cfq-iosched: reduce bit operations in cfq_choose_req()
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[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 rq_is_sync(rq1) ? rq1 : rq2;
672
673         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
674                 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
675
676         s1 = blk_rq_pos(rq1);
677         s2 = blk_rq_pos(rq2);
678
679         /*
680          * by definition, 1KiB is 2 sectors
681          */
682         back_max = cfqd->cfq_back_max * 2;
683
684         /*
685          * Strict one way elevator _except_ in the case where we allow
686          * short backward seeks which are biased as twice the cost of a
687          * similar forward seek.
688          */
689         if (s1 >= last)
690                 d1 = s1 - last;
691         else if (s1 + back_max >= last)
692                 d1 = (last - s1) * cfqd->cfq_back_penalty;
693         else
694                 wrap |= CFQ_RQ1_WRAP;
695
696         if (s2 >= last)
697                 d2 = s2 - last;
698         else if (s2 + back_max >= last)
699                 d2 = (last - s2) * cfqd->cfq_back_penalty;
700         else
701                 wrap |= CFQ_RQ2_WRAP;
702
703         /* Found required data */
704
705         /*
706          * By doing switch() on the bit mask "wrap" we avoid having to
707          * check two variables for all permutations: --> faster!
708          */
709         switch (wrap) {
710         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
711                 if (d1 < d2)
712                         return rq1;
713                 else if (d2 < d1)
714                         return rq2;
715                 else {
716                         if (s1 >= s2)
717                                 return rq1;
718                         else
719                                 return rq2;
720                 }
721
722         case CFQ_RQ2_WRAP:
723                 return rq1;
724         case CFQ_RQ1_WRAP:
725                 return rq2;
726         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727         default:
728                 /*
729                  * Since both rqs are wrapped,
730                  * start with the one that's further behind head
731                  * (--> only *one* back seek required),
732                  * since back seek takes more time than forward.
733                  */
734                 if (s1 <= s2)
735                         return rq1;
736                 else
737                         return rq2;
738         }
739 }
740
741 /*
742  * The below is leftmost cache rbtree addon
743  */
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
745 {
746         /* Service tree is empty */
747         if (!root->count)
748                 return NULL;
749
750         if (!root->left)
751                 root->left = rb_first(&root->rb);
752
753         if (root->left)
754                 return rb_entry(root->left, struct cfq_queue, rb_node);
755
756         return NULL;
757 }
758
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760 {
761         if (!root->left)
762                 root->left = rb_first(&root->rb);
763
764         if (root->left)
765                 return rb_entry_cfqg(root->left);
766
767         return NULL;
768 }
769
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
771 {
772         rb_erase(n, root);
773         RB_CLEAR_NODE(n);
774 }
775
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
777 {
778         if (root->left == n)
779                 root->left = NULL;
780         rb_erase_init(n, &root->rb);
781         --root->count;
782 }
783
784 /*
785  * would be nice to take fifo expire time into account as well
786  */
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789                   struct request *last)
790 {
791         struct rb_node *rbnext = rb_next(&last->rb_node);
792         struct rb_node *rbprev = rb_prev(&last->rb_node);
793         struct request *next = NULL, *prev = NULL;
794
795         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
796
797         if (rbprev)
798                 prev = rb_entry_rq(rbprev);
799
800         if (rbnext)
801                 next = rb_entry_rq(rbnext);
802         else {
803                 rbnext = rb_first(&cfqq->sort_list);
804                 if (rbnext && rbnext != &last->rb_node)
805                         next = rb_entry_rq(rbnext);
806         }
807
808         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
809 }
810
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812                                       struct cfq_queue *cfqq)
813 {
814         /*
815          * just an approximation, should be ok.
816          */
817         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
819 }
820
821 static inline s64
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
823 {
824         return cfqg->vdisktime - st->min_vdisktime;
825 }
826
827 static void
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
829 {
830         struct rb_node **node = &st->rb.rb_node;
831         struct rb_node *parent = NULL;
832         struct cfq_group *__cfqg;
833         s64 key = cfqg_key(st, cfqg);
834         int left = 1;
835
836         while (*node != NULL) {
837                 parent = *node;
838                 __cfqg = rb_entry_cfqg(parent);
839
840                 if (key < cfqg_key(st, __cfqg))
841                         node = &parent->rb_left;
842                 else {
843                         node = &parent->rb_right;
844                         left = 0;
845                 }
846         }
847
848         if (left)
849                 st->left = &cfqg->rb_node;
850
851         rb_link_node(&cfqg->rb_node, parent, node);
852         rb_insert_color(&cfqg->rb_node, &st->rb);
853 }
854
855 static void
856 cfq_update_group_weight(struct cfq_group *cfqg)
857 {
858         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
859         if (cfqg->needs_update) {
860                 cfqg->weight = cfqg->new_weight;
861                 cfqg->needs_update = false;
862         }
863 }
864
865 static void
866 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
867 {
868         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
869
870         cfq_update_group_weight(cfqg);
871         __cfq_group_service_tree_add(st, cfqg);
872         st->total_weight += cfqg->weight;
873 }
874
875 static void
876 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
877 {
878         struct cfq_rb_root *st = &cfqd->grp_service_tree;
879         struct cfq_group *__cfqg;
880         struct rb_node *n;
881
882         cfqg->nr_cfqq++;
883         if (!RB_EMPTY_NODE(&cfqg->rb_node))
884                 return;
885
886         /*
887          * Currently put the group at the end. Later implement something
888          * so that groups get lesser vtime based on their weights, so that
889          * if group does not loose all if it was not continuously backlogged.
