cfq-iosched: Fix a possible race with cfq cgroup removal code
[linux-3.10.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[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
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 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181         unsigned int new_weight;
182         bool needs_update;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /*
188          * Per group busy queus average. Useful for workload slice calc. We
189          * create the array for each prio class but at run time it is used
190          * only for RT and BE class and slot for IDLE class remains unused.
191          * This is primarily done to avoid confusion and a gcc warning.
192          */
193         unsigned int busy_queues_avg[CFQ_PRIO_NR];
194         /*
195          * rr lists of queues with requests. We maintain service trees for
196          * RT and BE classes. These trees are subdivided in subclasses
197          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198          * class there is no subclassification and all the cfq queues go on
199          * a single tree service_tree_idle.
200          * Counts are embedded in the cfq_rb_root
201          */
202         struct cfq_rb_root service_trees[2][3];
203         struct cfq_rb_root service_tree_idle;
204
205         unsigned long saved_workload_slice;
206         enum wl_type_t saved_workload;
207         enum wl_prio_t saved_serving_prio;
208         struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210         struct hlist_node cfqd_node;
211         int ref;
212 #endif
213         /* number of requests that are on the dispatch list or inside driver */
214         int dispatched;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_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
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303
304         /* Number of groups which are on blkcg->blkg_list */
305         unsigned int nr_blkcg_linked_grps;
306 };
307
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
309
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
311                                             enum wl_prio_t prio,
312                                             enum wl_type_t type)
313 {
314         if (!cfqg)
315                 return NULL;
316
317         if (prio == IDLE_WORKLOAD)
318                 return &cfqg->service_tree_idle;
319
320         return &cfqg->service_trees[prio][type];
321 }
322
323 enum cfqq_state_flags {
324         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
325         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
326         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
327         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
329         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
330         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
331         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
332         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
333         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
334         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
335         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
336         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
337 };
338
339 #define CFQ_CFQQ_FNS(name)                                              \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
341 {                                                                       \
342         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
343 }                                                                       \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
345 {                                                                       \
346         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
347 }                                                                       \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
349 {                                                                       \
350         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
351 }
352
353 CFQ_CFQQ_FNS(on_rr);
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(sync);
362 CFQ_CFQQ_FNS(coop);
363 CFQ_CFQQ_FNS(split_coop);
364 CFQ_CFQQ_FNS(deep);
365 CFQ_CFQQ_FNS(wait_busy);
366 #undef CFQ_CFQQ_FNS
367
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
370         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
373
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
375         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
376                                 blkg_path(&(cfqg)->blkg), ##args);      \
377
378 #else
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
380         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
382 #endif
383 #define cfq_log(cfqd, fmt, args...)     \
384         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
385
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388         for (i = 0; i <= IDLE_WORKLOAD; i++) \
389                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390                         : &cfqg->service_tree_idle; \
391                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392                         (i == IDLE_WORKLOAD && j == 0); \
393                         j++, st = i < IDLE_WORKLOAD ? \
394                         &cfqg->service_trees[i][j]: NULL) \
395
396
397 static inline bool iops_mode(struct cfq_data *cfqd)
398 {
399         /*
400          * If we are not idling on queues and it is a NCQ drive, parallel
401          * execution of requests is on and measuring time is not possible
402          * in most of the cases until and unless we drive shallower queue
403          * depths and that becomes a performance bottleneck. In such cases
404          * switch to start providing fairness in terms of number of IOs.
405          */
406         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
407                 return true;
408         else
409                 return false;
410 }
411
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
413 {
414         if (cfq_class_idle(cfqq))
415                 return IDLE_WORKLOAD;
416         if (cfq_class_rt(cfqq))
417                 return RT_WORKLOAD;
418         return BE_WORKLOAD;
419 }
420
421
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
423 {
424         if (!cfq_cfqq_sync(cfqq))
425                 return ASYNC_WORKLOAD;
426         if (!cfq_cfqq_idle_window(cfqq))
427                 return SYNC_NOIDLE_WORKLOAD;
428         return SYNC_WORKLOAD;
429 }
430
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432                                         struct cfq_data *cfqd,
433                                         struct cfq_group *cfqg)
434 {
435         if (wl == IDLE_WORKLOAD)
436                 return cfqg->service_tree_idle.count;
437
438         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
441 }
442
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444                                         struct cfq_group *cfqg)
445 {
446         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
448 }
449
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452                                        struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454                                                 struct io_context *);
455
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
457                                             bool is_sync)
458 {
459         return cic->cfqq[is_sync];
460 }
461
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463                                 struct cfq_queue *cfqq, bool is_sync)
464 {
465         cic->cfqq[is_sync] = cfqq;
466 }
467
468 #define CIC_DEAD_KEY    1ul
469 #define CIC_DEAD_INDEX_SHIFT    1
470
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
472 {
473         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
474 }
475
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
477 {
478         struct cfq_data *cfqd = cic->key;
479
480         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
481                 return NULL;
482
483         return cfqd;
484 }
485
486 /*
487  * We regard a request as SYNC, if it's either a read or has the SYNC bit
488  * set (in which case it could also be direct WRITE).
489  */
490 static inline bool cfq_bio_sync(struct bio *bio)
491 {
492         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
493 }
494
495 /*
496  * scheduler run of queue, if there are requests pending and no one in the
497  * driver that will restart queueing
498  */
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
500 {
501         if (cfqd->busy_queues) {
502                 cfq_log(cfqd, "schedule dispatch");
503                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
504         }
505 }
506
507 /*
508  * Scale schedule slice based on io priority. Use the sync time slice only
509  * if a queue is marked sync and has sync io queued. A sync queue with async
510  * io only, should not get full sync slice length.
511  */
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
513                                  unsigned short prio)
514 {
515         const int base_slice = cfqd->cfq_slice[sync];
516
517         WARN_ON(prio >= IOPRIO_BE_NR);
518
519         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
520 }
521
522 static inline int
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
524 {
525         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
526 }
527
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
529 {
530         u64 d = delta << CFQ_SERVICE_SHIFT;
531
532         d = d * BLKIO_WEIGHT_DEFAULT;
533         do_div(d, cfqg->weight);
534         return d;
535 }
536
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
538 {
539         s64 delta = (s64)(vdisktime - min_vdisktime);
540         if (delta > 0)
541                 min_vdisktime = vdisktime;
542
543         return min_vdisktime;
544 }
545
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
547 {
548         s64 delta = (s64)(vdisktime - min_vdisktime);
549         if (delta < 0)
550                 min_vdisktime = vdisktime;
551
552         return min_vdisktime;
553 }
554
555 static void update_min_vdisktime(struct cfq_rb_root *st)
556 {
557         struct cfq_group *cfqg;
558
559         if (st->left) {
560                 cfqg = rb_entry_cfqg(st->left);
561                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
562                                                   cfqg->vdisktime);
563         }
564 }
565
566 /*
567  * get averaged number of queues of RT/BE priority.
568  * average is updated, with a formula that gives more weight to higher numbers,
569  * to quickly follows sudden increases and decrease slowly
570  */
571
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573                                         struct cfq_group *cfqg, bool rt)
574 {
575         unsigned min_q, max_q;
576         unsigned mult  = cfq_hist_divisor - 1;
577         unsigned round = cfq_hist_divisor / 2;
578         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
579
580         min_q = min(cfqg->busy_queues_avg[rt], busy);
581         max_q = max(cfqg->busy_queues_avg[rt], busy);
582         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
583                 cfq_hist_divisor;
584         return cfqg->busy_queues_avg[rt];
585 }
586
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
589 {
590         struct cfq_rb_root *st = &cfqd->grp_service_tree;
591
592         return cfq_target_latency * cfqg->weight / st->total_weight;
593 }
594
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
597 {
598         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599         if (cfqd->cfq_latency) {
600                 /*
601                  * interested queues (we consider only the ones with the same
602                  * priority class in the cfq group)
603                  */
604                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
605                                                 cfq_class_rt(cfqq));
606                 unsigned sync_slice = cfqd->cfq_slice[1];
607                 unsigned expect_latency = sync_slice * iq;
608                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
609
610                 if (expect_latency > group_slice) {
611                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612                         /* scale low_slice according to IO priority
613                          * and sync vs async */
614                         unsigned low_slice =
615                                 min(slice, base_low_slice * slice / sync_slice);
616                         /* the adapted slice value is scaled to fit all iqs
617                          * into the target latency */
618                         slice = max(slice * group_slice / expect_latency,
619                                     low_slice);
620                 }
621         }
622         return slice;
623 }
624
625 static inline void
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
627 {
628         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
629
630         cfqq->slice_start = jiffies;
631         cfqq->slice_end = jiffies + slice;
632         cfqq->allocated_slice = slice;
633         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
634 }
635
636 /*
637  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638  * isn't valid until the first request from the dispatch is activated
639  * and the slice time set.
640  */
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
642 {
643         if (cfq_cfqq_slice_new(cfqq))
644                 return false;
645         if (time_before(jiffies, cfqq->slice_end))
646                 return false;
647
648         return true;
649 }
650
651 /*
652  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653  * We choose the request that is closest to the head right now. Distance
654  * behind the head is penalized and only allowed to a certain extent.
655  */
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
658 {
659         sector_t s1, s2, d1 = 0, d2 = 0;
660         unsigned long back_max;
661 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
663         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
664
665         if (rq1 == NULL || rq1 == rq2)
666                 return rq2;
667         if (rq2 == NULL)
668                 return rq1;
669
670         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
671                 return rq1;
672         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
673                 return rq2;
674         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
675                 return rq1;
676         else if ((rq2->cmd_flags & REQ_META) &&
677                  !(rq1->cmd_flags & REQ_META))
678                 return rq2;
679
680         s1 = blk_rq_pos(rq1);
681         s2 = blk_rq_pos(rq2);
682
683         /*
684          * by definition, 1KiB is 2 sectors
685          */
686         back_max = cfqd->cfq_back_max * 2;
687
688         /*
689          * Strict one way elevator _except_ in the case where we allow
690          * short backward seeks which are biased as twice the cost of a
691          * similar forward seek.
692          */
693         if (s1 >= last)
694                 d1 = s1 - last;
695         else if (s1 + back_max >= last)
696                 d1 = (last - s1) * cfqd->cfq_back_penalty;
697         else
698                 wrap |= CFQ_RQ1_WRAP;
699
700         if (s2 >= last)
701                 d2 = s2 - last;
702         else if (s2 + back_max >= last)
703                 d2 = (last - s2) * cfqd->cfq_back_penalty;
704         else
705                 wrap |= CFQ_RQ2_WRAP;
706
707         /* Found required data */
708
709         /*
710          * By doing switch() on the bit mask "wrap" we avoid having to
711          * check two variables for all permutations: --> faster!
712          */
713         switch (wrap) {
714         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
715                 if (d1 < d2)
716                         return rq1;
717                 else if (d2 < d1)
718                         return rq2;
719                 else {
720                         if (s1 >= s2)
721                                 return rq1;
722                         else
723                                 return rq2;
724                 }
725
726         case CFQ_RQ2_WRAP:
727                 return rq1;
728         case CFQ_RQ1_WRAP:
729                 return rq2;
730         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
731         default:
732                 /*
733                  * Since both rqs are wrapped,
734                  * start with the one that's further behind head
735                  * (--> only *one* back seek required),
736                  * since back seek takes more time than forward.
737                  */
738                 if (s1 <= s2)
739                         return rq1;
740                 else
741                         return rq2;
742         }
743 }
744
745 /*
746  * The below is leftmost cache rbtree addon
747  */
748 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
749 {
750         /* Service tree is empty */
751         if (!root->count)
752                 return NULL;
753
754         if (!root->left)
755                 root->left = rb_first(&root->rb);
756
757         if (root->left)
758                 return rb_entry(root->left, struct cfq_queue, rb_node);
759
760         return NULL;
761 }
762
763 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
764 {
765         if (!root->left)
766                 root->left = rb_first(&root->rb);
767
768         if (root->left)
769                 return rb_entry_cfqg(root->left);
770
771         return NULL;
772 }
773
774 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
775 {
776         rb_erase(n, root);
777         RB_CLEAR_NODE(n);
778 }
779
780 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
781 {
782         if (root->left == n)
783                 root->left = NULL;
784         rb_erase_init(n, &root->rb);
785         --root->count;
786 }
787
788 /*
789  * would be nice to take fifo expire time into account as well
790  */
791 static struct request *
792 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
793                   struct request *last)
794 {
795         struct rb_node *rbnext = rb_next(&last->rb_node);
796         struct rb_node *rbprev = rb_prev(&last->rb_node);
797         struct request *next = NULL, *prev = NULL;
798
799         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
800
801         if (rbprev)
802                 prev = rb_entry_rq(rbprev);
803
804         if (rbnext)
805                 next = rb_entry_rq(rbnext);
806         else {
807                 rbnext = rb_first(&cfqq->sort_list);
808                 if (rbnext && rbnext != &last->rb_node)
809                         next = rb_entry_rq(rbnext);
810         }
811
812         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
813 }
814
815 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
816                                       struct cfq_queue *cfqq)
817 {
818         /*
819          * just an approximation, should be ok.
