Merge branch 'for-2.6.37/core' of git://git.kernel.dk/linux-2.6-block
[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 int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private)
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct rb_node *active;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
93                         .count = 0, .min_vdisktime = 0, }
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         atomic_t ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* pending metadata requests */
134         int meta_pending;
135         /* number of requests that are on the dispatch list or inside driver */
136         int dispatched;
137
138         /* io prio of this group */
139         unsigned short ioprio, org_ioprio;
140         unsigned short ioprio_class, org_ioprio_class;
141
142         pid_t pid;
143
144         u32 seek_history;
145         sector_t last_request_pos;
146
147         struct cfq_rb_root *service_tree;
148         struct cfq_queue *new_cfqq;
149         struct cfq_group *cfqg;
150         struct cfq_group *orig_cfqg;
151         /* Number of sectors dispatched from queue in single dispatch round */
152         unsigned long nr_sectors;
153 };
154
155 /*
156  * First index in the service_trees.
157  * IDLE is handled separately, so it has negative index
158  */
159 enum wl_prio_t {
160         BE_WORKLOAD = 0,
161         RT_WORKLOAD = 1,
162         IDLE_WORKLOAD = 2,
163         CFQ_PRIO_NR,
164 };
165
166 /*
167  * Second index in the service_trees.
168  */
169 enum wl_type_t {
170         ASYNC_WORKLOAD = 0,
171         SYNC_NOIDLE_WORKLOAD = 1,
172         SYNC_WORKLOAD = 2
173 };
174
175 /* This is per cgroup per device grouping structure */
176 struct cfq_group {
177         /* group service_tree member */
178         struct rb_node rb_node;
179
180         /* group service_tree key */
181         u64 vdisktime;
182         unsigned int weight;
183         bool on_st;
184
185         /* number of cfqq currently on this group */
186         int nr_cfqq;
187
188         /*
189          * Per group busy queus average. Useful for workload slice calc. We
190          * create the array for each prio class but at run time it is used
191          * only for RT and BE class and slot for IDLE class remains unused.
192          * This is primarily done to avoid confusion and a gcc warning.
193          */
194         unsigned int busy_queues_avg[CFQ_PRIO_NR];
195         /*
196          * rr lists of queues with requests. We maintain service trees for
197          * RT and BE classes. These trees are subdivided in subclasses
198          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199          * class there is no subclassification and all the cfq queues go on
200          * a single tree service_tree_idle.
201          * Counts are embedded in the cfq_rb_root
202          */
203         struct cfq_rb_root service_trees[2][3];
204         struct cfq_rb_root service_tree_idle;
205
206         unsigned long saved_workload_slice;
207         enum wl_type_t saved_workload;
208         enum wl_prio_t saved_serving_prio;
209         struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211         struct hlist_node cfqd_node;
212         atomic_t ref;
213 #endif
214         /* number of requests that are on the dispatch list or inside driver */
215         int dispatched;
216 };
217
218 /*
219  * Per block device queue structure
220  */
221 struct cfq_data {
222         struct request_queue *queue;
223         /* Root service tree for cfq_groups */
224         struct cfq_rb_root grp_service_tree;
225         struct cfq_group root_group;
226
227         /*
228          * The priority currently being served
229          */
230         enum wl_prio_t serving_prio;
231         enum wl_type_t serving_type;
232         unsigned long workload_expires;
233         struct cfq_group *serving_group;
234
235         /*
236          * Each priority tree is sorted by next_request position.  These
237          * trees are used when determining if two or more queues are
238          * interleaving requests (see cfq_close_cooperator).
239          */
240         struct rb_root prio_trees[CFQ_PRIO_LISTS];
241
242         unsigned int busy_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290         unsigned int cfq_group_isolation;
291
292         unsigned int cic_index;
293         struct list_head cic_list;
294
295         /*
296          * Fallback dummy cfqq for extreme OOM conditions
297          */
298         struct cfq_queue oom_cfqq;
299
300         unsigned long last_delayed_sync;
301
302         /* List of cfq groups being managed on this device*/
303         struct hlist_head cfqg_list;
304         struct rcu_head rcu;
305 };
306
307 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
308
309 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
310                                             enum wl_prio_t prio,
311                                             enum wl_type_t type)
312 {
313         if (!cfqg)
314                 return NULL;
315
316         if (prio == IDLE_WORKLOAD)
317                 return &cfqg->service_tree_idle;
318
319         return &cfqg->service_trees[prio][type];
320 }
321
322 enum cfqq_state_flags {
323         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
324         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
325         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
326         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
327         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
328         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
329         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
330         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
331         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
332         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
333         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
334         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
335         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
336 };
337
338 #define CFQ_CFQQ_FNS(name)                                              \
339 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
340 {                                                                       \
341         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
342 }                                                                       \
343 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
344 {                                                                       \
345         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
346 }                                                                       \
347 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
348 {                                                                       \
349         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
350 }
351
352 CFQ_CFQQ_FNS(on_rr);
353 CFQ_CFQQ_FNS(wait_request);
354 CFQ_CFQQ_FNS(must_dispatch);
355 CFQ_CFQQ_FNS(must_alloc_slice);
356 CFQ_CFQQ_FNS(fifo_expire);
357 CFQ_CFQQ_FNS(idle_window);
358 CFQ_CFQQ_FNS(prio_changed);
359 CFQ_CFQQ_FNS(slice_new);
360 CFQ_CFQQ_FNS(sync);
361 CFQ_CFQQ_FNS(coop);
362 CFQ_CFQQ_FNS(split_coop);
363 CFQ_CFQQ_FNS(deep);
364 CFQ_CFQQ_FNS(wait_busy);
365 #undef CFQ_CFQQ_FNS
366
367 #ifdef CONFIG_CFQ_GROUP_IOSCHED
368 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
369         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
370                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
371                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
372
373 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
374         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
375                                 blkg_path(&(cfqg)->blkg), ##args);      \
376
377 #else
378 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
379         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
380 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
381 #endif
382 #define cfq_log(cfqd, fmt, args...)     \
383         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
384
385 /* Traverses through cfq group service trees */
386 #define for_each_cfqg_st(cfqg, i, j, st) \
387         for (i = 0; i <= IDLE_WORKLOAD; i++) \
388                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
389                         : &cfqg->service_tree_idle; \
390                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
391                         (i == IDLE_WORKLOAD && j == 0); \
392                         j++, st = i < IDLE_WORKLOAD ? \
393                         &cfqg->service_trees[i][j]: NULL) \
394
395
396 static inline bool iops_mode(struct cfq_data *cfqd)
397 {
398         /*
399          * If we are not idling on queues and it is a NCQ drive, parallel
400          * execution of requests is on and measuring time is not possible
401          * in most of the cases until and unless we drive shallower queue
402          * depths and that becomes a performance bottleneck. In such cases
403          * switch to start providing fairness in terms of number of IOs.
404          */
405         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
406                 return true;
407         else
408                 return false;
409 }
410
411 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
412 {
413         if (cfq_class_idle(cfqq))
414                 return IDLE_WORKLOAD;
415         if (cfq_class_rt(cfqq))
416                 return RT_WORKLOAD;
417         return BE_WORKLOAD;
418 }
419
420
421 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
422 {
423         if (!cfq_cfqq_sync(cfqq))
424                 return ASYNC_WORKLOAD;
425         if (!cfq_cfqq_idle_window(cfqq))
426                 return SYNC_NOIDLE_WORKLOAD;
427         return SYNC_WORKLOAD;
428 }
429
430 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
431                                         struct cfq_data *cfqd,
432                                         struct cfq_group *cfqg)
433 {
434         if (wl == IDLE_WORKLOAD)
435                 return cfqg->service_tree_idle.count;
436
437         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
438                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
439                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
440 }
441
442 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
443                                         struct cfq_group *cfqg)
444 {
445         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
446                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
447 }
448
449 static void cfq_dispatch_insert(struct request_queue *, struct request *);
450 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
451                                        struct io_context *, gfp_t);
452 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
453                                                 struct io_context *);
454
455 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
456                                             bool is_sync)
457 {
458         return cic->cfqq[is_sync];
459 }
460
461 static inline void cic_set_cfqq(struct cfq_io_context *cic,
462                                 struct cfq_queue *cfqq, bool is_sync)
463 {
464         cic->cfqq[is_sync] = cfqq;
465 }
466
467 #define CIC_DEAD_KEY    1ul
468 #define CIC_DEAD_INDEX_SHIFT    1
469
470 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
471 {
472         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
473 }
474
475 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
476 {
477         struct cfq_data *cfqd = cic->key;
478
479         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
480                 return NULL;
481
482         return cfqd;
483 }
484
485 /*
486  * We regard a request as SYNC, if it's either a read or has the SYNC bit
487  * set (in which case it could also be direct WRITE).
488  */
489 static inline bool cfq_bio_sync(struct bio *bio)
490 {
491         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
492 }
493
494 /*
495  * scheduler run of queue, if there are requests pending and no one in the
496  * driver that will restart queueing
497  */
498 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
499 {
500         if (cfqd->busy_queues) {
501                 cfq_log(cfqd, "schedule dispatch");
502                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
503         }
504 }
505
506 static int cfq_queue_empty(struct request_queue *q)
507 {
508         struct cfq_data *cfqd = q->elevator->elevator_data;
509
510         return !cfqd->rq_queued;
511 }
512
513 /*
514  * Scale schedule slice based on io priority. Use the sync time slice only
515  * if a queue is marked sync and has sync io queued. A sync queue with async
516  * io only, should not get full sync slice length.
517  */
518 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
519                                  unsigned short prio)
520 {
521         const int base_slice = cfqd->cfq_slice[sync];
522
523         WARN_ON(prio >= IOPRIO_BE_NR);
524
525         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
526 }
527
528 static inline int
529 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
530 {
531         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
532 }
533
534 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
535 {
536         u64 d = delta << CFQ_SERVICE_SHIFT;
537
538         d = d * BLKIO_WEIGHT_DEFAULT;
539         do_div(d, cfqg->weight);
540         return d;
541 }
542
543 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
544 {
545         s64 delta = (s64)(vdisktime - min_vdisktime);
546         if (delta > 0)
547                 min_vdisktime = vdisktime;
548
549         return min_vdisktime;
550 }
551
552 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
553 {
554         s64 delta = (s64)(vdisktime - min_vdisktime);
555         if (delta < 0)
556                 min_vdisktime = vdisktime;
557
558         return min_vdisktime;
559 }
560
561 static void update_min_vdisktime(struct cfq_rb_root *st)
562 {
563         u64 vdisktime = st->min_vdisktime;
564         struct cfq_group *cfqg;
565
566         if (st->active) {
567                 cfqg = rb_entry_cfqg(st->active);
568                 vdisktime = cfqg->vdisktime;
569         }
570
571         if (st->left) {
572                 cfqg = rb_entry_cfqg(st->left);
573                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
574         }
575
576         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
577 }
578
579 /*
580  * get averaged number of queues of RT/BE priority.
581  * average is updated, with a formula that gives more weight to higher numbers,
582  * to quickly follows sudden increases and decrease slowly
583  */
584
585 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
586                                         struct cfq_group *cfqg, bool rt)
587 {
588         unsigned min_q, max_q;
589         unsigned mult  = cfq_hist_divisor - 1;
590         unsigned round = cfq_hist_divisor / 2;
591         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
592
593         min_q = min(cfqg->busy_queues_avg[rt], busy);
594         max_q = max(cfqg->busy_queues_avg[rt], busy);
595         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
596                 cfq_hist_divisor;
597         return cfqg->busy_queues_avg[rt];
598 }
599
600 static inline unsigned
601 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
602 {
603         struct cfq_rb_root *st = &cfqd->grp_service_tree;
604
605         return cfq_target_latency * cfqg->weight / st->total_weight;
606 }
607
608 static inline void
609 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
610 {
611         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
612         if (cfqd->cfq_latency) {
613                 /*
614                  * interested queues (we consider only the ones with the same
615                  * priority class in the cfq group)
616                  */
617                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
618                                                 cfq_class_rt(cfqq));
619                 unsigned sync_slice = cfqd->cfq_slice[1];
620                 unsigned expect_latency = sync_slice * iq;
621                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
622
623                 if (expect_latency > group_slice) {
624                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
625                         /* scale low_slice according to IO priority
626                          * and sync vs async */
627                         unsigned low_slice =
628                                 min(slice, base_low_slice * slice / sync_slice);
629                         /* the adapted slice value is scaled to fit all iqs
630                          * into the target latency */
631                         slice = max(slice * group_slice / expect_latency,
632                                     low_slice);
633                 }
634         }
635         cfqq->slice_start = jiffies;
636         cfqq->slice_end = jiffies + slice;
637         cfqq->allocated_slice = slice;
638         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
639 }
640
641 /*
642  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
643  * isn't valid until the first request from the dispatch is activated
644  * and the slice time set.
645  */
646 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
647 {
648         if (cfq_cfqq_slice_new(cfqq))
649                 return 0;
650         if (time_before(jiffies, cfqq->slice_end))
651                 return 0;
652
653         return 1;
654 }
655
656 /*
657  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
658  * We choose the request that is closest to the head right now. Distance
659  * behind the head is penalized and only allowed to a certain extent.
660  */
661 static struct request *
662 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
663 {
664         sector_t s1, s2, d1 = 0, d2 = 0;
665         unsigned long back_max;
666 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
667 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
668         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
669
670         if (rq1 == NULL || rq1 == rq2)
671                 return rq2;
672         if (rq2 == NULL)
673                 return rq1;
674
675         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
676                 return rq1;
677         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
678                 return rq2;
679         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
680                 return rq1;
681         else if ((rq2->cmd_flags & REQ_META) &&
682                  !(rq1->cmd_flags & REQ_META))
683                 return rq2;
684
685         s1 = blk_rq_pos(rq1);
686         s2 = blk_rq_pos(rq2);
687
688         /*
689          * by definition, 1KiB is 2 sectors
690          */
691         back_max = cfqd->cfq_back_max * 2;
692
693         /*
694          * Strict one way elevator _except_ in the case where we allow
695          * short backward seeks which are biased as twice the cost of a
696          * similar forward seek.