890          */
891         n = rb_last(&st->rb);
892         if (n) {
893                 __cfqg = rb_entry_cfqg(n);
894                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
895         } else
896                 cfqg->vdisktime = st->min_vdisktime;
897         cfq_group_service_tree_add(st, cfqg);
898 }
899
900 static void
901 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
902 {
903         st->total_weight -= cfqg->weight;
904         if (!RB_EMPTY_NODE(&cfqg->rb_node))
905                 cfq_rb_erase(&cfqg->rb_node, st);
906 }
907
908 static void
909 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
910 {
911         struct cfq_rb_root *st = &cfqd->grp_service_tree;
912
913         BUG_ON(cfqg->nr_cfqq < 1);
914         cfqg->nr_cfqq--;
915
916         /* If there are other cfq queues under this group, don't delete it */
917         if (cfqg->nr_cfqq)
918                 return;
919
920         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
921         cfq_group_service_tree_del(st, cfqg);
922         cfqg->saved_workload_slice = 0;
923         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
924 }
925
926 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
927                                                 unsigned int *unaccounted_time)
928 {
929         unsigned int slice_used;
930
931         /*
932          * Queue got expired before even a single request completed or
933          * got expired immediately after first request completion.
934          */
935         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
936                 /*
937                  * Also charge the seek time incurred to the group, otherwise
938                  * if there are mutiple queues in the group, each can dispatch
939                  * a single request on seeky media and cause lots of seek time
940                  * and group will never know it.
941                  */
942                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
943                                         1);
944         } else {
945                 slice_used = jiffies - cfqq->slice_start;
946                 if (slice_used > cfqq->allocated_slice) {
947                         *unaccounted_time = slice_used - cfqq->allocated_slice;
948                         slice_used = cfqq->allocated_slice;
949                 }
950                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
951                         *unaccounted_time += cfqq->slice_start -
952                                         cfqq->dispatch_start;
953         }
954
955         return slice_used;
956 }
957
958 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
959                                 struct cfq_queue *cfqq)
960 {
961         struct cfq_rb_root *st = &cfqd->grp_service_tree;
962         unsigned int used_sl, charge, unaccounted_sl = 0;
963         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
964                         - cfqg->service_tree_idle.count;
965
966         BUG_ON(nr_sync < 0);
967         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
968
969         if (iops_mode(cfqd))
970                 charge = cfqq->slice_dispatch;
971         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
972                 charge = cfqq->allocated_slice;
973
974         /* Can't update vdisktime while group is on service tree */
975         cfq_group_service_tree_del(st, cfqg);
976         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
977         /* If a new weight was requested, update now, off tree */
978         cfq_group_service_tree_add(st, cfqg);
979
980         /* This group is being expired. Save the context */
981         if (time_after(cfqd->workload_expires, jiffies)) {
982                 cfqg->saved_workload_slice = cfqd->workload_expires
983                                                 - jiffies;
984                 cfqg->saved_workload = cfqd->serving_type;
985                 cfqg->saved_serving_prio = cfqd->serving_prio;
986         } else
987                 cfqg->saved_workload_slice = 0;
988
989         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
990                                         st->min_vdisktime);
991         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
992                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
993                         iops_mode(cfqd), cfqq->nr_sectors);
994         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
995                                           unaccounted_sl);
996         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
997 }
998
999 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1000 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1001 {
1002         if (blkg)
1003                 return container_of(blkg, struct cfq_group, blkg);
1004         return NULL;
1005 }
1006
1007 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1008                                         unsigned int weight)
1009 {
1010         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1011         cfqg->new_weight = weight;
1012         cfqg->needs_update = true;
1013 }
1014
1015 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1016                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1017 {
1018         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1019         unsigned int major, minor;
1020
1021         /*
1022          * Add group onto cgroup list. It might happen that bdi->dev is
1023          * not initialized yet. Initialize this new group without major
1024          * and minor info and this info will be filled in once a new thread
1025          * comes for IO.
1026          */
1027         if (bdi->dev) {
1028                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1029                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1030                                         (void *)cfqd, MKDEV(major, minor));
1031         } else
1032                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1033                                         (void *)cfqd, 0);
1034
1035         cfqd->nr_blkcg_linked_grps++;
1036         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1037
1038         /* Add group on cfqd list */
1039         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1040 }
1041
1042 /*
1043  * Should be called from sleepable context. No request queue lock as per
1044  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1045  * from sleepable context.
1046  */
1047 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1048 {
1049         struct cfq_group *cfqg = NULL;
1050         int i, j, ret;
1051         struct cfq_rb_root *st;
1052
1053         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1054         if (!cfqg)
1055                 return NULL;
1056
1057         for_each_cfqg_st(cfqg, i, j, st)
1058                 *st = CFQ_RB_ROOT;
1059         RB_CLEAR_NODE(&cfqg->rb_node);
1060
1061         /*
1062          * Take the initial reference that will be released on destroy
1063          * This can be thought of a joint reference by cgroup and
1064          * elevator which will be dropped by either elevator exit
1065          * or cgroup deletion path depending on who is exiting first.
1066          */
1067         cfqg->ref = 1;
1068
1069         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1070         if (ret) {
1071                 kfree(cfqg);
1072                 return NULL;
1073         }
1074
1075         return cfqg;
1076 }
1077
1078 static struct cfq_group *
1079 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1080 {
1081         struct cfq_group *cfqg = NULL;
1082         void *key = cfqd;
1083         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1084         unsigned int major, minor;
1085
1086         /*
1087          * This is the common case when there are no blkio cgroups.
1088          * Avoid lookup in this case
1089          */
1090         if (blkcg == &blkio_root_cgroup)
1091                 cfqg = &cfqd->root_group;
1092         else
1093                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1094
1095         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1096                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1097                 cfqg->blkg.dev = MKDEV(major, minor);
1098         }
1099
1100         return cfqg;
1101 }
1102
1103 /*
1104  * Search for the cfq group current task belongs to. request_queue lock must
1105  * be held.