820          */
821         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
822                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
823 }
824
825 static inline s64
826 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
827 {
828         return cfqg->vdisktime - st->min_vdisktime;
829 }
830
831 static void
832 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
833 {
834         struct rb_node **node = &st->rb.rb_node;
835         struct rb_node *parent = NULL;
836         struct cfq_group *__cfqg;
837         s64 key = cfqg_key(st, cfqg);
838         int left = 1;
839
840         while (*node != NULL) {
841                 parent = *node;
842                 __cfqg = rb_entry_cfqg(parent);
843
844                 if (key < cfqg_key(st, __cfqg))
845                         node = &parent->rb_left;
846                 else {
847                         node = &parent->rb_right;
848                         left = 0;
849                 }
850         }
851
852         if (left)
853                 st->left = &cfqg->rb_node;
854
855         rb_link_node(&cfqg->rb_node, parent, node);
856         rb_insert_color(&cfqg->rb_node, &st->rb);
857 }
858
859 static void
860 cfq_update_group_weight(struct cfq_group *cfqg)
861 {
862         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
863         if (cfqg->needs_update) {
864                 cfqg->weight = cfqg->new_weight;
865                 cfqg->needs_update = false;
866         }
867 }
868
869 static void
870 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
871 {
872         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
873
874         cfq_update_group_weight(cfqg);
875         __cfq_group_service_tree_add(st, cfqg);
876         st->total_weight += cfqg->weight;
877 }
878
879 static void
880 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
881 {
882         struct cfq_rb_root *st = &cfqd->grp_service_tree;
883         struct cfq_group *__cfqg;
884         struct rb_node *n;
885
886         cfqg->nr_cfqq++;
887         if (!RB_EMPTY_NODE(&cfqg->rb_node))
888                 return;
889
890         /*
891          * Currently put the group at the end. Later implement something
892          * so that groups get lesser vtime based on their weights, so that
893          * if group does not loose all if it was not continuously backlogged.
894          */
895         n = rb_last(&st->rb);
896         if (n) {
897                 __cfqg = rb_entry_cfqg(n);
898                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
899         } else
900                 cfqg->vdisktime = st->min_vdisktime;
901         cfq_group_service_tree_add(st, cfqg);
902 }
903
904 static void
905 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
906 {
907         st->total_weight -= cfqg->weight;
908         if (!RB_EMPTY_NODE(&cfqg->rb_node))
909                 cfq_rb_erase(&cfqg->rb_node, st);
910 }
911
912 static void
913 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
914 {
915         struct cfq_rb_root *st = &cfqd->grp_service_tree;
916
917         BUG_ON(cfqg->nr_cfqq < 1);
918         cfqg->nr_cfqq--;
919
920         /* If there are other cfq queues under this group, don't delete it */
921         if (cfqg->nr_cfqq)
922                 return;
923
924         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
925         cfq_group_service_tree_del(st, cfqg);
926         cfqg->saved_workload_slice = 0;
927         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
928 }
929
930 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
931                                                 unsigned int *unaccounted_time)
932 {
933         unsigned int slice_used;
934
935         /*
936          * Queue got expired before even a single request completed or
937          * got expired immediately after first request completion.
938          */
939         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
940                 /*
941                  * Also charge the seek time incurred to the group, otherwise
942                  * if there are mutiple queues in the group, each can dispatch
943                  * a single request on seeky media and cause lots of seek time
944                  * and group will never know it.
945                  */
946                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
947                                         1);
948         } else {
949                 slice_used = jiffies - cfqq->slice_start;
950                 if (slice_used > cfqq->allocated_slice) {
951                         *unaccounted_time = slice_used - cfqq->allocated_slice;
952                         slice_used = cfqq->allocated_slice;
953                 }
954                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
955                         *unaccounted_time += cfqq->slice_start -
956                                         cfqq->dispatch_start;
957         }
958
959         return slice_used;
960 }
961
962 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
963                                 struct cfq_queue *cfqq)
964 {
965         struct cfq_rb_root *st = &cfqd->grp_service_tree;
966         unsigned int used_sl, charge, unaccounted_sl = 0;
967         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
968                         - cfqg->service_tree_idle.count;
969
970         BUG_ON(nr_sync < 0);
971         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
972
973         if (iops_mode(cfqd))
974                 charge = cfqq->slice_dispatch;
975         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
976                 charge = cfqq->allocated_slice;
977
978         /* Can't update vdisktime while group is on service tree */
979         cfq_group_service_tree_del(st, cfqg);
980         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
981         /* If a new weight was requested, update now, off tree */
982         cfq_group_service_tree_add(st, cfqg);
983
984         /* This group is being expired. Save the context */
985         if (time_after(cfqd->workload_expires, jiffies)) {
986                 cfqg->saved_workload_slice = cfqd->workload_expires
987                                                 - jiffies;
988                 cfqg->saved_workload = cfqd->serving_type;
989                 cfqg->saved_serving_prio = cfqd->serving_prio;
990         } else
991                 cfqg->saved_workload_slice = 0;
992
993         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
994                                         st->min_vdisktime);
995         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
996                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
997                         iops_mode(cfqd), cfqq->nr_sectors);
998         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
999                                           unaccounted_sl);
1000         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1001 }
1002
1003 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1004 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1005 {
1006         if (blkg)
1007                 return container_of(blkg, struct cfq_group, blkg);
1008         return NULL;
1009 }
1010
1011 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1012                                         unsigned int weight)
1013 {
1014         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1015         cfqg->new_weight = weight;
1016         cfqg->needs_update = true;
1017 }
1018
1019 static struct cfq_group * cfq_find_alloc_cfqg(struct cfq_data *cfqd,
1020                 struct blkio_cgroup *blkcg)
1021 {
1022         struct cfq_group *cfqg = NULL;
1023         void *key = cfqd;
1024         int i, j;
1025         struct cfq_rb_root *st;
1026         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1027         unsigned int major, minor;
1028
1029         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1030         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1031                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1032                 cfqg->blkg.dev = MKDEV(major, minor);
1033                 goto done;
1034         }
1035         if (cfqg)
1036                 goto done;
1037
1038         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1039         if (!cfqg)
1040                 goto done;
1041
1042         for_each_cfqg_st(cfqg, i, j, st)
1043                 *st = CFQ_RB_ROOT;
1044         RB_CLEAR_NODE(&cfqg->rb_node);
1045
1046         /*
1047          * Take the initial reference that will be released on destroy
1048          * This can be thought of a joint reference by cgroup and
1049          * elevator which will be dropped by either elevator exit
1050          * or cgroup deletion path depending on who is exiting first.
1051          */
1052         cfqg->ref = 1;
1053
1054         /*
1055          * Add group onto cgroup list. It might happen that bdi->dev is
1056          * not initialized yet. Initialize this new group without major
1057          * and minor info and this info will be filled in once a new thread
1058          * comes for IO. See code above.
1059          */
1060         if (bdi->dev) {
1061                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1062                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1063                                         MKDEV(major, minor));
1064         } else
1065                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1066                                         0);
1067
1068         cfqd->nr_blkcg_linked_grps++;
1069         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1070
1071         /* Add group on cfqd list */
1072         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1073
1074 done:
1075         return cfqg;
1076 }
1077
1078 /*
1079  * Search for the cfq group current task belongs to. request_queue lock must
1080  * be held.
1081  */
1082 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1083 {
1084         struct blkio_cgroup *blkcg;
1085         struct cfq_group *cfqg = NULL;
1086
1087         rcu_read_lock();
1088         blkcg = task_blkio_cgroup(current);
1089         cfqg = cfq_find_alloc_cfqg(cfqd, blkcg);
1090         if (!cfqg)
1091                 cfqg = &cfqd->root_group;
1092         rcu_read_unlock();
1093         return cfqg;
1094 }
1095
1096 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1097 {
1098         cfqg->ref++;
1099         return cfqg;
1100 }
1101
1102 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1103 {
1104         /* Currently, all async queues are mapped to root group */
1105         if (!cfq_cfqq_sync(cfqq))
1106                 cfqg = &cfqq->cfqd->root_group;
1107
1108         cfqq->cfqg = cfqg;
1109         /* cfqq reference on cfqg */
1110         cfqq->cfqg->ref++;
1111 }
1112
1113 static void cfq_put_cfqg(struct cfq_group *cfqg)
1114 {
1115         struct cfq_rb_root *st;
1116         int i, j;
1117
1118         BUG_ON(cfqg->ref <= 0);
1119         cfqg->ref--;
1120         if (cfqg->ref)
1121                 return;
1122         for_each_cfqg_st(cfqg, i, j, st)
1123                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1124         kfree(cfqg);
1125 }
1126
1127 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1128 {
1129         /* Something wrong if we are trying to remove same group twice */
1130         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1131
1132         hlist_del_init(&cfqg->cfqd_node);
1133
1134         /*
1135          * Put the reference taken at the time of creation so that when all
1136          * queues are gone, group can be destroyed.
1137          */
1138         cfq_put_cfqg(cfqg);
1139 }
1140
1141 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1142 {
1143         struct hlist_node *pos, *n;
1144         struct cfq_group *cfqg;
1145
1146         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1147                 /*
1148                  * If cgroup removal path got to blk_group first and removed
1149                  * it from cgroup list, then it will take care of destroying
1150                  * cfqg also.
1151                  */
1152                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1153                         cfq_destroy_cfqg(cfqd, cfqg);
1154         }
1155 }
1156
1157 /*
1158  * Blk cgroup controller notification saying that blkio_group object is being
1159  * delinked as associated cgroup object is going away. That also means that
1160  * no new IO will come in this group. So get rid of this group as soon as
1161  * any pending IO in the group is finished.
1162  *
1163  * This function is called under rcu_read_lock(). key is the rcu protected
1164  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1165  * read lock.
1166  *
1167  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1168  * it should not be NULL as even if elevator was exiting, cgroup deltion
1169  * path got to it first.
1170  */
1171 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1172 {
1173         unsigned long  flags;
1174         struct cfq_data *cfqd = key;
1175
1176         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1177         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1178         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1179 }
1180
1181 #else /* GROUP_IOSCHED */
1182 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1183 {
1184         return &cfqd->root_group;
1185 }
1186
1187 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1188 {
1189         return cfqg;
1190 }
1191
1192 static inline void
1193 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1194         cfqq->cfqg = cfqg;
1195 }
1196
1197 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1198 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1199
1200 #endif /* GROUP_IOSCHED */
1201
1202 /*
1203  * The cfqd->service_trees holds all pending cfq_queue's that have
1204  * requests waiting to be processed. It is sorted in the order that
1205  * we will service the queues.
1206  */
1207 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1208                                  bool add_front)
1209 {
1210         struct rb_node **p, *parent;
1211         struct cfq_queue *__cfqq;
1212         unsigned long rb_key;
1213         struct cfq_rb_root *service_tree;
1214         int left;
1215         int new_cfqq = 1;
1216         int group_changed = 0;
1217
1218         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1219                                                 cfqq_type(cfqq));
1220         if (cfq_class_idle(cfqq)) {
1221                 rb_key = CFQ_IDLE_DELAY;
1222                 parent = rb_last(&service_tree->rb);
1223                 if (parent && parent != &cfqq->rb_node) {
1224                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1225                         rb_key += __cfqq->rb_key;
1226                 } else
1227                         rb_key += jiffies;
1228         } else if (!add_front) {
1229                 /*
1230                  * Get our rb key offset. Subtract any residual slice
1231                  * value carried from last service. A negative resid
1232                  * count indicates slice overrun, and this should position
1233                  * the next service time further away in the tree.
1234                  */
1235                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1236                 rb_key -= cfqq->slice_resid;
1237                 cfqq->slice_resid = 0;
1238         } else {
1239                 rb_key = -HZ;
1240                 __cfqq = cfq_rb_first(service_tree);
1241                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1242         }
1243
1244         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1245                 new_cfqq = 0;
1246                 /*
1247                  * same position, nothing more to do
1248                  */
1249                 if (rb_key == cfqq->rb_key &&
1250                     cfqq->service_tree == service_tree)
1251                         return;
1252
1253                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1254                 cfqq->service_tree = NULL;
1255         }
1256
1257         left = 1;
1258         parent = NULL;
1259         cfqq->service_tree = service_tree;
1260         p = &service_tree->rb.rb_node;
1261         while (*p) {
1262                 struct rb_node **n;
1263
1264                 parent = *p;
1265                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1266
1267                 /*
1268                  * sort by key, that represents service time.
1269                  */
1270                 if (time_before(rb_key, __cfqq->rb_key))
1271                         n = &(*p)->rb_left;
1272                 else {
1273                         n = &(*p)->rb_right;
1274                         left = 0;
1275                 }
1276
1277                 p = n;
1278         }
1279
1280         if (left)
1281                 service_tree->left = &cfqq->rb_node;
1282
1283         cfqq->rb_key = rb_key;
1284         rb_link_node(&cfqq->rb_node, parent, p);
1285         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1286         service_tree->count++;
1287         if ((add_front || !new_cfqq) && !group_changed)
1288                 return;
1289         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1290 }
1291
1292 static struct cfq_queue *
1293 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1294                      sector_t sector, struct rb_node **ret_parent,
1295                      struct rb_node ***rb_link)
1296 {
1297         struct rb_node **p, *parent;
1298         struct cfq_queue *cfqq = NULL;
1299
1300         parent = NULL;
1301         p = &root->rb_node;
1302         while (*p) {
1303                 struct rb_node **n;
1304
1305                 parent = *p;
1306                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1307
1308                 /*
1309                  * Sort strictly based on sector.  Smallest to the left,
1310                  * largest to the right.
1311                  */
1312                 if (sector > blk_rq_pos(cfqq->next_rq))
1313                         n = &(*p)->rb_right;
1314                 else if (sector < blk_rq_pos(cfqq->next_rq))
1315                         n = &(*p)->rb_left;
1316                 else
1317                         break;
1318                 p = n;
1319                 cfqq = NULL;
1320         }
1321
1322         *ret_parent = parent;
1323         if (rb_link)
1324                 *rb_link = p;
1325         return cfqq;
1326 }
1327
1328 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1329 {
1330         struct rb_node **p, *parent;
1331         struct cfq_queue *__cfqq;
1332
1333         if (cfqq->p_root) {
1334                 rb_erase(&cfqq->p_node, cfqq->p_root);
1335                 cfqq->p_root = NULL;
1336         }
1337
1338         if (cfq_class_idle(cfqq))
1339                 return;
1340         if (!cfqq->next_rq)
1341                 return;
1342
1343         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1344         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1345                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1346         if (!__cfqq) {
1347                 rb_link_node(&cfqq->p_node, parent, p);
1348                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1349         } else
1350                 cfqq->p_root = NULL;
1351 }
1352
1353 /*
1354  * Update cfqq's position in the service tree.
