving_prio = cfqd->serving_prio;
938         } else
939                 cfqg->saved_workload_slice = 0;
940
941         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
942                                         st->min_vdisktime);
943         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
944         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
945 }
946
947 #ifdef CONFIG_CFQ_GROUP_IOSCHED
948 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
949 {
950         if (blkg)
951                 return container_of(blkg, struct cfq_group, blkg);
952         return NULL;
953 }
954
955 void
956 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
957 {
958         cfqg_of_blkg(blkg)->weight = weight;
959 }
960
961 static struct cfq_group *
962 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
963 {
964         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
965         struct cfq_group *cfqg = NULL;
966         void *key = cfqd;
967         int i, j;
968         struct cfq_rb_root *st;
969         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
970         unsigned int major, minor;
971
972         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
973         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
974                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
975                 cfqg->blkg.dev = MKDEV(major, minor);
976                 goto done;
977         }
978         if (cfqg || !create)
979                 goto done;
980
981         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
982         if (!cfqg)
983                 goto done;
984
985         for_each_cfqg_st(cfqg, i, j, st)
986                 *st = CFQ_RB_ROOT;
987         RB_CLEAR_NODE(&cfqg->rb_node);
988
989         /*
990          * Take the initial reference that will be released on destroy
991          * This can be thought of a joint reference by cgroup and
992          * elevator which will be dropped by either elevator exit
993          * or cgroup deletion path depending on who is exiting first.
994          */
995         atomic_set(&cfqg->ref, 1);
996
997         /* Add group onto cgroup list */
998         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
999         cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1000                                         MKDEV(major, minor));
1001         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1002
1003         /* Add group on cfqd list */
1004         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1005
1006 done:
1007         return cfqg;
1008 }
1009
1010 /*
1011  * Search for the cfq group current task belongs to. If create = 1, then also
1012  * create the cfq group if it does not exist. request_queue lock must be held.
1013  */
1014 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1015 {
1016         struct cgroup *cgroup;
1017         struct cfq_group *cfqg = NULL;
1018
1019         rcu_read_lock();
1020         cgroup = task_cgroup(current, blkio_subsys_id);
1021         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1022         if (!cfqg && create)
1023                 cfqg = &cfqd->root_group;
1024         rcu_read_unlock();
1025         return cfqg;
1026 }
1027
1028 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1029 {
1030         atomic_inc(&cfqg->ref);
1031         return cfqg;
1032 }
1033
1034 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1035 {
1036         /* Currently, all async queues are mapped to root group */
1037         if (!cfq_cfqq_sync(cfqq))
1038                 cfqg = &cfqq->cfqd->root_group;
1039
1040         cfqq->cfqg = cfqg;
1041         /* cfqq reference on cfqg */
1042         atomic_inc(&cfqq->cfqg->ref);
1043 }
1044
1045 static void cfq_put_cfqg(struct cfq_group *cfqg)
1046 {
1047         struct cfq_rb_root *st;
1048         int i, j;
1049
1050         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1051         if (!atomic_dec_and_test(&cfqg->ref))
1052                 return;
1053         for_each_cfqg_st(cfqg, i, j, st)
1054                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1055         kfree(cfqg);
1056 }
1057
1058 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1059 {
1060         /* Something wrong if we are trying to remove same group twice */
1061         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1062
1063         hlist_del_init(&cfqg->cfqd_node);
1064
1065         /*
1066          * Put the reference taken at the time of creation so that when all
1067          * queues are gone, group can be destroyed.
1068          */
1069         cfq_put_cfqg(cfqg);
1070 }
1071
1072 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1073 {
1074         struct hlist_node *pos, *n;
1075         struct cfq_group *cfqg;
1076
1077         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1078                 /*
1079                  * If cgroup removal path got to blk_group first and removed
1080                  * it from cgroup list, then it will take care of destroying
1081                  * cfqg also.
1082                  */
1083                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1084                         cfq_destroy_cfqg(cfqd, cfqg);
1085         }
1086 }
1087
1088 /*
1089  * Blk cgroup controller notification saying that blkio_group object is being
1090  * delinked as associated cgroup object is going away. That also means that
1091  * no new IO will come in this group. So get rid of this group as soon as
1092  * any pending IO in the group is finished.
1093  *
1094  * This function is called under rcu_read_lock(). key is the rcu protected
1095  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1096  * read lock.
1097  *
1098  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1099  * it should not be NULL as even if elevator was exiting, cgroup deltion
1100  * path got to it first.
1101  */
1102 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1103 {
1104         unsigned long  flags;
1105         struct cfq_data *cfqd = key;
1106
1107         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1108         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1109         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1110 }
1111
1112 #else /* GROUP_IOSCHED */
1113 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1114 {
1115         return &cfqd->root_group;
1116 }
1117
1118 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1119 {
1120         return cfqg;
1121 }
1122
1123 static inline void
1124 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1125         cfqq->cfqg = cfqg;
1126 }
1127
1128 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1129 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1130
1131 #endif /* GROUP_IOSCHED */
1132
1133 /*
1134  * The cfqd->service_trees holds all pending cfq_queue's that have
1135  * requests waiting to be processed. It is sorted in the order that
1136  * we will service the queues.
1137  */
1138 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1139                                  bool add_front)
1140 {
1141         struct rb_node **p, *parent;
1142         struct cfq_queue *__cfqq;
1143         unsigned long rb_key;
1144         struct cfq_rb_root *service_tree;
1145         int left;
1146         int new_cfqq = 1;
1147         int group_changed = 0;
1148
1149 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1150         if (!cfqd->cfq_group_isolation
1151             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1152             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1153                 /* Move this cfq to root group */
1154                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1155                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1156                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1157                 cfqq->orig_cfqg = cfqq->cfqg;
1158                 cfqq->cfqg = &cfqd->root_group;
1159                 atomic_inc(&cfqd->root_group.ref);
1160                 group_changed = 1;
1161         } else if (!cfqd->cfq_group_isolation
1162                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1163                 /* cfqq is sequential now needs to go to its original group */
1164                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1165                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1166                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1167                 cfq_put_cfqg(cfqq->cfqg);
1168                 cfqq->cfqg = cfqq->orig_cfqg;
1169                 cfqq->orig_cfqg = NULL;
1170                 group_changed = 1;
1171                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1172         }
1173 #endif
1174
1175         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1176                                                 cfqq_type(cfqq));
1177         if (cfq_class_idle(cfqq)) {
1178                 rb_key = CFQ_IDLE_DELAY;
1179                 parent = rb_last(&service_tree->rb);
1180                 if (parent && parent != &cfqq->rb_node) {
1181                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1182                         rb_key += __cfqq->rb_key;
1183                 } else
1184                         rb_key += jiffies;
1185         } else if (!add_front) {
1186                 /*
1187                  * Get our rb key offset. Subtract any residual slice
1188                  * value carried from last service. A negative resid
1189                  * count indicates slice overrun, and this should position
1190                  * the next service time further away in the tree.
1191                  */
1192                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1193                 rb_key -= cfqq->slice_resid;
1194                 cfqq->slice_resid = 0;
1195         } else {
1196                 rb_key = -HZ;
1197                 __cfqq = cfq_rb_first(service_tree);
1198                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1199         }
1200
1201         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1202                 new_cfqq = 0;
1203                 /*
1204                  * same position, nothing more to do
1205                  */
1206                 if (rb_key == cfqq->rb_key &&
1207                     cfqq->service_tree == service_tree)
1208                         return;
1209
1210                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1211                 cfqq->service_tree = NULL;
1212         }
1213
1214         left = 1;
1215         parent = NULL;
1216         cfqq->service_tree = service_tree;
1217         p = &service_tree->rb.rb_node;
1218         while (*p) {
1219                 struct rb_node **n;
1220
1221                 parent = *p;
1222                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1223
1224                 /*
1225                  * sort by key, that represents service time.
1226                  */
1227                 if (time_before(rb_key, __cfqq->rb_key))
1228                         n = &(*p)->rb_left;
1229                 else {
1230                         n = &(*p)->rb_right;
1231                         left = 0;
1232                 }
1233
1234                 p = n;
1235         }
1236
1237         if (left)
1238                 service_tree->left = &cfqq->rb_node;
1239
1240         cfqq->rb_key = rb_key;
1241         rb_link_node(&cfqq->rb_node, parent, p);
1242         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1243         service_tree->count++;
1244         if ((add_front || !new_cfqq) && !group_changed)
1245                 return;
1246         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1247 }
1248
1249 static struct cfq_queue *
1250 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1251                      sector_t sector, struct rb_node **ret_parent,
1252                      struct rb_node ***rb_link)
1253 {
1254         struct rb_node **p, *parent;
1255         struct cfq_queue *cfqq = NULL;
1256
1257         parent = NULL;
1258         p = &root->rb_node;
1259         while (*p) {
1260                 struct rb_node **n;
1261
1262                 parent = *p;
1263                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1264
1265                 /*
1266                  * Sort strictly based on sector.  Smallest to the left,
1267                  * largest to the right.
1268                  */
1269                 if (sector > blk_rq_pos(cfqq->next_rq))
1270                         n = &(*p)->rb_right;
1271                 else if (sector < blk_rq_pos(cfqq->next_rq))
1272                         n = &(*p)->rb_left;
1273                 else
1274                         break;
1275                 p = n;
1276                 cfqq = NULL;
1277         }
1278
1279         *ret_parent = parent;
1280         if (rb_link)
1281                 *rb_link = p;
1282         return cfqq;
1283 }
1284
1285 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1286 {
1287         struct rb_node **p, *parent;
1288         struct cfq_queue *__cfqq;
1289
1290         if (cfqq->p_root) {
1291                 rb_erase(&cfqq->p_node, cfqq->p_root);
1292                 cfqq->p_root = NULL;
1293         }
1294
1295         if (cfq_class_idle(cfqq))
1296                 return;
1297         if (!cfqq->next_rq)
1298                 return;
1299
1300         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1301         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1302                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1303         if (!__cfqq) {
1304                 rb_link_node(&cfqq->p_node, parent, p);
1305                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1306         } else
1307                 cfqq->p_root = NULL;
1308 }
1309
1310 /*
1311  * Update cfqq's position in the service tree.
1312  */
1313 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1314 {
1315         /*
1316          * Resorting requires the cfqq to be on the RR list already.
1317          */
1318         if (cfq_cfqq_on_rr(cfqq)) {
1319                 cfq_service_tree_add(cfqd, cfqq, 0);
1320                 cfq_prio_tree_add(cfqd, cfqq);
1321         }
1322 }
1323
1324 /*
1325  * add to busy list of queues for service, trying to be fair in ordering
1326  * the pending list according to last request service
1327  */
1328 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1329 {
1330         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1331         BUG_ON(cfq_cfqq_on_rr(cfqq));
1332         cfq_mark_cfqq_on_rr(cfqq);
1333         cfqd->busy_queues++;
1334
1335         cfq_resort_rr_list(cfqd, cfqq);
1336 }
1337
1338 /*
1339  * Called when the cfqq no longer has requests pending, remove it from
1340  * the service tree.