697          */
698         if (s1 >= last)
699                 d1 = s1 - last;
700         else if (s1 + back_max >= last)
701                 d1 = (last - s1) * cfqd->cfq_back_penalty;
702         else
703                 wrap |= CFQ_RQ1_WRAP;
704
705         if (s2 >= last)
706                 d2 = s2 - last;
707         else if (s2 + back_max >= last)
708                 d2 = (last - s2) * cfqd->cfq_back_penalty;
709         else
710                 wrap |= CFQ_RQ2_WRAP;
711
712         /* Found required data */
713
714         /*
715          * By doing switch() on the bit mask "wrap" we avoid having to
716          * check two variables for all permutations: --> faster!
717          */
718         switch (wrap) {
719         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
720                 if (d1 < d2)
721                         return rq1;
722                 else if (d2 < d1)
723                         return rq2;
724                 else {
725                         if (s1 >= s2)
726                                 return rq1;
727                         else
728                                 return rq2;
729                 }
730
731         case CFQ_RQ2_WRAP:
732                 return rq1;
733         case CFQ_RQ1_WRAP:
734                 return rq2;
735         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
736         default:
737                 /*
738                  * Since both rqs are wrapped,
739                  * start with the one that's further behind head
740                  * (--> only *one* back seek required),
741                  * since back seek takes more time than forward.
742                  */
743                 if (s1 <= s2)
744                         return rq1;
745                 else
746                         return rq2;
747         }
748 }
749
750 /*
751  * The below is leftmost cache rbtree addon
752  */
753 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
754 {
755         /* Service tree is empty */
756         if (!root->count)
757                 return NULL;
758
759         if (!root->left)
760                 root->left = rb_first(&root->rb);
761
762         if (root->left)
763                 return rb_entry(root->left, struct cfq_queue, rb_node);
764
765         return NULL;
766 }
767
768 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
769 {
770         if (!root->left)
771                 root->left = rb_first(&root->rb);
772
773         if (root->left)
774                 return rb_entry_cfqg(root->left);
775
776         return NULL;
777 }
778
779 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
780 {
781         rb_erase(n, root);
782         RB_CLEAR_NODE(n);
783 }
784
785 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
786 {
787         if (root->left == n)
788                 root->left = NULL;
789         rb_erase_init(n, &root->rb);
790         --root->count;
791 }
792
793 /*
794  * would be nice to take fifo expire time into account as well
795  */
796 static struct request *
797 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
798                   struct request *last)
799 {
800         struct rb_node *rbnext = rb_next(&last->rb_node);
801         struct rb_node *rbprev = rb_prev(&last->rb_node);
802         struct request *next = NULL, *prev = NULL;
803
804         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
805
806         if (rbprev)
807                 prev = rb_entry_rq(rbprev);
808
809         if (rbnext)
810                 next = rb_entry_rq(rbnext);
811         else {
812                 rbnext = rb_first(&cfqq->sort_list);
813                 if (rbnext && rbnext != &last->rb_node)
814                         next = rb_entry_rq(rbnext);
815         }
816
817         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
818 }
819
820 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
821                                       struct cfq_queue *cfqq)
822 {
823         /*
824          * just an approximation, should be ok.
825          */
826         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
827                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
828 }
829
830 static inline s64
831 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
832 {
833         return cfqg->vdisktime - st->min_vdisktime;
834 }
835
836 static void
837 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
838 {
839         struct rb_node **node = &st->rb.rb_node;
840         struct rb_node *parent = NULL;
841         struct cfq_group *__cfqg;
842         s64 key = cfqg_key(st, cfqg);
843         int left = 1;
844
845         while (*node != NULL) {
846                 parent = *node;
847                 __cfqg = rb_entry_cfqg(parent);
848
849                 if (key < cfqg_key(st, __cfqg))
850                         node = &parent->rb_left;
851                 else {
852                         node = &parent->rb_right;
853                         left = 0;
854                 }
855         }
856
857         if (left)
858                 st->left = &cfqg->rb_node;
859
860         rb_link_node(&cfqg->rb_node, parent, node);
861         rb_insert_color(&cfqg->rb_node, &st->rb);
862 }
863
864 static void
865 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
866 {
867         struct cfq_rb_root *st = &cfqd->grp_service_tree;
868         struct cfq_group *__cfqg;
869         struct rb_node *n;
870
871         cfqg->nr_cfqq++;
872         if (cfqg->on_st)
873                 return;
874
875         /*
876          * Currently put the group at the end. Later implement something
877          * so that groups get lesser vtime based on their weights, so that
878          * if group does not loose all if it was not continously backlogged.
879          */
880         n = rb_last(&st->rb);
881         if (n) {
882                 __cfqg = rb_entry_cfqg(n);
883                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
884         } else
885                 cfqg->vdisktime = st->min_vdisktime;
886
887         __cfq_group_service_tree_add(st, cfqg);
888         cfqg->on_st = true;
889         st->total_weight += cfqg->weight;
890 }
891
892 static void
893 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
894 {
895         struct cfq_rb_root *st = &cfqd->grp_service_tree;
896
897         if (st->active == &cfqg->rb_node)
898                 st->active = NULL;
899
900         BUG_ON(cfqg->nr_cfqq < 1);
901         cfqg->nr_cfqq--;
902
903         /* If there are other cfq queues under this group, don't delete it */
904         if (cfqg->nr_cfqq)
905                 return;
906
907         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
908         cfqg->on_st = false;
909         st->total_weight -= cfqg->weight;
910         if (!RB_EMPTY_NODE(&cfqg->rb_node))
911                 cfq_rb_erase(&cfqg->rb_node, st);
912         cfqg->saved_workload_slice = 0;
913         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
914 }
915
916 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
917 {
918         unsigned int slice_used;
919
920         /*
921          * Queue got expired before even a single request completed or
922          * got expired immediately after first request completion.
923          */
924         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
925                 /*
926                  * Also charge the seek time incurred to the group, otherwise
927                  * if there are mutiple queues in the group, each can dispatch
928                  * a single request on seeky media and cause lots of seek time
929                  * and group will never know it.
930                  */
931                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
932                                         1);
933         } else {
934                 slice_used = jiffies - cfqq->slice_start;
935                 if (slice_used > cfqq->allocated_slice)
936                         slice_used = cfqq->allocated_slice;
937         }
938
939         return slice_used;
940 }
941
942 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
943                                 struct cfq_queue *cfqq)
944 {
945         struct cfq_rb_root *st = &cfqd->grp_service_tree;
946         unsigned int used_sl, charge;
947         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
948                         - cfqg->service_tree_idle.count;
949
950         BUG_ON(nr_sync < 0);
951         used_sl = charge = cfq_cfqq_slice_usage(cfqq);
952
953         if (iops_mode(cfqd))
954                 charge = cfqq->slice_dispatch;
955         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
956                 charge = cfqq->allocated_slice;
957
958         /* Can't update vdisktime while group is on service tree */
959         cfq_rb_erase(&cfqg->rb_node, st);
960         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
961         __cfq_group_service_tree_add(st, cfqg);
962
963         /* This group is being expired. Save the context */
964         if (time_after(cfqd->workload_expires, jiffies)) {
965                 cfqg->saved_workload_slice = cfqd->workload_expires
966                                                 - jiffies;
967                 cfqg->saved_workload = cfqd->serving_type;
968                 cfqg->saved_serving_prio = cfqd->serving_prio;
969         } else
970                 cfqg->saved_workload_slice = 0;
971
972         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
973                                         st->min_vdisktime);
974         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
975                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
976                         iops_mode(cfqd), cfqq->nr_sectors);
977         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
978         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
979 }
980
981 #ifdef CONFIG_CFQ_GROUP_IOSCHED
982 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
983 {
984         if (blkg)
985                 return container_of(blkg, struct cfq_group, blkg);
986         return NULL;
987 }
988
989 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
990                                         unsigned int weight)
991 {
992         cfqg_of_blkg(blkg)->weight = weight;
993 }
994
995 static struct cfq_group *
996 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
997 {
998         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
999         struct cfq_group *cfqg = NULL;
1000         void *key = cfqd;
1001         int i, j;
1002         struct cfq_rb_root *st;
1003         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1004         unsigned int major, minor;
1005
1006         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1007         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1008                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1009                 cfqg->blkg.dev = MKDEV(major, minor);
1010                 goto done;
1011         }
1012         if (cfqg || !create)
1013                 goto done;
1014
1015         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1016         if (!cfqg)
1017                 goto done;
1018
1019         for_each_cfqg_st(cfqg, i, j, st)
1020                 *st = CFQ_RB_ROOT;
1021         RB_CLEAR_NODE(&cfqg->rb_node);
1022
1023         /*
1024          * Take the initial reference that will be released on destroy
1025          * This can be thought of a joint reference by cgroup and
1026          * elevator which will be dropped by either elevator exit
1027          * or cgroup deletion path depending on who is exiting first.
1028          */
1029         atomic_set(&cfqg->ref, 1);
1030
1031         /*
1032          * Add group onto cgroup list. It might happen that bdi->dev is
1033          * not initiliazed yet. Initialize this new group without major
1034          * and minor info and this info will be filled in once a new thread
1035          * comes for IO. See code above.
1036          */
1037         if (bdi->dev) {
1038                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1039                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1040                                         MKDEV(major, minor));
1041         } else
1042                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1043                                         0);
1044
1045         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1046
1047         /* Add group on cfqd list */
1048         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1049
1050 done:
1051         return cfqg;
1052 }
1053
1054 /*
1055  * Search for the cfq group current task belongs to. If create = 1, then also
1056  * create the cfq group if it does not exist. request_queue lock must be held.
1057  */
1058 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1059 {
1060         struct cgroup *cgroup;
1061         struct cfq_group *cfqg = NULL;
1062
1063         rcu_read_lock();
1064         cgroup = task_cgroup(current, blkio_subsys_id);
1065         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1066         if (!cfqg && create)
1067                 cfqg = &cfqd->root_group;
1068         rcu_read_unlock();
1069         return cfqg;
1070 }
1071
1072 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1073 {
1074         atomic_inc(&cfqg->ref);
1075         return cfqg;
1076 }
1077
1078 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1079 {
1080         /* Currently, all async queues are mapped to root group */
1081         if (!cfq_cfqq_sync(cfqq))
1082                 cfqg = &cfqq->cfqd->root_group;
1083
1084         cfqq->cfqg = cfqg;
1085         /* cfqq reference on cfqg */
1086         atomic_inc(&cfqq->cfqg->ref);
1087 }
1088
1089 static void cfq_put_cfqg(struct cfq_group *cfqg)
1090 {
1091         struct cfq_rb_root *st;
1092         int i, j;
1093
1094         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1095         if (!atomic_dec_and_test(&cfqg->ref))
1096                 return;
1097         for_each_cfqg_st(cfqg, i, j, st)
1098                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1099         kfree(cfqg);
1100 }
1101
1102 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1103 {
1104         /* Something wrong if we are trying to remove same group twice */
1105         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1106
1107         hlist_del_init(&cfqg->cfqd_node);
1108
1109         /*
1110          * Put the reference taken at the time of creation so that when all
1111          * queues are gone, group can be destroyed.
1112          */
1113         cfq_put_cfqg(cfqg);
1114 }
1115
1116 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1117 {
1118         struct hlist_node *pos, *n;
1119         struct cfq_group *cfqg;
1120
1121         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1122                 /*
1123                  * If cgroup removal path got to blk_group first and removed
1124                  * it from cgroup list, then it will take care of destroying
1125                  * cfqg also.
1126                  */
1127                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1128                         cfq_destroy_cfqg(cfqd, cfqg);
1129         }
1130 }
1131
1132 /*
1133  * Blk cgroup controller notification saying that blkio_group object is being
1134  * delinked as associated cgroup object is going away. That also means that
1135  * no new IO will come in this group. So get rid of this group as soon as
1136  * any pending IO in the group is finished.
1137  *
1138  * This function is called under rcu_read_lock(). key is the rcu protected
1139  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1140  * read lock.
1141  *
1142  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1143  * it should not be NULL as even if elevator was exiting, cgroup deltion
1144  * path got to it first.
1145  */
1146 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1147 {
1148         unsigned long  flags;
1149         struct cfq_data *cfqd = key;
1150
1151         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1152         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1153         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1154 }
1155
1156 #else /* GROUP_IOSCHED */
1157 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1158 {
1159         return &cfqd->root_group;
1160 }
1161
1162 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1163 {
1164         return cfqg;
1165 }
1166
1167 static inline void
1168 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1169         cfqq->cfqg = cfqg;
1170 }
1171
1172 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1173 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1174
1175 #endif /* GROUP_IOSCHED */
1176
1177 /*
1178  * The cfqd->service_trees holds all pending cfq_queue's that have
1179  * requests waiting to be processed. It is sorted in the order that
1180  * we will service the queues.
1181  */
1182 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1183                                  bool add_front)
1184 {
1185         struct rb_node **p, *parent;
1186         struct cfq_queue *__cfqq;
1187         unsigned long rb_key;
1188         struct cfq_rb_root *service_tree;
1189         int left;
1190         int new_cfqq = 1;
1191         int group_changed = 0;
1192
1193 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1194         if (!cfqd->cfq_group_isolation
1195             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1196             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1197                 /* Move this cfq to root group */
1198                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1199                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1200                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1201                 cfqq->orig_cfqg = cfqq->cfqg;
1202                 cfqq->cfqg = &cfqd->root_group;
1203                 atomic_inc(&cfqd->root_group.ref);
1204                 group_changed = 1;
1205         } else if (!cfqd->cfq_group_isolation
1206                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1207                 /* cfqq is sequential now needs to go to its original group */
1208                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1209                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1210                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1211                 cfq_put_cfqg(cfqq->cfqg);
1212                 cfqq->cfqg = cfqq->orig_cfqg;
1213                 cfqq->orig_cfqg = NULL;
1214                 group_changed = 1;
1215                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1216         }
1217 #endif
1218
1219         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1220                                                 cfqq_type(cfqq));
1221         if (cfq_class_idle(cfqq)) {
1222                 rb_key = CFQ_IDLE_DELAY;
1223                 parent = rb_last(&service_tree->rb);
1224                 if (parent && parent != &cfqq->rb_node) {
1225                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1226                         rb_key += __cfqq->rb_key;
1227                 } else
1228                         rb_key += jiffies;
1229         } else if (!add_front) {
1230                 /*
1231                  * Get our rb key offset. Subtract any residual slice
1232                  * value carried from last service. A negative resid
1233                  * count indicates slice overrun, and this should position
1234                  * the next service time further away in the tree.