1355  */
1356 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1357 {
1358         /*
1359          * Resorting requires the cfqq to be on the RR list already.
1360          */
1361         if (cfq_cfqq_on_rr(cfqq)) {
1362                 cfq_service_tree_add(cfqd, cfqq, 0);
1363                 cfq_prio_tree_add(cfqd, cfqq);
1364         }
1365 }
1366
1367 /*
1368  * add to busy list of queues for service, trying to be fair in ordering
1369  * the pending list according to last request service
1370  */
1371 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1372 {
1373         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1374         BUG_ON(cfq_cfqq_on_rr(cfqq));
1375         cfq_mark_cfqq_on_rr(cfqq);
1376         cfqd->busy_queues++;
1377         if (cfq_cfqq_sync(cfqq))
1378                 cfqd->busy_sync_queues++;
1379
1380         cfq_resort_rr_list(cfqd, cfqq);
1381 }
1382
1383 /*
1384  * Called when the cfqq no longer has requests pending, remove it from
1385  * the service tree.
1386  */
1387 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1388 {
1389         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1390         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1391         cfq_clear_cfqq_on_rr(cfqq);
1392
1393         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1394                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1395                 cfqq->service_tree = NULL;
1396         }
1397         if (cfqq->p_root) {
1398                 rb_erase(&cfqq->p_node, cfqq->p_root);
1399                 cfqq->p_root = NULL;
1400         }
1401
1402         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1403         BUG_ON(!cfqd->busy_queues);
1404         cfqd->busy_queues--;
1405         if (cfq_cfqq_sync(cfqq))
1406                 cfqd->busy_sync_queues--;
1407 }
1408
1409 /*
1410  * rb tree support functions
1411  */
1412 static void cfq_del_rq_rb(struct request *rq)
1413 {
1414         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1415         const int sync = rq_is_sync(rq);
1416
1417         BUG_ON(!cfqq->queued[sync]);
1418         cfqq->queued[sync]--;
1419
1420         elv_rb_del(&cfqq->sort_list, rq);
1421
1422         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1423                 /*
1424                  * Queue will be deleted from service tree when we actually
1425                  * expire it later. Right now just remove it from prio tree
1426                  * as it is empty.
1427                  */
1428                 if (cfqq->p_root) {
1429                         rb_erase(&cfqq->p_node, cfqq->p_root);
1430                         cfqq->p_root = NULL;
1431                 }
1432         }
1433 }
1434
1435 static void cfq_add_rq_rb(struct request *rq)
1436 {
1437         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1438         struct cfq_data *cfqd = cfqq->cfqd;
1439         struct request *__alias, *prev;
1440
1441         cfqq->queued[rq_is_sync(rq)]++;
1442
1443         /*
1444          * looks a little odd, but the first insert might return an alias.
1445          * if that happens, put the alias on the dispatch list
1446          */
1447         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1448                 cfq_dispatch_insert(cfqd->queue, __alias);
1449
1450         if (!cfq_cfqq_on_rr(cfqq))
1451                 cfq_add_cfqq_rr(cfqd, cfqq);
1452
1453         /*
1454          * check if this request is a better next-serve candidate
1455          */
1456         prev = cfqq->next_rq;
1457         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1458
1459         /*
1460          * adjust priority tree position, if ->next_rq changes
1461          */
1462         if (prev != cfqq->next_rq)
1463                 cfq_prio_tree_add(cfqd, cfqq);
1464
1465         BUG_ON(!cfqq->next_rq);
1466 }
1467
1468 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1469 {
1470         elv_rb_del(&cfqq->sort_list, rq);
1471         cfqq->queued[rq_is_sync(rq)]--;
1472         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1473                                         rq_data_dir(rq), rq_is_sync(rq));
1474         cfq_add_rq_rb(rq);
1475         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1476                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1477                         rq_is_sync(rq));
1478 }
1479
1480 static struct request *
1481 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1482 {
1483         struct task_struct *tsk = current;
1484         struct cfq_io_context *cic;
1485         struct cfq_queue *cfqq;
1486
1487         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1488         if (!cic)
1489                 return NULL;
1490
1491         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1492         if (cfqq) {
1493                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1494
1495                 return elv_rb_find(&cfqq->sort_list, sector);
1496         }
1497
1498         return NULL;
1499 }
1500
1501 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1502 {
1503         struct cfq_data *cfqd = q->elevator->elevator_data;
1504
1505         cfqd->rq_in_driver++;
1506         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1507                                                 cfqd->rq_in_driver);
1508
1509         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1510 }
1511
1512 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1513 {
1514         struct cfq_data *cfqd = q->elevator->elevator_data;
1515
1516         WARN_ON(!cfqd->rq_in_driver);
1517         cfqd->rq_in_driver--;
1518         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1519                                                 cfqd->rq_in_driver);
1520 }
1521
1522 static void cfq_remove_request(struct request *rq)
1523 {
1524         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1525
1526         if (cfqq->next_rq == rq)
1527                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1528
1529         list_del_init(&rq->queuelist);
1530         cfq_del_rq_rb(rq);
1531
1532         cfqq->cfqd->rq_queued--;
1533         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1534                                         rq_data_dir(rq), rq_is_sync(rq));
1535         if (rq->cmd_flags & REQ_META) {
1536                 WARN_ON(!cfqq->meta_pending);
1537                 cfqq->meta_pending--;
1538         }
1539 }
1540
1541 static int cfq_merge(struct request_queue *q, struct request **req,
1542                      struct bio *bio)
1543 {
1544         struct cfq_data *cfqd = q->elevator->elevator_data;
1545         struct request *__rq;
1546
1547         __rq = cfq_find_rq_fmerge(cfqd, bio);
1548         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1549                 *req = __rq;
1550                 return ELEVATOR_FRONT_MERGE;
1551         }
1552
1553         return ELEVATOR_NO_MERGE;
1554 }
1555
1556 static void cfq_merged_request(struct request_queue *q, struct request *req,
1557                                int type)
1558 {
1559         if (type == ELEVATOR_FRONT_MERGE) {
1560                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1561
1562                 cfq_reposition_rq_rb(cfqq, req);
1563         }
1564 }
1565
1566 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1567                                 struct bio *bio)
1568 {
1569         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1570                                         bio_data_dir(bio), cfq_bio_sync(bio));
1571 }
1572
1573 static void
1574 cfq_merged_requests(struct request_queue *q, struct request *rq,
1575                     struct request *next)
1576 {
1577         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1578         /*
1579          * reposition in fifo if next is older than rq
1580          */
1581         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1582             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1583                 list_move(&rq->queuelist, &next->queuelist);
1584                 rq_set_fifo_time(rq, rq_fifo_time(next));
1585         }
1586
1587         if (cfqq->next_rq == next)
1588                 cfqq->next_rq = rq;
1589         cfq_remove_request(next);
1590         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1591                                         rq_data_dir(next), rq_is_sync(next));
1592 }
1593
1594 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1595                            struct bio *bio)
1596 {
1597         struct cfq_data *cfqd = q->elevator->elevator_data;
1598         struct cfq_io_context *cic;
1599         struct cfq_queue *cfqq;
1600
1601         /*
1602          * Disallow merge of a sync bio into an async request.
1603          */
1604         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1605                 return false;
1606
1607         /*
1608          * Lookup the cfqq that this bio will be queued with. Allow
1609          * merge only if rq is queued there.
1610          */
1611         cic = cfq_cic_lookup(cfqd, current->io_context);
1612         if (!cic)
1613                 return false;
1614
1615         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1616         return cfqq == RQ_CFQQ(rq);
1617 }
1618
1619 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1620 {
1621         del_timer(&cfqd->idle_slice_timer);
1622         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1623 }
1624
1625 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1626                                    struct cfq_queue *cfqq)
1627 {
1628         if (cfqq) {
1629                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1630                                 cfqd->serving_prio, cfqd->serving_type);
1631                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1632                 cfqq->slice_start = 0;
1633                 cfqq->dispatch_start = jiffies;
1634                 cfqq->allocated_slice = 0;
1635                 cfqq->slice_end = 0;
1636                 cfqq->slice_dispatch = 0;
1637                 cfqq->nr_sectors = 0;
1638
1639                 cfq_clear_cfqq_wait_request(cfqq);
1640                 cfq_clear_cfqq_must_dispatch(cfqq);
1641                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1642                 cfq_clear_cfqq_fifo_expire(cfqq);
1643                 cfq_mark_cfqq_slice_new(cfqq);
1644
1645                 cfq_del_timer(cfqd, cfqq);
1646         }
1647
1648         cfqd->active_queue = cfqq;
1649 }
1650
1651 /*
1652  * current cfqq expired its slice (or was too idle), select new one
1653  */
1654 static void
1655 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1656                     bool timed_out)
1657 {
1658         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1659
1660         if (cfq_cfqq_wait_request(cfqq))
1661                 cfq_del_timer(cfqd, cfqq);
1662
1663         cfq_clear_cfqq_wait_request(cfqq);
1664         cfq_clear_cfqq_wait_busy(cfqq);
1665
1666         /*
1667          * If this cfqq is shared between multiple processes, check to
1668          * make sure that those processes are still issuing I/Os within
1669          * the mean seek distance.  If not, it may be time to break the
1670          * queues apart again.
1671          */
1672         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1673                 cfq_mark_cfqq_split_coop(cfqq);
1674
1675         /*
1676          * store what was left of this slice, if the queue idled/timed out
1677          */
1678         if (timed_out) {
1679                 if (cfq_cfqq_slice_new(cfqq))
1680                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1681                 else
1682                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1683                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1684         }
1685
1686         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1687
1688         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1689                 cfq_del_cfqq_rr(cfqd, cfqq);
1690
1691         cfq_resort_rr_list(cfqd, cfqq);
1692
1693         if (cfqq == cfqd->active_queue)
1694                 cfqd->active_queue = NULL;
1695
1696         if (cfqd->active_cic) {
1697                 put_io_context(cfqd->active_cic->ioc);
1698                 cfqd->active_cic = NULL;
1699         }
1700 }
1701
1702 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1703 {
1704         struct cfq_queue *cfqq = cfqd->active_queue;
1705
1706         if (cfqq)
1707                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1708 }
1709
1710 /*
1711  * Get next queue for service. Unless we have a queue preemption,
1712  * we'll simply select the first cfqq in the service tree.
1713  */
1714 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1715 {
1716         struct cfq_rb_root *service_tree =
1717                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1718                                         cfqd->serving_type);
1719
1720         if (!cfqd->rq_queued)
1721                 return NULL;
1722
1723         /* There is nothing to dispatch */
1724         if (!service_tree)
1725                 return NULL;
1726         if (RB_EMPTY_ROOT(&service_tree->rb))
1727                 return NULL;
1728         return cfq_rb_first(service_tree);
1729 }
1730
1731 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1732 {
1733         struct cfq_group *cfqg;
1734         struct cfq_queue *cfqq;
1735         int i, j;
1736         struct cfq_rb_root *st;
1737
1738         if (!cfqd->rq_queued)
1739                 return NULL;
1740
1741         cfqg = cfq_get_next_cfqg(cfqd);
1742         if (!cfqg)
1743                 return NULL;
1744
1745         for_each_cfqg_st(cfqg, i, j, st)
1746                 if ((cfqq = cfq_rb_first(st)) != NULL)
1747                         return cfqq;
1748         return NULL;
1749 }
1750
1751 /*
1752  * Get and set a new active queue for service.
1753  */
1754 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1755                                               struct cfq_queue *cfqq)
1756 {
1757         if (!cfqq)
1758                 cfqq = cfq_get_next_queue(cfqd);
1759
1760         __cfq_set_active_queue(cfqd, cfqq);
1761         return cfqq;
1762 }
1763
1764 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1765                                           struct request *rq)
1766 {
1767         if (blk_rq_pos(rq) >= cfqd->last_position)
1768                 return blk_rq_pos(rq) - cfqd->last_position;
1769         else
1770                 return cfqd->last_position - blk_rq_pos(rq);
1771 }
1772
1773 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1774                                struct request *rq)
1775 {
1776         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1777 }
1778
1779 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1780                                     struct cfq_queue *cur_cfqq)
1781 {
1782         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1783         struct rb_node *parent, *node;
1784         struct cfq_queue *__cfqq;
1785         sector_t sector = cfqd->last_position;
1786
1787         if (RB_EMPTY_ROOT(root))
1788                 return NULL;
1789
1790         /*
1791          * First, if we find a request starting at the end of the last
1792          * request, choose it.
1793          */
1794         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1795         if (__cfqq)
1796                 return __cfqq;
1797
1798         /*
1799          * If the exact sector wasn't found, the parent of the NULL leaf
1800          * will contain the closest sector.
1801          */
1802         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1803         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1804                 return __cfqq;
1805
1806         if (blk_rq_pos(__cfqq->next_rq) < sector)
1807                 node = rb_next(&__cfqq->p_node);
1808         else
1809                 node = rb_prev(&__cfqq->p_node);
1810         if (!node)
1811                 return NULL;
1812
1813         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1814         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1815                 return __cfqq;
1816
1817         return NULL;
1818 }
1819
1820 /*
1821  * cfqd - obvious
1822  * cur_cfqq - passed in so that we don't decide that the current queue is
1823  *            closely cooperating with itself.
1824  *
1825  * So, basically we're assuming that that cur_cfqq has dispatched at least
1826  * one request, and that cfqd->last_position reflects a position on the disk
1827  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1828  * assumption.