1106  */
1107 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1108 {
1109         struct blkio_cgroup *blkcg;
1110         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1111         struct request_queue *q = cfqd->queue;
1112
1113         rcu_read_lock();
1114         blkcg = task_blkio_cgroup(current);
1115         cfqg = cfq_find_cfqg(cfqd, blkcg);
1116         if (cfqg) {
1117                 rcu_read_unlock();
1118                 return cfqg;
1119         }
1120
1121         /*
1122          * Need to allocate a group. Allocation of group also needs allocation
1123          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1124          * we need to drop rcu lock and queue_lock before we call alloc.
1125          *
1126          * Not taking any queue reference here and assuming that queue is
1127          * around by the time we return. CFQ queue allocation code does
1128          * the same. It might be racy though.
1129          */
1130
1131         rcu_read_unlock();
1132         spin_unlock_irq(q->queue_lock);
1133
1134         cfqg = cfq_alloc_cfqg(cfqd);
1135
1136         spin_lock_irq(q->queue_lock);
1137
1138         rcu_read_lock();
1139         blkcg = task_blkio_cgroup(current);
1140
1141         /*
1142          * If some other thread already allocated the group while we were
1143          * not holding queue lock, free up the group
1144          */
1145         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1146
1147         if (__cfqg) {
1148                 kfree(cfqg);
1149                 rcu_read_unlock();
1150                 return __cfqg;
1151         }
1152
1153         if (!cfqg)
1154                 cfqg = &cfqd->root_group;
1155
1156         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1157         rcu_read_unlock();
1158         return cfqg;
1159 }
1160
1161 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1162 {
1163         cfqg->ref++;
1164         return cfqg;
1165 }
1166
1167 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1168 {
1169         /* Currently, all async queues are mapped to root group */
1170         if (!cfq_cfqq_sync(cfqq))
1171                 cfqg = &cfqq->cfqd->root_group;
1172
1173         cfqq->cfqg = cfqg;
1174         /* cfqq reference on cfqg */
1175         cfqq->cfqg->ref++;
1176 }
1177
1178 static void cfq_put_cfqg(struct cfq_group *cfqg)
1179 {
1180         struct cfq_rb_root *st;
1181         int i, j;
1182
1183         BUG_ON(cfqg->ref <= 0);
1184         cfqg->ref--;
1185         if (cfqg->ref)
1186                 return;
1187         for_each_cfqg_st(cfqg, i, j, st)
1188                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1189         free_percpu(cfqg->blkg.stats_cpu);
1190         kfree(cfqg);
1191 }
1192
1193 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1194 {
1195         /* Something wrong if we are trying to remove same group twice */
1196         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1197
1198         hlist_del_init(&cfqg->cfqd_node);
1199
1200         /*
1201          * Put the reference taken at the time of creation so that when all
1202          * queues are gone, group can be destroyed.
1203          */
1204         cfq_put_cfqg(cfqg);
1205 }
1206
1207 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1208 {
1209         struct hlist_node *pos, *n;
1210         struct cfq_group *cfqg;
1211
1212         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1213                 /*
1214                  * If cgroup removal path got to blk_group first and removed
1215                  * it from cgroup list, then it will take care of destroying
1216                  * cfqg also.
1217                  */
1218                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1219                         cfq_destroy_cfqg(cfqd, cfqg);
1220         }
1221 }
1222
1223 /*
1224  * Blk cgroup controller notification saying that blkio_group object is being
1225  * delinked as associated cgroup object is going away. That also means that
1226  * no new IO will come in this group. So get rid of this group as soon as
1227  * any pending IO in the group is finished.
1228  *
1229  * This function is called under rcu_read_lock(). key is the rcu protected
1230  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1231  * read lock.
1232  *
1233  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1234  * it should not be NULL as even if elevator was exiting, cgroup deltion
1235  * path got to it first.
1236  */
1237 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1238 {
1239         unsigned long  flags;
1240         struct cfq_data *cfqd = key;
1241
1242         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1243         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1244         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1245 }
1246
1247 #else /* GROUP_IOSCHED */
1248 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1249 {
1250         return &cfqd->root_group;
1251 }
1252
1253 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1254 {
1255         return cfqg;
1256 }
1257
1258 static inline void
1259 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1260         cfqq->cfqg = cfqg;
1261 }
1262
1263 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1264 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1265
1266 #endif /* GROUP_IOSCHED */
1267
1268 /*
1269  * The cfqd->service_trees holds all pending cfq_queue's that have
1270  * requests waiting to be processed. It is sorted in the order that
1271  * we will service the queues.
1272  */
1273 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1274                                  bool add_front)
1275 {
1276         struct rb_node **p, *parent;
1277         struct cfq_queue *__cfqq;
1278         unsigned long rb_key;
1279         struct cfq_rb_root *service_tree;
1280         int left;
1281         int new_cfqq = 1;
1282         int group_changed = 0;
1283
1284         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1285                                                 cfqq_type(cfqq));
1286         if (cfq_class_idle(cfqq)) {
1287                 rb_key = CFQ_IDLE_DELAY;
1288                 parent = rb_last(&service_tree->rb);
1289                 if (parent && parent != &cfqq->rb_node) {
1290                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1291                         rb_key += __cfqq->rb_key;
1292                 } else
1293                         rb_key += jiffies;
1294         } else if (!add_front) {
1295                 /*
1296                  * Get our rb key offset. Subtract any residual slice
1297                  * value carried from last service. A negative resid
1298                  * count indicates slice overrun, and this should position
1299                  * the next service time further away in the tree.