1341  */
1342 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1343 {
1344         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1345         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1346         cfq_clear_cfqq_on_rr(cfqq);
1347
1348         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1349                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1350                 cfqq->service_tree = NULL;
1351         }
1352         if (cfqq->p_root) {
1353                 rb_erase(&cfqq->p_node, cfqq->p_root);
1354                 cfqq->p_root = NULL;
1355         }
1356
1357         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1358         BUG_ON(!cfqd->busy_queues);
1359         cfqd->busy_queues--;
1360 }
1361
1362 /*
1363  * rb tree support functions
1364  */
1365 static void cfq_del_rq_rb(struct request *rq)
1366 {
1367         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1368         const int sync = rq_is_sync(rq);
1369
1370         BUG_ON(!cfqq->queued[sync]);
1371         cfqq->queued[sync]--;
1372
1373         elv_rb_del(&cfqq->sort_list, rq);
1374
1375         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1376                 /*
1377                  * Queue will be deleted from service tree when we actually
1378                  * expire it later. Right now just remove it from prio tree
1379                  * as it is empty.
1380                  */
1381                 if (cfqq->p_root) {
1382                         rb_erase(&cfqq->p_node, cfqq->p_root);
1383                         cfqq->p_root = NULL;
1384                 }
1385         }
1386 }
1387
1388 static void cfq_add_rq_rb(struct request *rq)
1389 {
1390         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1391         struct cfq_data *cfqd = cfqq->cfqd;
1392         struct request *__alias, *prev;
1393
1394         cfqq->queued[rq_is_sync(rq)]++;
1395
1396         /*
1397          * looks a little odd, but the first insert might return an alias.
1398          * if that happens, put the alias on the dispatch list
1399          */
1400         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1401                 cfq_dispatch_insert(cfqd->queue, __alias);
1402
1403         if (!cfq_cfqq_on_rr(cfqq))
1404                 cfq_add_cfqq_rr(cfqd, cfqq);
1405
1406         /*
1407          * check if this request is a better next-serve candidate
1408          */
1409         prev = cfqq->next_rq;
1410         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1411
1412         /*
1413          * adjust priority tree position, if ->next_rq changes
1414          */
1415         if (prev != cfqq->next_rq)
1416                 cfq_prio_tree_add(cfqd, cfqq);
1417
1418         BUG_ON(!cfqq->next_rq);
1419 }
1420
1421 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1422 {
1423         elv_rb_del(&cfqq->sort_list, rq);
1424         cfqq->queued[rq_is_sync(rq)]--;
1425         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1426                                         rq_data_dir(rq), rq_is_sync(rq));
1427         cfq_add_rq_rb(rq);
1428         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1429                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1430                         rq_is_sync(rq));
1431 }
1432
1433 static struct request *
1434 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1435 {
1436         struct task_struct *tsk = current;
1437         struct cfq_io_context *cic;
1438         struct cfq_queue *cfqq;
1439
1440         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1441         if (!cic)
1442                 return NULL;
1443
1444         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1445         if (cfqq) {
1446                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1447
1448                 return elv_rb_find(&cfqq->sort_list, sector);
1449         }
1450
1451         return NULL;
1452 }
1453
1454 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1455 {
1456         struct cfq_data *cfqd = q->elevator->elevator_data;
1457
1458         cfqd->rq_in_driver++;
1459         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1460                                                 cfqd->rq_in_driver);
1461
1462         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1463 }
1464
1465 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1466 {
1467         struct cfq_data *cfqd = q->elevator->elevator_data;
1468
1469         WARN_ON(!cfqd->rq_in_driver);
1470         cfqd->rq_in_driver--;
1471         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1472                                                 cfqd->rq_in_driver);
1473 }
1474
1475 static void cfq_remove_request(struct request *rq)
1476 {
1477         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1478
1479         if (cfqq->next_rq == rq)
1480                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1481
1482         list_del_init(&rq->queuelist);
1483         cfq_del_rq_rb(rq);
1484
1485         cfqq->cfqd->rq_queued--;
1486         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1487                                         rq_data_dir(rq), rq_is_sync(rq));
1488         if (rq->cmd_flags & REQ_META) {
1489                 WARN_ON(!cfqq->meta_pending);
1490                 cfqq->meta_pending--;
1491         }
1492 }
1493
1494 static int cfq_merge(struct request_queue *q, struct request **req,
1495                      struct bio *bio)
1496 {
1497         struct cfq_data *cfqd = q->elevator->elevator_data;
1498         struct request *__rq;
1499
1500         __rq = cfq_find_rq_fmerge(cfqd, bio);
1501         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1502                 *req = __rq;
1503                 return ELEVATOR_FRONT_MERGE;
1504         }
1505
1506         return ELEVATOR_NO_MERGE;
1507 }
1508
1509 static void cfq_merged_request(struct request_queue *q, struct request *req,
1510                                int type)
1511 {
1512         if (type == ELEVATOR_FRONT_MERGE) {
1513                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1514
1515                 cfq_reposition_rq_rb(cfqq, req);
1516         }
1517 }
1518
1519 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1520                                 struct bio *bio)
1521 {
1522         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1523                                         bio_data_dir(bio), cfq_bio_sync(bio));
1524 }
1525
1526 static void
1527 cfq_merged_requests(struct request_queue *q, struct request *rq,
1528                     struct request *next)
1529 {
1530         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1531         /*
1532          * reposition in fifo if next is older than rq
1533          */
1534         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1535             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1536                 list_move(&rq->queuelist, &next->queuelist);
1537                 rq_set_fifo_time(rq, rq_fifo_time(next));
1538         }
1539
1540         if (cfqq->next_rq == next)
1541                 cfqq->next_rq = rq;
1542         cfq_remove_request(next);
1543         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1544                                         rq_data_dir(next), rq_is_sync(next));
1545 }
1546
1547 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1548                            struct bio *bio)
1549 {
1550         struct cfq_data *cfqd = q->elevator->elevator_data;
1551         struct cfq_io_context *cic;
1552         struct cfq_queue *cfqq;
1553
1554         /*
1555          * Disallow merge of a sync bio into an async request.
1556          */
1557         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1558                 return false;
1559
1560         /*
1561          * Lookup the cfqq that this bio will be queued with. Allow
1562          * merge only if rq is queued there.
1563          */
1564         cic = cfq_cic_lookup(cfqd, current->io_context);
1565         if (!cic)
1566                 return false;
1567
1568         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1569         return cfqq == RQ_CFQQ(rq);
1570 }
1571
1572 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1573 {
1574         del_timer(&cfqd->idle_slice_timer);
1575         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1576 }
1577
1578 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1579                                    struct cfq_queue *cfqq)
1580 {
1581         if (cfqq) {
1582                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1583                                 cfqd->serving_prio, cfqd->serving_type);
1584                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1585                 cfqq->slice_start = 0;
1586                 cfqq->dispatch_start = jiffies;
1587                 cfqq->allocated_slice = 0;
1588                 cfqq->slice_end = 0;
1589                 cfqq->slice_dispatch = 0;
1590
1591                 cfq_clear_cfqq_wait_request(cfqq);
1592                 cfq_clear_cfqq_must_dispatch(cfqq);
1593                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1594                 cfq_clear_cfqq_fifo_expire(cfqq);
1595                 cfq_mark_cfqq_slice_new(cfqq);
1596
1597                 cfq_del_timer(cfqd, cfqq);
1598         }
1599
1600         cfqd->active_queue = cfqq;
1601 }
1602
1603 /*
1604  * current cfqq expired its slice (or was too idle), select new one
1605  */
1606 static void
1607 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1608                     bool timed_out)
1609 {
1610         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1611
1612         if (cfq_cfqq_wait_request(cfqq))
1613                 cfq_del_timer(cfqd, cfqq);
1614
1615         cfq_clear_cfqq_wait_request(cfqq);
1616         cfq_clear_cfqq_wait_busy(cfqq);
1617
1618         /*
1619          * If this cfqq is shared between multiple processes, check to
1620          * make sure that those processes are still issuing I/Os within
1621          * the mean seek distance.  If not, it may be time to break the
1622          * queues apart again.
1623          */
1624         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1625                 cfq_mark_cfqq_split_coop(cfqq);
1626
1627         /*
1628          * store what was left of this slice, if the queue idled/timed out
1629          */
1630         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1631                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1632                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1633         }
1634
1635         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1636
1637         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1638                 cfq_del_cfqq_rr(cfqd, cfqq);
1639
1640         cfq_resort_rr_list(cfqd, cfqq);
1641
1642         if (cfqq == cfqd->active_queue)
1643                 cfqd->active_queue = NULL;
1644
1645         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1646                 cfqd->grp_service_tree.active = NULL;
1647
1648         if (cfqd->active_cic) {
1649                 put_io_context(cfqd->active_cic->ioc);
1650                 cfqd->active_cic = NULL;
1651         }
1652 }
1653
1654 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1655 {
1656         struct cfq_queue *cfqq = cfqd->active_queue;
1657
1658         if (cfqq)
1659                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1660 }
1661
1662 /*
1663  * Get next queue for service. Unless we have a queue preemption,
1664  * we'll simply select the first cfqq in the service tree.
1665  */
1666 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1667 {
1668         struct cfq_rb_root *service_tree =
1669                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1670                                         cfqd->serving_type);
1671
1672         if (!cfqd->rq_queued)
1673                 return NULL;
1674
1675         /* There is nothing to dispatch */
1676         if (!service_tree)
1677                 return NULL;
1678         if (RB_EMPTY_ROOT(&service_tree->rb))
1679                 return NULL;
1680         return cfq_rb_first(service_tree);
1681 }
1682
1683 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1684 {
1685         struct cfq_group *cfqg;
1686         struct cfq_queue *cfqq;
1687         int i, j;
1688         struct cfq_rb_root *st;
1689
1690         if (!cfqd->rq_queued)
1691                 return NULL;
1692
1693         cfqg = cfq_get_next_cfqg(cfqd);
1694         if (!cfqg)
1695                 return NULL;
1696
1697         for_each_cfqg_st(cfqg, i, j, st)
1698                 if ((cfqq = cfq_rb_first(st)) != NULL)
1699                         return cfqq;
1700         return NULL;
1701 }
1702
1703 /*
1704  * Get and set a new active queue for service.
1705  */
1706 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1707                                               struct cfq_queue *cfqq)
1708 {
1709         if (!cfqq)
1710                 cfqq = cfq_get_next_queue(cfqd);
1711
1712         __cfq_set_active_queue(cfqd, cfqq);
1713         return cfqq;
1714 }
1715
1716 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1717                                           struct request *rq)
1718 {
1719         if (blk_rq_pos(rq) >= cfqd->last_position)
1720                 return blk_rq_pos(rq) - cfqd->last_position;
1721         else
1722                 return cfqd->last_position - blk_rq_pos(rq);
1723 }
1724
1725 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1726                                struct request *rq)
1727 {
1728         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1729 }
1730
1731 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1732                                     struct cfq_queue *cur_cfqq)
1733 {
1734         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1735         struct rb_node *parent, *node;
1736         struct cfq_queue *__cfqq;
1737         sector_t sector = cfqd->last_position;
1738
1739         if (RB_EMPTY_ROOT(root))
1740                 return NULL;
1741
1742         /*
1743          * First, if we find a request starting at the end of the last
1744          * request, choose it.
1745          */
1746         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1747         if (__cfqq)
1748                 return __cfqq;
1749
1750         /*
1751          * If the exact sector wasn't found, the parent of the NULL leaf
1752          * will contain the closest sector.