1829  */
1830 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1831                                               struct cfq_queue *cur_cfqq)
1832 {
1833         struct cfq_queue *cfqq;
1834
1835         if (cfq_class_idle(cur_cfqq))
1836                 return NULL;
1837         if (!cfq_cfqq_sync(cur_cfqq))
1838                 return NULL;
1839         if (CFQQ_SEEKY(cur_cfqq))
1840                 return NULL;
1841
1842         /*
1843          * Don't search priority tree if it's the only queue in the group.
1844          */
1845         if (cur_cfqq->cfqg->nr_cfqq == 1)
1846                 return NULL;
1847
1848         /*
1849          * We should notice if some of the queues are cooperating, eg
1850          * working closely on the same area of the disk. In that case,
1851          * we can group them together and don't waste time idling.
1852          */
1853         cfqq = cfqq_close(cfqd, cur_cfqq);
1854         if (!cfqq)
1855                 return NULL;
1856
1857         /* If new queue belongs to different cfq_group, don't choose it */
1858         if (cur_cfqq->cfqg != cfqq->cfqg)
1859                 return NULL;
1860
1861         /*
1862          * It only makes sense to merge sync queues.
1863          */
1864         if (!cfq_cfqq_sync(cfqq))
1865                 return NULL;
1866         if (CFQQ_SEEKY(cfqq))
1867                 return NULL;
1868
1869         /*
1870          * Do not merge queues of different priority classes
1871          */
1872         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1873                 return NULL;
1874
1875         return cfqq;
1876 }
1877
1878 /*
1879  * Determine whether we should enforce idle window for this queue.
1880  */
1881
1882 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1883 {
1884         enum wl_prio_t prio = cfqq_prio(cfqq);
1885         struct cfq_rb_root *service_tree = cfqq->service_tree;
1886
1887         BUG_ON(!service_tree);
1888         BUG_ON(!service_tree->count);
1889
1890         if (!cfqd->cfq_slice_idle)
1891                 return false;
1892
1893         /* We never do for idle class queues. */
1894         if (prio == IDLE_WORKLOAD)
1895                 return false;
1896
1897         /* We do for queues that were marked with idle window flag. */
1898         if (cfq_cfqq_idle_window(cfqq) &&
1899            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1900                 return true;
1901
1902         /*
1903          * Otherwise, we do only if they are the last ones
1904          * in their service tree.
1905          */
1906         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1907                 return true;
1908         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1909                         service_tree->count);
1910         return false;
1911 }
1912
1913 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1914 {
1915         struct cfq_queue *cfqq = cfqd->active_queue;
1916         struct cfq_io_context *cic;
1917         unsigned long sl, group_idle = 0;
1918
1919         /*
1920          * SSD device without seek penalty, disable idling. But only do so
1921          * for devices that support queuing, otherwise we still have a problem
1922          * with sync vs async workloads.
1923          */
1924         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1925                 return;
1926
1927         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1928         WARN_ON(cfq_cfqq_slice_new(cfqq));
1929
1930         /*
1931          * idle is disabled, either manually or by past process history
1932          */
1933         if (!cfq_should_idle(cfqd, cfqq)) {
1934                 /* no queue idling. Check for group idling */
1935                 if (cfqd->cfq_group_idle)
1936                         group_idle = cfqd->cfq_group_idle;
1937                 else
1938                         return;
1939         }
1940
1941         /*
1942          * still active requests from this queue, don't idle
1943          */
1944         if (cfqq->dispatched)
1945                 return;
1946
1947         /*
1948          * task has exited, don't wait
1949          */
1950         cic = cfqd->active_cic;
1951         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1952                 return;
1953
1954         /*
1955          * If our average think time is larger than the remaining time
1956          * slice, then don't idle. This avoids overrunning the allotted
1957          * time slice.
1958          */
1959         if (sample_valid(cic->ttime_samples) &&
1960             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1961                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1962                                 cic->ttime_mean);
1963                 return;
1964         }
1965
1966         /* There are other queues in the group, don't do group idle */
1967         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1968                 return;
1969
1970         cfq_mark_cfqq_wait_request(cfqq);
1971
1972         if (group_idle)
1973                 sl = cfqd->cfq_group_idle;
1974         else
1975                 sl = cfqd->cfq_slice_idle;
1976
1977         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1978         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1979         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1980                         group_idle ? 1 : 0);
1981 }
1982
1983 /*
1984  * Move request from internal lists to the request queue dispatch list.
1985  */
1986 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1987 {
1988         struct cfq_data *cfqd = q->elevator->elevator_data;
1989         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1990
1991         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1992
1993         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1994         cfq_remove_request(rq);
1995         cfqq->dispatched++;
1996         (RQ_CFQG(rq))->dispatched++;
1997         elv_dispatch_sort(q, rq);
1998
1999         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2000         cfqq->nr_sectors += blk_rq_sectors(rq);
2001         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2002                                         rq_data_dir(rq), rq_is_sync(rq));
2003 }
2004
2005 /*
2006  * return expired entry, or NULL to just start from scratch in rbtree
2007  */
2008 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2009 {
2010         struct request *rq = NULL;
2011
2012         if (cfq_cfqq_fifo_expire(cfqq))
2013                 return NULL;
2014
2015         cfq_mark_cfqq_fifo_expire(cfqq);
2016
2017         if (list_empty(&cfqq->fifo))
2018                 return NULL;
2019
2020         rq = rq_entry_fifo(cfqq->fifo.next);
2021         if (time_before(jiffies, rq_fifo_time(rq)))
2022                 rq = NULL;
2023
2024         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2025         return rq;
2026 }
2027
2028 static inline int
2029 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2030 {
2031         const int base_rq = cfqd->cfq_slice_async_rq;
2032
2033         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2034
2035         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2036 }
2037
2038 /*
2039  * Must be called with the queue_lock held.
2040  */
2041 static int cfqq_process_refs(struct cfq_queue *cfqq)
2042 {
2043         int process_refs, io_refs;
2044
2045         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2046         process_refs = cfqq->ref - io_refs;
2047         BUG_ON(process_refs < 0);
2048         return process_refs;
2049 }
2050
2051 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2052 {
2053         int process_refs, new_process_refs;
2054         struct cfq_queue *__cfqq;
2055
2056         /*
2057          * If there are no process references on the new_cfqq, then it is
2058          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2059          * chain may have dropped their last reference (not just their
2060          * last process reference).
2061          */
2062         if (!cfqq_process_refs(new_cfqq))
2063                 return;
2064
2065         /* Avoid a circular list and skip interim queue merges */
2066         while ((__cfqq = new_cfqq->new_cfqq)) {
2067                 if (__cfqq == cfqq)
2068                         return;
2069                 new_cfqq = __cfqq;
2070         }
2071
2072         process_refs = cfqq_process_refs(cfqq);
2073         new_process_refs = cfqq_process_refs(new_cfqq);
2074         /*
2075          * If the process for the cfqq has gone away, there is no
2076          * sense in merging the queues.
2077          */
2078         if (process_refs == 0 || new_process_refs == 0)
2079                 return;
2080
2081         /*
2082          * Merge in the direction of the lesser amount of work.
2083          */
2084         if (new_process_refs >= process_refs) {
2085                 cfqq->new_cfqq = new_cfqq;
2086                 new_cfqq->ref += process_refs;
2087         } else {
2088                 new_cfqq->new_cfqq = cfqq;
2089                 cfqq->ref += new_process_refs;
2090         }
2091 }
2092
2093 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2094                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2095 {
2096         struct cfq_queue *queue;
2097         int i;
2098         bool key_valid = false;
2099         unsigned long lowest_key = 0;
2100         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2101
2102         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2103                 /* select the one with lowest rb_key */
2104                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2105                 if (queue &&
2106                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2107                         lowest_key = queue->rb_key;
2108                         cur_best = i;
2109                         key_valid = true;
2110                 }
2111         }
2112
2113         return cur_best;
2114 }
2115
2116 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2117 {
2118         unsigned slice;
2119         unsigned count;
2120         struct cfq_rb_root *st;
2121         unsigned group_slice;
2122         enum wl_prio_t original_prio = cfqd->serving_prio;
2123
2124         /* Choose next priority. RT > BE > IDLE */
2125         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2126                 cfqd->serving_prio = RT_WORKLOAD;
2127         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2128                 cfqd->serving_prio = BE_WORKLOAD;
2129         else {
2130                 cfqd->serving_prio = IDLE_WORKLOAD;
2131                 cfqd->workload_expires = jiffies + 1;
2132                 return;
2133         }
2134
2135         if (original_prio != cfqd->serving_prio)
2136                 goto new_workload;
2137
2138         /*
2139          * For RT and BE, we have to choose also the type
2140          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2141          * expiration time
2142          */
2143         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2144         count = st->count;
2145
2146         /*
2147          * check workload expiration, and that we still have other queues ready
2148          */
2149         if (count && !time_after(jiffies, cfqd->workload_expires))
2150                 return;
2151
2152 new_workload:
2153         /* otherwise select new workload type */
2154         cfqd->serving_type =
2155                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2156         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2157         count = st->count;
2158
2159         /*
2160          * the workload slice is computed as a fraction of target latency
2161          * proportional to the number of queues in that workload, over
2162          * all the queues in the same priority class
2163          */
2164         group_slice = cfq_group_slice(cfqd, cfqg);
2165
2166         slice = group_slice * count /
2167                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2168                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2169
2170         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2171                 unsigned int tmp;
2172
2173                 /*
2174                  * Async queues are currently system wide. Just taking
2175                  * proportion of queues with-in same group will lead to higher
2176                  * async ratio system wide as generally root group is going
2177                  * to have higher weight. A more accurate thing would be to
2178                  * calculate system wide asnc/sync ratio.
2179                  */
2180                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2181                 tmp = tmp/cfqd->busy_queues;
2182                 slice = min_t(unsigned, slice, tmp);
2183
2184                 /* async workload slice is scaled down according to
2185                  * the sync/async slice ratio. */
2186                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2187         } else
2188                 /* sync workload slice is at least 2 * cfq_slice_idle */
2189                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2190
2191         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2192         cfq_log(cfqd, "workload slice:%d", slice);
2193         cfqd->workload_expires = jiffies + slice;
2194 }
2195
2196 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2197 {
2198         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2199         struct cfq_group *cfqg;
2200
2201         if (RB_EMPTY_ROOT(&st->rb))
2202                 return NULL;
2203         cfqg = cfq_rb_first_group(st);
2204         update_min_vdisktime(st);
2205         return cfqg;
2206 }
2207
2208 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2209 {
2210         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2211
2212         cfqd->serving_group = cfqg;
2213
2214         /* Restore the workload type data */
2215         if (cfqg->saved_workload_slice) {
2216                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2217                 cfqd->serving_type = cfqg->saved_workload;
2218                 cfqd->serving_prio = cfqg->saved_serving_prio;
2219         } else
2220                 cfqd->workload_expires = jiffies - 1;
2221
2222         choose_service_tree(cfqd, cfqg);
2223 }
2224
2225 /*
2226  * Select a queue for service. If we have a current active queue,
2227  * check whether to continue servicing it, or retrieve and set a new one.
2228  */
2229 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2230 {
2231         struct cfq_queue *cfqq, *new_cfqq = NULL;
2232
2233         cfqq = cfqd->active_queue;
2234         if (!cfqq)
2235                 goto new_queue;
2236
2237         if (!cfqd->rq_queued)
2238                 return NULL;
2239
2240         /*
2241          * We were waiting for group to get backlogged. Expire the queue
2242          */
2243         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2244                 goto expire;
2245
2246         /*
2247          * The active queue has run out of time, expire it and select new.
2248          */
2249         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2250                 /*
2251                  * If slice had not expired at the completion of last request
2252                  * we might not have turned on wait_busy flag. Don't expire
2253                  * the queue yet. Allow the group to get backlogged.
2254                  *
2255                  * The very fact that we have used the slice, that means we
2256                  * have been idling all along on this queue and it should be
2257                  * ok to wait for this request to complete.
2258                  */
2259                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2260                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2261                         cfqq = NULL;
2262                         goto keep_queue;
2263                 } else
2264                         goto check_group_idle;
2265         }
2266
2267         /*
2268          * The active queue has requests and isn't expired, allow it to
2269          * dispatch.
2270          */
2271         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2272                 goto keep_queue;
2273
2274         /*
2275          * If another queue has a request waiting within our mean seek
2276          * distance, let it run.  The expire code will check for close
2277          * cooperators and put the close queue at the front of the service
2278          * tree.  If possible, merge the expiring queue with the new cfqq.
2279          */
2280         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2281         if (new_cfqq) {
2282                 if (!cfqq->new_cfqq)
2283                         cfq_setup_merge(cfqq, new_cfqq);
2284                 goto expire;
2285         }
2286
2287         /*
2288          * No requests pending. If the active queue still has requests in
2289          * flight or is idling for a new request, allow either of these
2290          * conditions to happen (or time out) before selecting a new queue.
2291          */
2292         if (timer_pending(&cfqd->idle_slice_timer)) {
2293                 cfqq = NULL;
2294                 goto keep_queue;
2295         }
2296
2297         /*
2298          * This is a deep seek queue, but the device is much faster than
2299          * the queue can deliver, don't idle
2300          **/
2301         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2302             (cfq_cfqq_slice_new(cfqq) ||
2303             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2304                 cfq_clear_cfqq_deep(cfqq);
2305                 cfq_clear_cfqq_idle_window(cfqq);
2306         }
2307
2308         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2309                 cfqq = NULL;
2310                 goto keep_queue;
2311         }
2312
2313         /*
2314          * If group idle is enabled and there are requests dispatched from
2315          * this group, wait for requests to complete.