1300                  */
1301                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1302                 rb_key -= cfqq->slice_resid;
1303                 cfqq->slice_resid = 0;
1304         } else {
1305                 rb_key = -HZ;
1306                 __cfqq = cfq_rb_first(service_tree);
1307                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1308         }
1309
1310         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1311                 new_cfqq = 0;
1312                 /*
1313                  * same position, nothing more to do
1314                  */
1315                 if (rb_key == cfqq->rb_key &&
1316                     cfqq->service_tree == service_tree)
1317                         return;
1318
1319                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1320                 cfqq->service_tree = NULL;
1321         }
1322
1323         left = 1;
1324         parent = NULL;
1325         cfqq->service_tree = service_tree;
1326         p = &service_tree->rb.rb_node;
1327         while (*p) {
1328                 struct rb_node **n;
1329
1330                 parent = *p;
1331                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1332
1333                 /*
1334                  * sort by key, that represents service time.
1335                  */
1336                 if (time_before(rb_key, __cfqq->rb_key))
1337                         n = &(*p)->rb_left;
1338                 else {
1339                         n = &(*p)->rb_right;
1340                         left = 0;
1341                 }
1342
1343                 p = n;
1344         }
1345
1346         if (left)
1347                 service_tree->left = &cfqq->rb_node;
1348
1349         cfqq->rb_key = rb_key;
1350         rb_link_node(&cfqq->rb_node, parent, p);
1351         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1352         service_tree->count++;
1353         if ((add_front || !new_cfqq) && !group_changed)
1354                 return;
1355         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1356 }
1357
1358 static struct cfq_queue *
1359 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1360                      sector_t sector, struct rb_node **ret_parent,
1361                      struct rb_node ***rb_link)
1362 {
1363         struct rb_node **p, *parent;
1364         struct cfq_queue *cfqq = NULL;
1365
1366         parent = NULL;
1367         p = &root->rb_node;
1368         while (*p) {
1369                 struct rb_node **n;
1370
1371                 parent = *p;
1372                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1373
1374                 /*
1375                  * Sort strictly based on sector.  Smallest to the left,
1376                  * largest to the right.
1377                  */
1378                 if (sector > blk_rq_pos(cfqq->next_rq))
1379                         n = &(*p)->rb_right;
1380                 else if (sector < blk_rq_pos(cfqq->next_rq))
1381                         n = &(*p)->rb_left;
1382                 else
1383                         break;
1384                 p = n;
1385                 cfqq = NULL;
1386         }
1387
1388         *ret_parent = parent;
1389         if (rb_link)
1390                 *rb_link = p;
1391         return cfqq;
1392 }
1393
1394 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1395 {
1396         struct rb_node **p, *parent;
1397         struct cfq_queue *__cfqq;
1398
1399         if (cfqq->p_root) {
1400                 rb_erase(&cfqq->p_node, cfqq->p_root);
1401                 cfqq->p_root = NULL;
1402         }
1403
1404         if (cfq_class_idle(cfqq))
1405                 return;
1406         if (!cfqq->next_rq)
1407                 return;
1408
1409         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1410         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1411                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1412         if (!__cfqq) {
1413                 rb_link_node(&cfqq->p_node, parent, p);
1414                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1415         } else
1416                 cfqq->p_root = NULL;
1417 }
1418
1419 /*
1420  * Update cfqq's position in the service tree.
1421  */
1422 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1423 {
1424         /*
1425          * Resorting requires the cfqq to be on the RR list already.
1426          */
1427         if (cfq_cfqq_on_rr(cfqq)) {
1428                 cfq_service_tree_add(cfqd, cfqq, 0);
1429                 cfq_prio_tree_add(cfqd, cfqq);
1430         }
1431 }
1432
1433 /*
1434  * add to busy list of queues for service, trying to be fair in ordering
1435  * the pending list according to last request service
1436  */
1437 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1438 {
1439         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1440         BUG_ON(cfq_cfqq_on_rr(cfqq));
1441         cfq_mark_cfqq_on_rr(cfqq);
1442         cfqd->busy_queues++;
1443         if (cfq_cfqq_sync(cfqq))
1444                 cfqd->busy_sync_queues++;
1445
1446         cfq_resort_rr_list(cfqd, cfqq);
1447 }
1448
1449 /*
1450  * Called when the cfqq no longer has requests pending, remove it from
1451  * the service tree.
1452  */
1453 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1454 {
1455         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1456         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1457         cfq_clear_cfqq_on_rr(cfqq);
1458
1459         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1460                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1461                 cfqq->service_tree = NULL;
1462         }
1463         if (cfqq->p_root) {
1464                 rb_erase(&cfqq->p_node, cfqq->p_root);
1465                 cfqq->p_root = NULL;
1466         }
1467
1468         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1469         BUG_ON(!cfqd->busy_queues);
1470         cfqd->busy_queues--;
1471         if (cfq_cfqq_sync(cfqq))
1472                 cfqd->busy_sync_queues--;
1473 }
1474
1475 /*
1476  * rb tree support functions
1477  */
1478 static void cfq_del_rq_rb(struct request *rq)
1479 {
1480         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1481         const int sync = rq_is_sync(rq);
1482
1483         BUG_ON(!cfqq->queued[sync]);
1484         cfqq->queued[sync]--;
1485
1486         elv_rb_del(&cfqq->sort_list, rq);
1487
1488         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1489                 /*
1490                  * Queue will be deleted from service tree when we actually
1491                  * expire it later. Right now just remove it from prio tree
1492                  * as it is empty.
1493                  */
1494                 if (cfqq->p_root) {
1495                         rb_erase(&cfqq->p_node, cfqq->p_root);
1496                         cfqq->p_root = NULL;
1497                 }
1498         }
1499 }
1500
1501 static void cfq_add_rq_rb(struct request *rq)
1502 {
1503         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1504         struct cfq_data *cfqd = cfqq->cfqd;
1505         struct request *__alias, *prev;
1506
1507         cfqq->queued[rq_is_sync(rq)]++;
1508
1509         /*
1510          * looks a little odd, but the first insert might return an alias.