1753          */
1754         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1755         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1756                 return __cfqq;
1757
1758         if (blk_rq_pos(__cfqq->next_rq) < sector)
1759                 node = rb_next(&__cfqq->p_node);
1760         else
1761                 node = rb_prev(&__cfqq->p_node);
1762         if (!node)
1763                 return NULL;
1764
1765         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1766         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1767                 return __cfqq;
1768
1769         return NULL;
1770 }
1771
1772 /*
1773  * cfqd - obvious
1774  * cur_cfqq - passed in so that we don't decide that the current queue is
1775  *            closely cooperating with itself.
1776  *
1777  * So, basically we're assuming that that cur_cfqq has dispatched at least
1778  * one request, and that cfqd->last_position reflects a position on the disk
1779  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1780  * assumption.
1781  */
1782 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1783                                               struct cfq_queue *cur_cfqq)
1784 {
1785         struct cfq_queue *cfqq;
1786
1787         if (cfq_class_idle(cur_cfqq))
1788                 return NULL;
1789         if (!cfq_cfqq_sync(cur_cfqq))
1790                 return NULL;
1791         if (CFQQ_SEEKY(cur_cfqq))
1792                 return NULL;
1793
1794         /*
1795          * Don't search priority tree if it's the only queue in the group.
1796          */
1797         if (cur_cfqq->cfqg->nr_cfqq == 1)
1798                 return NULL;
1799
1800         /*
1801          * We should notice if some of the queues are cooperating, eg
1802          * working closely on the same area of the disk. In that case,
1803          * we can group them together and don't waste time idling.
1804          */
1805         cfqq = cfqq_close(cfqd, cur_cfqq);
1806         if (!cfqq)
1807                 return NULL;
1808
1809         /* If new queue belongs to different cfq_group, don't choose it */
1810         if (cur_cfqq->cfqg != cfqq->cfqg)
1811                 return NULL;
1812
1813         /*
1814          * It only makes sense to merge sync queues.
1815          */
1816         if (!cfq_cfqq_sync(cfqq))
1817                 return NULL;
1818         if (CFQQ_SEEKY(cfqq))
1819                 return NULL;
1820
1821         /*
1822          * Do not merge queues of different priority classes
1823          */
1824         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1825                 return NULL;
1826
1827         return cfqq;
1828 }
1829
1830 /*
1831  * Determine whether we should enforce idle window for this queue.
1832  */
1833
1834 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1835 {
1836         enum wl_prio_t prio = cfqq_prio(cfqq);
1837         struct cfq_rb_root *service_tree = cfqq->service_tree;
1838
1839         BUG_ON(!service_tree);
1840         BUG_ON(!service_tree->count);
1841
1842         /* We never do for idle class queues. */
1843         if (prio == IDLE_WORKLOAD)
1844                 return false;
1845
1846         /* We do for queues that were marked with idle window flag. */
1847         if (cfq_cfqq_idle_window(cfqq) &&
1848            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1849                 return true;
1850
1851         /*
1852          * Otherwise, we do only if they are the last ones
1853          * in their service tree.
1854          */
1855         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1856                 return 1;
1857         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1858                         service_tree->count);
1859         return 0;
1860 }
1861
1862 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1863 {
1864         struct cfq_queue *cfqq = cfqd->active_queue;
1865         struct cfq_io_context *cic;
1866         unsigned long sl;
1867
1868         /*
1869          * SSD device without seek penalty, disable idling. But only do so
1870          * for devices that support queuing, otherwise we still have a problem
1871          * with sync vs async workloads.
1872          */
1873         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1874                 return;
1875
1876         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1877         WARN_ON(cfq_cfqq_slice_new(cfqq));
1878
1879         /*
1880          * idle is disabled, either manually or by past process history
1881          */
1882         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1883                 return;
1884
1885         /*
1886          * still active requests from this queue, don't idle
1887          */
1888         if (cfqq->dispatched)
1889                 return;
1890
1891         /*
1892          * task has exited, don't wait
1893          */
1894         cic = cfqd->active_cic;
1895         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1896                 return;
1897
1898         /*
1899          * If our average think time is larger than the remaining time
1900          * slice, then don't idle. This avoids overrunning the allotted
1901          * time slice.
1902          */
1903         if (sample_valid(cic->ttime_samples) &&
1904             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1905                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1906                                 cic->ttime_mean);
1907                 return;
1908         }
1909
1910         cfq_mark_cfqq_wait_request(cfqq);
1911
1912         sl = cfqd->cfq_slice_idle;
1913
1914         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1915         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1916         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1917 }
1918
1919 /*
1920  * Move request from internal lists to the request queue dispatch list.
1921  */
1922 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1923 {
1924         struct cfq_data *cfqd = q->elevator->elevator_data;
1925         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1926
1927         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1928
1929         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1930         cfq_remove_request(rq);
1931         cfqq->dispatched++;
1932         elv_dispatch_sort(q, rq);
1933
1934         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1935         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1936                                         rq_data_dir(rq), rq_is_sync(rq));
1937 }
1938
1939 /*
1940  * return expired entry, or NULL to just start from scratch in rbtree
1941  */
1942 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1943 {
1944         struct request *rq = NULL;
1945
1946         if (cfq_cfqq_fifo_expire(cfqq))
1947                 return NULL;
1948
1949         cfq_mark_cfqq_fifo_expire(cfqq);
1950
1951         if (list_empty(&cfqq->fifo))
1952                 return NULL;
1953
1954         rq = rq_entry_fifo(cfqq->fifo.next);
1955         if (time_before(jiffies, rq_fifo_time(rq)))
1956                 rq = NULL;
1957
1958         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1959         return rq;
1960 }
1961
1962 static inline int
1963 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1964 {
1965         const int base_rq = cfqd->cfq_slice_async_rq;
1966
1967         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1968
1969         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1970 }
1971
1972 /*
1973  * Must be called with the queue_lock held.
1974  */
1975 static int cfqq_process_refs(struct cfq_queue *cfqq)
1976 {
1977         int process_refs, io_refs;
1978
1979         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1980         process_refs = atomic_read(&cfqq->ref) - io_refs;
1981         BUG_ON(process_refs < 0);
1982         return process_refs;
1983 }
1984
1985 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1986 {
1987         int process_refs, new_process_refs;
1988         struct cfq_queue *__cfqq;
1989
1990         /*
1991          * If there are no process references on the new_cfqq, then it is
1992          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1993          * chain may have dropped their last reference (not just their
1994          * last process reference).
1995          */
1996         if (!cfqq_process_refs(new_cfqq))
1997                 return;
1998
1999         /* Avoid a circular list and skip interim queue merges */
2000         while ((__cfqq = new_cfqq->new_cfqq)) {
2001                 if (__cfqq == cfqq)
2002                         return;
2003                 new_cfqq = __cfqq;
2004         }
2005
2006         process_refs = cfqq_process_refs(cfqq);
2007         new_process_refs = cfqq_process_refs(new_cfqq);
2008         /*
2009          * If the process for the cfqq has gone away, there is no
2010          * sense in merging the queues.
2011          */
2012         if (process_refs == 0 || new_process_refs == 0)
2013                 return;
2014
2015         /*
2016          * Merge in the direction of the lesser amount of work.
2017          */
2018         if (new_process_refs >= process_refs) {
2019                 cfqq->new_cfqq = new_cfqq;
2020                 atomic_add(process_refs, &new_cfqq->ref);
2021         } else {
2022                 new_cfqq->new_cfqq = cfqq;
2023                 atomic_add(new_process_refs, &cfqq->ref);
2024         }
2025 }
2026
2027 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2028                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2029 {
2030         struct cfq_queue *queue;
2031         int i;
2032         bool key_valid = false;
2033         unsigned long lowest_key = 0;
2034         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2035
2036         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2037                 /* select the one with lowest rb_key */
2038                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2039                 if (queue &&
2040                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2041                         lowest_key = queue->rb_key;
2042                         cur_best = i;
2043                         key_valid = true;
2044                 }
2045         }
2046
2047         return cur_best;
2048 }
2049
2050 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2051 {
2052         unsigned slice;
2053         unsigned count;
2054         struct cfq_rb_root *st;
2055         unsigned group_slice;
2056
2057         if (!cfqg) {
2058                 cfqd->serving_prio = IDLE_WORKLOAD;
2059                 cfqd->workload_expires = jiffies + 1;
2060                 return;
2061         }
2062
2063         /* Choose next priority. RT > BE > IDLE */
2064         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2065                 cfqd->serving_prio = RT_WORKLOAD;
2066         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2067                 cfqd->serving_prio = BE_WORKLOAD;
2068         else {
2069                 cfqd->serving_prio = IDLE_WORKLOAD;
2070                 cfqd->workload_expires = jiffies + 1;
2071                 return;
2072         }
2073
2074         /*
2075          * For RT and BE, we have to choose also the type
2076          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2077          * expiration time
2078          */
2079         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2080         count = st->count;
2081
2082         /*
2083          * check workload expiration, and that we still have other queues ready
2084          */
2085         if (count && !time_after(jiffies, cfqd->workload_expires))
2086                 return;
2087
2088         /* otherwise select new workload type */
2089         cfqd->serving_type =
2090                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2091         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2092         count = st->count;
2093
2094         /*
2095          * the workload slice is computed as a fraction of target latency
2096          * proportional to the number of queues in that workload, over
2097          * all the queues in the same priority class
2098          */
2099         group_slice = cfq_group_slice(cfqd, cfqg);
2100
2101         slice = group_slice * count /
2102                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2103                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2104
2105         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2106                 unsigned int tmp;
2107
2108                 /*
2109                  * Async queues are currently system wide. Just taking
2110                  * proportion of queues with-in same group will lead to higher
2111                  * async ratio system wide as generally root group is going
2112                  * to have higher weight. A more accurate thing would be to
2113                  * calculate system wide asnc/sync ratio.
2114                  */
2115                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2116                 tmp = tmp/cfqd->busy_queues;
2117                 slice = min_t(unsigned, slice, tmp);
2118
2119                 /* async workload slice is scaled down according to
2120                  * the sync/async slice ratio. */
2121                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2122         } else
2123                 /* sync workload slice is at least 2 * cfq_slice_idle */
2124                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2125
2126         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2127         cfq_log(cfqd, "workload slice:%d", slice);
2128         cfqd->workload_expires = jiffies + slice;
2129         cfqd->noidle_tree_requires_idle = false;
2130 }
2131
2132 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2133 {
2134         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2135         struct cfq_group *cfqg;
2136
2137         if (RB_EMPTY_ROOT(&st->rb))
2138                 return NULL;
2139         cfqg = cfq_rb_first_group(st);
2140         st->active = &cfqg->rb_node;
2141         update_min_vdisktime(st);
2142         return cfqg;
2143 }
2144
2145 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2146 {
2147         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2148
2149         cfqd->serving_group = cfqg;
2150
2151         /* Restore the workload type data */
2152         if (cfqg->saved_workload_slice) {
2153                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2154                 cfqd->serving_type = cfqg->saved_workload;
2155                 cfqd->serving_prio = cfqg->saved_serving_prio;
2156         } else
2157                 cfqd->workload_expires = jiffies - 1;
2158
2159         choose_service_tree(cfqd, cfqg);
2160 }
2161
2162 /*
2163  * Select a queue for service. If we have a current active queue,
2164  * check whether to continue servicing it, or retrieve and set a new one.