2316          */
2317 check_group_idle:
2318         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2319             && cfqq->cfqg->dispatched) {
2320                 cfqq = NULL;
2321                 goto keep_queue;
2322         }
2323
2324 expire:
2325         cfq_slice_expired(cfqd, 0);
2326 new_queue:
2327         /*
2328          * Current queue expired. Check if we have to switch to a new
2329          * service tree
2330          */
2331         if (!new_cfqq)
2332                 cfq_choose_cfqg(cfqd);
2333
2334         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2335 keep_queue:
2336         return cfqq;
2337 }
2338
2339 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2340 {
2341         int dispatched = 0;
2342
2343         while (cfqq->next_rq) {
2344                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2345                 dispatched++;
2346         }
2347
2348         BUG_ON(!list_empty(&cfqq->fifo));
2349
2350         /* By default cfqq is not expired if it is empty. Do it explicitly */
2351         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2352         return dispatched;
2353 }
2354
2355 /*
2356  * Drain our current requests. Used for barriers and when switching
2357  * io schedulers on-the-fly.
2358  */
2359 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2360 {
2361         struct cfq_queue *cfqq;
2362         int dispatched = 0;
2363
2364         /* Expire the timeslice of the current active queue first */
2365         cfq_slice_expired(cfqd, 0);
2366         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2367                 __cfq_set_active_queue(cfqd, cfqq);
2368                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2369         }
2370
2371         BUG_ON(cfqd->busy_queues);
2372
2373         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2374         return dispatched;
2375 }
2376
2377 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2378         struct cfq_queue *cfqq)
2379 {
2380         /* the queue hasn't finished any request, can't estimate */
2381         if (cfq_cfqq_slice_new(cfqq))
2382                 return true;
2383         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2384                 cfqq->slice_end))
2385                 return true;
2386
2387         return false;
2388 }
2389
2390 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2391 {
2392         unsigned int max_dispatch;
2393
2394         /*
2395          * Drain async requests before we start sync IO
2396          */
2397         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2398                 return false;
2399
2400         /*
2401          * If this is an async queue and we have sync IO in flight, let it wait
2402          */
2403         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2404                 return false;
2405
2406         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2407         if (cfq_class_idle(cfqq))
2408                 max_dispatch = 1;
2409
2410         /*
2411          * Does this cfqq already have too much IO in flight?
2412          */
2413         if (cfqq->dispatched >= max_dispatch) {
2414                 bool promote_sync = false;
2415                 /*
2416                  * idle queue must always only have a single IO in flight
2417                  */
2418                 if (cfq_class_idle(cfqq))
2419                         return false;
2420
2421                 /*
2422                  * If there is only one sync queue
2423                  * we can ignore async queue here and give the sync
2424                  * queue no dispatch limit. The reason is a sync queue can
2425                  * preempt async queue, limiting the sync queue doesn't make
2426                  * sense. This is useful for aiostress test.
2427                  */
2428                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2429                         promote_sync = true;
2430
2431                 /*
2432                  * We have other queues, don't allow more IO from this one
2433                  */
2434                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2435                                 !promote_sync)
2436                         return false;
2437
2438                 /*
2439                  * Sole queue user, no limit
2440                  */
2441                 if (cfqd->busy_queues == 1 || promote_sync)
2442                         max_dispatch = -1;
2443                 else
2444                         /*
2445                          * Normally we start throttling cfqq when cfq_quantum/2
2446                          * requests have been dispatched. But we can drive
2447                          * deeper queue depths at the beginning of slice
2448                          * subjected to upper limit of cfq_quantum.
2449                          * */
2450                         max_dispatch = cfqd->cfq_quantum;
2451         }
2452
2453         /*
2454          * Async queues must wait a bit before being allowed dispatch.
2455          * We also ramp up the dispatch depth gradually for async IO,
2456          * based on the last sync IO we serviced
2457          */
2458         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2459                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2460                 unsigned int depth;
2461
2462                 depth = last_sync / cfqd->cfq_slice[1];
2463                 if (!depth && !cfqq->dispatched)
2464                         depth = 1;
2465                 if (depth < max_dispatch)
2466                         max_dispatch = depth;
2467         }
2468
2469         /*
2470          * If we're below the current max, allow a dispatch
2471          */
2472         return cfqq->dispatched < max_dispatch;
2473 }
2474
2475 /*
2476  * Dispatch a request from cfqq, moving them to the request queue
2477  * dispatch list.
2478  */
2479 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2480 {
2481         struct request *rq;
2482
2483         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2484
2485         if (!cfq_may_dispatch(cfqd, cfqq))
2486                 return false;
2487
2488         /*
2489          * follow expired path, else get first next available
2490          */
2491         rq = cfq_check_fifo(cfqq);
2492         if (!rq)
2493                 rq = cfqq->next_rq;
2494
2495         /*
2496          * insert request into driver dispatch list
2497          */
2498         cfq_dispatch_insert(cfqd->queue, rq);
2499
2500         if (!cfqd->active_cic) {
2501                 struct cfq_io_context *cic = RQ_CIC(rq);
2502
2503                 atomic_long_inc(&cic->ioc->refcount);
2504                 cfqd->active_cic = cic;
2505         }
2506
2507         return true;
2508 }
2509
2510 /*
2511  * Find the cfqq that we need to service and move a request from that to the
2512  * dispatch list
2513  */
2514 static int cfq_dispatch_requests(struct request_queue *q, int force)
2515 {
2516         struct cfq_data *cfqd = q->elevator->elevator_data;
2517         struct cfq_queue *cfqq;
2518
2519         if (!cfqd->busy_queues)
2520                 return 0;
2521
2522         if (unlikely(force))
2523                 return cfq_forced_dispatch(cfqd);
2524
2525         cfqq = cfq_select_queue(cfqd);
2526         if (!cfqq)
2527                 return 0;
2528
2529         /*
2530          * Dispatch a request from this cfqq, if it is allowed
2531          */
2532         if (!cfq_dispatch_request(cfqd, cfqq))
2533                 return 0;
2534
2535         cfqq->slice_dispatch++;
2536         cfq_clear_cfqq_must_dispatch(cfqq);
2537
2538         /*
2539          * expire an async queue immediately if it has used up its slice. idle
2540          * queue always expire after 1 dispatch round.
2541          */
2542         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2543             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2544             cfq_class_idle(cfqq))) {
2545                 cfqq->slice_end = jiffies + 1;
2546                 cfq_slice_expired(cfqd, 0);
2547         }
2548
2549         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2550         return 1;
2551 }
2552
2553 /*
2554  * task holds one reference to the queue, dropped when task exits. each rq
2555  * in-flight on this queue also holds a reference, dropped when rq is freed.
2556  *
2557  * Each cfq queue took a reference on the parent group. Drop it now.
2558  * queue lock must be held here.
2559  */
2560 static void cfq_put_queue(struct cfq_queue *cfqq)
2561 {
2562         struct cfq_data *cfqd = cfqq->cfqd;
2563         struct cfq_group *cfqg;
2564
2565         BUG_ON(cfqq->ref <= 0);
2566
2567         cfqq->ref--;
2568         if (cfqq->ref)
2569                 return;
2570
2571         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2572         BUG_ON(rb_first(&cfqq->sort_list));
2573         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2574         cfqg = cfqq->cfqg;
2575
2576         if (unlikely(cfqd->active_queue == cfqq)) {
2577                 __cfq_slice_expired(cfqd, cfqq, 0);
2578                 cfq_schedule_dispatch(cfqd);
2579         }
2580
2581         BUG_ON(cfq_cfqq_on_rr(cfqq));
2582         kmem_cache_free(cfq_pool, cfqq);
2583         cfq_put_cfqg(cfqg);
2584 }
2585
2586 /*
2587  * Call func for each cic attached to this ioc.
2588  */
2589 static void
2590 call_for_each_cic(struct io_context *ioc,
2591                   void (*func)(struct io_context *, struct cfq_io_context *))
2592 {
2593         struct cfq_io_context *cic;
2594         struct hlist_node *n;
2595
2596         rcu_read_lock();
2597
2598         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2599                 func(ioc, cic);
2600
2601         rcu_read_unlock();
2602 }
2603
2604 static void cfq_cic_free_rcu(struct rcu_head *head)
2605 {
2606         struct cfq_io_context *cic;
2607
2608         cic = container_of(head, struct cfq_io_context, rcu_head);
2609
2610         kmem_cache_free(cfq_ioc_pool, cic);
2611         elv_ioc_count_dec(cfq_ioc_count);
2612
2613         if (ioc_gone) {
2614                 /*
2615                  * CFQ scheduler is exiting, grab exit lock and check
2616                  * the pending io context count. If it hits zero,
2617                  * complete ioc_gone and set it back to NULL
2618                  */
2619                 spin_lock(&ioc_gone_lock);
2620                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2621                         complete(ioc_gone);
2622                         ioc_gone = NULL;
2623                 }
2624                 spin_unlock(&ioc_gone_lock);
2625         }
2626 }
2627
2628 static void cfq_cic_free(struct cfq_io_context *cic)
2629 {
2630         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2631 }
2632
2633 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2634 {
2635         unsigned long flags;
2636         unsigned long dead_key = (unsigned long) cic->key;
2637
2638         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2639
2640         spin_lock_irqsave(&ioc->lock, flags);
2641         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2642         hlist_del_rcu(&cic->cic_list);
2643         spin_unlock_irqrestore(&ioc->lock, flags);
2644
2645         cfq_cic_free(cic);
2646 }
2647
2648 /*
2649  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2650  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2651  * and ->trim() which is called with the task lock held
2652  */
2653 static void cfq_free_io_context(struct io_context *ioc)
2654 {
2655         /*
2656          * ioc->refcount is zero here, or we are called from elv_unregister(),
2657          * so no more cic's are allowed to be linked into this ioc.  So it
2658          * should be ok to iterate over the known list, we will see all cic's
2659          * since no new ones are added.
2660          */
2661         call_for_each_cic(ioc, cic_free_func);
2662 }
2663
2664 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2665 {
2666         struct cfq_queue *__cfqq, *next;
2667
2668         /*
2669          * If this queue was scheduled to merge with another queue, be
2670          * sure to drop the reference taken on that queue (and others in
2671          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2672          */
2673         __cfqq = cfqq->new_cfqq;
2674         while (__cfqq) {
2675                 if (__cfqq == cfqq) {
2676                         WARN(1, "cfqq->new_cfqq loop detected\n");
2677                         break;
2678                 }
2679                 next = __cfqq->new_cfqq;
2680                 cfq_put_queue(__cfqq);
2681                 __cfqq = next;
2682         }
2683 }
2684
2685 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2686 {
2687         if (unlikely(cfqq == cfqd->active_queue)) {
2688                 __cfq_slice_expired(cfqd, cfqq, 0);
2689                 cfq_schedule_dispatch(cfqd);
2690         }
2691
2692         cfq_put_cooperator(cfqq);
2693
2694         cfq_put_queue(cfqq);
2695 }
2696
2697 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2698                                          struct cfq_io_context *cic)
2699 {
2700         struct io_context *ioc = cic->ioc;
2701
2702         list_del_init(&cic->queue_list);
2703
2704         /*
2705          * Make sure dead mark is seen for dead queues
2706          */
2707         smp_wmb();
2708         cic->key = cfqd_dead_key(cfqd);
2709
2710         if (ioc->ioc_data == cic)
2711                 rcu_assign_pointer(ioc->ioc_data, NULL);
2712
2713         if (cic->cfqq[BLK_RW_ASYNC]) {
2714                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2715                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2716         }
2717
2718         if (cic->cfqq[BLK_RW_SYNC]) {
2719                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2720                 cic->cfqq[BLK_RW_SYNC] = NULL;
2721         }
2722 }
2723
2724 static void cfq_exit_single_io_context(struct io_context *ioc,
2725                                        struct cfq_io_context *cic)
2726 {
2727         struct cfq_data *cfqd = cic_to_cfqd(cic);
2728
2729         if (cfqd) {
2730                 struct request_queue *q = cfqd->queue;
2731                 unsigned long flags;
2732
2733                 spin_lock_irqsave(q->queue_lock, flags);
2734
2735                 /*
2736                  * Ensure we get a fresh copy of the ->key to prevent
2737                  * race between exiting task and queue
2738                  */
2739                 smp_read_barrier_depends();
2740                 if (cic->key == cfqd)
2741                         __cfq_exit_single_io_context(cfqd, cic);
2742
2743                 spin_unlock_irqrestore(q->queue_lock, flags);
2744         }
2745 }
2746
2747 /*
2748  * The process that ioc belongs to has exited, we need to clean up
2749  * and put the internal structures we have that belongs to that process.