1235                  */
1236                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1237                 rb_key -= cfqq->slice_resid;
1238                 cfqq->slice_resid = 0;
1239         } else {
1240                 rb_key = -HZ;
1241                 __cfqq = cfq_rb_first(service_tree);
1242                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1243         }
1244
1245         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1246                 new_cfqq = 0;
1247                 /*
1248                  * same position, nothing more to do
1249                  */
1250                 if (rb_key == cfqq->rb_key &&
1251                     cfqq->service_tree == service_tree)
1252                         return;
1253
1254                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1255                 cfqq->service_tree = NULL;
1256         }
1257
1258         left = 1;
1259         parent = NULL;
1260         cfqq->service_tree = service_tree;
1261         p = &service_tree->rb.rb_node;
1262         while (*p) {
1263                 struct rb_node **n;
1264
1265                 parent = *p;
1266                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1267
1268                 /*
1269                  * sort by key, that represents service time.
1270                  */
1271                 if (time_before(rb_key, __cfqq->rb_key))
1272                         n = &(*p)->rb_left;
1273                 else {
1274                         n = &(*p)->rb_right;
1275                         left = 0;
1276                 }
1277
1278                 p = n;
1279         }
1280
1281         if (left)
1282                 service_tree->left = &cfqq->rb_node;
1283
1284         cfqq->rb_key = rb_key;
1285         rb_link_node(&cfqq->rb_node, parent, p);
1286         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1287         service_tree->count++;
1288         if ((add_front || !new_cfqq) && !group_changed)
1289                 return;
1290         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1291 }
1292
1293 static struct cfq_queue *
1294 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1295                      sector_t sector, struct rb_node **ret_parent,
1296                      struct rb_node ***rb_link)
1297 {
1298         struct rb_node **p, *parent;
1299         struct cfq_queue *cfqq = NULL;
1300
1301         parent = NULL;
1302         p = &root->rb_node;
1303         while (*p) {
1304                 struct rb_node **n;
1305
1306                 parent = *p;
1307                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1308
1309                 /*
1310                  * Sort strictly based on sector.  Smallest to the left,
1311                  * largest to the right.
1312                  */
1313                 if (sector > blk_rq_pos(cfqq->next_rq))
1314                         n = &(*p)->rb_right;
1315                 else if (sector < blk_rq_pos(cfqq->next_rq))
1316                         n = &(*p)->rb_left;
1317                 else
1318                         break;
1319                 p = n;
1320                 cfqq = NULL;
1321         }
1322
1323         *ret_parent = parent;
1324         if (rb_link)
1325                 *rb_link = p;
1326         return cfqq;
1327 }
1328
1329 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1330 {
1331         struct rb_node **p, *parent;
1332         struct cfq_queue *__cfqq;
1333
1334         if (cfqq->p_root) {
1335                 rb_erase(&cfqq->p_node, cfqq->p_root);
1336                 cfqq->p_root = NULL;
1337         }
1338
1339         if (cfq_class_idle(cfqq))
1340                 return;
1341         if (!cfqq->next_rq)
1342                 return;
1343
1344         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1345         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1346                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1347         if (!__cfqq) {
1348                 rb_link_node(&cfqq->p_node, parent, p);
1349                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1350         } else
1351                 cfqq->p_root = NULL;
1352 }
1353
1354 /*
1355  * Update cfqq's position in the service tree.
1356  */
1357 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1358 {
1359         /*
1360          * Resorting requires the cfqq to be on the RR list already.
1361          */
1362         if (cfq_cfqq_on_rr(cfqq)) {
1363                 cfq_service_tree_add(cfqd, cfqq, 0);
1364                 cfq_prio_tree_add(cfqd, cfqq);
1365         }
1366 }
1367
1368 /*
1369  * add to busy list of queues for service, trying to be fair in ordering
1370  * the pending list according to last request service
1371  */
1372 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1373 {
1374         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1375         BUG_ON(cfq_cfqq_on_rr(cfqq));
1376         cfq_mark_cfqq_on_rr(cfqq);
1377         cfqd->busy_queues++;
1378
1379         cfq_resort_rr_list(cfqd, cfqq);
1380 }
1381
1382 /*
1383  * Called when the cfqq no longer has requests pending, remove it from
1384  * the service tree.
1385  */
1386 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1387 {
1388         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1389         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1390         cfq_clear_cfqq_on_rr(cfqq);
1391
1392         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1393                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1394                 cfqq->service_tree = NULL;
1395         }
1396         if (cfqq->p_root) {
1397                 rb_erase(&cfqq->p_node, cfqq->p_root);
1398                 cfqq->p_root = NULL;
1399         }
1400
1401         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1402         BUG_ON(!cfqd->busy_queues);
1403         cfqd->busy_queues--;
1404 }
1405
1406 /*
1407  * rb tree support functions
1408  */
1409 static void cfq_del_rq_rb(struct request *rq)
1410 {
1411         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1412         const int sync = rq_is_sync(rq);
1413
1414         BUG_ON(!cfqq->queued[sync]);
1415         cfqq->queued[sync]--;
1416
1417         elv_rb_del(&cfqq->sort_list, rq);
1418
1419         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1420                 /*
1421                  * Queue will be deleted from service tree when we actually
1422                  * expire it later. Right now just remove it from prio tree
1423                  * as it is empty.
1424                  */
1425                 if (cfqq->p_root) {
1426                         rb_erase(&cfqq->p_node, cfqq->p_root);
1427                         cfqq->p_root = NULL;
1428                 }
1429         }
1430 }
1431
1432 static void cfq_add_rq_rb(struct request *rq)
1433 {
1434         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1435         struct cfq_data *cfqd = cfqq->cfqd;
1436         struct request *__alias, *prev;
1437
1438         cfqq->queued[rq_is_sync(rq)]++;
1439
1440         /*
1441          * looks a little odd, but the first insert might return an alias.
1442          * if that happens, put the alias on the dispatch list
1443          */
1444         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1445                 cfq_dispatch_insert(cfqd->queue, __alias);
1446
1447         if (!cfq_cfqq_on_rr(cfqq))
1448                 cfq_add_cfqq_rr(cfqd, cfqq);
1449
1450         /*
1451          * check if this request is a better next-serve candidate
1452          */
1453         prev = cfqq->next_rq;
1454         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1455
1456         /*
1457          * adjust priority tree position, if ->next_rq changes
1458          */
1459         if (prev != cfqq->next_rq)
1460                 cfq_prio_tree_add(cfqd, cfqq);
1461
1462         BUG_ON(!cfqq->next_rq);
1463 }
1464
1465 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1466 {
1467         elv_rb_del(&cfqq->sort_list, rq);
1468         cfqq->queued[rq_is_sync(rq)]--;
1469         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1470                                         rq_data_dir(rq), rq_is_sync(rq));
1471         cfq_add_rq_rb(rq);
1472         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1473                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1474                         rq_is_sync(rq));
1475 }
1476
1477 static struct request *
1478 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1479 {
1480         struct task_struct *tsk = current;
1481         struct cfq_io_context *cic;
1482         struct cfq_queue *cfqq;
1483
1484         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1485         if (!cic)
1486                 return NULL;
1487
1488         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1489         if (cfqq) {
1490                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1491
1492                 return elv_rb_find(&cfqq->sort_list, sector);
1493         }
1494
1495         return NULL;
1496 }
1497
1498 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1499 {
1500         struct cfq_data *cfqd = q->elevator->elevator_data;
1501
1502         cfqd->rq_in_driver++;
1503         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1504                                                 cfqd->rq_in_driver);
1505
1506         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1507 }
1508
1509 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1510 {
1511         struct cfq_data *cfqd = q->elevator->elevator_data;
1512
1513         WARN_ON(!cfqd->rq_in_driver);
1514         cfqd->rq_in_driver--;
1515         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1516                                                 cfqd->rq_in_driver);
1517 }
1518
1519 static void cfq_remove_request(struct request *rq)
1520 {
1521         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1522
1523         if (cfqq->next_rq == rq)
1524                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1525
1526         list_del_init(&rq->queuelist);
1527         cfq_del_rq_rb(rq);
1528
1529         cfqq->cfqd->rq_queued--;
1530         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1531                                         rq_data_dir(rq), rq_is_sync(rq));
1532         if (rq->cmd_flags & REQ_META) {
1533                 WARN_ON(!cfqq->meta_pending);
1534                 cfqq->meta_pending--;
1535         }
1536 }
1537
1538 static int cfq_merge(struct request_queue *q, struct request **req,
1539                      struct bio *bio)
1540 {
1541         struct cfq_data *cfqd = q->elevator->elevator_data;
1542         struct request *__rq;
1543
1544         __rq = cfq_find_rq_fmerge(cfqd, bio);
1545         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1546                 *req = __rq;
1547                 return ELEVATOR_FRONT_MERGE;
1548         }
1549
1550         return ELEVATOR_NO_MERGE;
1551 }
1552
1553 static void cfq_merged_request(struct request_queue *q, struct request *req,
1554                                int type)
1555 {
1556         if (type == ELEVATOR_FRONT_MERGE) {
1557                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1558
1559                 cfq_reposition_rq_rb(cfqq, req);
1560         }
1561 }
1562
1563 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1564                                 struct bio *bio)
1565 {
1566         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1567                                         bio_data_dir(bio), cfq_bio_sync(bio));
1568 }
1569
1570 static void
1571 cfq_merged_requests(struct request_queue *q, struct request *rq,
1572                     struct request *next)
1573 {
1574         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1575         /*
1576          * reposition in fifo if next is older than rq
1577          */
1578         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1579             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1580                 list_move(&rq->queuelist, &next->queuelist);
1581                 rq_set_fifo_time(rq, rq_fifo_time(next));
1582         }
1583
1584         if (cfqq->next_rq == next)
1585                 cfqq->next_rq = rq;
1586         cfq_remove_request(next);
1587         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1588                                         rq_data_dir(next), rq_is_sync(next));
1589 }
1590
1591 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1592                            struct bio *bio)
1593 {
1594         struct cfq_data *cfqd = q->elevator->elevator_data;
1595         struct cfq_io_context *cic;
1596         struct cfq_queue *cfqq;
1597
1598         /*
1599          * Disallow merge of a sync bio into an async request.
1600          */
1601         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1602                 return false;
1603
1604         /*
1605          * Lookup the cfqq that this bio will be queued with. Allow
1606          * merge only if rq is queued there.
1607          */
1608         cic = cfq_cic_lookup(cfqd, current->io_context);
1609         if (!cic)
1610                 return false;
1611
1612         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1613         return cfqq == RQ_CFQQ(rq);
1614 }
1615
1616 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1617 {
1618         del_timer(&cfqd->idle_slice_timer);
1619         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1620 }
1621
1622 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1623                                    struct cfq_queue *cfqq)
1624 {
1625         if (cfqq) {
1626                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1627                                 cfqd->serving_prio, cfqd->serving_type);
1628                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1629                 cfqq->slice_start = 0;
1630                 cfqq->dispatch_start = jiffies;
1631                 cfqq->allocated_slice = 0;
1632                 cfqq->slice_end = 0;
1633                 cfqq->slice_dispatch = 0;
1634                 cfqq->nr_sectors = 0;
1635
1636                 cfq_clear_cfqq_wait_request(cfqq);
1637                 cfq_clear_cfqq_must_dispatch(cfqq);
1638                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1639                 cfq_clear_cfqq_fifo_expire(cfqq);
1640                 cfq_mark_cfqq_slice_new(cfqq);
1641
1642                 cfq_del_timer(cfqd, cfqq);
1643         }
1644
1645         cfqd->active_queue = cfqq;
1646 }
1647
1648 /*
1649  * current cfqq expired its slice (or was too idle), select new one
1650  */
1651 static void
1652 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1653                     bool timed_out)
1654 {
1655         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1656
1657         if (cfq_cfqq_wait_request(cfqq))
1658                 cfq_del_timer(cfqd, cfqq);
1659
1660         cfq_clear_cfqq_wait_request(cfqq);
1661         cfq_clear_cfqq_wait_busy(cfqq);
1662
1663         /*
1664          * If this cfqq is shared between multiple processes, check to
1665          * make sure that those processes are still issuing I/Os within
1666          * the mean seek distance.  If not, it may be time to break the
1667          * queues apart again.
1668          */
1669         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1670                 cfq_mark_cfqq_split_coop(cfqq);
1671
1672         /*
1673          * store what was left of this slice, if the queue idled/timed out
1674          */
1675         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1676                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1677                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1678         }
1679
1680         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1681
1682         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1683                 cfq_del_cfqq_rr(cfqd, cfqq);
1684
1685         cfq_resort_rr_list(cfqd, cfqq);
1686
1687         if (cfqq == cfqd->active_queue)
1688                 cfqd->active_queue = NULL;
1689
1690         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1691                 cfqd->grp_service_tree.active = NULL;
1692
1693         if (cfqd->active_cic) {
1694                 put_io_context(cfqd->active_cic->ioc);
1695                 cfqd->active_cic = NULL;
1696         }
1697 }
1698
1699 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1700 {
1701         struct cfq_queue *cfqq = cfqd->active_queue;
1702
1703         if (cfqq)
1704                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1705 }
1706
1707 /*
1708  * Get next queue for service. Unless we have a queue preemption,
1709  * we'll simply select the first cfqq in the service tree.