2165  */
2166 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2167 {
2168         struct cfq_queue *cfqq, *new_cfqq = NULL;
2169
2170         cfqq = cfqd->active_queue;
2171         if (!cfqq)
2172                 goto new_queue;
2173
2174         if (!cfqd->rq_queued)
2175                 return NULL;
2176
2177         /*
2178          * We were waiting for group to get backlogged. Expire the queue
2179          */
2180         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2181                 goto expire;
2182
2183         /*
2184          * The active queue has run out of time, expire it and select new.
2185          */
2186         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2187                 /*
2188                  * If slice had not expired at the completion of last request
2189                  * we might not have turned on wait_busy flag. Don't expire
2190                  * the queue yet. Allow the group to get backlogged.
2191                  *
2192                  * The very fact that we have used the slice, that means we
2193                  * have been idling all along on this queue and it should be
2194                  * ok to wait for this request to complete.
2195                  */
2196                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2197                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2198                         cfqq = NULL;
2199                         goto keep_queue;
2200                 } else
2201                         goto expire;
2202         }
2203
2204         /*
2205          * The active queue has requests and isn't expired, allow it to
2206          * dispatch.
2207          */
2208         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2209                 goto keep_queue;
2210
2211         /*
2212          * If another queue has a request waiting within our mean seek
2213          * distance, let it run.  The expire code will check for close
2214          * cooperators and put the close queue at the front of the service
2215          * tree.  If possible, merge the expiring queue with the new cfqq.
2216          */
2217         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2218         if (new_cfqq) {
2219                 if (!cfqq->new_cfqq)
2220                         cfq_setup_merge(cfqq, new_cfqq);
2221                 goto expire;
2222         }
2223
2224         /*
2225          * No requests pending. If the active queue still has requests in
2226          * flight or is idling for a new request, allow either of these
2227          * conditions to happen (or time out) before selecting a new queue.
2228          */
2229         if (timer_pending(&cfqd->idle_slice_timer) ||
2230             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2231                 cfqq = NULL;
2232                 goto keep_queue;
2233         }
2234
2235 expire:
2236         cfq_slice_expired(cfqd, 0);
2237 new_queue:
2238         /*
2239          * Current queue expired. Check if we have to switch to a new
2240          * service tree
2241          */
2242         if (!new_cfqq)
2243                 cfq_choose_cfqg(cfqd);
2244
2245         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2246 keep_queue:
2247         return cfqq;
2248 }
2249
2250 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2251 {
2252         int dispatched = 0;
2253
2254         while (cfqq->next_rq) {
2255                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2256                 dispatched++;
2257         }
2258
2259         BUG_ON(!list_empty(&cfqq->fifo));
2260
2261         /* By default cfqq is not expired if it is empty. Do it explicitly */
2262         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2263         return dispatched;
2264 }
2265
2266 /*
2267  * Drain our current requests. Used for barriers and when switching
2268  * io schedulers on-the-fly.
2269  */
2270 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2271 {
2272         struct cfq_queue *cfqq;
2273         int dispatched = 0;
2274
2275         /* Expire the timeslice of the current active queue first */
2276         cfq_slice_expired(cfqd, 0);
2277         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2278                 __cfq_set_active_queue(cfqd, cfqq);
2279                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2280         }
2281
2282         BUG_ON(cfqd->busy_queues);
2283
2284         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2285         return dispatched;
2286 }
2287
2288 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2289         struct cfq_queue *cfqq)
2290 {
2291         /* the queue hasn't finished any request, can't estimate */
2292         if (cfq_cfqq_slice_new(cfqq))
2293                 return 1;
2294         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2295                 cfqq->slice_end))
2296                 return 1;
2297
2298         return 0;
2299 }
2300
2301 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2302 {
2303         unsigned int max_dispatch;
2304
2305         /*
2306          * Drain async requests before we start sync IO
2307          */
2308         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2309                 return false;
2310
2311         /*
2312          * If this is an async queue and we have sync IO in flight, let it wait
2313          */
2314         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2315                 return false;
2316
2317         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2318         if (cfq_class_idle(cfqq))
2319                 max_dispatch = 1;
2320
2321         /*
2322          * Does this cfqq already have too much IO in flight?
2323          */
2324         if (cfqq->dispatched >= max_dispatch) {
2325                 /*
2326                  * idle queue must always only have a single IO in flight
2327                  */
2328                 if (cfq_class_idle(cfqq))
2329                         return false;
2330
2331                 /*
2332                  * We have other queues, don't allow more IO from this one
2333                  */
2334                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2335                         return false;
2336
2337                 /*
2338                  * Sole queue user, no limit
2339                  */
2340                 if (cfqd->busy_queues == 1)
2341                         max_dispatch = -1;
2342                 else
2343                         /*
2344                          * Normally we start throttling cfqq when cfq_quantum/2
2345                          * requests have been dispatched. But we can drive
2346                          * deeper queue depths at the beginning of slice
2347                          * subjected to upper limit of cfq_quantum.
2348                          * */
2349                         max_dispatch = cfqd->cfq_quantum;
2350         }
2351
2352         /*
2353          * Async queues must wait a bit before being allowed dispatch.
2354          * We also ramp up the dispatch depth gradually for async IO,
2355          * based on the last sync IO we serviced
2356          */
2357         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2358                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2359                 unsigned int depth;
2360
2361                 depth = last_sync / cfqd->cfq_slice[1];
2362                 if (!depth && !cfqq->dispatched)
2363                         depth = 1;
2364                 if (depth < max_dispatch)
2365                         max_dispatch = depth;
2366         }
2367
2368         /*
2369          * If we're below the current max, allow a dispatch
2370          */
2371         return cfqq->dispatched < max_dispatch;
2372 }
2373
2374 /*
2375  * Dispatch a request from cfqq, moving them to the request queue
2376  * dispatch list.
2377  */
2378 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2379 {
2380         struct request *rq;
2381
2382         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2383
2384         if (!cfq_may_dispatch(cfqd, cfqq))
2385                 return false;
2386
2387         /*
2388          * follow expired path, else get first next available
2389          */
2390         rq = cfq_check_fifo(cfqq);
2391         if (!rq)
2392                 rq = cfqq->next_rq;
2393
2394         /*
2395          * insert request into driver dispatch list
2396          */
2397         cfq_dispatch_insert(cfqd->queue, rq);
2398
2399         if (!cfqd->active_cic) {
2400                 struct cfq_io_context *cic = RQ_CIC(rq);
2401
2402                 atomic_long_inc(&cic->ioc->refcount);
2403                 cfqd->active_cic = cic;
2404         }
2405
2406         return true;
2407 }
2408
2409 /*
2410  * Find the cfqq that we need to service and move a request from that to the
2411  * dispatch list
2412  */
2413 static int cfq_dispatch_requests(struct request_queue *q, int force)
2414 {
2415         struct cfq_data *cfqd = q->elevator->elevator_data;
2416         struct cfq_queue *cfqq;
2417
2418         if (!cfqd->busy_queues)
2419                 return 0;
2420
2421         if (unlikely(force))
2422                 return cfq_forced_dispatch(cfqd);
2423
2424         cfqq = cfq_select_queue(cfqd);
2425         if (!cfqq)
2426                 return 0;
2427
2428         /*
2429          * Dispatch a request from this cfqq, if it is allowed
2430          */
2431         if (!cfq_dispatch_request(cfqd, cfqq))
2432                 return 0;
2433
2434         cfqq->slice_dispatch++;
2435         cfq_clear_cfqq_must_dispatch(cfqq);
2436
2437         /*
2438          * expire an async queue immediately if it has used up its slice. idle
2439          * queue always expire after 1 dispatch round.
2440          */
2441         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2442             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2443             cfq_class_idle(cfqq))) {
2444                 cfqq->slice_end = jiffies + 1;
2445                 cfq_slice_expired(cfqd, 0);
2446         }
2447
2448         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2449         return 1;
2450 }
2451
2452 /*
2453  * task holds one reference to the queue, dropped when task exits. each rq
2454  * in-flight on this queue also holds a reference, dropped when rq is freed.
2455  *
2456  * Each cfq queue took a reference on the parent group. Drop it now.
2457  * queue lock must be held here.
2458  */
2459 static void cfq_put_queue(struct cfq_queue *cfqq)
2460 {
2461         struct cfq_data *cfqd = cfqq->cfqd;
2462         struct cfq_group *cfqg, *orig_cfqg;
2463
2464         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2465
2466         if (!atomic_dec_and_test(&cfqq->ref))
2467                 return;
2468
2469         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2470         BUG_ON(rb_first(&cfqq->sort_list));
2471         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2472         cfqg = cfqq->cfqg;
2473         orig_cfqg = cfqq->orig_cfqg;
2474
2475         if (unlikely(cfqd->active_queue == cfqq)) {
2476                 __cfq_slice_expired(cfqd, cfqq, 0);
2477                 cfq_schedule_dispatch(cfqd);
2478         }
2479
2480         BUG_ON(cfq_cfqq_on_rr(cfqq));
2481         kmem_cache_free(cfq_pool, cfqq);
2482         cfq_put_cfqg(cfqg);
2483         if (orig_cfqg)
2484                 cfq_put_cfqg(orig_cfqg);
2485 }
2486
2487 /*
2488  * Must always be called with the rcu_read_lock() held
2489  */
2490 static void
2491 __call_for_each_cic(struct io_context *ioc,
2492                     void (*func)(struct io_context *, struct cfq_io_context *))
2493 {
2494         struct cfq_io_context *cic;
2495         struct hlist_node *n;
2496
2497         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2498                 func(ioc, cic);
2499 }
2500
2501 /*
2502  * Call func for each cic attached to this ioc.
2503  */
2504 static void
2505 call_for_each_cic(struct io_context *ioc,
2506                   void (*func)(struct io_context *, struct cfq_io_context *))
2507 {
2508         rcu_read_lock();
2509         __call_for_each_cic(ioc, func);
2510         rcu_read_unlock();
2511 }
2512
2513 static void cfq_cic_free_rcu(struct rcu_head *head)
2514 {
2515         struct cfq_io_context *cic;
2516
2517         cic = container_of(head, struct cfq_io_context, rcu_head);
2518
2519         kmem_cache_free(cfq_ioc_pool, cic);
2520         elv_ioc_count_dec(cfq_ioc_count);
2521
2522         if (ioc_gone) {
2523                 /*
2524                  * CFQ scheduler is exiting, grab exit lock and check
2525                  * the pending io context count. If it hits zero,
2526                  * complete ioc_gone and set it back to NULL
2527                  */
2528                 spin_lock(&ioc_gone_lock);
2529                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2530                         complete(ioc_gone);
2531                         ioc_gone = NULL;
2532                 }
2533                 spin_unlock(&ioc_gone_lock);
2534         }
2535 }
2536
2537 static void cfq_cic_free(struct cfq_io_context *cic)
2538 {
2539         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2540 }
2541
2542 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2543 {
2544         unsigned long flags;
2545         unsigned long dead_key = (unsigned long) cic->key;
2546
2547         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2548
2549         spin_lock_irqsave(&ioc->lock, flags);
2550         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2551         hlist_del_rcu(&cic->cic_list);
2552         spin_unlock_irqrestore(&ioc->lock, flags);
2553
2554         cfq_cic_free(cic);
2555 }
2556
2557 /*
2558  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2559  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2560  * and ->trim() which is called with the task lock held
2561  */
2562 static void cfq_free_io_context(struct io_context *ioc)
2563 {
2564         /*
2565          * ioc->refcount is zero here, or we are called from elv_unregister(),
2566          * so no more cic's are allowed to be linked into this ioc.  So it
2567          * should be ok to iterate over the known list, we will see all cic's
2568          * since no new ones are added.