2750  */
2751 static void cfq_exit_io_context(struct io_context *ioc)
2752 {
2753         call_for_each_cic(ioc, cfq_exit_single_io_context);
2754 }
2755
2756 static struct cfq_io_context *
2757 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2758 {
2759         struct cfq_io_context *cic;
2760
2761         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2762                                                         cfqd->queue->node);
2763         if (cic) {
2764                 cic->last_end_request = jiffies;
2765                 INIT_LIST_HEAD(&cic->queue_list);
2766                 INIT_HLIST_NODE(&cic->cic_list);
2767                 cic->dtor = cfq_free_io_context;
2768                 cic->exit = cfq_exit_io_context;
2769                 elv_ioc_count_inc(cfq_ioc_count);
2770         }
2771
2772         return cic;
2773 }
2774
2775 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2776 {
2777         struct task_struct *tsk = current;
2778         int ioprio_class;
2779
2780         if (!cfq_cfqq_prio_changed(cfqq))
2781                 return;
2782
2783         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2784         switch (ioprio_class) {
2785         default:
2786                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2787         case IOPRIO_CLASS_NONE:
2788                 /*
2789                  * no prio set, inherit CPU scheduling settings
2790                  */
2791                 cfqq->ioprio = task_nice_ioprio(tsk);
2792                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2793                 break;
2794         case IOPRIO_CLASS_RT:
2795                 cfqq->ioprio = task_ioprio(ioc);
2796                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2797                 break;
2798         case IOPRIO_CLASS_BE:
2799                 cfqq->ioprio = task_ioprio(ioc);
2800                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2801                 break;
2802         case IOPRIO_CLASS_IDLE:
2803                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2804                 cfqq->ioprio = 7;
2805                 cfq_clear_cfqq_idle_window(cfqq);
2806                 break;
2807         }
2808
2809         /*
2810          * keep track of original prio settings in case we have to temporarily
2811          * elevate the priority of this queue
2812          */
2813         cfqq->org_ioprio = cfqq->ioprio;
2814         cfqq->org_ioprio_class = cfqq->ioprio_class;
2815         cfq_clear_cfqq_prio_changed(cfqq);
2816 }
2817
2818 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2819 {
2820         struct cfq_data *cfqd = cic_to_cfqd(cic);
2821         struct cfq_queue *cfqq;
2822         unsigned long flags;
2823
2824         if (unlikely(!cfqd))
2825                 return;
2826
2827         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2828
2829         cfqq = cic->cfqq[BLK_RW_ASYNC];
2830         if (cfqq) {
2831                 struct cfq_queue *new_cfqq;
2832                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2833                                                 GFP_ATOMIC);
2834                 if (new_cfqq) {
2835                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2836                         cfq_put_queue(cfqq);
2837                 }
2838         }
2839
2840         cfqq = cic->cfqq[BLK_RW_SYNC];
2841         if (cfqq)
2842                 cfq_mark_cfqq_prio_changed(cfqq);
2843
2844         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2845 }
2846
2847 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2848 {
2849         call_for_each_cic(ioc, changed_ioprio);
2850         ioc->ioprio_changed = 0;
2851 }
2852
2853 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2854                           pid_t pid, bool is_sync)
2855 {
2856         RB_CLEAR_NODE(&cfqq->rb_node);
2857         RB_CLEAR_NODE(&cfqq->p_node);
2858         INIT_LIST_HEAD(&cfqq->fifo);
2859
2860         cfqq->ref = 0;
2861         cfqq->cfqd = cfqd;
2862
2863         cfq_mark_cfqq_prio_changed(cfqq);
2864
2865         if (is_sync) {
2866                 if (!cfq_class_idle(cfqq))
2867                         cfq_mark_cfqq_idle_window(cfqq);
2868                 cfq_mark_cfqq_sync(cfqq);
2869         }
2870         cfqq->pid = pid;
2871 }
2872
2873 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2874 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2875 {
2876         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2877         struct cfq_data *cfqd = cic_to_cfqd(cic);
2878         unsigned long flags;
2879         struct request_queue *q;
2880
2881         if (unlikely(!cfqd))
2882                 return;
2883
2884         q = cfqd->queue;
2885
2886         spin_lock_irqsave(q->queue_lock, flags);
2887
2888         if (sync_cfqq) {
2889                 /*
2890                  * Drop reference to sync queue. A new sync queue will be
2891                  * assigned in new group upon arrival of a fresh request.
2892                  */
2893                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2894                 cic_set_cfqq(cic, NULL, 1);
2895                 cfq_put_queue(sync_cfqq);
2896         }
2897
2898         spin_unlock_irqrestore(q->queue_lock, flags);
2899 }
2900
2901 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2902 {
2903         call_for_each_cic(ioc, changed_cgroup);
2904         ioc->cgroup_changed = 0;
2905 }
2906 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2907
2908 static struct cfq_queue *
2909 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2910                      struct io_context *ioc, gfp_t gfp_mask)
2911 {
2912         struct cfq_queue *cfqq, *new_cfqq = NULL;
2913         struct cfq_io_context *cic;
2914         struct cfq_group *cfqg;
2915
2916 retry:
2917         cfqg = cfq_get_cfqg(cfqd);
2918         cic = cfq_cic_lookup(cfqd, ioc);
2919         /* cic always exists here */
2920         cfqq = cic_to_cfqq(cic, is_sync);
2921
2922         /*
2923          * Always try a new alloc if we fell back to the OOM cfqq
2924          * originally, since it should just be a temporary situation.
2925          */
2926         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2927                 cfqq = NULL;
2928                 if (new_cfqq) {
2929                         cfqq = new_cfqq;
2930                         new_cfqq = NULL;
2931                 } else if (gfp_mask & __GFP_WAIT) {
2932                         spin_unlock_irq(cfqd->queue->queue_lock);
2933                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2934                                         gfp_mask | __GFP_ZERO,
2935                                         cfqd->queue->node);
2936                         spin_lock_irq(cfqd->queue->queue_lock);
2937                         if (new_cfqq)
2938                                 goto retry;
2939                 } else {
2940                         cfqq = kmem_cache_alloc_node(cfq_pool,
2941                                         gfp_mask | __GFP_ZERO,
2942                                         cfqd->queue->node);
2943                 }
2944
2945                 if (cfqq) {
2946                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2947                         cfq_init_prio_data(cfqq, ioc);
2948                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2949                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2950                 } else
2951                         cfqq = &cfqd->oom_cfqq;
2952         }
2953
2954         if (new_cfqq)
2955                 kmem_cache_free(cfq_pool, new_cfqq);
2956
2957         return cfqq;
2958 }
2959
2960 static struct cfq_queue **
2961 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2962 {
2963         switch (ioprio_class) {
2964         case IOPRIO_CLASS_RT:
2965                 return &cfqd->async_cfqq[0][ioprio];
2966         case IOPRIO_CLASS_BE:
2967                 return &cfqd->async_cfqq[1][ioprio];
2968         case IOPRIO_CLASS_IDLE:
2969                 return &cfqd->async_idle_cfqq;
2970         default:
2971                 BUG();
2972         }
2973 }
2974
2975 static struct cfq_queue *
2976 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2977               gfp_t gfp_mask)
2978 {
2979         const int ioprio = task_ioprio(ioc);
2980         const int ioprio_class = task_ioprio_class(ioc);
2981         struct cfq_queue **async_cfqq = NULL;
2982         struct cfq_queue *cfqq = NULL;
2983
2984         if (!is_sync) {
2985                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2986                 cfqq = *async_cfqq;
2987         }
2988
2989         if (!cfqq)
2990                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2991
2992         /*
2993          * pin the queue now that it's allocated, scheduler exit will prune it
2994          */
2995         if (!is_sync && !(*async_cfqq)) {
2996                 cfqq->ref++;
2997                 *async_cfqq = cfqq;
2998         }
2999
3000         cfqq->ref++;
3001         return cfqq;
3002 }
3003
3004 /*
3005  * We drop cfq io contexts lazily, so we may find a dead one.
3006  */
3007 static void
3008 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3009                   struct cfq_io_context *cic)
3010 {
3011         unsigned long flags;
3012
3013         WARN_ON(!list_empty(&cic->queue_list));
3014         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3015
3016         spin_lock_irqsave(&ioc->lock, flags);
3017
3018         BUG_ON(ioc->ioc_data == cic);
3019
3020         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3021         hlist_del_rcu(&cic->cic_list);
3022         spin_unlock_irqrestore(&ioc->lock, flags);
3023
3024         cfq_cic_free(cic);
3025 }
3026
3027 static struct cfq_io_context *
3028 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3029 {
3030         struct cfq_io_context *cic;
3031         unsigned long flags;
3032
3033         if (unlikely(!ioc))
3034                 return NULL;
3035
3036         rcu_read_lock();
3037
3038         /*
3039          * we maintain a last-hit cache, to avoid browsing over the tree
3040          */
3041         cic = rcu_dereference(ioc->ioc_data);
3042         if (cic && cic->key == cfqd) {
3043                 rcu_read_unlock();
3044                 return cic;
3045         }
3046
3047         do {
3048                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3049                 rcu_read_unlock();
3050                 if (!cic)
3051                         break;
3052                 if (unlikely(cic->key != cfqd)) {
3053                         cfq_drop_dead_cic(cfqd, ioc, cic);
3054                         rcu_read_lock();
3055                         continue;
3056                 }
3057
3058                 spin_lock_irqsave(&ioc->lock, flags);
3059                 rcu_assign_pointer(ioc->ioc_data, cic);
3060                 spin_unlock_irqrestore(&ioc->lock, flags);
3061                 break;
3062         } while (1);
3063
3064         return cic;
3065 }
3066
3067 /*
3068  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3069  * the process specific cfq io context when entered from the block layer.
3070  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3071  */
3072 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3073                         struct cfq_io_context *cic, gfp_t gfp_mask)
3074 {
3075         unsigned long flags;
3076         int ret;
3077
3078         ret = radix_tree_preload(gfp_mask);
3079         if (!ret) {
3080                 cic->ioc = ioc;
3081                 cic->key = cfqd;
3082
3083                 spin_lock_irqsave(&ioc->lock, flags);
3084                 ret = radix_tree_insert(&ioc->radix_root,
3085                                                 cfqd->cic_index, cic);
3086                 if (!ret)
3087                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3088                 spin_unlock_irqrestore(&ioc->lock, flags);
3089
3090                 radix_tree_preload_end();
3091
3092                 if (!ret) {
3093                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3094                         list_add(&cic->queue_list, &cfqd->cic_list);
3095                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3096                 }
3097         }
3098
3099         if (ret)
3100                 printk(KERN_ERR "cfq: cic link failed!\n");
3101
3102         return ret;
3103 }
3104
3105 /*
3106  * Setup general io context and cfq io context. There can be several cfq
3107  * io contexts per general io context, if this process is doing io to more
3108  * than one device managed by cfq.
3109  */
3110 static struct cfq_io_context *
3111 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3112 {
3113         struct io_context *ioc = NULL;
3114         struct cfq_io_context *cic;
3115
3116         might_sleep_if(gfp_mask & __GFP_WAIT);
3117
3118         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3119         if (!ioc)
3120                 return NULL;
3121
3122         cic = cfq_cic_lookup(cfqd, ioc);
3123         if (cic)
3124                 goto out;
3125
3126         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3127         if (cic == NULL)
3128                 goto err;
3129
3130         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3131                 goto err_free;
3132
3133 out:
3134         smp_read_barrier_depends();
3135         if (unlikely(ioc->ioprio_changed))
3136                 cfq_ioc_set_ioprio(ioc);
3137
3138 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3139         if (unlikely(ioc->cgroup_changed))
3140                 cfq_ioc_set_cgroup(ioc);
3141 #endif
3142         return cic;
3143 err_free:
3144         cfq_cic_free(cic);
3145 err:
3146         put_io_context(ioc);
3147         return NULL;
3148 }
3149
3150 static void
3151 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3152 {
3153         unsigned long elapsed = jiffies - cic->last_end_request;
3154         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3155
3156         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3157         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3158         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3159 }
3160
3161 static void
3162 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3163                        struct request *rq)
3164 {
3165         sector_t sdist = 0;
3166         sector_t n_sec = blk_rq_sectors(rq);
3167         if (cfqq->last_request_pos) {
3168                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3169                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3170                 else
3171                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3172         }
3173
3174         cfqq->seek_history <<= 1;
3175         if (blk_queue_nonrot(cfqd->queue))
3176                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3177         else
3178                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3179 }
3180
3181 /*
3182  * Disable idle window if the process thinks too long or seeks so much that
3183  * it doesn't matter
3184  */
3185 static void
3186 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3187                        struct cfq_io_context *cic)
3188 {
3189         int old_idle, enable_idle;
3190
3191         /*
3192          * Don't idle for async or idle io prio class
3193          */
3194         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3195                 return;
3196
3197         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3198
3199         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3200                 cfq_mark_cfqq_deep(cfqq);
3201
3202         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3203                 enable_idle = 0;
3204         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3205             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3206                 enable_idle = 0;
3207         else if (sample_valid(cic->ttime_samples)) {
3208                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3209                         enable_idle = 0;
3210                 else
3211                         enable_idle = 1;
3212         }
3213
3214         if (old_idle != enable_idle) {
3215                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3216                 if (enable_idle)
3217                         cfq_mark_cfqq_idle_window(cfqq);
3218                 else
3219                         cfq_clear_cfqq_idle_window(cfqq);
3220         }
3221 }
3222
3223 /*
3224  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3225  * no or if we aren't sure, a 1 will cause a preempt.
3226  */
3227 static bool
3228 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3229                    struct request *rq)
3230 {
3231         struct cfq_queue *cfqq;
3232
3233         cfqq = cfqd->active_queue;
3234         if (!cfqq)
3235                 return false;
3236
3237         if (cfq_class_idle(new_cfqq))
3238                 return false;
3239
3240         if (cfq_class_idle(cfqq))
3241                 return true;
3242
3243         /*
3244          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3245          */
3246         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3247                 return false;
3248
3249         /*
3250          * if the new request is sync, but the currently running queue is
3251          * not, let the sync request have priority.
3252          */
3253         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3254                 return true;
3255
3256         if (new_cfqq->cfqg != cfqq->cfqg)
3257                 return false;
3258
3259         if (cfq_slice_used(cfqq))
3260                 return true;
3261
3262         /* Allow preemption only if we are idling on sync-noidle tree */
3263         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3264             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3265             new_cfqq->service_tree->count == 2 &&
3266             RB_EMPTY_ROOT(&cfqq->sort_list))
3267                 return true;
3268
3269         /*
3270          * So both queues are sync. Let the new request get disk time if
3271          * it's a metadata request and the current queue is doing regular IO.
3272          */
3273         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3274                 return true;
3275
3276         /*
3277          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3278          */
3279         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3280                 return true;
3281
3282         /* An idle queue should not be idle now for some reason */
3283         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3284                 return true;
3285
3286         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3287                 return false;
3288
3289         /*
3290          * if this request is as-good as one we would expect from the
3291          * current cfqq, let it preempt
3292          */
3293         if (cfq_rq_close(cfqd, cfqq, rq))
3294                 return true;
3295
3296         return false;
3297 }
3298
3299 /*
3300  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3301  * let it have half of its nominal slice.