1511          * if that happens, put the alias on the dispatch list
1512          */
1513         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1514                 cfq_dispatch_insert(cfqd->queue, __alias);
1515
1516         if (!cfq_cfqq_on_rr(cfqq))
1517                 cfq_add_cfqq_rr(cfqd, cfqq);
1518
1519         /*
1520          * check if this request is a better next-serve candidate
1521          */
1522         prev = cfqq->next_rq;
1523         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1524
1525         /*
1526          * adjust priority tree position, if ->next_rq changes
1527          */
1528         if (prev != cfqq->next_rq)
1529                 cfq_prio_tree_add(cfqd, cfqq);
1530
1531         BUG_ON(!cfqq->next_rq);
1532 }
1533
1534 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1535 {
1536         elv_rb_del(&cfqq->sort_list, rq);
1537         cfqq->queued[rq_is_sync(rq)]--;
1538         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1539                                         rq_data_dir(rq), rq_is_sync(rq));
1540         cfq_add_rq_rb(rq);
1541         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1542                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1543                         rq_is_sync(rq));
1544 }
1545
1546 static struct request *
1547 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1548 {
1549         struct task_struct *tsk = current;
1550         struct cfq_io_context *cic;
1551         struct cfq_queue *cfqq;
1552
1553         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1554         if (!cic)
1555                 return NULL;
1556
1557         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1558         if (cfqq) {
1559                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1560
1561                 return elv_rb_find(&cfqq->sort_list, sector);
1562         }
1563
1564         return NULL;
1565 }
1566
1567 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1568 {
1569         struct cfq_data *cfqd = q->elevator->elevator_data;
1570
1571         cfqd->rq_in_driver++;
1572         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1573                                                 cfqd->rq_in_driver);
1574
1575         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1576 }
1577
1578 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1579 {
1580         struct cfq_data *cfqd = q->elevator->elevator_data;
1581
1582         WARN_ON(!cfqd->rq_in_driver);
1583         cfqd->rq_in_driver--;
1584         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1585                                                 cfqd->rq_in_driver);
1586 }
1587
1588 static void cfq_remove_request(struct request *rq)
1589 {
1590         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1591
1592         if (cfqq->next_rq == rq)
1593                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1594
1595         list_del_init(&rq->queuelist);
1596         cfq_del_rq_rb(rq);
1597
1598         cfqq->cfqd->rq_queued--;
1599         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1600                                         rq_data_dir(rq), rq_is_sync(rq));
1601         if (rq->cmd_flags & REQ_META) {
1602                 WARN_ON(!cfqq->meta_pending);
1603                 cfqq->meta_pending--;
1604         }
1605 }
1606
1607 static int cfq_merge(struct request_queue *q, struct request **req,
1608                      struct bio *bio)
1609 {
1610         struct cfq_data *cfqd = q->elevator->elevator_data;
1611         struct request *__rq;
1612
1613         __rq = cfq_find_rq_fmerge(cfqd, bio);
1614         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1615                 *req = __rq;
1616                 return ELEVATOR_FRONT_MERGE;
1617         }
1618
1619         return ELEVATOR_NO_MERGE;
1620 }
1621
1622 static void cfq_merged_request(struct request_queue *q, struct request *req,
1623                                int type)
1624 {
1625         if (type == ELEVATOR_FRONT_MERGE) {
1626                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1627
1628                 cfq_reposition_rq_rb(cfqq, req);
1629         }
1630 }
1631
1632 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1633                                 struct bio *bio)
1634 {
1635         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1636                                         bio_data_dir(bio), cfq_bio_sync(bio));
1637 }
1638
1639 static void
1640 cfq_merged_requests(struct request_queue *q, struct request *rq,
1641                     struct request *next)
1642 {
1643         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1644         /*
1645          * reposition in fifo if next is older than rq
1646          */
1647         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1648             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1649                 list_move(&rq->queuelist, &next->queuelist);
1650                 rq_set_fifo_time(rq, rq_fifo_time(next));
1651         }
1652
1653         if (cfqq->next_rq == next)
1654                 cfqq->next_rq = rq;
1655         cfq_remove_request(next);
1656         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1657                                         rq_data_dir(next), rq_is_sync(next));
1658 }
1659
1660 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1661                            struct bio *bio)
1662 {
1663         struct cfq_data *cfqd = q->elevator->elevator_data;
1664         struct cfq_io_context *cic;
1665         struct cfq_queue *cfqq;
1666
1667         /*
1668          * Disallow merge of a sync bio into an async request.
1669          */
1670         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1671                 return false;
1672
1673         /*
1674          * Lookup the cfqq that this bio will be queued with. Allow
1675          * merge only if rq is queued there.
1676          */
1677         cic = cfq_cic_lookup(cfqd, current->io_context);
1678         if (!cic)
1679                 return false;
1680
1681         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1682         return cfqq == RQ_CFQQ(rq);
1683 }
1684
1685 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1686 {
1687         del_timer(&cfqd->idle_slice_timer);
1688         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1689 }
1690
1691 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1692                                    struct cfq_queue *cfqq)
1693 {
1694         if (cfqq) {
1695                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1696                                 cfqd->serving_prio, cfqd->serving_type);
1697                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1698                 cfqq->slice_start = 0;
1699                 cfqq->dispatch_start = jiffies;
1700                 cfqq->allocated_slice = 0;
1701                 cfqq->slice_end = 0;
1702                 cfqq->slice_dispatch = 0;
1703                 cfqq->nr_sectors = 0;
1704
1705                 cfq_clear_cfqq_wait_request(cfqq);
1706                 cfq_clear_cfqq_must_dispatch(cfqq);
1707                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1708                 cfq_clear_cfqq_fifo_expire(cfqq);
1709                 cfq_mark_cfqq_slice_new(cfqq);
1710
1711                 cfq_del_timer(cfqd, cfqq);
1712         }
1713
1714         cfqd->active_queue = cfqq;
1715 }
1716
1717 /*
1718  * current cfqq expired its slice (or was too idle), select new one
1719  */
1720 static void
1721 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1722                     bool timed_out)
1723 {
1724         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1725
1726         if (cfq_cfqq_wait_request(cfqq))
1727                 cfq_del_timer(cfqd, cfqq);
1728
1729         cfq_clear_cfqq_wait_request(cfqq);
1730         cfq_clear_cfqq_wait_busy(cfqq);
1731
1732         /*
1733          * If this cfqq is shared between multiple processes, check to
1734          * make sure that those processes are still issuing I/Os within
1735          * the mean seek distance.  If not, it may be time to break the
1736          * queues apart again.