2569          */
2570         __call_for_each_cic(ioc, cic_free_func);
2571 }
2572
2573 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2574 {
2575         struct cfq_queue *__cfqq, *next;
2576
2577         /*
2578          * If this queue was scheduled to merge with another queue, be
2579          * sure to drop the reference taken on that queue (and others in
2580          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2581          */
2582         __cfqq = cfqq->new_cfqq;
2583         while (__cfqq) {
2584                 if (__cfqq == cfqq) {
2585                         WARN(1, "cfqq->new_cfqq loop detected\n");
2586                         break;
2587                 }
2588                 next = __cfqq->new_cfqq;
2589                 cfq_put_queue(__cfqq);
2590                 __cfqq = next;
2591         }
2592 }
2593
2594 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2595 {
2596         if (unlikely(cfqq == cfqd->active_queue)) {
2597                 __cfq_slice_expired(cfqd, cfqq, 0);
2598                 cfq_schedule_dispatch(cfqd);
2599         }
2600
2601         cfq_put_cooperator(cfqq);
2602
2603         cfq_put_queue(cfqq);
2604 }
2605
2606 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2607                                          struct cfq_io_context *cic)
2608 {
2609         struct io_context *ioc = cic->ioc;
2610
2611         list_del_init(&cic->queue_list);
2612
2613         /*
2614          * Make sure dead mark is seen for dead queues
2615          */
2616         smp_wmb();
2617         cic->key = cfqd_dead_key(cfqd);
2618
2619         if (ioc->ioc_data == cic)
2620                 rcu_assign_pointer(ioc->ioc_data, NULL);
2621
2622         if (cic->cfqq[BLK_RW_ASYNC]) {
2623                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2624                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2625         }
2626
2627         if (cic->cfqq[BLK_RW_SYNC]) {
2628                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2629                 cic->cfqq[BLK_RW_SYNC] = NULL;
2630         }
2631 }
2632
2633 static void cfq_exit_single_io_context(struct io_context *ioc,
2634                                        struct cfq_io_context *cic)
2635 {
2636         struct cfq_data *cfqd = cic_to_cfqd(cic);
2637
2638         if (cfqd) {
2639                 struct request_queue *q = cfqd->queue;
2640                 unsigned long flags;
2641
2642                 spin_lock_irqsave(q->queue_lock, flags);
2643
2644                 /*
2645                  * Ensure we get a fresh copy of the ->key to prevent
2646                  * race between exiting task and queue
2647                  */
2648                 smp_read_barrier_depends();
2649                 if (cic->key == cfqd)
2650                         __cfq_exit_single_io_context(cfqd, cic);
2651
2652                 spin_unlock_irqrestore(q->queue_lock, flags);
2653         }
2654 }
2655
2656 /*
2657  * The process that ioc belongs to has exited, we need to clean up
2658  * and put the internal structures we have that belongs to that process.
2659  */
2660 static void cfq_exit_io_context(struct io_context *ioc)
2661 {
2662         call_for_each_cic(ioc, cfq_exit_single_io_context);
2663 }
2664
2665 static struct cfq_io_context *
2666 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2667 {
2668         struct cfq_io_context *cic;
2669
2670         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2671                                                         cfqd->queue->node);
2672         if (cic) {
2673                 cic->last_end_request = jiffies;
2674                 INIT_LIST_HEAD(&cic->queue_list);
2675                 INIT_HLIST_NODE(&cic->cic_list);
2676                 cic->dtor = cfq_free_io_context;
2677                 cic->exit = cfq_exit_io_context;
2678                 elv_ioc_count_inc(cfq_ioc_count);
2679         }
2680
2681         return cic;
2682 }
2683
2684 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2685 {
2686         struct task_struct *tsk = current;
2687         int ioprio_class;
2688
2689         if (!cfq_cfqq_prio_changed(cfqq))
2690                 return;
2691
2692         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2693         switch (ioprio_class) {
2694         default:
2695                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2696         case IOPRIO_CLASS_NONE:
2697                 /*
2698                  * no prio set, inherit CPU scheduling settings
2699                  */
2700                 cfqq->ioprio = task_nice_ioprio(tsk);
2701                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2702                 break;
2703         case IOPRIO_CLASS_RT:
2704                 cfqq->ioprio = task_ioprio(ioc);
2705                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2706                 break;
2707         case IOPRIO_CLASS_BE:
2708                 cfqq->ioprio = task_ioprio(ioc);
2709                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2710                 break;
2711         case IOPRIO_CLASS_IDLE:
2712                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2713                 cfqq->ioprio = 7;
2714                 cfq_clear_cfqq_idle_window(cfqq);
2715                 break;
2716         }
2717
2718         /*
2719          * keep track of original prio settings in case we have to temporarily
2720          * elevate the priority of this queue
2721          */
2722         cfqq->org_ioprio = cfqq->ioprio;
2723         cfqq->org_ioprio_class = cfqq->ioprio_class;
2724         cfq_clear_cfqq_prio_changed(cfqq);
2725 }
2726
2727 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2728 {
2729         struct cfq_data *cfqd = cic_to_cfqd(cic);
2730         struct cfq_queue *cfqq;
2731         unsigned long flags;
2732
2733         if (unlikely(!cfqd))
2734                 return;
2735
2736         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2737
2738         cfqq = cic->cfqq[BLK_RW_ASYNC];
2739         if (cfqq) {
2740                 struct cfq_queue *new_cfqq;
2741                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2742                                                 GFP_ATOMIC);
2743                 if (new_cfqq) {
2744                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2745                         cfq_put_queue(cfqq);
2746                 }
2747         }
2748
2749         cfqq = cic->cfqq[BLK_RW_SYNC];
2750         if (cfqq)
2751                 cfq_mark_cfqq_prio_changed(cfqq);
2752
2753         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2754 }
2755
2756 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2757 {
2758         call_for_each_cic(ioc, changed_ioprio);
2759         ioc->ioprio_changed = 0;
2760 }
2761
2762 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2763                           pid_t pid, bool is_sync)
2764 {
2765         RB_CLEAR_NODE(&cfqq->rb_node);
2766         RB_CLEAR_NODE(&cfqq->p_node);
2767         INIT_LIST_HEAD(&cfqq->fifo);
2768
2769         atomic_set(&cfqq->ref, 0);
2770         cfqq->cfqd = cfqd;
2771
2772         cfq_mark_cfqq_prio_changed(cfqq);
2773
2774         if (is_sync) {
2775                 if (!cfq_class_idle(cfqq))
2776                         cfq_mark_cfqq_idle_window(cfqq);
2777                 cfq_mark_cfqq_sync(cfqq);
2778         }
2779         cfqq->pid = pid;
2780 }
2781
2782 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2783 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2784 {
2785         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2786         struct cfq_data *cfqd = cic_to_cfqd(cic);
2787         unsigned long flags;
2788         struct request_queue *q;
2789
2790         if (unlikely(!cfqd))
2791                 return;
2792
2793         q = cfqd->queue;
2794
2795         spin_lock_irqsave(q->queue_lock, flags);
2796
2797         if (sync_cfqq) {
2798                 /*
2799                  * Drop reference to sync queue. A new sync queue will be
2800                  * assigned in new group upon arrival of a fresh request.
2801                  */
2802                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2803                 cic_set_cfqq(cic, NULL, 1);
2804                 cfq_put_queue(sync_cfqq);
2805         }
2806
2807         spin_unlock_irqrestore(q->queue_lock, flags);
2808 }
2809
2810 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2811 {
2812         call_for_each_cic(ioc, changed_cgroup);
2813         ioc->cgroup_changed = 0;
2814 }
2815 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2816
2817 static struct cfq_queue *
2818 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2819                      struct io_context *ioc, gfp_t gfp_mask)
2820 {
2821         struct cfq_queue *cfqq, *new_cfqq = NULL;
2822         struct cfq_io_context *cic;
2823         struct cfq_group *cfqg;
2824
2825 retry:
2826         cfqg = cfq_get_cfqg(cfqd, 1);
2827         cic = cfq_cic_lookup(cfqd, ioc);
2828         /* cic always exists here */
2829         cfqq = cic_to_cfqq(cic, is_sync);
2830
2831         /*
2832          * Always try a new alloc if we fell back to the OOM cfqq
2833          * originally, since it should just be a temporary situation.
2834          */
2835         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2836                 cfqq = NULL;
2837                 if (new_cfqq) {
2838                         cfqq = new_cfqq;
2839                         new_cfqq = NULL;
2840                 } else if (gfp_mask & __GFP_WAIT) {
2841                         spin_unlock_irq(cfqd->queue->queue_lock);
2842                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2843                                         gfp_mask | __GFP_ZERO,
2844                                         cfqd->queue->node);
2845                         spin_lock_irq(cfqd->queue->queue_lock);
2846                         if (new_cfqq)
2847                                 goto retry;
2848                 } else {
2849                         cfqq = kmem_cache_alloc_node(cfq_pool,
2850                                         gfp_mask | __GFP_ZERO,
2851                                         cfqd->queue->node);
2852                 }
2853
2854                 if (cfqq) {
2855                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2856                         cfq_init_prio_data(cfqq, ioc);
2857                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2858                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2859                 } else
2860                         cfqq = &cfqd->oom_cfqq;
2861         }
2862
2863         if (new_cfqq)
2864                 kmem_cache_free(cfq_pool, new_cfqq);
2865
2866         return cfqq;
2867 }
2868
2869 static struct cfq_queue **
2870 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2871 {
2872         switch (ioprio_class) {
2873         case IOPRIO_CLASS_RT:
2874                 return &cfqd->async_cfqq[0][ioprio];
2875         case IOPRIO_CLASS_BE:
2876                 return &cfqd->async_cfqq[1][ioprio];
2877         case IOPRIO_CLASS_IDLE:
2878                 return &cfqd->async_idle_cfqq;
2879         default:
2880                 BUG();
2881         }
2882 }
2883
2884 static struct cfq_queue *
2885 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2886               gfp_t gfp_mask)
2887 {
2888         const int ioprio = task_ioprio(ioc);
2889         const int ioprio_class = task_ioprio_class(ioc);
2890         struct cfq_queue **async_cfqq = NULL;
2891         struct cfq_queue *cfqq = NULL;
2892
2893         if (!is_sync) {
2894                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2895                 cfqq = *async_cfqq;
2896         }
2897
2898         if (!cfqq)
2899                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2900
2901         /*
2902          * pin the queue now that it's allocated, scheduler exit will prune it
2903          */
2904         if (!is_sync && !(*async_cfqq)) {
2905                 atomic_inc(&cfqq->ref);
2906                 *async_cfqq = cfqq;
2907         }
2908
2909         atomic_inc(&cfqq->ref);
2910         return cfqq;
2911 }
2912
2913 /*
2914  * We drop cfq io contexts lazily, so we may find a dead one.