3302  */
3303 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3304 {
3305         struct cfq_queue *old_cfqq = cfqd->active_queue;
3306
3307         cfq_log_cfqq(cfqd, cfqq, "preempt");
3308         cfq_slice_expired(cfqd, 1);
3309
3310         /*
3311          * workload type is changed, don't save slice, otherwise preempt
3312          * doesn't happen
3313          */
3314         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3315                 cfqq->cfqg->saved_workload_slice = 0;
3316
3317         /*
3318          * Put the new queue at the front of the of the current list,
3319          * so we know that it will be selected next.
3320          */
3321         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3322
3323         cfq_service_tree_add(cfqd, cfqq, 1);
3324
3325         cfqq->slice_end = 0;
3326         cfq_mark_cfqq_slice_new(cfqq);
3327 }
3328
3329 /*
3330  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3331  * something we should do about it
3332  */
3333 static void
3334 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3335                 struct request *rq)
3336 {
3337         struct cfq_io_context *cic = RQ_CIC(rq);
3338
3339         cfqd->rq_queued++;
3340         if (rq->cmd_flags & REQ_META)
3341                 cfqq->meta_pending++;
3342
3343         cfq_update_io_thinktime(cfqd, cic);
3344         cfq_update_io_seektime(cfqd, cfqq, rq);
3345         cfq_update_idle_window(cfqd, cfqq, cic);
3346
3347         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3348
3349         if (cfqq == cfqd->active_queue) {
3350                 /*
3351                  * Remember that we saw a request from this process, but
3352                  * don't start queuing just yet. Otherwise we risk seeing lots
3353                  * of tiny requests, because we disrupt the normal plugging
3354                  * and merging. If the request is already larger than a single
3355                  * page, let it rip immediately. For that case we assume that
3356                  * merging is already done. Ditto for a busy system that
3357                  * has other work pending, don't risk delaying until the
3358                  * idle timer unplug to continue working.
3359                  */
3360                 if (cfq_cfqq_wait_request(cfqq)) {
3361                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3362                             cfqd->busy_queues > 1) {
3363                                 cfq_del_timer(cfqd, cfqq);
3364                                 cfq_clear_cfqq_wait_request(cfqq);
3365                                 __blk_run_queue(cfqd->queue);
3366                         } else {
3367                                 cfq_blkiocg_update_idle_time_stats(
3368                                                 &cfqq->cfqg->blkg);
3369                                 cfq_mark_cfqq_must_dispatch(cfqq);
3370                         }
3371                 }
3372         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3373                 /*
3374                  * not the active queue - expire current slice if it is
3375                  * idle and has expired it's mean thinktime or this new queue
3376                  * has some old slice time left and is of higher priority or
3377                  * this new queue is RT and the current one is BE
3378                  */
3379                 cfq_preempt_queue(cfqd, cfqq);
3380                 __blk_run_queue(cfqd->queue);
3381         }
3382 }
3383
3384 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3385 {
3386         struct cfq_data *cfqd = q->elevator->elevator_data;
3387         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3388
3389         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3390         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3391
3392         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3393         list_add_tail(&rq->queuelist, &cfqq->fifo);
3394         cfq_add_rq_rb(rq);
3395         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3396                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3397                         rq_is_sync(rq));
3398         cfq_rq_enqueued(cfqd, cfqq, rq);
3399 }
3400
3401 /*
3402  * Update hw_tag based on peak queue depth over 50 samples under
3403  * sufficient load.
3404  */
3405 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3406 {
3407         struct cfq_queue *cfqq = cfqd->active_queue;
3408
3409         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3410                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3411
3412         if (cfqd->hw_tag == 1)
3413                 return;
3414
3415         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3416             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3417                 return;
3418
3419         /*
3420          * If active queue hasn't enough requests and can idle, cfq might not
3421          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3422          * case
3423          */
3424         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3425             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3426             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3427                 return;
3428
3429         if (cfqd->hw_tag_samples++ < 50)
3430                 return;
3431
3432         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3433                 cfqd->hw_tag = 1;
3434         else
3435                 cfqd->hw_tag = 0;
3436 }
3437
3438 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3439 {
3440         struct cfq_io_context *cic = cfqd->active_cic;
3441
3442         /* If the queue already has requests, don't wait */
3443         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3444                 return false;
3445
3446         /* If there are other queues in the group, don't wait */
3447         if (cfqq->cfqg->nr_cfqq > 1)
3448                 return false;
3449
3450         if (cfq_slice_used(cfqq))
3451                 return true;
3452
3453         /* if slice left is less than think time, wait busy */
3454         if (cic && sample_valid(cic->ttime_samples)
3455             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3456                 return true;
3457
3458         /*
3459          * If think times is less than a jiffy than ttime_mean=0 and above
3460          * will not be true. It might happen that slice has not expired yet
3461          * but will expire soon (4-5 ns) during select_queue(). To cover the
3462          * case where think time is less than a jiffy, mark the queue wait
3463          * busy if only 1 jiffy is left in the slice.
3464          */
3465         if (cfqq->slice_end - jiffies == 1)
3466                 return true;
3467
3468         return false;
3469 }
3470
3471 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3472 {
3473         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3474         struct cfq_data *cfqd = cfqq->cfqd;
3475         const int sync = rq_is_sync(rq);
3476         unsigned long now;
3477
3478         now = jiffies;
3479         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3480                      !!(rq->cmd_flags & REQ_NOIDLE));
3481
3482         cfq_update_hw_tag(cfqd);
3483
3484         WARN_ON(!cfqd->rq_in_driver);
3485         WARN_ON(!cfqq->dispatched);
3486         cfqd->rq_in_driver--;
3487         cfqq->dispatched--;
3488         (RQ_CFQG(rq))->dispatched--;
3489         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3490                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3491                         rq_data_dir(rq), rq_is_sync(rq));
3492
3493         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3494
3495         if (sync) {
3496                 RQ_CIC(rq)->last_end_request = now;
3497                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3498                         cfqd->last_delayed_sync = now;
3499         }
3500
3501         /*
3502          * If this is the active queue, check if it needs to be expired,
3503          * or if we want to idle in case it has no pending requests.
3504          */
3505         if (cfqd->active_queue == cfqq) {
3506                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3507
3508                 if (cfq_cfqq_slice_new(cfqq)) {
3509                         cfq_set_prio_slice(cfqd, cfqq);
3510                         cfq_clear_cfqq_slice_new(cfqq);
3511                 }
3512
3513                 /*
3514                  * Should we wait for next request to come in before we expire
3515                  * the queue.
3516                  */
3517                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3518                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3519                         if (!cfqd->cfq_slice_idle)
3520                                 extend_sl = cfqd->cfq_group_idle;
3521                         cfqq->slice_end = jiffies + extend_sl;
3522                         cfq_mark_cfqq_wait_busy(cfqq);
3523                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3524                 }
3525
3526                 /*
3527                  * Idling is not enabled on:
3528                  * - expired queues
3529                  * - idle-priority queues
3530                  * - async queues
3531                  * - queues with still some requests queued
3532                  * - when there is a close cooperator
3533                  */
3534                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3535                         cfq_slice_expired(cfqd, 1);
3536                 else if (sync && cfqq_empty &&
3537                          !cfq_close_cooperator(cfqd, cfqq)) {
3538                         cfq_arm_slice_timer(cfqd);
3539                 }
3540         }
3541
3542         if (!cfqd->rq_in_driver)
3543                 cfq_schedule_dispatch(cfqd);
3544 }
3545
3546 /*
3547  * we temporarily boost lower priority queues if they are holding fs exclusive
3548  * resources. they are boosted to normal prio (CLASS_BE/4)
3549  */
3550 static void cfq_prio_boost(struct cfq_queue *cfqq)
3551 {
3552         if (has_fs_excl()) {
3553                 /*
3554                  * boost idle prio on transactions that would lock out other
3555                  * users of the filesystem
3556                  */
3557                 if (cfq_class_idle(cfqq))
3558                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3559                 if (cfqq->ioprio > IOPRIO_NORM)
3560                         cfqq->ioprio = IOPRIO_NORM;
3561         } else {
3562                 /*
3563                  * unboost the queue (if needed)
3564                  */
3565                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3566                 cfqq->ioprio = cfqq->org_ioprio;
3567         }
3568 }
3569
3570 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3571 {
3572         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3573                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3574                 return ELV_MQUEUE_MUST;
3575         }
3576
3577         return ELV_MQUEUE_MAY;
3578 }
3579
3580 static int cfq_may_queue(struct request_queue *q, int rw)
3581 {
3582         struct cfq_data *cfqd = q->elevator->elevator_data;
3583         struct task_struct *tsk = current;
3584         struct cfq_io_context *cic;
3585         struct cfq_queue *cfqq;
3586
3587         /*
3588          * don't force setup of a queue from here, as a call to may_queue
3589          * does not necessarily imply that a request actually will be queued.
3590          * so just lookup a possibly existing queue, or return 'may queue'
3591          * if that fails
3592          */
3593         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3594         if (!cic)
3595                 return ELV_MQUEUE_MAY;
3596
3597         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3598         if (cfqq) {
3599                 cfq_init_prio_data(cfqq, cic->ioc);
3600                 cfq_prio_boost(cfqq);
3601
3602                 return __cfq_may_queue(cfqq);
3603         }
3604
3605         return ELV_MQUEUE_MAY;
3606 }
3607
3608 /*
3609  * queue lock held here
3610  */
3611 static void cfq_put_request(struct request *rq)
3612 {
3613         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3614
3615         if (cfqq) {
3616                 const int rw = rq_data_dir(rq);
3617
3618                 BUG_ON(!cfqq->allocated[rw]);
3619                 cfqq->allocated[rw]--;
3620
3621                 put_io_context(RQ_CIC(rq)->ioc);
3622
3623                 rq->elevator_private[0] = NULL;
3624                 rq->elevator_private[1] = NULL;
3625
3626                 /* Put down rq reference on cfqg */
3627                 cfq_put_cfqg(RQ_CFQG(rq));
3628                 rq->elevator_private[2] = NULL;
3629
3630                 cfq_put_queue(cfqq);
3631         }
3632 }
3633
3634 static struct cfq_queue *
3635 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3636                 struct cfq_queue *cfqq)
3637 {
3638         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3639         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3640         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3641         cfq_put_queue(cfqq);
3642         return cic_to_cfqq(cic, 1);
3643 }
3644
3645 /*
3646  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3647  * was the last process referring to said cfqq.
3648  */
3649 static struct cfq_queue *
3650 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3651 {
3652         if (cfqq_process_refs(cfqq) == 1) {
3653                 cfqq->pid = current->pid;
3654                 cfq_clear_cfqq_coop(cfqq);
3655                 cfq_clear_cfqq_split_coop(cfqq);
3656                 return cfqq;
3657         }
3658
3659         cic_set_cfqq(cic, NULL, 1);
3660
3661         cfq_put_cooperator(cfqq);
3662
3663         cfq_put_queue(cfqq);
3664         return NULL;
3665 }
3666 /*
3667  * Allocate cfq data structures associated with this request.
3668  */
3669 static int
3670 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3671 {
3672         struct cfq_data *cfqd = q->elevator->elevator_data;
3673         struct cfq_io_context *cic;
3674         const int rw = rq_data_dir(rq);
3675         const bool is_sync = rq_is_sync(rq);
3676         struct cfq_queue *cfqq;
3677         unsigned long flags;
3678
3679         might_sleep_if(gfp_mask & __GFP_WAIT);
3680
3681         cic = cfq_get_io_context(cfqd, gfp_mask);
3682
3683         spin_lock_irqsave(q->queue_lock, flags);
3684
3685         if (!cic)
3686                 goto queue_fail;
3687
3688 new_queue:
3689         cfqq = cic_to_cfqq(cic, is_sync);
3690         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3691                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3692                 cic_set_cfqq(cic, cfqq, is_sync);
3693         } else {
3694                 /*
3695                  * If the queue was seeky for too long, break it apart.
3696                  */
3697                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3698                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3699                         cfqq = split_cfqq(cic, cfqq);
3700                         if (!cfqq)
3701                                 goto new_queue;
3702                 }
3703
3704                 /*
3705                  * Check to see if this queue is scheduled to merge with
3706                  * another, closely cooperating queue.  The merging of
3707                  * queues happens here as it must be done in process context.
3708                  * The reference on new_cfqq was taken in merge_cfqqs.