1737          */
1738         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1739                 cfq_mark_cfqq_split_coop(cfqq);
1740
1741         /*
1742          * store what was left of this slice, if the queue idled/timed out
1743          */
1744         if (timed_out) {
1745                 if (cfq_cfqq_slice_new(cfqq))
1746                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1747                 else
1748                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1749                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1750         }
1751
1752         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1753
1754         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1755                 cfq_del_cfqq_rr(cfqd, cfqq);
1756
1757         cfq_resort_rr_list(cfqd, cfqq);
1758
1759         if (cfqq == cfqd->active_queue)
1760                 cfqd->active_queue = NULL;
1761
1762         if (cfqd->active_cic) {
1763                 put_io_context(cfqd->active_cic->ioc);
1764                 cfqd->active_cic = NULL;
1765         }
1766 }
1767
1768 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1769 {
1770         struct cfq_queue *cfqq = cfqd->active_queue;
1771
1772         if (cfqq)
1773                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1774 }
1775
1776 /*
1777  * Get next queue for service. Unless we have a queue preemption,
1778  * we'll simply select the first cfqq in the service tree.
1779  */
1780 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1781 {
1782         struct cfq_rb_root *service_tree =
1783                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1784                                         cfqd->serving_type);
1785
1786         if (!cfqd->rq_queued)
1787                 return NULL;
1788
1789         /* There is nothing to dispatch */
1790         if (!service_tree)
1791                 return NULL;
1792         if (RB_EMPTY_ROOT(&service_tree->rb))
1793                 return NULL;
1794         return cfq_rb_first(service_tree);
1795 }
1796
1797 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1798 {
1799         struct cfq_group *cfqg;
1800         struct cfq_queue *cfqq;
1801         int i, j;
1802         struct cfq_rb_root *st;
1803
1804         if (!cfqd->rq_queued)
1805                 return NULL;
1806
1807         cfqg = cfq_get_next_cfqg(cfqd);
1808         if (!cfqg)
1809                 return NULL;
1810
1811         for_each_cfqg_st(cfqg, i, j, st)
1812                 if ((cfqq = cfq_rb_first(st)) != NULL)
1813                         return cfqq;
1814         return NULL;
1815 }
1816
1817 /*
1818  * Get and set a new active queue for service.
1819  */
1820 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1821                                               struct cfq_queue *cfqq)
1822 {
1823         if (!cfqq)
1824                 cfqq = cfq_get_next_queue(cfqd);
1825
1826         __cfq_set_active_queue(cfqd, cfqq);
1827         return cfqq;
1828 }
1829
1830 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1831                                           struct request *rq)
1832 {
1833         if (blk_rq_pos(rq) >= cfqd->last_position)
1834                 return blk_rq_pos(rq) - cfqd->last_position;
1835         else
1836                 return cfqd->last_position - blk_rq_pos(rq);
1837 }
1838
1839 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1840                                struct request *rq)
1841 {
1842         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1843 }
1844
1845 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1846                                     struct cfq_queue *cur_cfqq)
1847 {
1848         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1849         struct rb_node *parent, *node;
1850         struct cfq_queue *__cfqq;
1851         sector_t sector = cfqd->last_position;
1852
1853         if (RB_EMPTY_ROOT(root))
1854                 return NULL;
1855
1856         /*
1857          * First, if we find a request starting at the end of the last
1858          * request, choose it.
1859          */
1860         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1861         if (__cfqq)
1862                 return __cfqq;
1863
1864         /*
1865          * If the exact sector wasn't found, the parent of the NULL leaf
1866          * will contain the closest sector.
1867          */
1868         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1869         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1870                 return __cfqq;
1871
1872         if (blk_rq_pos(__cfqq->next_rq) < sector)
1873                 node = rb_next(&__cfqq->p_node);
1874         else
1875                 node = rb_prev(&__cfqq->p_node);
1876         if (!node)
1877                 return NULL;
1878
1879         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1880         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1881                 return __cfqq;
1882
1883         return NULL;
1884 }
1885
1886 /*
1887  * cfqd - obvious
1888  * cur_cfqq - passed in so that we don't decide that the current queue is
1889  *            closely cooperating with itself.
1890  *
1891  * So, basically we're assuming that that cur_cfqq has dispatched at least
1892  * one request, and that cfqd->last_position reflects a position on the disk
1893  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1894  * assumption.
1895  */
1896 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1897                                               struct cfq_queue *cur_cfqq)
1898 {
1899         struct cfq_queue *cfqq;
1900
1901         if (cfq_class_idle(cur_cfqq))
1902                 return NULL;
1903         if (!cfq_cfqq_sync(cur_cfqq))
1904                 return NULL;
1905         if (CFQQ_SEEKY(cur_cfqq))
1906                 return NULL;
1907
1908         /*
1909          * Don't search priority tree if it's the only queue in the group.
1910          */
1911         if (cur_cfqq->cfqg->nr_cfqq == 1)
1912                 return NULL;
1913
1914         /*
1915          * We should notice if some of the queues are cooperating, eg
1916          * working closely on the same area of the disk. In that case,
1917          * we can group them together and don't waste time idling.