1710  */
1711 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1712 {
1713         struct cfq_rb_root *service_tree =
1714                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1715                                         cfqd->serving_type);
1716
1717         if (!cfqd->rq_queued)
1718                 return NULL;
1719
1720         /* There is nothing to dispatch */
1721         if (!service_tree)
1722                 return NULL;
1723         if (RB_EMPTY_ROOT(&service_tree->rb))
1724                 return NULL;
1725         return cfq_rb_first(service_tree);
1726 }
1727
1728 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1729 {
1730         struct cfq_group *cfqg;
1731         struct cfq_queue *cfqq;
1732         int i, j;
1733         struct cfq_rb_root *st;
1734
1735         if (!cfqd->rq_queued)
1736                 return NULL;
1737
1738         cfqg = cfq_get_next_cfqg(cfqd);
1739         if (!cfqg)
1740                 return NULL;
1741
1742         for_each_cfqg_st(cfqg, i, j, st)
1743                 if ((cfqq = cfq_rb_first(st)) != NULL)
1744                         return cfqq;
1745         return NULL;
1746 }
1747
1748 /*
1749  * Get and set a new active queue for service.
1750  */
1751 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1752                                               struct cfq_queue *cfqq)
1753 {
1754         if (!cfqq)
1755                 cfqq = cfq_get_next_queue(cfqd);
1756
1757         __cfq_set_active_queue(cfqd, cfqq);
1758         return cfqq;
1759 }
1760
1761 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1762                                           struct request *rq)
1763 {
1764         if (blk_rq_pos(rq) >= cfqd->last_position)
1765                 return blk_rq_pos(rq) - cfqd->last_position;
1766         else
1767                 return cfqd->last_position - blk_rq_pos(rq);
1768 }
1769
1770 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1771                                struct request *rq)
1772 {
1773         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1774 }
1775
1776 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1777                                     struct cfq_queue *cur_cfqq)
1778 {
1779         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1780         struct rb_node *parent, *node;
1781         struct cfq_queue *__cfqq;
1782         sector_t sector = cfqd->last_position;
1783
1784         if (RB_EMPTY_ROOT(root))
1785                 return NULL;
1786
1787         /*
1788          * First, if we find a request starting at the end of the last
1789          * request, choose it.
1790          */
1791         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1792         if (__cfqq)
1793                 return __cfqq;
1794
1795         /*
1796          * If the exact sector wasn't found, the parent of the NULL leaf
1797          * will contain the closest sector.
1798          */
1799         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1800         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1801                 return __cfqq;
1802
1803         if (blk_rq_pos(__cfqq->next_rq) < sector)
1804                 node = rb_next(&__cfqq->p_node);
1805         else
1806                 node = rb_prev(&__cfqq->p_node);
1807         if (!node)
1808                 return NULL;
1809
1810         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1811         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1812                 return __cfqq;
1813
1814         return NULL;
1815 }
1816
1817 /*
1818  * cfqd - obvious
1819  * cur_cfqq - passed in so that we don't decide that the current queue is
1820  *            closely cooperating with itself.
1821  *
1822  * So, basically we're assuming that that cur_cfqq has dispatched at least
1823  * one request, and that cfqd->last_position reflects a position on the disk
1824  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1825  * assumption.
1826  */
1827 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1828                                               struct cfq_queue *cur_cfqq)
1829 {
1830         struct cfq_queue *cfqq;
1831
1832         if (cfq_class_idle(cur_cfqq))
1833                 return NULL;
1834         if (!cfq_cfqq_sync(cur_cfqq))
1835                 return NULL;
1836         if (CFQQ_SEEKY(cur_cfqq))
1837                 return NULL;
1838
1839         /*
1840          * Don't search priority tree if it's the only queue in the group.
1841          */
1842         if (cur_cfqq->cfqg->nr_cfqq == 1)
1843                 return NULL;
1844
1845         /*
1846          * We should notice if some of the queues are cooperating, eg
1847          * working closely on the same area of the disk. In that case,
1848          * we can group them together and don't waste time idling.
1849          */
1850         cfqq = cfqq_close(cfqd, cur_cfqq);
1851         if (!cfqq)
1852                 return NULL;
1853
1854         /* If new queue belongs to different cfq_group, don't choose it */
1855         if (cur_cfqq->cfqg != cfqq->cfqg)
1856                 return NULL;
1857
1858         /*
1859          * It only makes sense to merge sync queues.
1860          */
1861         if (!cfq_cfqq_sync(cfqq))
1862                 return NULL;
1863         if (CFQQ_SEEKY(cfqq))
1864                 return NULL;
1865
1866         /*
1867          * Do not merge queues of different priority classes
1868          */
1869         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1870                 return NULL;
1871
1872         return cfqq;
1873 }
1874
1875 /*
1876  * Determine whether we should enforce idle window for this queue.
1877  */
1878
1879 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1880 {
1881         enum wl_prio_t prio = cfqq_prio(cfqq);
1882         struct cfq_rb_root *service_tree = cfqq->service_tree;
1883
1884         BUG_ON(!service_tree);
1885         BUG_ON(!service_tree->count);
1886
1887         if (!cfqd->cfq_slice_idle)
1888                 return false;
1889
1890         /* We never do for idle class queues. */
1891         if (prio == IDLE_WORKLOAD)
1892                 return false;
1893
1894         /* We do for queues that were marked with idle window flag. */
1895         if (cfq_cfqq_idle_window(cfqq) &&
1896            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1897                 return true;
1898
1899         /*
1900          * Otherwise, we do only if they are the last ones
1901          * in their service tree.
1902          */
1903         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1904                 return 1;
1905         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1906                         service_tree->count);
1907         return 0;
1908 }
1909
1910 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1911 {
1912         struct cfq_queue *cfqq = cfqd->active_queue;
1913         struct cfq_io_context *cic;
1914         unsigned long sl, group_idle = 0;
1915
1916         /*
1917          * SSD device without seek penalty, disable idling. But only do so
1918          * for devices that support queuing, otherwise we still have a problem
1919          * with sync vs async workloads.
1920          */
1921         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1922                 return;
1923
1924         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1925         WARN_ON(cfq_cfqq_slice_new(cfqq));
1926
1927         /*
1928          * idle is disabled, either manually or by past process history
1929          */
1930         if (!cfq_should_idle(cfqd, cfqq)) {
1931                 /* no queue idling. Check for group idling */
1932                 if (cfqd->cfq_group_idle)
1933                         group_idle = cfqd->cfq_group_idle;
1934                 else
1935                         return;
1936         }
1937
1938         /*
1939          * still active requests from this queue, don't idle
1940          */
1941         if (cfqq->dispatched)
1942                 return;
1943
1944         /*
1945          * task has exited, don't wait
1946          */
1947         cic = cfqd->active_cic;
1948         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1949                 return;
1950
1951         /*
1952          * If our average think time is larger than the remaining time
1953          * slice, then don't idle. This avoids overrunning the allotted
1954          * time slice.
1955          */
1956         if (sample_valid(cic->ttime_samples) &&
1957             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1958                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1959                                 cic->ttime_mean);
1960                 return;
1961         }
1962
1963         /* There are other queues in the group, don't do group idle */
1964         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1965                 return;
1966
1967         cfq_mark_cfqq_wait_request(cfqq);
1968
1969         if (group_idle)
1970                 sl = cfqd->cfq_group_idle;
1971         else
1972                 sl = cfqd->cfq_slice_idle;
1973
1974         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1975         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1976         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1977                         group_idle ? 1 : 0);
1978 }
1979
1980 /*
1981  * Move request from internal lists to the request queue dispatch list.
1982  */
1983 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1984 {
1985         struct cfq_data *cfqd = q->elevator->elevator_data;
1986         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1987
1988         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1989
1990         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1991         cfq_remove_request(rq);
1992         cfqq->dispatched++;
1993         (RQ_CFQG(rq))->dispatched++;
1994         elv_dispatch_sort(q, rq);
1995
1996         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1997         cfqq->nr_sectors += blk_rq_sectors(rq);
1998         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1999                                         rq_data_dir(rq), rq_is_sync(rq));
2000 }
2001
2002 /*
2003  * return expired entry, or NULL to just start from scratch in rbtree
2004  */
2005 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2006 {
2007         struct request *rq = NULL;
2008
2009         if (cfq_cfqq_fifo_expire(cfqq))
2010                 return NULL;
2011
2012         cfq_mark_cfqq_fifo_expire(cfqq);
2013
2014         if (list_empty(&cfqq->fifo))
2015                 return NULL;
2016
2017         rq = rq_entry_fifo(cfqq->fifo.next);
2018         if (time_before(jiffies, rq_fifo_time(rq)))
2019                 rq = NULL;
2020
2021         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2022         return rq;
2023 }
2024
2025 static inline int
2026 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2027 {
2028         const int base_rq = cfqd->cfq_slice_async_rq;
2029
2030         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2031
2032         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2033 }
2034
2035 /*
2036  * Must be called with the queue_lock held.
2037  */
2038 static int cfqq_process_refs(struct cfq_queue *cfqq)
2039 {
2040         int process_refs, io_refs;
2041
2042         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2043         process_refs = atomic_read(&cfqq->ref) - io_refs;
2044         BUG_ON(process_refs < 0);
2045         return process_refs;
2046 }
2047
2048 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2049 {
2050         int process_refs, new_process_refs;
2051         struct cfq_queue *__cfqq;
2052
2053         /*
2054          * If there are no process references on the new_cfqq, then it is
2055          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2056          * chain may have dropped their last reference (not just their
2057          * last process reference).
2058          */
2059         if (!cfqq_process_refs(new_cfqq))
2060                 return;
2061
2062         /* Avoid a circular list and skip interim queue merges */
2063         while ((__cfqq = new_cfqq->new_cfqq)) {
2064                 if (__cfqq == cfqq)
2065                         return;
2066                 new_cfqq = __cfqq;
2067         }
2068
2069         process_refs = cfqq_process_refs(cfqq);
2070         new_process_refs = cfqq_process_refs(new_cfqq);
2071         /*
2072          * If the process for the cfqq has gone away, there is no
2073          * sense in merging the queues.
2074          */
2075         if (process_refs == 0 || new_process_refs == 0)
2076                 return;
2077
2078         /*
2079          * Merge in the direction of the lesser amount of work.
2080          */
2081         if (new_process_refs >= process_refs) {
2082                 cfqq->new_cfqq = new_cfqq;
2083                 atomic_add(process_refs, &new_cfqq->ref);
2084         } else {
2085                 new_cfqq->new_cfqq = cfqq;
2086                 atomic_add(new_process_refs, &cfqq->ref);
2087         }
2088 }
2089
2090 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2091                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2092 {
2093         struct cfq_queue *queue;
2094         int i;
2095         bool key_valid = false;
2096         unsigned long lowest_key = 0;
2097         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2098
2099         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2100                 /* select the one with lowest rb_key */
2101                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2102                 if (queue &&
2103                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2104                         lowest_key = queue->rb_key;
2105                         cur_best = i;
2106                         key_valid = true;
2107                 }
2108         }
2109
2110         return cur_best;
2111 }
2112
2113 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2114 {
2115         unsigned slice;
2116         unsigned count;
2117         struct cfq_rb_root *st;
2118         unsigned group_slice;
2119
2120         if (!cfqg) {
2121                 cfqd->serving_prio = IDLE_WORKLOAD;
2122                 cfqd->workload_expires = jiffies + 1;
2123                 return;
2124         }
2125
2126         /* Choose next priority. RT > BE > IDLE */
2127         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2128                 cfqd->serving_prio = RT_WORKLOAD;
2129         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2130                 cfqd->serving_prio = BE_WORKLOAD;
2131         else {
2132                 cfqd->serving_prio = IDLE_WORKLOAD;
2133                 cfqd->workload_expires = jiffies + 1;
2134                 return;
2135         }
2136
2137         /*
2138          * For RT and BE, we have to choose also the type
2139          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2140          * expiration time
2141          */
2142         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2143         count = st->count;
2144
2145         /*
2146          * check workload expiration, and that we still have other queues ready
2147          */
2148         if (count && !time_after(jiffies, cfqd->workload_expires))
2149                 return;
2150
2151         /* otherwise select new workload type */
2152         cfqd->serving_type =
2153                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2154         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2155         count = st->count;
2156
2157         /*
2158          * the workload slice is computed as a fraction of target latency
2159          * proportional to the number of queues in that workload, over
2160          * all the queues in the same priority class
2161          */
2162         group_slice = cfq_group_slice(cfqd, cfqg);
2163
2164         slice = group_slice * count /
2165                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2166                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2167
2168         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2169                 unsigned int tmp;
2170
2171                 /*
2172                  * Async queues are currently system wide. Just taking
2173                  * proportion of queues with-in same group will lead to higher
2174                  * async ratio system wide as generally root group is going
2175                  * to have higher weight. A more accurate thing would be to
2176                  * calculate system wide asnc/sync ratio.
2177                  */
2178                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2179                 tmp = tmp/cfqd->busy_queues;
2180                 slice = min_t(unsigned, slice, tmp);
2181
2182                 /* async workload slice is scaled down according to
2183                  * the sync/async slice ratio. */
2184                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2185         } else
2186                 /* sync workload slice is at least 2 * cfq_slice_idle */
2187                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2188
2189         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2190         cfq_log(cfqd, "workload slice:%d", slice);
2191         cfqd->workload_expires = jiffies + slice;
2192 }
2193
2194 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2195 {
2196         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2197         struct cfq_group *cfqg;
2198
2199         if (RB_EMPTY_ROOT(&st->rb))
2200                 return NULL;
2201         cfqg = cfq_rb_first_group(st);
2202         st->active = &cfqg->rb_node;
2203         update_min_vdisktime(st);
2204         return cfqg;
2205 }
2206
2207 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2208 {
2209         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2210
2211         cfqd->serving_group = cfqg;
2212
2213         /* Restore the workload type data */
2214         if (cfqg->saved_workload_slice) {
2215                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2216                 cfqd->serving_type = cfqg->saved_workload;
2217                 cfqd->serving_prio = cfqg->saved_serving_prio;
2218         } else
2219                 cfqd->workload_expires = jiffies - 1;
2220
2221         choose_service_tree(cfqd, cfqg);
2222 }
2223
2224 /*
2225  * Select a queue for service. If we have a current active queue,
2226  * check whether to continue servicing it, or retrieve and set a new one.