3709                  */
3710                 if (cfqq->new_cfqq)
3711                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3712         }
3713
3714         cfqq->allocated[rw]++;
3715
3716         cfqq->ref++;
3717         rq->elevator_private[0] = cic;
3718         rq->elevator_private[1] = cfqq;
3719         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3720         spin_unlock_irqrestore(q->queue_lock, flags);
3721         return 0;
3722
3723 queue_fail:
3724         if (cic)
3725                 put_io_context(cic->ioc);
3726
3727         cfq_schedule_dispatch(cfqd);
3728         spin_unlock_irqrestore(q->queue_lock, flags);
3729         cfq_log(cfqd, "set_request fail");
3730         return 1;
3731 }
3732
3733 static void cfq_kick_queue(struct work_struct *work)
3734 {
3735         struct cfq_data *cfqd =
3736                 container_of(work, struct cfq_data, unplug_work);
3737         struct request_queue *q = cfqd->queue;
3738
3739         spin_lock_irq(q->queue_lock);
3740         __blk_run_queue(cfqd->queue);
3741         spin_unlock_irq(q->queue_lock);
3742 }
3743
3744 /*
3745  * Timer running if the active_queue is currently idling inside its time slice
3746  */
3747 static void cfq_idle_slice_timer(unsigned long data)
3748 {
3749         struct cfq_data *cfqd = (struct cfq_data *) data;
3750         struct cfq_queue *cfqq;
3751         unsigned long flags;
3752         int timed_out = 1;
3753
3754         cfq_log(cfqd, "idle timer fired");
3755
3756         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3757
3758         cfqq = cfqd->active_queue;
3759         if (cfqq) {
3760                 timed_out = 0;
3761
3762                 /*
3763                  * We saw a request before the queue expired, let it through
3764                  */
3765                 if (cfq_cfqq_must_dispatch(cfqq))
3766                         goto out_kick;
3767
3768                 /*
3769                  * expired
3770                  */
3771                 if (cfq_slice_used(cfqq))
3772                         goto expire;
3773
3774                 /*
3775                  * only expire and reinvoke request handler, if there are
3776                  * other queues with pending requests
3777                  */
3778                 if (!cfqd->busy_queues)
3779                         goto out_cont;
3780
3781                 /*
3782                  * not expired and it has a request pending, let it dispatch
3783                  */
3784                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3785                         goto out_kick;
3786
3787                 /*
3788                  * Queue depth flag is reset only when the idle didn't succeed
3789                  */
3790                 cfq_clear_cfqq_deep(cfqq);
3791         }
3792 expire:
3793         cfq_slice_expired(cfqd, timed_out);
3794 out_kick:
3795         cfq_schedule_dispatch(cfqd);
3796 out_cont:
3797         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3798 }
3799
3800 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3801 {
3802         del_timer_sync(&cfqd->idle_slice_timer);
3803         cancel_work_sync(&cfqd->unplug_work);
3804 }
3805
3806 static void cfq_put_async_queues(struct cfq_data *cfqd)
3807 {
3808         int i;
3809
3810         for (i = 0; i < IOPRIO_BE_NR; i++) {
3811                 if (cfqd->async_cfqq[0][i])
3812                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3813                 if (cfqd->async_cfqq[1][i])
3814                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3815         }
3816
3817         if (cfqd->async_idle_cfqq)
3818                 cfq_put_queue(cfqd->async_idle_cfqq);
3819 }
3820
3821 static void cfq_exit_queue(struct elevator_queue *e)
3822 {
3823         struct cfq_data *cfqd = e->elevator_data;
3824         struct request_queue *q = cfqd->queue;
3825         bool wait = false;
3826
3827         cfq_shutdown_timer_wq(cfqd);
3828
3829         spin_lock_irq(q->queue_lock);
3830
3831         if (cfqd->active_queue)
3832                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3833
3834         while (!list_empty(&cfqd->cic_list)) {
3835                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3836                                                         struct cfq_io_context,
3837                                                         queue_list);
3838
3839                 __cfq_exit_single_io_context(cfqd, cic);
3840         }
3841
3842         cfq_put_async_queues(cfqd);
3843         cfq_release_cfq_groups(cfqd);
3844
3845         /*
3846          * If there are groups which we could not unlink from blkcg list,
3847          * wait for a rcu period for them to be freed.
3848          */
3849         if (cfqd->nr_blkcg_linked_grps)
3850                 wait = true;
3851
3852         spin_unlock_irq(q->queue_lock);
3853
3854         cfq_shutdown_timer_wq(cfqd);
3855
3856         spin_lock(&cic_index_lock);
3857         ida_remove(&cic_index_ida, cfqd->cic_index);
3858         spin_unlock(&cic_index_lock);
3859
3860         /*
3861          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3862          * Do this wait only if there are other unlinked groups out
3863          * there. This can happen if cgroup deletion path claimed the
3864          * responsibility of cleaning up a group before queue cleanup code
3865          * get to the group.
3866          *
3867          * Do not call synchronize_rcu() unconditionally as there are drivers
3868          * which create/delete request queue hundreds of times during scan/boot
3869          * and synchronize_rcu() can take significant time and slow down boot.
3870          */
3871         if (wait)
3872                 synchronize_rcu();
3873         kfree(cfqd);
3874 }
3875
3876 static int cfq_alloc_cic_index(void)
3877 {
3878         int index, error;
3879
3880         do {
3881                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3882                         return -ENOMEM;
3883
3884                 spin_lock(&cic_index_lock);
3885                 error = ida_get_new(&cic_index_ida, &index);
3886                 spin_unlock(&cic_index_lock);
3887                 if (error && error != -EAGAIN)
3888                         return error;
3889         } while (error);
3890
3891         return index;
3892 }
3893
3894 static void *cfq_init_queue(struct request_queue *q)
3895 {
3896         struct cfq_data *cfqd;
3897         int i, j;
3898         struct cfq_group *cfqg;
3899         struct cfq_rb_root *st;
3900
3901         i = cfq_alloc_cic_index();
3902         if (i < 0)
3903                 return NULL;
3904
3905         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3906         if (!cfqd)
3907                 return NULL;
3908
3909         /*
3910          * Don't need take queue_lock in the routine, since we are
3911          * initializing the ioscheduler, and nobody is using cfqd
3912          */
3913         cfqd->cic_index = i;
3914
3915         /* Init root service tree */
3916         cfqd->grp_service_tree = CFQ_RB_ROOT;
3917
3918         /* Init root group */
3919         cfqg = &cfqd->root_group;
3920         for_each_cfqg_st(cfqg, i, j, st)
3921                 *st = CFQ_RB_ROOT;
3922         RB_CLEAR_NODE(&cfqg->rb_node);
3923
3924         /* Give preference to root group over other groups */
3925         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3926
3927 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3928         /*
3929          * Set root group reference to 2. One reference will be dropped when
3930          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3931          * Other reference will remain there as we don't want to delete this
3932          * group as it is statically allocated and gets destroyed when
3933          * throtl_data goes away.
3934          */
3935         cfqg->ref = 2;
3936         rcu_read_lock();
3937         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3938                                         (void *)cfqd, 0);
3939         rcu_read_unlock();
3940         cfqd->nr_blkcg_linked_grps++;
3941
3942         /* Add group on cfqd->cfqg_list */
3943         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3944 #endif
3945         /*
3946          * Not strictly needed (since RB_ROOT just clears the node and we
3947          * zeroed cfqd on alloc), but better be safe in case someone decides
3948          * to add magic to the rb code
3949          */
3950         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3951                 cfqd->prio_trees[i] = RB_ROOT;
3952
3953         /*
3954          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3955          * Grab a permanent reference to it, so that the normal code flow
3956          * will not attempt to free it.
3957          */
3958         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3959         cfqd->oom_cfqq.ref++;
3960         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3961
3962         INIT_LIST_HEAD(&cfqd->cic_list);
3963
3964         cfqd->queue = q;
3965
3966         init_timer(&cfqd->idle_slice_timer);
3967         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3968         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3969
3970         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3971
3972         cfqd->cfq_quantum = cfq_quantum;
3973         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3974         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3975         cfqd->cfq_back_max = cfq_back_max;
3976         cfqd->cfq_back_penalty = cfq_back_penalty;
3977         cfqd->cfq_slice[0] = cfq_slice_async;
3978         cfqd->cfq_slice[1] = cfq_slice_sync;
3979         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3980         cfqd->cfq_slice_idle = cfq_slice_idle;
3981         cfqd->cfq_group_idle = cfq_group_idle;
3982         cfqd->cfq_latency = 1;
3983         cfqd->hw_tag = -1;
3984         /*
3985          * we optimistically start assuming sync ops weren't delayed in last
3986          * second, in order to have larger depth for async operations.
3987          */
3988         cfqd->last_delayed_sync = jiffies - HZ;
3989         return cfqd;
3990 }
3991
3992 static void cfq_slab_kill(void)
3993 {
3994         /*
3995          * Caller already ensured that pending RCU callbacks are completed,
3996          * so we should have no busy allocations at this point.
3997          */
3998         if (cfq_pool)
3999                 kmem_cache_destroy(cfq_pool);
4000         if (cfq_ioc_pool)
4001                 kmem_cache_destroy(cfq_ioc_pool);
4002 }
4003
4004 static int __init cfq_slab_setup(void)
4005 {
4006         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4007         if (!cfq_pool)
4008                 goto fail;
4009
4010         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4011         if (!cfq_ioc_pool)
4012                 goto fail;
4013
4014         return 0;
4015 fail:
4016         cfq_slab_kill();
4017         return -ENOMEM;
4018 }
4019
4020 /*
4021  * sysfs parts below -->
4022  */
4023 static ssize_t
4024 cfq_var_show(unsigned int var, char *page)
4025 {
4026         return sprintf(page, "%d\n", var);
4027 }
4028
4029 static ssize_t
4030 cfq_var_store(unsigned int *var, const char *page, size_t count)
4031 {
4032         char *p = (char *) page;
4033
4034         *var = simple_strtoul(p, &p, 10);
4035         return count;
4036 }
4037
4038 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4039 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4040 {                                                                       \
4041         struct cfq_data *cfqd = e->elevator_data;                       \
4042         unsigned int __data = __VAR;                                    \
4043         if (__CONV)                                                     \
4044                 __data = jiffies_to_msecs(__data);                      \
4045         return cfq_var_show(__data, (page));                            \
4046 }
4047 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4048 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4049 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4050 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4051 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4052 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4053 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4054 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4055 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4056 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4057 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4058 #undef SHOW_FUNCTION
4059
4060 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4061 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4062 {                                                                       \
4063         struct cfq_data *cfqd = e->elevator_data;                       \
4064         unsigned int __data;                                            \
4065         int ret = cfq_var_store(&__data, (page), count);                \
4066         if (__data < (MIN))                                             \
4067                 __data = (MIN);                                         \
4068         else if (__data > (MAX))                                        \
4069                 __data = (MAX);                                         \
4070         if (__CONV)                                                     \
4071                 *(__PTR) = msecs_to_jiffies(__data);                    \
4072         else                                                            \
4073                 *(__PTR) = __data;                                      \
4074         return ret;                                                     \
4075 }
4076 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4077 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4078                 UINT_MAX, 1);
4079 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4080                 UINT_MAX, 1);
4081 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4082 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4083                 UINT_MAX, 0);
4084 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4085 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4086 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4087 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4088 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4089                 UINT_MAX, 0);
4090 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4091 #undef STORE_FUNCTION
4092
4093 #define CFQ_ATTR(name) \
4094         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4095
4096 static struct elv_fs_entry cfq_attrs[] = {
4097         CFQ_ATTR(quantum),
4098         CFQ_ATTR(fifo_expire_sync),
4099         CFQ_ATTR(fifo_expire_async),
4100         CFQ_ATTR(back_seek_max),
4101         CFQ_ATTR(back_seek_penalty),
4102         CFQ_ATTR(slice_sync),
4103         CFQ_ATTR(slice_async),
4104         CFQ_ATTR(slice_async_rq),
4105         CFQ_ATTR(slice_idle),
4106         CFQ_ATTR(group_idle),
4107         CFQ_ATTR(low_latency),
4108         __ATTR_NULL
4109 };
4110
4111 static struct elevator_type iosched_cfq = {
4112         .ops = {
4113                 .elevator_merge_fn =            cfq_merge,
4114                 .elevator_merged_fn =           cfq_merged_request,
4115                 .elevator_merge_req_fn =        cfq_merged_requests,
4116                 .elevator_allow_merge_fn =      cfq_allow_merge,
4117                 .elevator_bio_merged_fn =       cfq_bio_merged,
4118                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4119                 .elevator_add_req_fn =          cfq_insert_request,
4120                 .elevator_activate_req_fn =     cfq_activate_request,
4121                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4122                 .elevator_completed_req_fn =    cfq_completed_request,
4123                 .elevator_former_req_fn =       elv_rb_former_request,
4124                 .elevator_latter_req_fn =       elv_rb_latter_request,
4125                 .elevator_set_req_fn =          cfq_set_request,
4126                 .elevator_put_req_fn =          cfq_put_request,
4127                 .elevator_may_queue_fn =        cfq_may_queue,
4128                 .elevator_init_fn =             cfq_init_queue,
4129                 .elevator_exit_fn =             cfq_exit_queue,
4130                 .trim =                         cfq_free_io_context,
4131         },
4132         .elevator_attrs =       cfq_attrs,
4133         .elevator_name =        "cfq",
4134         .elevator_owner =       THIS_MODULE,
4135 };
4136
4137 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4138 static struct blkio_policy_type blkio_policy_cfq = {
4139         .ops = {
4140                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4141                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4142         },
4143         .plid = BLKIO_POLICY_PROP,
4144 };
4145 #else
4146 static struct blkio_policy_type blkio_policy_cfq;
4147 #endif
4148
4149 static int __init cfq_init(void)
4150 {
4151         /*
4152          * could be 0 on HZ < 1000 setups
4153          */
4154         if (!cfq_slice_async)
4155                 cfq_slice_async = 1;
4156         if (!cfq_slice_idle)
4157                 cfq_slice_idle = 1;
4158
4159 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4160         if (!cfq_group_idle)
4161                 cfq_group_idle = 1;
4162 #else
4163                 cfq_group_idle = 0;
4164 #endif
4165         if (cfq_slab_setup())
4166                 return -ENOMEM;
4167
4168         elv_register(&iosched_cfq);
4169         blkio_policy_register(&blkio_policy_cfq);
4170
4171         return 0;
4172 }
4173
4174 static void __exit cfq_exit(void)
4175 {
4176         DECLARE_COMPLETION_ONSTACK(all_gone);
4177         blkio_policy_unregister(&blkio_policy_cfq);
4178         elv_unregister(&iosched_cfq);
4179         ioc_gone = &all_gone;
4180         /* ioc_gone's update must be visible before reading ioc_count */
4181         smp_wmb();
4182
4183         /*
4184          * this also protects us from entering cfq_slab_kill() with
4185          * pending RCU callbacks
4186          */
4187         if (elv_ioc_count_read(cfq_ioc_count))
4188                 wait_for_completion(&all_gone);
4189         ida_destroy(&cic_index_ida);
4190         cfq_slab_kill();
4191 }
4192
4193 module_init(cfq_init);
4194 module_exit(cfq_exit);
4195
4196 MODULE_AUTHOR("Jens Axboe");
4197 MODULE_LICENSE("GPL");
4198 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");