2915  */
2916 static void
2917 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2918                   struct cfq_io_context *cic)
2919 {
2920         unsigned long flags;
2921
2922         WARN_ON(!list_empty(&cic->queue_list));
2923         BUG_ON(cic->key != cfqd_dead_key(cfqd));
2924
2925         spin_lock_irqsave(&ioc->lock, flags);
2926
2927         BUG_ON(ioc->ioc_data == cic);
2928
2929         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2930         hlist_del_rcu(&cic->cic_list);
2931         spin_unlock_irqrestore(&ioc->lock, flags);
2932
2933         cfq_cic_free(cic);
2934 }
2935
2936 static struct cfq_io_context *
2937 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2938 {
2939         struct cfq_io_context *cic;
2940         unsigned long flags;
2941
2942         if (unlikely(!ioc))
2943                 return NULL;
2944
2945         rcu_read_lock();
2946
2947         /*
2948          * we maintain a last-hit cache, to avoid browsing over the tree
2949          */
2950         cic = rcu_dereference(ioc->ioc_data);
2951         if (cic && cic->key == cfqd) {
2952                 rcu_read_unlock();
2953                 return cic;
2954         }
2955
2956         do {
2957                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
2958                 rcu_read_unlock();
2959                 if (!cic)
2960                         break;
2961                 if (unlikely(cic->key != cfqd)) {
2962                         cfq_drop_dead_cic(cfqd, ioc, cic);
2963                         rcu_read_lock();
2964                         continue;
2965                 }
2966
2967                 spin_lock_irqsave(&ioc->lock, flags);
2968                 rcu_assign_pointer(ioc->ioc_data, cic);
2969                 spin_unlock_irqrestore(&ioc->lock, flags);
2970                 break;
2971         } while (1);
2972
2973         return cic;
2974 }
2975
2976 /*
2977  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2978  * the process specific cfq io context when entered from the block layer.
2979  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2980  */
2981 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2982                         struct cfq_io_context *cic, gfp_t gfp_mask)
2983 {
2984         unsigned long flags;
2985         int ret;
2986
2987         ret = radix_tree_preload(gfp_mask);
2988         if (!ret) {
2989                 cic->ioc = ioc;
2990                 cic->key = cfqd;
2991
2992                 spin_lock_irqsave(&ioc->lock, flags);
2993                 ret = radix_tree_insert(&ioc->radix_root,
2994                                                 cfqd->cic_index, cic);
2995                 if (!ret)
2996                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2997                 spin_unlock_irqrestore(&ioc->lock, flags);
2998
2999                 radix_tree_preload_end();
3000
3001                 if (!ret) {
3002                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3003                         list_add(&cic->queue_list, &cfqd->cic_list);
3004                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3005                 }
3006         }
3007
3008         if (ret)
3009                 printk(KERN_ERR "cfq: cic link failed!\n");
3010
3011         return ret;
3012 }
3013
3014 /*
3015  * Setup general io context and cfq io context. There can be several cfq
3016  * io contexts per general io context, if this process is doing io to more
3017  * than one device managed by cfq.
3018  */
3019 static struct cfq_io_context *
3020 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3021 {
3022         struct io_context *ioc = NULL;
3023         struct cfq_io_context *cic;
3024
3025         might_sleep_if(gfp_mask & __GFP_WAIT);
3026
3027         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3028         if (!ioc)
3029                 return NULL;
3030
3031         cic = cfq_cic_lookup(cfqd, ioc);
3032         if (cic)
3033                 goto out;
3034
3035         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3036         if (cic == NULL)
3037                 goto err;
3038
3039         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3040                 goto err_free;
3041
3042 out:
3043         smp_read_barrier_depends();
3044         if (unlikely(ioc->ioprio_changed))
3045                 cfq_ioc_set_ioprio(ioc);
3046
3047 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3048         if (unlikely(ioc->cgroup_changed))
3049                 cfq_ioc_set_cgroup(ioc);
3050 #endif
3051         return cic;
3052 err_free:
3053         cfq_cic_free(cic);
3054 err:
3055         put_io_context(ioc);
3056         return NULL;
3057 }
3058
3059 static void
3060 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3061 {
3062         unsigned long elapsed = jiffies - cic->last_end_request;
3063         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3064
3065         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3066         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3067         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3068 }
3069
3070 static void
3071 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3072                        struct request *rq)
3073 {
3074         sector_t sdist = 0;
3075         sector_t n_sec = blk_rq_sectors(rq);
3076         if (cfqq->last_request_pos) {
3077                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3078                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3079                 else
3080                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3081         }
3082
3083         cfqq->seek_history <<= 1;
3084         if (blk_queue_nonrot(cfqd->queue))
3085                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3086         else
3087                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3088 }
3089
3090 /*
3091  * Disable idle window if the process thinks too long or seeks so much that
3092  * it doesn't matter
3093  */
3094 static void
3095 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3096                        struct cfq_io_context *cic)
3097 {
3098         int old_idle, enable_idle;
3099
3100         /*
3101          * Don't idle for async or idle io prio class
3102          */
3103         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3104                 return;
3105
3106         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3107
3108         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3109                 cfq_mark_cfqq_deep(cfqq);
3110
3111         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3112             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3113                 enable_idle = 0;
3114         else if (sample_valid(cic->ttime_samples)) {
3115                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3116                         enable_idle = 0;
3117                 else
3118                         enable_idle = 1;
3119         }
3120
3121         if (old_idle != enable_idle) {
3122                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3123                 if (enable_idle)
3124                         cfq_mark_cfqq_idle_window(cfqq);
3125                 else
3126                         cfq_clear_cfqq_idle_window(cfqq);
3127         }
3128 }
3129
3130 /*
3131  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3132  * no or if we aren't sure, a 1 will cause a preempt.
3133  */
3134 static bool
3135 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3136                    struct request *rq)
3137 {
3138         struct cfq_queue *cfqq;
3139
3140         cfqq = cfqd->active_queue;
3141         if (!cfqq)
3142                 return false;
3143
3144         if (cfq_class_idle(new_cfqq))
3145                 return false;
3146
3147         if (cfq_class_idle(cfqq))
3148                 return true;
3149
3150         /*
3151          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3152          */
3153         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3154                 return false;
3155
3156         /*
3157          * if the new request is sync, but the currently running queue is
3158          * not, let the sync request have priority.
3159          */
3160         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3161                 return true;
3162
3163         if (new_cfqq->cfqg != cfqq->cfqg)
3164                 return false;
3165
3166         if (cfq_slice_used(cfqq))
3167                 return true;
3168
3169         /* Allow preemption only if we are idling on sync-noidle tree */
3170         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3171             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3172             new_cfqq->service_tree->count == 2 &&
3173             RB_EMPTY_ROOT(&cfqq->sort_list))
3174                 return true;
3175
3176         /*
3177          * So both queues are sync. Let the new request get disk time if
3178          * it's a metadata request and the current queue is doing regular IO.
3179          */
3180         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3181                 return true;
3182
3183         /*
3184          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3185          */
3186         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3187                 return true;
3188
3189         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3190                 return false;
3191
3192         /*
3193          * if this request is as-good as one we would expect from the
3194          * current cfqq, let it preempt
3195          */
3196         if (cfq_rq_close(cfqd, cfqq, rq))
3197                 return true;
3198
3199         return false;
3200 }
3201
3202 /*
3203  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3204  * let it have half of its nominal slice.
3205  */
3206 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3207 {
3208         cfq_log_cfqq(cfqd, cfqq, "preempt");
3209         cfq_slice_expired(cfqd, 1);
3210
3211         /*
3212          * Put the new queue at the front of the of the current list,
3213          * so we know that it will be selected next.
3214          */
3215         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3216
3217         cfq_service_tree_add(cfqd, cfqq, 1);
3218
3219         cfqq->slice_end = 0;
3220         cfq_mark_cfqq_slice_new(cfqq);
3221 }
3222
3223 /*
3224  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3225  * something we should do about it
3226  */
3227 static void
3228 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3229                 struct request *rq)
3230 {
3231         struct cfq_io_context *cic = RQ_CIC(rq);
3232
3233         cfqd->rq_queued++;
3234         if (rq->cmd_flags & REQ_META)
3235                 cfqq->meta_pending++;
3236
3237         cfq_update_io_thinktime(cfqd, cic);
3238         cfq_update_io_seektime(cfqd, cfqq, rq);
3239         cfq_update_idle_window(cfqd, cfqq, cic);
3240
3241         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3242
3243         if (cfqq == cfqd->active_queue) {
3244                 /*
3245                  * Remember that we saw a request from this process, but
3246                  * don't start queuing just yet. Otherwise we risk seeing lots
3247                  * of tiny requests, because we disrupt the normal plugging
3248                  * and merging. If the request is already larger than a single
3249                  * page, let it rip immediately. For that case we assume that
3250                  * merging is already done. Ditto for a busy system that
3251                  * has other work pending, don't risk delaying until the
3252                  * idle timer unplug to continue working.
3253                  */
3254                 if (cfq_cfqq_wait_request(cfqq)) {
3255                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3256                             cfqd->busy_queues > 1) {
3257                                 cfq_del_timer(cfqd, cfqq);
3258                                 cfq_clear_cfqq_wait_request(cfqq);
3259                                 __blk_run_queue(cfqd->queue);
3260                         } else {
3261                                 cfq_blkiocg_update_idle_time_stats(
3262                                                 &cfqq->cfqg->blkg);
3263                                 cfq_mark_cfqq_must_dispatch(cfqq);
3264                         }
3265                 }
3266         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3267                 /*
3268                  * not the active queue - expire current slice if it is
3269                  * idle and has expired it's mean thinktime or this new queue
3270                  * has some old slice time left and is of higher priority or
3271                  * this new queue is RT and the current one is BE
3272                  */
3273                 cfq_preempt_queue(cfqd, cfqq);
3274                 __blk_run_queue(cfqd->queue);
3275         }
3276 }
3277
3278 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3279 {
3280         struct cfq_data *cfqd = q->elevator->elevator_data;
3281         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3282
3283         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3284         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3285
3286         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3287         list_add_tail(&rq->queuelist, &cfqq->fifo);
3288         cfq_add_rq_rb(rq);
3289         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3290                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3291                         rq_is_sync(rq));
3292         cfq_rq_enqueued(cfqd, cfqq, rq);
3293 }
3294
3295 /*
3296  * Update hw_tag based on peak queue depth over 50 samples under
3297  * sufficient load.
3298  */
3299 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3300 {
3301         struct cfq_queue *cfqq = cfqd->active_queue;
3302
3303         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3304                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3305
3306         if (cfqd->hw_tag == 1)
3307                 return;
3308
3309         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3310             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3311                 return;
3312
3313         /*
3314          * If active queue hasn't enough requests and can idle, cfq might not
3315          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3316          * case
3317          */
3318         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3319             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3320             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3321                 return;
3322
3323         if (cfqd->hw_tag_samples++ < 50)
3324                 return;
3325
3326         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3327                 cfqd->hw_tag = 1;
3328         else
3329                 cfqd->hw_tag = 0;
3330 }
3331
3332 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3333 {
3334         struct cfq_io_context *cic = cfqd->active_cic;
3335
3336         /* If there are other queues in the group, don't wait */
3337         if (cfqq->cfqg->nr_cfqq > 1)
3338                 return false;
3339
3340         if (cfq_slice_used(cfqq))
3341                 return true;
3342
3343         /* if slice left is less than think time, wait busy */
3344         if (cic && sample_valid(cic->ttime_samples)
3345             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3346                 return true;
3347
3348         /*
3349          * If think times is less than a jiffy than ttime_mean=0 and above
3350          * will not be true. It might happen that slice has not expired yet
3351          * but will expire soon (4-5 ns) during select_queue(). To cover the
3352          * case where think time is less than a jiffy, mark the queue wait
3353          * busy if only 1 jiffy is left in the slice.