2227  */
2228 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2229 {
2230         struct cfq_queue *cfqq, *new_cfqq = NULL;
2231
2232         cfqq = cfqd->active_queue;
2233         if (!cfqq)
2234                 goto new_queue;
2235
2236         if (!cfqd->rq_queued)
2237                 return NULL;
2238
2239         /*
2240          * We were waiting for group to get backlogged. Expire the queue
2241          */
2242         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2243                 goto expire;
2244
2245         /*
2246          * The active queue has run out of time, expire it and select new.
2247          */
2248         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2249                 /*
2250                  * If slice had not expired at the completion of last request
2251                  * we might not have turned on wait_busy flag. Don't expire
2252                  * the queue yet. Allow the group to get backlogged.
2253                  *
2254                  * The very fact that we have used the slice, that means we
2255                  * have been idling all along on this queue and it should be
2256                  * ok to wait for this request to complete.
2257                  */
2258                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2259                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2260                         cfqq = NULL;
2261                         goto keep_queue;
2262                 } else
2263                         goto check_group_idle;
2264         }
2265
2266         /*
2267          * The active queue has requests and isn't expired, allow it to
2268          * dispatch.
2269          */
2270         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2271                 goto keep_queue;
2272
2273         /*
2274          * If another queue has a request waiting within our mean seek
2275          * distance, let it run.  The expire code will check for close
2276          * cooperators and put the close queue at the front of the service
2277          * tree.  If possible, merge the expiring queue with the new cfqq.
2278          */
2279         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2280         if (new_cfqq) {
2281                 if (!cfqq->new_cfqq)
2282                         cfq_setup_merge(cfqq, new_cfqq);
2283                 goto expire;
2284         }
2285
2286         /*
2287          * No requests pending. If the active queue still has requests in
2288          * flight or is idling for a new request, allow either of these
2289          * conditions to happen (or time out) before selecting a new queue.
2290          */
2291         if (timer_pending(&cfqd->idle_slice_timer)) {
2292                 cfqq = NULL;
2293                 goto keep_queue;
2294         }
2295
2296         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2297                 cfqq = NULL;
2298                 goto keep_queue;
2299         }
2300
2301         /*
2302          * If group idle is enabled and there are requests dispatched from
2303          * this group, wait for requests to complete.
2304          */
2305 check_group_idle:
2306         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2307             && cfqq->cfqg->dispatched) {
2308                 cfqq = NULL;
2309                 goto keep_queue;
2310         }
2311
2312 expire:
2313         cfq_slice_expired(cfqd, 0);
2314 new_queue:
2315         /*
2316          * Current queue expired. Check if we have to switch to a new
2317          * service tree
2318          */
2319         if (!new_cfqq)
2320                 cfq_choose_cfqg(cfqd);
2321
2322         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2323 keep_queue:
2324         return cfqq;
2325 }
2326
2327 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2328 {
2329         int dispatched = 0;
2330
2331         while (cfqq->next_rq) {
2332                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2333                 dispatched++;
2334         }
2335
2336         BUG_ON(!list_empty(&cfqq->fifo));
2337
2338         /* By default cfqq is not expired if it is empty. Do it explicitly */
2339         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2340         return dispatched;
2341 }
2342
2343 /*
2344  * Drain our current requests. Used for barriers and when switching
2345  * io schedulers on-the-fly.
2346  */
2347 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2348 {
2349         struct cfq_queue *cfqq;
2350         int dispatched = 0;
2351
2352         /* Expire the timeslice of the current active queue first */
2353         cfq_slice_expired(cfqd, 0);
2354         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2355                 __cfq_set_active_queue(cfqd, cfqq);
2356                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2357         }
2358
2359         BUG_ON(cfqd->busy_queues);
2360
2361         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2362         return dispatched;
2363 }
2364
2365 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2366         struct cfq_queue *cfqq)
2367 {
2368         /* the queue hasn't finished any request, can't estimate */
2369         if (cfq_cfqq_slice_new(cfqq))
2370                 return 1;
2371         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2372                 cfqq->slice_end))
2373                 return 1;
2374
2375         return 0;
2376 }
2377
2378 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2379 {
2380         unsigned int max_dispatch;
2381
2382         /*
2383          * Drain async requests before we start sync IO
2384          */
2385         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2386                 return false;
2387
2388         /*
2389          * If this is an async queue and we have sync IO in flight, let it wait
2390          */
2391         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2392                 return false;
2393
2394         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2395         if (cfq_class_idle(cfqq))
2396                 max_dispatch = 1;
2397
2398         /*
2399          * Does this cfqq already have too much IO in flight?
2400          */
2401         if (cfqq->dispatched >= max_dispatch) {
2402                 /*
2403                  * idle queue must always only have a single IO in flight
2404                  */
2405                 if (cfq_class_idle(cfqq))
2406                         return false;
2407
2408                 /*
2409                  * We have other queues, don't allow more IO from this one
2410                  */
2411                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2412                         return false;
2413
2414                 /*
2415                  * Sole queue user, no limit
2416                  */
2417                 if (cfqd->busy_queues == 1)
2418                         max_dispatch = -1;
2419                 else
2420                         /*
2421                          * Normally we start throttling cfqq when cfq_quantum/2
2422                          * requests have been dispatched. But we can drive
2423                          * deeper queue depths at the beginning of slice
2424                          * subjected to upper limit of cfq_quantum.
2425                          * */
2426                         max_dispatch = cfqd->cfq_quantum;
2427         }
2428
2429         /*
2430          * Async queues must wait a bit before being allowed dispatch.
2431          * We also ramp up the dispatch depth gradually for async IO,
2432          * based on the last sync IO we serviced
2433          */
2434         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2435                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2436                 unsigned int depth;
2437
2438                 depth = last_sync / cfqd->cfq_slice[1];
2439                 if (!depth && !cfqq->dispatched)
2440                         depth = 1;
2441                 if (depth < max_dispatch)
2442                         max_dispatch = depth;
2443         }
2444
2445         /*
2446          * If we're below the current max, allow a dispatch
2447          */
2448         return cfqq->dispatched < max_dispatch;
2449 }
2450
2451 /*
2452  * Dispatch a request from cfqq, moving them to the request queue
2453  * dispatch list.
2454  */
2455 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2456 {
2457         struct request *rq;
2458
2459         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2460
2461         if (!cfq_may_dispatch(cfqd, cfqq))
2462                 return false;
2463
2464         /*
2465          * follow expired path, else get first next available
2466          */
2467         rq = cfq_check_fifo(cfqq);
2468         if (!rq)
2469                 rq = cfqq->next_rq;
2470
2471         /*
2472          * insert request into driver dispatch list
2473          */
2474         cfq_dispatch_insert(cfqd->queue, rq);
2475
2476         if (!cfqd->active_cic) {
2477                 struct cfq_io_context *cic = RQ_CIC(rq);
2478
2479                 atomic_long_inc(&cic->ioc->refcount);
2480                 cfqd->active_cic = cic;
2481         }
2482
2483         return true;
2484 }
2485
2486 /*
2487  * Find the cfqq that we need to service and move a request from that to the
2488  * dispatch list
2489  */
2490 static int cfq_dispatch_requests(struct request_queue *q, int force)
2491 {
2492         struct cfq_data *cfqd = q->elevator->elevator_data;
2493         struct cfq_queue *cfqq;
2494
2495         if (!cfqd->busy_queues)
2496                 return 0;
2497
2498         if (unlikely(force))
2499                 return cfq_forced_dispatch(cfqd);
2500
2501         cfqq = cfq_select_queue(cfqd);
2502         if (!cfqq)
2503                 return 0;
2504
2505         /*
2506          * Dispatch a request from this cfqq, if it is allowed
2507          */
2508         if (!cfq_dispatch_request(cfqd, cfqq))
2509                 return 0;
2510
2511         cfqq->slice_dispatch++;
2512         cfq_clear_cfqq_must_dispatch(cfqq);
2513
2514         /*
2515          * expire an async queue immediately if it has used up its slice. idle
2516          * queue always expire after 1 dispatch round.
2517          */
2518         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2519             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2520             cfq_class_idle(cfqq))) {
2521                 cfqq->slice_end = jiffies + 1;
2522                 cfq_slice_expired(cfqd, 0);
2523         }
2524
2525         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2526         return 1;
2527 }
2528
2529 /*
2530  * task holds one reference to the queue, dropped when task exits. each rq
2531  * in-flight on this queue also holds a reference, dropped when rq is freed.
2532  *
2533  * Each cfq queue took a reference on the parent group. Drop it now.
2534  * queue lock must be held here.
2535  */
2536 static void cfq_put_queue(struct cfq_queue *cfqq)
2537 {
2538         struct cfq_data *cfqd = cfqq->cfqd;
2539         struct cfq_group *cfqg, *orig_cfqg;
2540
2541         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2542
2543         if (!atomic_dec_and_test(&cfqq->ref))
2544                 return;
2545
2546         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2547         BUG_ON(rb_first(&cfqq->sort_list));
2548         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2549         cfqg = cfqq->cfqg;
2550         orig_cfqg = cfqq->orig_cfqg;
2551
2552         if (unlikely(cfqd->active_queue == cfqq)) {
2553                 __cfq_slice_expired(cfqd, cfqq, 0);
2554                 cfq_schedule_dispatch(cfqd);
2555         }
2556
2557         BUG_ON(cfq_cfqq_on_rr(cfqq));
2558         kmem_cache_free(cfq_pool, cfqq);
2559         cfq_put_cfqg(cfqg);
2560         if (orig_cfqg)
2561                 cfq_put_cfqg(orig_cfqg);
2562 }
2563
2564 /*
2565  * Must always be called with the rcu_read_lock() held
2566  */
2567 static void
2568 __call_for_each_cic(struct io_context *ioc,
2569                     void (*func)(struct io_context *, struct cfq_io_context *))
2570 {
2571         struct cfq_io_context *cic;
2572         struct hlist_node *n;
2573
2574         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2575                 func(ioc, cic);
2576 }
2577
2578 /*
2579  * Call func for each cic attached to this ioc.
2580  */
2581 static void
2582 call_for_each_cic(struct io_context *ioc,
2583                   void (*func)(struct io_context *, struct cfq_io_context *))
2584 {
2585         rcu_read_lock();
2586         __call_for_each_cic(ioc, func);
2587         rcu_read_unlock();
2588 }
2589
2590 static void cfq_cic_free_rcu(struct rcu_head *head)
2591 {
2592         struct cfq_io_context *cic;
2593
2594         cic = container_of(head, struct cfq_io_context, rcu_head);
2595
2596         kmem_cache_free(cfq_ioc_pool, cic);
2597         elv_ioc_count_dec(cfq_ioc_count);
2598
2599         if (ioc_gone) {
2600                 /*
2601                  * CFQ scheduler is exiting, grab exit lock and check
2602                  * the pending io context count. If it hits zero,
2603                  * complete ioc_gone and set it back to NULL
2604                  */
2605                 spin_lock(&ioc_gone_lock);
2606                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2607                         complete(ioc_gone);
2608                         ioc_gone = NULL;
2609                 }
2610                 spin_unlock(&ioc_gone_lock);
2611         }
2612 }
2613
2614 static void cfq_cic_free(struct cfq_io_context *cic)
2615 {
2616         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2617 }
2618
2619 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2620 {
2621         unsigned long flags;
2622         unsigned long dead_key = (unsigned long) cic->key;
2623
2624         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2625
2626         spin_lock_irqsave(&ioc->lock, flags);
2627         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2628         hlist_del_rcu(&cic->cic_list);
2629         spin_unlock_irqrestore(&ioc->lock, flags);
2630
2631         cfq_cic_free(cic);
2632 }
2633
2634 /*
2635  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2636  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2637  * and ->trim() which is called with the task lock held
2638  */
2639 static void cfq_free_io_context(struct io_context *ioc)
2640 {
2641         /*
2642          * ioc->refcount is zero here, or we are called from elv_unregister(),
2643          * so no more cic's are allowed to be linked into this ioc.  So it
2644          * should be ok to iterate over the known list, we will see all cic's
2645          * since no new ones are added.
2646          */
2647         __call_for_each_cic(ioc, cic_free_func);
2648 }
2649
2650 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2651 {
2652         struct cfq_queue *__cfqq, *next;
2653
2654         /*
2655          * If this queue was scheduled to merge with another queue, be
2656          * sure to drop the reference taken on that queue (and others in
2657          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2658          */
2659         __cfqq = cfqq->new_cfqq;
2660         while (__cfqq) {
2661                 if (__cfqq == cfqq) {
2662                         WARN(1, "cfqq->new_cfqq loop detected\n");
2663                         break;
2664                 }
2665                 next = __cfqq->new_cfqq;
2666                 cfq_put_queue(__cfqq);
2667                 __cfqq = next;
2668         }
2669 }
2670
2671 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2672 {
2673         if (unlikely(cfqq == cfqd->active_queue)) {
2674                 __cfq_slice_expired(cfqd, cfqq, 0);
2675                 cfq_schedule_dispatch(cfqd);
2676         }
2677
2678         cfq_put_cooperator(cfqq);
2679
2680         cfq_put_queue(cfqq);
2681 }
2682
2683 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2684                                          struct cfq_io_context *cic)
2685 {
2686         struct io_context *ioc = cic->ioc;
2687
2688         list_del_init(&cic->queue_list);
2689
2690         /*
2691          * Make sure dead mark is seen for dead queues
2692          */
2693         smp_wmb();
2694         cic->key = cfqd_dead_key(cfqd);
2695
2696         if (ioc->ioc_data == cic)
2697                 rcu_assign_pointer(ioc->ioc_data, NULL);
2698
2699         if (cic->cfqq[BLK_RW_ASYNC]) {
2700                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2701                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2702         }
2703
2704         if (cic->cfqq[BLK_RW_SYNC]) {
2705                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2706                 cic->cfqq[BLK_RW_SYNC] = NULL;
2707         }
2708 }
2709
2710 static void cfq_exit_single_io_context(struct io_context *ioc,
2711                                        struct cfq_io_context *cic)
2712 {
2713         struct cfq_data *cfqd = cic_to_cfqd(cic);
2714
2715         if (cfqd) {
2716                 struct request_queue *q = cfqd->queue;
2717                 unsigned long flags;
2718
2719                 spin_lock_irqsave(q->queue_lock, flags);
2720
2721                 /*
2722                  * Ensure we get a fresh copy of the ->key to prevent
2723                  * race between exiting task and queue
2724                  */
2725                 smp_read_barrier_depends();
2726                 if (cic->key == cfqd)
2727                         __cfq_exit_single_io_context(cfqd, cic);
2728
2729                 spin_unlock_irqrestore(q->queue_lock, flags);
2730         }
2731 }
2732
2733 /*
2734  * The process that ioc belongs to has exited, we need to clean up
2735  * and put the internal structures we have that belongs to that process.