1918          */
1919         cfqq = cfqq_close(cfqd, cur_cfqq);
1920         if (!cfqq)
1921                 return NULL;
1922
1923         /* If new queue belongs to different cfq_group, don't choose it */
1924         if (cur_cfqq->cfqg != cfqq->cfqg)
1925                 return NULL;
1926
1927         /*
1928          * It only makes sense to merge sync queues.
1929          */
1930         if (!cfq_cfqq_sync(cfqq))
1931                 return NULL;
1932         if (CFQQ_SEEKY(cfqq))
1933                 return NULL;
1934
1935         /*
1936          * Do not merge queues of different priority classes
1937          */
1938         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1939                 return NULL;
1940
1941         return cfqq;
1942 }
1943
1944 /*
1945  * Determine whether we should enforce idle window for this queue.
1946  */
1947
1948 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1949 {
1950         enum wl_prio_t prio = cfqq_prio(cfqq);
1951         struct cfq_rb_root *service_tree = cfqq->service_tree;
1952
1953         BUG_ON(!service_tree);
1954         BUG_ON(!service_tree->count);
1955
1956         if (!cfqd->cfq_slice_idle)
1957                 return false;
1958
1959         /* We never do for idle class queues. */
1960         if (prio == IDLE_WORKLOAD)
1961                 return false;
1962
1963         /* We do for queues that were marked with idle window flag. */
1964         if (cfq_cfqq_idle_window(cfqq) &&
1965            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1966                 return true;
1967
1968         /*
1969          * Otherwise, we do only if they are the last ones
1970          * in their service tree.
1971          */
1972         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1973                 return true;
1974         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1975                         service_tree->count);
1976         return false;
1977 }
1978
1979 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1980 {
1981         struct cfq_queue *cfqq = cfqd->active_queue;
1982         struct cfq_io_context *cic;
1983         unsigned long sl, group_idle = 0;
1984
1985         /*
1986          * SSD device without seek penalty, disable idling. But only do so
1987          * for devices that support queuing, otherwise we still have a problem
1988          * with sync vs async workloads.
1989          */
1990         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1991                 return;
1992
1993         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1994         WARN_ON(cfq_cfqq_slice_new(cfqq));
1995
1996         /*
1997          * idle is disabled, either manually or by past process history
1998          */
1999         if (!cfq_should_idle(cfqd, cfqq)) {
2000                 /* no queue idling. Check for group idling */
2001                 if (cfqd->cfq_group_idle)
2002                         group_idle = cfqd->cfq_group_idle;
2003                 else
2004                         return;
2005         }
2006
2007         /*
2008          * still active requests from this queue, don't idle
2009          */
2010         if (cfqq->dispatched)
2011                 return;
2012
2013         /*
2014          * task has exited, don't wait
2015          */
2016         cic = cfqd->active_cic;
2017         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2018                 return;
2019
2020         /*
2021          * If our average think time is larger than the remaining time
2022          * slice, then don't idle. This avoids overrunning the allotted
2023          * time slice.
2024          */
2025         if (sample_valid(cic->ttime_samples) &&
2026             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2027                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
2028                                 cic->ttime_mean);
2029                 return;
2030         }
2031
2032         /* There are other queues in the group, don't do group idle */
2033         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2034                 return;
2035
2036         cfq_mark_cfqq_wait_request(cfqq);
2037
2038         if (group_idle)
2039                 sl = cfqd->cfq_group_idle;
2040         else
2041                 sl = cfqd->cfq_slice_idle;
2042
2043         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2044         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2045         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2046                         group_idle ? 1 : 0);
2047 }
2048
2049 /*
2050  * Move request from internal lists to the request queue dispatch list.
2051  */
2052 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2053 {
2054         struct cfq_data *cfqd = q->elevator->elevator_data;
2055         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2056
2057         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2058
2059         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2060         cfq_remove_request(rq);
2061         cfqq->dispatched++;
2062         (RQ_CFQG(rq))->dispatched++;
2063         elv_dispatch_sort(q, rq);
2064
2065         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2066         cfqq->nr_sectors += blk_rq_sectors(rq);
2067         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2068                                         rq_data_dir(rq), rq_is_sync(rq));
2069 }
2070
2071 /*
2072  * return expired entry, or NULL to just start from scratch in rbtree
2073  */
2074 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2075 {
2076         struct request *rq = NULL;
2077
2078         if (cfq_cfqq_fifo_expire(cfqq))
2079                 return NULL;
2080
2081         cfq_mark_cfqq_fifo_expire(cfqq);
2082
2083         if (list_empty(&cfqq->fifo))
2084                 return NULL;
2085
2086         rq = rq_entry_fifo(cfqq->fifo.next);
2087         if (time_before(jiffies, rq_fifo_time(rq)))
2088                 rq = NULL;
2089
2090         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2091         return rq;
2092 }
2093
2094 static inline int
2095 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2096 {
2097         const int base_rq = cfqd->cfq_slice_async_rq;
2098
2099         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2100
2101         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2102 }
2103
2104 /*
2105  * Must be called with the queue_lock held.
2106  */
2107 static int cfqq_process_refs(struct cfq_queue *cfqq)
2108 {
2109         int process_refs, io_refs;
2110
2111         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2112         process_refs = cfqq->ref - io_refs;
2113         BUG_ON(process_refs < 0);
2114         return process_refs;
2115 }
2116
2117 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2118 {
2119         int process_refs, new_process_refs;
2120         struct cfq_queue *__cfqq;
2121
2122         /*
2123          * If there are no process references on the new_cfqq, then it is
2124          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2125          * chain may have dropped their last reference (not just their
2126          * last process reference).
2127          */
2128         if (!cfqq_process_refs(new_cfqq))
2129                 return;
2130
2131         /* Avoid a circular list and skip interim queue merges */
2132         while ((__cfqq = new_cfqq->new_cfqq)) {
2133                 if (__cfqq == cfqq)
2134                         return;
2135                 new_cfqq = __cfqq;
2136         }
2137
2138         process_refs = cfqq_process_refs(cfqq);
2139         new_process_refs = cfqq_process_refs(new_cfqq);
2140         /*
2141          * If the process for the cfqq has gone away, there is no
2142          * sense in merging the queues.