3354          */
3355         if (cfqq->slice_end - jiffies == 1)
3356                 return true;
3357
3358         return false;
3359 }
3360
3361 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3362 {
3363         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3364         struct cfq_data *cfqd = cfqq->cfqd;
3365         const int sync = rq_is_sync(rq);
3366         unsigned long now;
3367
3368         now = jiffies;
3369         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3370                      !!(rq->cmd_flags & REQ_NOIDLE));
3371
3372         cfq_update_hw_tag(cfqd);
3373
3374         WARN_ON(!cfqd->rq_in_driver);
3375         WARN_ON(!cfqq->dispatched);
3376         cfqd->rq_in_driver--;
3377         cfqq->dispatched--;
3378         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3379                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3380                         rq_data_dir(rq), rq_is_sync(rq));
3381
3382         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3383
3384         if (sync) {
3385                 RQ_CIC(rq)->last_end_request = now;
3386                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3387                         cfqd->last_delayed_sync = now;
3388         }
3389
3390         /*
3391          * If this is the active queue, check if it needs to be expired,
3392          * or if we want to idle in case it has no pending requests.
3393          */
3394         if (cfqd->active_queue == cfqq) {
3395                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3396
3397                 if (cfq_cfqq_slice_new(cfqq)) {
3398                         cfq_set_prio_slice(cfqd, cfqq);
3399                         cfq_clear_cfqq_slice_new(cfqq);
3400                 }
3401
3402                 /*
3403                  * Should we wait for next request to come in before we expire
3404                  * the queue.
3405                  */
3406                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3407                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3408                         cfq_mark_cfqq_wait_busy(cfqq);
3409                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3410                 }
3411
3412                 /*
3413                  * Idling is not enabled on:
3414                  * - expired queues
3415                  * - idle-priority queues
3416                  * - async queues
3417                  * - queues with still some requests queued
3418                  * - when there is a close cooperator
3419                  */
3420                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3421                         cfq_slice_expired(cfqd, 1);
3422                 else if (sync && cfqq_empty &&
3423                          !cfq_close_cooperator(cfqd, cfqq)) {
3424                         cfqd->noidle_tree_requires_idle |=
3425                                 !(rq->cmd_flags & REQ_NOIDLE);
3426                         /*
3427                          * Idling is enabled for SYNC_WORKLOAD.
3428                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3429                          * only if we processed at least one !REQ_NOIDLE request
3430                          */
3431                         if (cfqd->serving_type == SYNC_WORKLOAD
3432                             || cfqd->noidle_tree_requires_idle
3433                             || cfqq->cfqg->nr_cfqq == 1)
3434                                 cfq_arm_slice_timer(cfqd);
3435                 }
3436         }
3437
3438         if (!cfqd->rq_in_driver)
3439                 cfq_schedule_dispatch(cfqd);
3440 }
3441
3442 /*
3443  * we temporarily boost lower priority queues if they are holding fs exclusive
3444  * resources. they are boosted to normal prio (CLASS_BE/4)
3445  */
3446 static void cfq_prio_boost(struct cfq_queue *cfqq)
3447 {
3448         if (has_fs_excl()) {
3449                 /*
3450                  * boost idle prio on transactions that would lock out other
3451                  * users of the filesystem
3452                  */
3453                 if (cfq_class_idle(cfqq))
3454                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3455                 if (cfqq->ioprio > IOPRIO_NORM)
3456                         cfqq->ioprio = IOPRIO_NORM;
3457         } else {
3458                 /*
3459                  * unboost the queue (if needed)
3460                  */
3461                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3462                 cfqq->ioprio = cfqq->org_ioprio;
3463         }
3464 }
3465
3466 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3467 {
3468         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3469                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3470                 return ELV_MQUEUE_MUST;
3471         }
3472
3473         return ELV_MQUEUE_MAY;
3474 }
3475
3476 static int cfq_may_queue(struct request_queue *q, int rw)
3477 {
3478         struct cfq_data *cfqd = q->elevator->elevator_data;
3479         struct task_struct *tsk = current;
3480         struct cfq_io_context *cic;
3481         struct cfq_queue *cfqq;
3482
3483         /*
3484          * don't force setup of a queue from here, as a call to may_queue
3485          * does not necessarily imply that a request actually will be queued.
3486          * so just lookup a possibly existing queue, or return 'may queue'
3487          * if that fails
3488          */
3489         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3490         if (!cic)
3491                 return ELV_MQUEUE_MAY;
3492
3493         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3494         if (cfqq) {
3495                 cfq_init_prio_data(cfqq, cic->ioc);
3496                 cfq_prio_boost(cfqq);
3497
3498                 return __cfq_may_queue(cfqq);
3499         }
3500
3501         return ELV_MQUEUE_MAY;
3502 }
3503
3504 /*
3505  * queue lock held here
3506  */
3507 static void cfq_put_request(struct request *rq)
3508 {
3509         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3510
3511         if (cfqq) {
3512                 const int rw = rq_data_dir(rq);
3513
3514                 BUG_ON(!cfqq->allocated[rw]);
3515                 cfqq->allocated[rw]--;
3516
3517                 put_io_context(RQ_CIC(rq)->ioc);
3518
3519                 rq->elevator_private = NULL;
3520                 rq->elevator_private2 = NULL;
3521
3522                 /* Put down rq reference on cfqg */
3523                 cfq_put_cfqg(RQ_CFQG(rq));
3524                 rq->elevator_private3 = NULL;
3525
3526                 cfq_put_queue(cfqq);
3527         }
3528 }
3529
3530 static struct cfq_queue *
3531 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3532                 struct cfq_queue *cfqq)
3533 {
3534         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3535         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3536         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3537         cfq_put_queue(cfqq);
3538         return cic_to_cfqq(cic, 1);
3539 }
3540
3541 /*
3542  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3543  * was the last process referring to said cfqq.
3544  */
3545 static struct cfq_queue *
3546 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3547 {
3548         if (cfqq_process_refs(cfqq) == 1) {
3549                 cfqq->pid = current->pid;
3550                 cfq_clear_cfqq_coop(cfqq);
3551                 cfq_clear_cfqq_split_coop(cfqq);
3552                 return cfqq;
3553         }
3554
3555         cic_set_cfqq(cic, NULL, 1);
3556
3557         cfq_put_cooperator(cfqq);
3558
3559         cfq_put_queue(cfqq);
3560         return NULL;
3561 }
3562 /*
3563  * Allocate cfq data structures associated with this request.
3564  */
3565 static int
3566 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3567 {
3568         struct cfq_data *cfqd = q->elevator->elevator_data;
3569         struct cfq_io_context *cic;
3570         const int rw = rq_data_dir(rq);
3571         const bool is_sync = rq_is_sync(rq);
3572         struct cfq_queue *cfqq;
3573         unsigned long flags;
3574
3575         might_sleep_if(gfp_mask & __GFP_WAIT);
3576
3577         cic = cfq_get_io_context(cfqd, gfp_mask);
3578
3579         spin_lock_irqsave(q->queue_lock, flags);
3580
3581         if (!cic)
3582                 goto queue_fail;
3583
3584 new_queue:
3585         cfqq = cic_to_cfqq(cic, is_sync);
3586         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3587                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3588                 cic_set_cfqq(cic, cfqq, is_sync);
3589         } else {
3590                 /*
3591                  * If the queue was seeky for too long, break it apart.
3592                  */
3593                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3594                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3595                         cfqq = split_cfqq(cic, cfqq);
3596                         if (!cfqq)
3597                                 goto new_queue;
3598                 }
3599
3600                 /*
3601                  * Check to see if this queue is scheduled to merge with
3602                  * another, closely cooperating queue.  The merging of
3603                  * queues happens here as it must be done in process context.
3604                  * The reference on new_cfqq was taken in merge_cfqqs.