2736  */
2737 static void cfq_exit_io_context(struct io_context *ioc)
2738 {
2739         call_for_each_cic(ioc, cfq_exit_single_io_context);
2740 }
2741
2742 static struct cfq_io_context *
2743 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2744 {
2745         struct cfq_io_context *cic;
2746
2747         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2748                                                         cfqd->queue->node);
2749         if (cic) {
2750                 cic->last_end_request = jiffies;
2751                 INIT_LIST_HEAD(&cic->queue_list);
2752                 INIT_HLIST_NODE(&cic->cic_list);
2753                 cic->dtor = cfq_free_io_context;
2754                 cic->exit = cfq_exit_io_context;
2755                 elv_ioc_count_inc(cfq_ioc_count);
2756         }
2757
2758         return cic;
2759 }
2760
2761 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2762 {
2763         struct task_struct *tsk = current;
2764         int ioprio_class;
2765
2766         if (!cfq_cfqq_prio_changed(cfqq))
2767                 return;
2768
2769         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2770         switch (ioprio_class) {
2771         default:
2772                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2773         case IOPRIO_CLASS_NONE:
2774                 /*
2775                  * no prio set, inherit CPU scheduling settings
2776                  */
2777                 cfqq->ioprio = task_nice_ioprio(tsk);
2778                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2779                 break;
2780         case IOPRIO_CLASS_RT:
2781                 cfqq->ioprio = task_ioprio(ioc);
2782                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2783                 break;
2784         case IOPRIO_CLASS_BE:
2785                 cfqq->ioprio = task_ioprio(ioc);
2786                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2787                 break;
2788         case IOPRIO_CLASS_IDLE:
2789                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2790                 cfqq->ioprio = 7;
2791                 cfq_clear_cfqq_idle_window(cfqq);
2792                 break;
2793         }
2794
2795         /*
2796          * keep track of original prio settings in case we have to temporarily
2797          * elevate the priority of this queue
2798          */
2799         cfqq->org_ioprio = cfqq->ioprio;
2800         cfqq->org_ioprio_class = cfqq->ioprio_class;
2801         cfq_clear_cfqq_prio_changed(cfqq);
2802 }
2803
2804 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2805 {
2806         struct cfq_data *cfqd = cic_to_cfqd(cic);
2807         struct cfq_queue *cfqq;
2808         unsigned long flags;
2809
2810         if (unlikely(!cfqd))
2811                 return;
2812
2813         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2814
2815         cfqq = cic->cfqq[BLK_RW_ASYNC];
2816         if (cfqq) {
2817                 struct cfq_queue *new_cfqq;
2818                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2819                                                 GFP_ATOMIC);
2820                 if (new_cfqq) {
2821                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2822                         cfq_put_queue(cfqq);
2823                 }
2824         }
2825
2826         cfqq = cic->cfqq[BLK_RW_SYNC];
2827         if (cfqq)
2828                 cfq_mark_cfqq_prio_changed(cfqq);
2829
2830         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2831 }
2832
2833 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2834 {
2835         call_for_each_cic(ioc, changed_ioprio);
2836         ioc->ioprio_changed = 0;
2837 }
2838
2839 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2840                           pid_t pid, bool is_sync)
2841 {
2842         RB_CLEAR_NODE(&cfqq->rb_node);
2843         RB_CLEAR_NODE(&cfqq->p_node);
2844         INIT_LIST_HEAD(&cfqq->fifo);
2845
2846         atomic_set(&cfqq->ref, 0);
2847         cfqq->cfqd = cfqd;
2848
2849         cfq_mark_cfqq_prio_changed(cfqq);
2850
2851         if (is_sync) {
2852                 if (!cfq_class_idle(cfqq))
2853                         cfq_mark_cfqq_idle_window(cfqq);
2854                 cfq_mark_cfqq_sync(cfqq);
2855         }
2856         cfqq->pid = pid;
2857 }
2858
2859 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2860 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2861 {
2862         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2863         struct cfq_data *cfqd = cic_to_cfqd(cic);
2864         unsigned long flags;
2865         struct request_queue *q;
2866
2867         if (unlikely(!cfqd))
2868                 return;
2869
2870         q = cfqd->queue;
2871
2872         spin_lock_irqsave(q->queue_lock, flags);
2873
2874         if (sync_cfqq) {
2875                 /*
2876                  * Drop reference to sync queue. A new sync queue will be
2877                  * assigned in new group upon arrival of a fresh request.
2878                  */
2879                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2880                 cic_set_cfqq(cic, NULL, 1);
2881                 cfq_put_queue(sync_cfqq);
2882         }
2883
2884         spin_unlock_irqrestore(q->queue_lock, flags);
2885 }
2886
2887 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2888 {
2889         call_for_each_cic(ioc, changed_cgroup);
2890         ioc->cgroup_changed = 0;
2891 }
2892 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2893
2894 static struct cfq_queue *
2895 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2896                      struct io_context *ioc, gfp_t gfp_mask)
2897 {
2898         struct cfq_queue *cfqq, *new_cfqq = NULL;
2899         struct cfq_io_context *cic;
2900         struct cfq_group *cfqg;
2901
2902 retry:
2903         cfqg = cfq_get_cfqg(cfqd, 1);
2904         cic = cfq_cic_lookup(cfqd, ioc);
2905         /* cic always exists here */
2906         cfqq = cic_to_cfqq(cic, is_sync);
2907
2908         /*
2909          * Always try a new alloc if we fell back to the OOM cfqq
2910          * originally, since it should just be a temporary situation.
2911          */
2912         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2913                 cfqq = NULL;
2914                 if (new_cfqq) {
2915                         cfqq = new_cfqq;
2916                         new_cfqq = NULL;
2917                 } else if (gfp_mask & __GFP_WAIT) {
2918                         spin_unlock_irq(cfqd->queue->queue_lock);
2919                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2920                                         gfp_mask | __GFP_ZERO,
2921                                         cfqd->queue->node);
2922                         spin_lock_irq(cfqd->queue->queue_lock);
2923                         if (new_cfqq)
2924                                 goto retry;
2925                 } else {
2926                         cfqq = kmem_cache_alloc_node(cfq_pool,
2927                                         gfp_mask | __GFP_ZERO,
2928                                         cfqd->queue->node);
2929                 }
2930
2931                 if (cfqq) {
2932                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2933                         cfq_init_prio_data(cfqq, ioc);
2934                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2935                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2936                 } else
2937                         cfqq = &cfqd->oom_cfqq;
2938         }
2939
2940         if (new_cfqq)
2941                 kmem_cache_free(cfq_pool, new_cfqq);
2942
2943         return cfqq;
2944 }
2945
2946 static struct cfq_queue **
2947 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2948 {
2949         switch (ioprio_class) {
2950         case IOPRIO_CLASS_RT:
2951                 return &cfqd->async_cfqq[0][ioprio];
2952         case IOPRIO_CLASS_BE:
2953                 return &cfqd->async_cfqq[1][ioprio];
2954         case IOPRIO_CLASS_IDLE:
2955                 return &cfqd->async_idle_cfqq;
2956         default:
2957                 BUG();
2958         }
2959 }
2960
2961 static struct cfq_queue *
2962 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2963               gfp_t gfp_mask)
2964 {
2965         const int ioprio = task_ioprio(ioc);
2966         const int ioprio_class = task_ioprio_class(ioc);
2967         struct cfq_queue **async_cfqq = NULL;
2968         struct cfq_queue *cfqq = NULL;
2969
2970         if (!is_sync) {
2971                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2972                 cfqq = *async_cfqq;
2973         }
2974
2975         if (!cfqq)
2976                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2977
2978         /*
2979          * pin the queue now that it's allocated, scheduler exit will prune it
2980          */
2981         if (!is_sync && !(*async_cfqq)) {
2982                 atomic_inc(&cfqq->ref);
2983                 *async_cfqq = cfqq;
2984         }
2985
2986         atomic_inc(&cfqq->ref);
2987         return cfqq;
2988 }
2989
2990 /*
2991  * We drop cfq io contexts lazily, so we may find a dead one.
2992  */
2993 static void
2994 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2995                   struct cfq_io_context *cic)
2996 {
2997         unsigned long flags;
2998
2999         WARN_ON(!list_empty(&cic->queue_list));
3000         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3001
3002         spin_lock_irqsave(&ioc->lock, flags);
3003
3004         BUG_ON(ioc->ioc_data == cic);
3005
3006         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3007         hlist_del_rcu(&cic->cic_list);
3008         spin_unlock_irqrestore(&ioc->lock, flags);
3009
3010         cfq_cic_free(cic);
3011 }
3012
3013 static struct cfq_io_context *
3014 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3015 {
3016         struct cfq_io_context *cic;
3017         unsigned long flags;
3018
3019         if (unlikely(!ioc))
3020                 return NULL;
3021
3022         rcu_read_lock();
3023
3024         /*
3025          * we maintain a last-hit cache, to avoid browsing over the tree
3026          */
3027         cic = rcu_dereference(ioc->ioc_data);
3028         if (cic && cic->key == cfqd) {
3029                 rcu_read_unlock();
3030                 return cic;
3031         }
3032
3033         do {
3034                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3035                 rcu_read_unlock();
3036                 if (!cic)
3037                         break;
3038                 if (unlikely(cic->key != cfqd)) {
3039                         cfq_drop_dead_cic(cfqd, ioc, cic);
3040                         rcu_read_lock();
3041                         continue;
3042                 }
3043
3044                 spin_lock_irqsave(&ioc->lock, flags);
3045                 rcu_assign_pointer(ioc->ioc_data, cic);
3046                 spin_unlock_irqrestore(&ioc->lock, flags);
3047                 break;
3048         } while (1);
3049
3050         return cic;
3051 }
3052
3053 /*
3054  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3055  * the process specific cfq io context when entered from the block layer.
3056  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3057  */
3058 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3059                         struct cfq_io_context *cic, gfp_t gfp_mask)
3060 {
3061         unsigned long flags;
3062         int ret;
3063
3064         ret = radix_tree_preload(gfp_mask);
3065         if (!ret) {
3066                 cic->ioc = ioc;
3067                 cic->key = cfqd;
3068
3069                 spin_lock_irqsave(&ioc->lock, flags);
3070                 ret = radix_tree_insert(&ioc->radix_root,
3071                                                 cfqd->cic_index, cic);
3072                 if (!ret)
3073                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3074                 spin_unlock_irqrestore(&ioc->lock, flags);
3075
3076                 radix_tree_preload_end();
3077
3078                 if (!ret) {
3079                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3080                         list_add(&cic->queue_list, &cfqd->cic_list);
3081                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3082                 }
3083         }
3084
3085         if (ret)
3086                 printk(KERN_ERR "cfq: cic link failed!\n");
3087
3088         return ret;
3089 }
3090
3091 /*
3092  * Setup general io context and cfq io context. There can be several cfq
3093  * io contexts per general io context, if this process is doing io to more
3094  * than one device managed by cfq.
3095  */
3096 static struct cfq_io_context *
3097 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3098 {
3099         struct io_context *ioc = NULL;
3100         struct cfq_io_context *cic;
3101
3102         might_sleep_if(gfp_mask & __GFP_WAIT);
3103
3104         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3105         if (!ioc)
3106                 return NULL;
3107
3108         cic = cfq_cic_lookup(cfqd, ioc);
3109         if (cic)
3110                 goto out;
3111
3112         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3113         if (cic == NULL)
3114                 goto err;
3115
3116         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3117                 goto err_free;
3118
3119 out:
3120         smp_read_barrier_depends();
3121         if (unlikely(ioc->ioprio_changed))
3122                 cfq_ioc_set_ioprio(ioc);
3123
3124 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3125         if (unlikely(ioc->cgroup_changed))
3126                 cfq_ioc_set_cgroup(ioc);
3127 #endif
3128         return cic;
3129 err_free:
3130         cfq_cic_free(cic);
3131 err:
3132         put_io_context(ioc);
3133         return NULL;
3134 }
3135
3136 static void
3137 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3138 {
3139         unsigned long elapsed = jiffies - cic->last_end_request;
3140         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3141
3142         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3143         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3144         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3145 }
3146
3147 static void
3148 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3149                        struct request *rq)
3150 {
3151         sector_t sdist = 0;
3152         sector_t n_sec = blk_rq_sectors(rq);
3153         if (cfqq->last_request_pos) {
3154                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3155                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3156                 else
3157                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3158         }
3159
3160         cfqq->seek_history <<= 1;
3161         if (blk_queue_nonrot(cfqd->queue))
3162                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3163         else
3164                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3165 }
3166
3167 /*
3168  * Disable idle window if the process thinks too long or seeks so much that
3169  * it doesn't matter
3170  */
3171 static void
3172 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3173                        struct cfq_io_context *cic)
3174 {
3175         int old_idle, enable_idle;
3176
3177         /*
3178          * Don't idle for async or idle io prio class
3179          */
3180         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3181                 return;
3182
3183         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3184
3185         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3186                 cfq_mark_cfqq_deep(cfqq);
3187
3188         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3189                 enable_idle = 0;
3190         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3191             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3192                 enable_idle = 0;
3193         else if (sample_valid(cic->ttime_samples)) {
3194                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3195                         enable_idle = 0;
3196                 else
3197                         enable_idle = 1;
3198         }
3199
3200         if (old_idle != enable_idle) {
3201                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3202                 if (enable_idle)
3203                         cfq_mark_cfqq_idle_window(cfqq);
3204                 else
3205                         cfq_clear_cfqq_idle_window(cfqq);
3206         }
3207 }
3208
3209 /*
3210  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3211  * no or if we aren't sure, a 1 will cause a preempt.
3212  */
3213 static bool
3214 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3215                    struct request *rq)
3216 {
3217         struct cfq_queue *cfqq;
3218
3219         cfqq = cfqd->active_queue;
3220         if (!cfqq)
3221                 return false;
3222
3223         if (cfq_class_idle(new_cfqq))
3224                 return false;
3225
3226         if (cfq_class_idle(cfqq))
3227                 return true;
3228
3229         /*
3230          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3231          */
3232         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3233                 return false;
3234
3235         /*
3236          * if the new request is sync, but the currently running queue is
3237          * not, let the sync request have priority.
3238          */
3239         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3240                 return true;
3241
3242         if (new_cfqq->cfqg != cfqq->cfqg)
3243                 return false;
3244
3245         if (cfq_slice_used(cfqq))
3246                 return true;
3247
3248         /* Allow preemption only if we are idling on sync-noidle tree */
3249         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3250             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3251             new_cfqq->service_tree->count == 2 &&
3252             RB_EMPTY_ROOT(&cfqq->sort_list))
3253                 return true;
3254
3255         /*
3256          * So both queues are sync. Let the new request get disk time if
3257          * it's a metadata request and the current queue is doing regular IO.
3258          */
3259         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3260                 return true;
3261
3262         /*
3263          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3264          */
3265         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3266                 return true;
3267
3268         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3269                 return false;
3270
3271         /*
3272          * if this request is as-good as one we would expect from the
3273          * current cfqq, let it preempt
3274          */
3275         if (cfq_rq_close(cfqd, cfqq, rq))
3276                 return true;
3277
3278         return false;
3279 }
3280
3281 /*
3282  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3283  * let it have half of its nominal slice.
3284  */
3285 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3286 {
3287         cfq_log_cfqq(cfqd, cfqq, "preempt");
3288         cfq_slice_expired(cfqd, 1);
3289
3290         /*
3291          * Put the new queue at the front of the of the current list,
3292          * so we know that it will be selected next.
3293          */
3294         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3295
3296         cfq_service_tree_add(cfqd, cfqq, 1);
3297
3298         cfqq->slice_end = 0;
3299         cfq_mark_cfqq_slice_new(cfqq);
3300 }
3301
3302 /*
3303  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3304  * something we should do about it
3305  */
3306 static void
3307 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3308                 struct request *rq)
3309 {
3310         struct cfq_io_context *cic = RQ_CIC(rq);
3311
3312         cfqd->rq_queued++;
3313         if (rq->cmd_flags & REQ_META)
3314                 cfqq->meta_pending++;
3315
3316         cfq_update_io_thinktime(cfqd, cic);
3317         cfq_update_io_seektime(cfqd, cfqq, rq);
3318         cfq_update_idle_window(cfqd, cfqq, cic);
3319
3320         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3321
3322         if (cfqq == cfqd->active_queue) {
3323                 /*
3324                  * Remember that we saw a request from this process, but
3325                  * don't start queuing just yet. Otherwise we risk seeing lots
3326                  * of tiny requests, because we disrupt the normal plugging
3327                  * and merging. If the request is already larger than a single
3328                  * page, let it rip immediately. For that case we assume that
3329                  * merging is already done. Ditto for a busy system that
3330                  * has other work pending, don't risk delaying until the
3331                  * idle timer unplug to continue working.
3332                  */
3333                 if (cfq_cfqq_wait_request(cfqq)) {
3334                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3335                             cfqd->busy_queues > 1) {
3336                                 cfq_del_timer(cfqd, cfqq);
3337                                 cfq_clear_cfqq_wait_request(cfqq);
3338                                 __blk_run_queue(cfqd->queue);
3339                         } else {
3340                                 cfq_blkiocg_update_idle_time_stats(
3341                                                 &cfqq->cfqg->blkg);
3342                                 cfq_mark_cfqq_must_dispatch(cfqq);
3343                         }
3344                 }
3345         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3346                 /*
3347                  * not the active queue - expire current slice if it is
3348                  * idle and has expired it's mean thinktime or this new queue
3349                  * has some old slice time left and is of higher priority or
3350                  * this new queue is RT and the current one is BE
3351                  */
3352                 cfq_preempt_queue(cfqd, cfqq);
3353                 __blk_run_queue(cfqd->queue);
3354         }
3355 }
3356
3357 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3358 {
3359         struct cfq_data *cfqd = q->elevator->elevator_data;
3360         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3361
3362         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3363         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3364
3365         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3366         list_add_tail(&rq->queuelist, &cfqq->fifo);
3367         cfq_add_rq_rb(rq);
3368         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3369                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3370                         rq_is_sync(rq));
3371         cfq_rq_enqueued(cfqd, cfqq, rq);
3372 }
3373
3374 /*
3375  * Update hw_tag based on peak queue depth over 50 samples under
3376  * sufficient load.
3377  */
3378 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3379 {
3380         struct cfq_queue *cfqq = cfqd->active_queue;
3381
3382         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3383                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3384
3385         if (cfqd->hw_tag == 1)
3386                 return;
3387
3388         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3389             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3390                 return;
3391
3392         /*
3393          * If active queue hasn't enough requests and can idle, cfq might not
3394          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3395          * case
3396          */
3397         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3398             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3399             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3400                 return;
3401
3402         if (cfqd->hw_tag_samples++ < 50)
3403                 return;
3404
3405         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3406                 cfqd->hw_tag = 1;
3407         else
3408                 cfqd->hw_tag = 0;
3409 }
3410
3411 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3412 {
3413         struct cfq_io_context *cic = cfqd->active_cic;
3414
3415         /* If there are other queues in the group, don't wait */
3416         if (cfqq->cfqg->nr_cfqq > 1)
3417                 return false;
3418
3419         if (cfq_slice_used(cfqq))
3420                 return true;
3421
3422         /* if slice left is less than think time, wait busy */
3423         if (cic && sample_valid(cic->ttime_samples)
3424             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3425                 return true;
3426
3427         /*
3428          * If think times is less than a jiffy than ttime_mean=0 and above
3429          * will not be true. It might happen that slice has not expired yet
3430          * but will expire soon (4-5 ns) during select_queue(). To cover the
3431          * case where think time is less than a jiffy, mark the queue wait
3432          * busy if only 1 jiffy is left in the slice.
3433          */
3434         if (cfqq->slice_end - jiffies == 1)
3435                 return true;
3436
3437         return false;
3438 }
3439
3440 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3441 {
3442         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3443         struct cfq_data *cfqd = cfqq->cfqd;
3444         const int sync = rq_is_sync(rq);
3445         unsigned long now;
3446
3447         now = jiffies;
3448         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3449                      !!(rq->cmd_flags & REQ_NOIDLE));
3450
3451         cfq_update_hw_tag(cfqd);
3452
3453         WARN_ON(!cfqd->rq_in_driver);
3454         WARN_ON(!cfqq->dispatched);
3455         cfqd->rq_in_driver--;
3456         cfqq->dispatched--;
3457         (RQ_CFQG(rq))->dispatched--;
3458         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3459                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3460                         rq_data_dir(rq), rq_is_sync(rq));
3461
3462         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3463
3464         if (sync) {
3465                 RQ_CIC(rq)->last_end_request = now;
3466                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3467                         cfqd->last_delayed_sync = now;
3468         }
3469
3470         /*
3471          * If this is the active queue, check if it needs to be expired,
3472          * or if we want to idle in case it has no pending requests.
3473          */
3474         if (cfqd->active_queue == cfqq) {
3475                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3476
3477                 if (cfq_cfqq_slice_new(cfqq)) {
3478                         cfq_set_prio_slice(cfqd, cfqq);
3479                         cfq_clear_cfqq_slice_new(cfqq);
3480                 }
3481
3482                 /*
3483                  * Should we wait for next request to come in before we expire
3484                  * the queue.
3485                  */
3486                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3487                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3488                         if (!cfqd->cfq_slice_idle)
3489                                 extend_sl = cfqd->cfq_group_idle;
3490                         cfqq->slice_end = jiffies + extend_sl;
3491                         cfq_mark_cfqq_wait_busy(cfqq);
3492                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3493                 }
3494
3495                 /*
3496                  * Idling is not enabled on:
3497                  * - expired queues
3498                  * - idle-priority queues
3499                  * - async queues
3500                  * - queues with still some requests queued
3501                  * - when there is a close cooperator
3502                  */
3503                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3504                         cfq_slice_expired(cfqd, 1);
3505                 else if (sync && cfqq_empty &&
3506                          !cfq_close_cooperator(cfqd, cfqq)) {
3507                         cfq_arm_slice_timer(cfqd);
3508                 }
3509         }
3510
3511         if (!cfqd->rq_in_driver)
3512                 cfq_schedule_dispatch(cfqd);
3513 }
3514
3515 /*
3516  * we temporarily boost lower priority queues if they are holding fs exclusive
3517  * resources. they are boosted to normal prio (CLASS_BE/4)
3518  */
3519 static void cfq_prio_boost(struct cfq_queue *cfqq)
3520 {
3521         if (has_fs_excl()) {
3522                 /*
3523                  * boost idle prio on transactions that would lock out other
3524                  * users of the filesystem
3525                  */
3526                 if (cfq_class_idle(cfqq))
3527                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3528                 if (cfqq->ioprio > IOPRIO_NORM)
3529                         cfqq->ioprio = IOPRIO_NORM;
3530         } else {
3531                 /*
3532                  * unboost the queue (if needed)
3533                  */
3534                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3535                 cfqq->ioprio = cfqq->org_ioprio;
3536         }
3537 }
3538
3539 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3540 {
3541         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3542                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3543                 return ELV_MQUEUE_MUST;
3544         }
3545
3546         return ELV_MQUEUE_MAY;
3547 }
3548
3549 static int cfq_may_queue(struct request_queue *q, int rw)
3550 {
3551         struct cfq_data *cfqd = q->elevator->elevator_data;
3552         struct task_struct *tsk = current;
3553         struct cfq_io_context *cic;
3554         struct cfq_queue *cfqq;
3555
3556         /*
3557          * don't force setup of a queue from here, as a call to may_queue
3558          * does not necessarily imply that a request actually will be queued.
3559          * so just lookup a possibly existing queue, or return 'may queue'
3560          * if that fails
3561          */
3562         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3563         if (!cic)
3564                 return ELV_MQUEUE_MAY;
3565
3566         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3567         if (cfqq) {
3568                 cfq_init_prio_data(cfqq, cic->ioc);
3569                 cfq_prio_boost(cfqq);
3570
3571                 return __cfq_may_queue(cfqq);
3572         }
3573
3574         return ELV_MQUEUE_MAY;
3575 }
3576
3577 /*
3578  * queue lock held here
3579  */
3580 static void cfq_put_request(struct request *rq)
3581 {
3582         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3583
3584         if (cfqq) {
3585                 const int rw = rq_data_dir(rq);
3586
3587                 BUG_ON(!cfqq->allocated[rw]);
3588                 cfqq->allocated[rw]--;
3589
3590                 put_io_context(RQ_CIC(rq)->ioc);
3591
3592                 rq->elevator_private = NULL;
3593                 rq->elevator_private2 = NULL;
3594
3595                 /* Put down rq reference on cfqg */
3596                 cfq_put_cfqg(RQ_CFQG(rq));
3597                 rq->elevator_private3 = NULL;
3598
3599                 cfq_put_queue(cfqq);
3600         }
3601 }
3602
3603 static struct cfq_queue *
3604 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3605                 struct cfq_queue *cfqq)
3606 {
3607         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3608         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3609         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3610         cfq_put_queue(cfqq);
3611         return cic_to_cfqq(cic, 1);
3612 }
3613
3614 /*
3615  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3616  * was the last process referring to said cfqq.
3617  */
3618 static struct cfq_queue *
3619 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3620 {
3621         if (cfqq_process_refs(cfqq) == 1) {
3622                 cfqq->pid = current->pid;
3623                 cfq_clear_cfqq_coop(cfqq);
3624                 cfq_clear_cfqq_split_coop(cfqq);
3625                 return cfqq;
3626         }
3627
3628         cic_set_cfqq(cic, NULL, 1);
3629
3630         cfq_put_cooperator(cfqq);
3631
3632         cfq_put_queue(cfqq);
3633         return NULL;
3634 }
3635 /*
3636  * Allocate cfq data structures associated with this request.
3637  */
3638 static int
3639 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3640 {
3641         struct cfq_data *cfqd = q->elevator->elevator_data;
3642         struct cfq_io_context *cic;
3643         const int rw = rq_data_dir(rq);
3644         const bool is_sync = rq_is_sync(rq);
3645         struct cfq_queue *cfqq;
3646         unsigned long flags;
3647
3648         might_sleep_if(gfp_mask & __GFP_WAIT);
3649
3650         cic = cfq_get_io_context(cfqd, gfp_mask);
3651
3652         spin_lock_irqsave(q->queue_lock, flags);
3653
3654         if (!cic)
3655                 goto queue_fail;
3656
3657 new_queue:
3658         cfqq = cic_to_cfqq(cic, is_sync);
3659         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3660                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3661                 cic_set_cfqq(cic, cfqq, is_sync);
3662         } else {
3663                 /*
3664                  * If the queue was seeky for too long, break it apart.
3665                  */
3666                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3667                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3668                         cfqq = split_cfqq(cic, cfqq);
3669                         if (!cfqq)
3670                                 goto new_queue;
3671                 }
3672
3673                 /*
3674                  * Check to see if this queue is scheduled to merge with
3675                  * another, closely cooperating queue.  The merging of
3676                  * queues happens here as it must be done in process context.
3677                  * The reference on new_cfqq was taken in merge_cfqqs.