3605                  */
3606                 if (cfqq->new_cfqq)
3607                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3608         }
3609
3610         cfqq->allocated[rw]++;
3611         atomic_inc(&cfqq->ref);
3612
3613         spin_unlock_irqrestore(q->queue_lock, flags);
3614
3615         rq->elevator_private = cic;
3616         rq->elevator_private2 = cfqq;
3617         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3618         return 0;
3619
3620 queue_fail:
3621         if (cic)
3622                 put_io_context(cic->ioc);
3623
3624         cfq_schedule_dispatch(cfqd);
3625         spin_unlock_irqrestore(q->queue_lock, flags);
3626         cfq_log(cfqd, "set_request fail");
3627         return 1;
3628 }
3629
3630 static void cfq_kick_queue(struct work_struct *work)
3631 {
3632         struct cfq_data *cfqd =
3633                 container_of(work, struct cfq_data, unplug_work);
3634         struct request_queue *q = cfqd->queue;
3635
3636         spin_lock_irq(q->queue_lock);
3637         __blk_run_queue(cfqd->queue);
3638         spin_unlock_irq(q->queue_lock);
3639 }
3640
3641 /*
3642  * Timer running if the active_queue is currently idling inside its time slice
3643  */
3644 static void cfq_idle_slice_timer(unsigned long data)
3645 {
3646         struct cfq_data *cfqd = (struct cfq_data *) data;
3647         struct cfq_queue *cfqq;
3648         unsigned long flags;
3649         int timed_out = 1;
3650
3651         cfq_log(cfqd, "idle timer fired");
3652
3653         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3654
3655         cfqq = cfqd->active_queue;
3656         if (cfqq) {
3657                 timed_out = 0;
3658
3659                 /*
3660                  * We saw a request before the queue expired, let it through
3661                  */
3662                 if (cfq_cfqq_must_dispatch(cfqq))
3663                         goto out_kick;
3664
3665                 /*
3666                  * expired
3667                  */
3668                 if (cfq_slice_used(cfqq))
3669                         goto expire;
3670
3671                 /*
3672                  * only expire and reinvoke request handler, if there are
3673                  * other queues with pending requests
3674                  */
3675                 if (!cfqd->busy_queues)
3676                         goto out_cont;
3677
3678                 /*
3679                  * not expired and it has a request pending, let it dispatch
3680                  */
3681                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3682                         goto out_kick;
3683
3684                 /*
3685                  * Queue depth flag is reset only when the idle didn't succeed
3686                  */
3687                 cfq_clear_cfqq_deep(cfqq);
3688         }
3689 expire:
3690         cfq_slice_expired(cfqd, timed_out);
3691 out_kick:
3692         cfq_schedule_dispatch(cfqd);
3693 out_cont:
3694         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3695 }
3696
3697 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3698 {
3699         del_timer_sync(&cfqd->idle_slice_timer);
3700         cancel_work_sync(&cfqd->unplug_work);
3701 }
3702
3703 static void cfq_put_async_queues(struct cfq_data *cfqd)
3704 {
3705         int i;
3706
3707         for (i = 0; i < IOPRIO_BE_NR; i++) {
3708                 if (cfqd->async_cfqq[0][i])
3709                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3710                 if (cfqd->async_cfqq[1][i])
3711                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3712         }
3713
3714         if (cfqd->async_idle_cfqq)
3715                 cfq_put_queue(cfqd->async_idle_cfqq);
3716 }
3717
3718 static void cfq_cfqd_free(struct rcu_head *head)
3719 {
3720         kfree(container_of(head, struct cfq_data, rcu));
3721 }
3722
3723 static void cfq_exit_queue(struct elevator_queue *e)
3724 {
3725         struct cfq_data *cfqd = e->elevator_data;
3726         struct request_queue *q = cfqd->queue;
3727
3728         cfq_shutdown_timer_wq(cfqd);
3729
3730         spin_lock_irq(q->queue_lock);
3731
3732         if (cfqd->active_queue)
3733                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3734
3735         while (!list_empty(&cfqd->cic_list)) {
3736                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3737                                                         struct cfq_io_context,
3738                                                         queue_list);
3739
3740                 __cfq_exit_single_io_context(cfqd, cic);
3741         }
3742
3743         cfq_put_async_queues(cfqd);
3744         cfq_release_cfq_groups(cfqd);
3745         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3746
3747         spin_unlock_irq(q->queue_lock);
3748
3749         cfq_shutdown_timer_wq(cfqd);
3750
3751         spin_lock(&cic_index_lock);
3752         ida_remove(&cic_index_ida, cfqd->cic_index);
3753         spin_unlock(&cic_index_lock);
3754
3755         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3756         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3757 }
3758
3759 static int cfq_alloc_cic_index(void)
3760 {
3761         int index, error;
3762
3763         do {
3764                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3765                         return -ENOMEM;
3766
3767                 spin_lock(&cic_index_lock);
3768                 error = ida_get_new(&cic_index_ida, &index);
3769                 spin_unlock(&cic_index_lock);
3770                 if (error && error != -EAGAIN)
3771                         return error;
3772         } while (error);
3773
3774         return index;
3775 }
3776
3777 static void *cfq_init_queue(struct request_queue *q)
3778 {
3779         struct cfq_data *cfqd;
3780         int i, j;
3781         struct cfq_group *cfqg;
3782         struct cfq_rb_root *st;
3783
3784         i = cfq_alloc_cic_index();
3785         if (i < 0)
3786                 return NULL;
3787
3788         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3789         if (!cfqd)
3790                 return NULL;
3791
3792         cfqd->cic_index = i;
3793
3794         /* Init root service tree */
3795         cfqd->grp_service_tree = CFQ_RB_ROOT;
3796
3797         /* Init root group */
3798         cfqg = &cfqd->root_group;
3799         for_each_cfqg_st(cfqg, i, j, st)
3800                 *st = CFQ_RB_ROOT;
3801         RB_CLEAR_NODE(&cfqg->rb_node);
3802
3803         /* Give preference to root group over other groups */
3804         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3805
3806 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3807         /*
3808          * Take a reference to root group which we never drop. This is just
3809          * to make sure that cfq_put_cfqg() does not try to kfree root group
3810          */
3811         atomic_set(&cfqg->ref, 1);
3812         rcu_read_lock();
3813         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3814                                         (void *)cfqd, 0);
3815         rcu_read_unlock();
3816 #endif
3817         /*
3818          * Not strictly needed (since RB_ROOT just clears the node and we
3819          * zeroed cfqd on alloc), but better be safe in case someone decides
3820          * to add magic to the rb code
3821          */
3822         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3823                 cfqd->prio_trees[i] = RB_ROOT;
3824
3825         /*
3826          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3827          * Grab a permanent reference to it, so that the normal code flow
3828          * will not attempt to free it.
3829          */
3830         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3831         atomic_inc(&cfqd->oom_cfqq.ref);
3832         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3833
3834         INIT_LIST_HEAD(&cfqd->cic_list);
3835
3836         cfqd->queue = q;
3837
3838         init_timer(&cfqd->idle_slice_timer);
3839         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3840         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3841
3842         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3843
3844         cfqd->cfq_quantum = cfq_quantum;
3845         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3846         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3847         cfqd->cfq_back_max = cfq_back_max;
3848         cfqd->cfq_back_penalty = cfq_back_penalty;
3849         cfqd->cfq_slice[0] = cfq_slice_async;
3850         cfqd->cfq_slice[1] = cfq_slice_sync;
3851         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3852         cfqd->cfq_slice_idle = cfq_slice_idle;
3853         cfqd->cfq_latency = 1;
3854         cfqd->cfq_group_isolation = 0;
3855         cfqd->hw_tag = -1;
3856         /*
3857          * we optimistically start assuming sync ops weren't delayed in last
3858          * second, in order to have larger depth for async operations.
3859          */
3860         cfqd->last_delayed_sync = jiffies - HZ;
3861         return cfqd;
3862 }
3863
3864 static void cfq_slab_kill(void)
3865 {
3866         /*
3867          * Caller already ensured that pending RCU callbacks are completed,
3868          * so we should have no busy allocations at this point.
3869          */
3870         if (cfq_pool)
3871                 kmem_cache_destroy(cfq_pool);
3872         if (cfq_ioc_pool)
3873                 kmem_cache_destroy(cfq_ioc_pool);
3874 }
3875
3876 static int __init cfq_slab_setup(void)
3877 {
3878         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3879         if (!cfq_pool)
3880                 goto fail;
3881
3882         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3883         if (!cfq_ioc_pool)
3884                 goto fail;
3885
3886         return 0;
3887 fail:
3888         cfq_slab_kill();
3889         return -ENOMEM;
3890 }
3891
3892 /*
3893  * sysfs parts below -->
3894  */
3895 static ssize_t
3896 cfq_var_show(unsigned int var, char *page)
3897 {
3898         return sprintf(page, "%d\n", var);
3899 }
3900
3901 static ssize_t
3902 cfq_var_store(unsigned int *var, const char *page, size_t count)
3903 {
3904         char *p = (char *) page;
3905
3906         *var = simple_strtoul(p, &p, 10);
3907         return count;
3908 }
3909
3910 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3911 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3912 {                                                                       \
3913         struct cfq_data *cfqd = e->elevator_data;                       \
3914         unsigned int __data = __VAR;                                    \
3915         if (__CONV)                                                     \
3916                 __data = jiffies_to_msecs(__data);                      \
3917         return cfq_var_show(__data, (page));                            \
3918 }
3919 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3920 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3921 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3922 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3923 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3924 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3925 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3926 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3927 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3928 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3929 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3930 #undef SHOW_FUNCTION
3931
3932 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3933 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3934 {                                                                       \
3935         struct cfq_data *cfqd = e->elevator_data;                       \
3936         unsigned int __data;                                            \
3937         int ret = cfq_var_store(&__data, (page), count);                \
3938         if (__data < (MIN))                                             \
3939                 __data = (MIN);                                         \
3940         else if (__data > (MAX))                                        \
3941                 __data = (MAX);                                         \
3942         if (__CONV)                                                     \
3943                 *(__PTR) = msecs_to_jiffies(__data);                    \
3944         else                                                            \
3945                 *(__PTR) = __data;                                      \
3946         return ret;                                                     \
3947 }
3948 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3949 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3950                 UINT_MAX, 1);
3951 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3952                 UINT_MAX, 1);
3953 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3954 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3955                 UINT_MAX, 0);
3956 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3957 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3958 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3959 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3960                 UINT_MAX, 0);
3961 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3962 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3963 #undef STORE_FUNCTION
3964
3965 #define CFQ_ATTR(name) \
3966         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3967
3968 static struct elv_fs_entry cfq_attrs[] = {
3969         CFQ_ATTR(quantum),
3970         CFQ_ATTR(fifo_expire_sync),
3971         CFQ_ATTR(fifo_expire_async),
3972         CFQ_ATTR(back_seek_max),
3973         CFQ_ATTR(back_seek_penalty),
3974         CFQ_ATTR(slice_sync),
3975         CFQ_ATTR(slice_async),
3976         CFQ_ATTR(slice_async_rq),
3977         CFQ_ATTR(slice_idle),
3978         CFQ_ATTR(low_latency),
3979         CFQ_ATTR(group_isolation),
3980         __ATTR_NULL
3981 };
3982
3983 static struct elevator_type iosched_cfq = {
3984         .ops = {
3985                 .elevator_merge_fn =            cfq_merge,
3986                 .elevator_merged_fn =           cfq_merged_request,
3987                 .elevator_merge_req_fn =        cfq_merged_requests,
3988                 .elevator_allow_merge_fn =      cfq_allow_merge,
3989                 .elevator_bio_merged_fn =       cfq_bio_merged,
3990                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3991                 .elevator_add_req_fn =          cfq_insert_request,
3992                 .elevator_activate_req_fn =     cfq_activate_request,
3993                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3994                 .elevator_queue_empty_fn =      cfq_queue_empty,
3995                 .elevator_completed_req_fn =    cfq_completed_request,
3996                 .elevator_former_req_fn =       elv_rb_former_request,
3997                 .elevator_latter_req_fn =       elv_rb_latter_request,
3998                 .elevator_set_req_fn =          cfq_set_request,
3999                 .elevator_put_req_fn =          cfq_put_request,
4000                 .elevator_may_queue_fn =        cfq_may_queue,
4001                 .elevator_init_fn =             cfq_init_queue,
4002                 .elevator_exit_fn =             cfq_exit_queue,
4003                 .trim =                         cfq_free_io_context,
4004         },
4005         .elevator_attrs =       cfq_attrs,
4006         .elevator_name =        "cfq",
4007         .elevator_owner =       THIS_MODULE,
4008 };
4009
4010 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4011 static struct blkio_policy_type blkio_policy_cfq = {
4012         .ops = {
4013                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4014                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4015         },
4016         .plid = BLKIO_POLICY_PROP,
4017 };
4018 #else
4019 static struct blkio_policy_type blkio_policy_cfq;
4020 #endif
4021
4022 static int __init cfq_init(void)
4023 {
4024         /*
4025          * could be 0 on HZ < 1000 setups
4026          */
4027         if (!cfq_slice_async)
4028                 cfq_slice_async = 1;
4029         if (!cfq_slice_idle)
4030                 cfq_slice_idle = 1;
4031
4032         if (cfq_slab_setup())
4033                 return -ENOMEM;
4034
4035         elv_register(&iosched_cfq);
4036         blkio_policy_register(&blkio_policy_cfq);
4037
4038         return 0;
4039 }
4040
4041 static void __exit cfq_exit(void)
4042 {
4043         DECLARE_COMPLETION_ONSTACK(all_gone);
4044         blkio_policy_unregister(&blkio_policy_cfq);
4045         elv_unregister(&iosched_cfq);
4046         ioc_gone = &all_gone;
4047         /* ioc_gone's update must be visible before reading ioc_count */
4048         smp_wmb();
4049
4050         /*
4051          * this also protects us from entering cfq_slab_kill() with
4052          * pending RCU callbacks
4053          */
4054         if (elv_ioc_count_read(cfq_ioc_count))
4055                 wait_for_completion(&all_gone);
4056         ida_destroy(&cic_index_ida);
4057         cfq_slab_kill();
4058 }
4059
4060 module_init(cfq_init);
4061 module_exit(cfq_exit);
4062
4063 MODULE_AUTHOR("Jens Axboe");
4064 MODULE_LICENSE("GPL");
4065 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");