3678                  */
3679                 if (cfqq->new_cfqq)
3680                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3681         }
3682
3683         cfqq->allocated[rw]++;
3684         atomic_inc(&cfqq->ref);
3685
3686         spin_unlock_irqrestore(q->queue_lock, flags);
3687
3688         rq->elevator_private = cic;
3689         rq->elevator_private2 = cfqq;
3690         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3691         return 0;
3692
3693 queue_fail:
3694         if (cic)
3695                 put_io_context(cic->ioc);
3696
3697         cfq_schedule_dispatch(cfqd);
3698         spin_unlock_irqrestore(q->queue_lock, flags);
3699         cfq_log(cfqd, "set_request fail");
3700         return 1;
3701 }
3702
3703 static void cfq_kick_queue(struct work_struct *work)
3704 {
3705         struct cfq_data *cfqd =
3706                 container_of(work, struct cfq_data, unplug_work);
3707         struct request_queue *q = cfqd->queue;
3708
3709         spin_lock_irq(q->queue_lock);
3710         __blk_run_queue(cfqd->queue);
3711         spin_unlock_irq(q->queue_lock);
3712 }
3713
3714 /*
3715  * Timer running if the active_queue is currently idling inside its time slice
3716  */
3717 static void cfq_idle_slice_timer(unsigned long data)
3718 {
3719         struct cfq_data *cfqd = (struct cfq_data *) data;
3720         struct cfq_queue *cfqq;
3721         unsigned long flags;
3722         int timed_out = 1;
3723
3724         cfq_log(cfqd, "idle timer fired");
3725
3726         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3727
3728         cfqq = cfqd->active_queue;
3729         if (cfqq) {
3730                 timed_out = 0;
3731
3732                 /*
3733                  * We saw a request before the queue expired, let it through
3734                  */
3735                 if (cfq_cfqq_must_dispatch(cfqq))
3736                         goto out_kick;
3737
3738                 /*
3739                  * expired
3740                  */
3741                 if (cfq_slice_used(cfqq))
3742                         goto expire;
3743
3744                 /*
3745                  * only expire and reinvoke request handler, if there are
3746                  * other queues with pending requests
3747                  */
3748                 if (!cfqd->busy_queues)
3749                         goto out_cont;
3750
3751                 /*
3752                  * not expired and it has a request pending, let it dispatch
3753                  */
3754                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3755                         goto out_kick;
3756
3757                 /*
3758                  * Queue depth flag is reset only when the idle didn't succeed
3759                  */
3760                 cfq_clear_cfqq_deep(cfqq);
3761         }
3762 expire:
3763         cfq_slice_expired(cfqd, timed_out);
3764 out_kick:
3765         cfq_schedule_dispatch(cfqd);
3766 out_cont:
3767         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3768 }
3769
3770 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3771 {
3772         del_timer_sync(&cfqd->idle_slice_timer);
3773         cancel_work_sync(&cfqd->unplug_work);
3774 }
3775
3776 static void cfq_put_async_queues(struct cfq_data *cfqd)
3777 {
3778         int i;
3779
3780         for (i = 0; i < IOPRIO_BE_NR; i++) {
3781                 if (cfqd->async_cfqq[0][i])
3782                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3783                 if (cfqd->async_cfqq[1][i])
3784                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3785         }
3786
3787         if (cfqd->async_idle_cfqq)
3788                 cfq_put_queue(cfqd->async_idle_cfqq);
3789 }
3790
3791 static void cfq_cfqd_free(struct rcu_head *head)
3792 {
3793         kfree(container_of(head, struct cfq_data, rcu));
3794 }
3795
3796 static void cfq_exit_queue(struct elevator_queue *e)
3797 {
3798         struct cfq_data *cfqd = e->elevator_data;
3799         struct request_queue *q = cfqd->queue;
3800
3801         cfq_shutdown_timer_wq(cfqd);
3802
3803         spin_lock_irq(q->queue_lock);
3804
3805         if (cfqd->active_queue)
3806                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3807
3808         while (!list_empty(&cfqd->cic_list)) {
3809                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3810                                                         struct cfq_io_context,
3811                                                         queue_list);
3812
3813                 __cfq_exit_single_io_context(cfqd, cic);
3814         }
3815
3816         cfq_put_async_queues(cfqd);
3817         cfq_release_cfq_groups(cfqd);
3818         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3819
3820         spin_unlock_irq(q->queue_lock);
3821
3822         cfq_shutdown_timer_wq(cfqd);
3823
3824         spin_lock(&cic_index_lock);
3825         ida_remove(&cic_index_ida, cfqd->cic_index);
3826         spin_unlock(&cic_index_lock);
3827
3828         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3829         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3830 }
3831
3832 static int cfq_alloc_cic_index(void)
3833 {
3834         int index, error;
3835
3836         do {
3837                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3838                         return -ENOMEM;
3839
3840                 spin_lock(&cic_index_lock);
3841                 error = ida_get_new(&cic_index_ida, &index);
3842                 spin_unlock(&cic_index_lock);
3843                 if (error && error != -EAGAIN)
3844                         return error;
3845         } while (error);
3846
3847         return index;
3848 }
3849
3850 static void *cfq_init_queue(struct request_queue *q)
3851 {
3852         struct cfq_data *cfqd;
3853         int i, j;
3854         struct cfq_group *cfqg;
3855         struct cfq_rb_root *st;
3856
3857         i = cfq_alloc_cic_index();
3858         if (i < 0)
3859                 return NULL;
3860
3861         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3862         if (!cfqd)
3863                 return NULL;
3864
3865         cfqd->cic_index = i;
3866
3867         /* Init root service tree */
3868         cfqd->grp_service_tree = CFQ_RB_ROOT;
3869
3870         /* Init root group */
3871         cfqg = &cfqd->root_group;
3872         for_each_cfqg_st(cfqg, i, j, st)
3873                 *st = CFQ_RB_ROOT;
3874         RB_CLEAR_NODE(&cfqg->rb_node);
3875
3876         /* Give preference to root group over other groups */
3877         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3878
3879 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3880         /*
3881          * Take a reference to root group which we never drop. This is just
3882          * to make sure that cfq_put_cfqg() does not try to kfree root group
3883          */
3884         atomic_set(&cfqg->ref, 1);
3885         rcu_read_lock();
3886         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3887                                         (void *)cfqd, 0);
3888         rcu_read_unlock();
3889 #endif
3890         /*
3891          * Not strictly needed (since RB_ROOT just clears the node and we
3892          * zeroed cfqd on alloc), but better be safe in case someone decides
3893          * to add magic to the rb code
3894          */
3895         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3896                 cfqd->prio_trees[i] = RB_ROOT;
3897
3898         /*
3899          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3900          * Grab a permanent reference to it, so that the normal code flow
3901          * will not attempt to free it.
3902          */
3903         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3904         atomic_inc(&cfqd->oom_cfqq.ref);
3905         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3906
3907         INIT_LIST_HEAD(&cfqd->cic_list);
3908
3909         cfqd->queue = q;
3910
3911         init_timer(&cfqd->idle_slice_timer);
3912         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3913         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3914
3915         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3916
3917         cfqd->cfq_quantum = cfq_quantum;
3918         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3919         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3920         cfqd->cfq_back_max = cfq_back_max;
3921         cfqd->cfq_back_penalty = cfq_back_penalty;
3922         cfqd->cfq_slice[0] = cfq_slice_async;
3923         cfqd->cfq_slice[1] = cfq_slice_sync;
3924         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3925         cfqd->cfq_slice_idle = cfq_slice_idle;
3926         cfqd->cfq_group_idle = cfq_group_idle;
3927         cfqd->cfq_latency = 1;
3928         cfqd->cfq_group_isolation = 0;
3929         cfqd->hw_tag = -1;
3930         /*
3931          * we optimistically start assuming sync ops weren't delayed in last
3932          * second, in order to have larger depth for async operations.
3933          */
3934         cfqd->last_delayed_sync = jiffies - HZ;
3935         return cfqd;
3936 }
3937
3938 static void cfq_slab_kill(void)
3939 {
3940         /*
3941          * Caller already ensured that pending RCU callbacks are completed,
3942          * so we should have no busy allocations at this point.
3943          */
3944         if (cfq_pool)
3945                 kmem_cache_destroy(cfq_pool);
3946         if (cfq_ioc_pool)
3947                 kmem_cache_destroy(cfq_ioc_pool);
3948 }
3949
3950 static int __init cfq_slab_setup(void)
3951 {
3952         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3953         if (!cfq_pool)
3954                 goto fail;
3955
3956         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3957         if (!cfq_ioc_pool)
3958                 goto fail;
3959
3960         return 0;
3961 fail:
3962         cfq_slab_kill();
3963         return -ENOMEM;
3964 }
3965
3966 /*
3967  * sysfs parts below -->
3968  */
3969 static ssize_t
3970 cfq_var_show(unsigned int var, char *page)
3971 {
3972         return sprintf(page, "%d\n", var);
3973 }
3974
3975 static ssize_t
3976 cfq_var_store(unsigned int *var, const char *page, size_t count)
3977 {
3978         char *p = (char *) page;
3979
3980         *var = simple_strtoul(p, &p, 10);
3981         return count;
3982 }
3983
3984 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3985 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3986 {                                                                       \
3987         struct cfq_data *cfqd = e->elevator_data;                       \
3988         unsigned int __data = __VAR;                                    \
3989         if (__CONV)                                                     \
3990                 __data = jiffies_to_msecs(__data);                      \
3991         return cfq_var_show(__data, (page));                            \
3992 }
3993 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3994 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3995 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3996 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3997 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3998 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3999 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4000 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4001 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4002 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4003 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4004 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4005 #undef SHOW_FUNCTION
4006
4007 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4008 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4009 {                                                                       \
4010         struct cfq_data *cfqd = e->elevator_data;                       \
4011         unsigned int __data;                                            \
4012         int ret = cfq_var_store(&__data, (page), count);                \
4013         if (__data < (MIN))                                             \
4014                 __data = (MIN);                                         \
4015         else if (__data > (MAX))                                        \
4016                 __data = (MAX);                                         \
4017         if (__CONV)                                                     \
4018                 *(__PTR) = msecs_to_jiffies(__data);                    \
4019         else                                                            \
4020                 *(__PTR) = __data;                                      \
4021         return ret;                                                     \
4022 }
4023 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4024 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4025                 UINT_MAX, 1);
4026 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4027                 UINT_MAX, 1);
4028 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4029 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4030                 UINT_MAX, 0);
4031 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4032 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4033 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4034 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4035 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4036                 UINT_MAX, 0);
4037 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4038 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4039 #undef STORE_FUNCTION
4040
4041 #define CFQ_ATTR(name) \
4042         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4043
4044 static struct elv_fs_entry cfq_attrs[] = {
4045         CFQ_ATTR(quantum),
4046         CFQ_ATTR(fifo_expire_sync),
4047         CFQ_ATTR(fifo_expire_async),
4048         CFQ_ATTR(back_seek_max),
4049         CFQ_ATTR(back_seek_penalty),
4050         CFQ_ATTR(slice_sync),
4051         CFQ_ATTR(slice_async),
4052         CFQ_ATTR(slice_async_rq),
4053         CFQ_ATTR(slice_idle),
4054         CFQ_ATTR(group_idle),
4055         CFQ_ATTR(low_latency),
4056         CFQ_ATTR(group_isolation),
4057         __ATTR_NULL
4058 };
4059
4060 static struct elevator_type iosched_cfq = {
4061         .ops = {
4062                 .elevator_merge_fn =            cfq_merge,
4063                 .elevator_merged_fn =           cfq_merged_request,
4064                 .elevator_merge_req_fn =        cfq_merged_requests,
4065                 .elevator_allow_merge_fn =      cfq_allow_merge,
4066                 .elevator_bio_merged_fn =       cfq_bio_merged,
4067                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4068                 .elevator_add_req_fn =          cfq_insert_request,
4069                 .elevator_activate_req_fn =     cfq_activate_request,
4070                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4071                 .elevator_queue_empty_fn =      cfq_queue_empty,
4072                 .elevator_completed_req_fn =    cfq_completed_request,
4073                 .elevator_former_req_fn =       elv_rb_former_request,
4074                 .elevator_latter_req_fn =       elv_rb_latter_request,
4075                 .elevator_set_req_fn =          cfq_set_request,
4076                 .elevator_put_req_fn =          cfq_put_request,
4077                 .elevator_may_queue_fn =        cfq_may_queue,
4078                 .elevator_init_fn =             cfq_init_queue,
4079                 .elevator_exit_fn =             cfq_exit_queue,
4080                 .trim =                         cfq_free_io_context,
4081         },
4082         .elevator_attrs =       cfq_attrs,
4083         .elevator_name =        "cfq",
4084         .elevator_owner =       THIS_MODULE,
4085 };
4086
4087 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4088 static struct blkio_policy_type blkio_policy_cfq = {
4089         .ops = {
4090                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4091                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4092         },
4093         .plid = BLKIO_POLICY_PROP,
4094 };
4095 #else
4096 static struct blkio_policy_type blkio_policy_cfq;
4097 #endif
4098
4099 static int __init cfq_init(void)
4100 {
4101         /*
4102          * could be 0 on HZ < 1000 setups
4103          */
4104         if (!cfq_slice_async)
4105                 cfq_slice_async = 1;
4106         if (!cfq_slice_idle)
4107                 cfq_slice_idle = 1;
4108
4109 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4110         if (!cfq_group_idle)
4111                 cfq_group_idle = 1;
4112 #else
4113                 cfq_group_idle = 0;
4114 #endif
4115         if (cfq_slab_setup())
4116                 return -ENOMEM;
4117
4118         elv_register(&iosched_cfq);
4119         blkio_policy_register(&blkio_policy_cfq);
4120
4121         return 0;
4122 }
4123
4124 static void __exit cfq_exit(void)
4125 {
4126         DECLARE_COMPLETION_ONSTACK(all_gone);
4127         blkio_policy_unregister(&blkio_policy_cfq);
4128         elv_unregister(&iosched_cfq);
4129         ioc_gone = &all_gone;
4130         /* ioc_gone's update must be visible before reading ioc_count */
4131         smp_wmb();
4132
4133         /*
4134          * this also protects us from entering cfq_slab_kill() with
4135          * pending RCU callbacks
4136          */
4137         if (elv_ioc_count_read(cfq_ioc_count))
4138                 wait_for_completion(&all_gone);
4139         ida_destroy(&cic_index_ida);
4140         cfq_slab_kill();
4141 }
4142
4143 module_init(cfq_init);
4144 module_exit(cfq_exit);
4145
4146 MODULE_AUTHOR("Jens Axboe");
4147 MODULE_LICENSE("GPL");
4148 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");