2143          */
2144         if (process_refs == 0 || new_process_refs == 0)
2145                 return;
2146
2147         /*
2148          * Merge in the direction of the lesser amount of work.
2149          */
2150         if (new_process_refs >= process_refs) {
2151                 cfqq->new_cfqq = new_cfqq;
2152                 new_cfqq->ref += process_refs;
2153         } else {
2154                 new_cfqq->new_cfqq = cfqq;
2155                 cfqq->ref += new_process_refs;
2156         }
2157 }
2158
2159 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2160                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2161 {
2162         struct cfq_queue *queue;
2163         int i;
2164         bool key_valid = false;
2165         unsigned long lowest_key = 0;
2166         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2167
2168         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2169                 /* select the one with lowest rb_key */
2170                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2171                 if (queue &&
2172                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2173                         lowest_key = queue->rb_key;
2174                         cur_best = i;
2175                         key_valid = true;
2176                 }
2177         }
2178
2179         return cur_best;
2180 }
2181
2182 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2183 {
2184         unsigned slice;
2185         unsigned count;
2186         struct cfq_rb_root *st;
2187         unsigned group_slice;
2188         enum wl_prio_t original_prio = cfqd->serving_prio;
2189
2190         /* Choose next priority. RT > BE > IDLE */
2191         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2192                 cfqd->serving_prio = RT_WORKLOAD;
2193         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2194                 cfqd->serving_prio = BE_WORKLOAD;
2195         else {
2196                 cfqd->serving_prio = IDLE_WORKLOAD;
2197                 cfqd->workload_expires = jiffies + 1;
2198                 return;
2199         }
2200
2201         if (original_prio != cfqd->serving_prio)
2202                 goto new_workload;
2203
2204         /*
2205          * For RT and BE, we have to choose also the type
2206          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2207          * expiration time
2208          */
2209         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2210         count = st->count;
2211
2212         /*
2213          * check workload expiration, and that we still have other queues ready
2214          */
2215         if (count && !time_after(jiffies, cfqd->workload_expires))
2216                 return;
2217
2218 new_workload:
2219         /* otherwise select new workload type */
2220         cfqd->serving_type =
2221                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2222         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2223         count = st->count;
2224
2225         /*
2226          * the workload slice is computed as a fraction of target latency
2227          * proportional to the number of queues in that workload, over
2228          * all the queues in the same priority class
2229          */
2230         group_slice = cfq_group_slice(cfqd, cfqg);
2231
2232         slice = group_slice * count /
2233                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2234                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2235
2236         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2237                 unsigned int tmp;
2238
2239                 /*
2240                  * Async queues are currently system wide. Just taking
2241                  * proportion of queues with-in same group will lead to higher
2242                  * async ratio system wide as generally root group is going
2243                  * to have higher weight. A more accurate thing would be to
2244                  * calculate system wide asnc/sync ratio.
2245                  */
2246                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2247                 tmp = tmp/cfqd->busy_queues;
2248                 slice = min_t(unsigned, slice, tmp);
2249
2250                 /* async workload slice is scaled down according to
2251                  * the sync/async slice ratio. */
2252                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2253         } else
2254                 /* sync workload slice is at least 2 * cfq_slice_idle */
2255                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2256
2257         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2258         cfq_log(cfqd, "workload slice:%d", slice);
2259         cfqd->workload_expires = jiffies + slice;
2260 }
2261
2262 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2263 {
2264         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2265         struct cfq_group *cfqg;
2266
2267         if (RB_EMPTY_ROOT(&st->rb))
2268                 return NULL;
2269         cfqg = cfq_rb_first_group(st);
2270         update_min_vdisktime(st);
2271         return cfqg;
2272 }
2273
2274 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2275 {
2276         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2277
2278         cfqd->serving_group = cfqg;
2279
2280         /* Restore the workload type data */
2281         if (cfqg->saved_workload_slice) {
2282                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2283                 cfqd->serving_type = cfqg->saved_workload;
2284                 cfqd->serving_prio = cfqg->saved_serving_prio;
2285         } else
2286                 cfqd->workload_expires = jiffies - 1;
2287
2288         choose_service_tree(cfqd, cfqg);
2289 }
2290
2291 /*
2292  * Select a queue for service. If we have a current active queue,
2293  * check whether to continue servicing it, or retrieve and set a new one.
2294  */
2295 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2296 {
2297         struct cfq_queue *cfqq, *new_cfqq = NULL;
2298
2299         cfqq = cfqd->active_queue;
2300         if (!cfqq)
2301                 goto new_queue;
2302
2303         if (!cfqd->rq_queued)
2304                 return NULL;
2305
2306         /*
2307          * We were waiting for group to get backlogged. Expire the queue
2308          */
2309         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2310                 goto expire;
2311
2312         /*
2313          * The active queue has run out of time, expire it and select new.
2314          */
2315         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2316                 /*
2317                  * If slice had not expired at the completion of last request
2318                  * we might not have turned on wait_busy flag. Don't expire
2319                  * the queue yet. Allow the group to get backlogged.
2320                  *
2321                  * The very fact that we have used the slice, that means we
2322                  * have been idling all along on this queue and it should be
2323                  * ok to wait for this request to complete.
2324                  */
2325                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2326                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2327                         cfqq = NULL;
2328                         goto keep_queue;
2329                 } else
2330                         goto check_group_idle;
2331         }
2332
2333         /*
2334          * The active queue has requests and isn't expired, allow it to
2335          * dispatch.
2336          */
2337         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2338                 goto keep_queue;
2339
2340         /*
2341          * If another queue has a request waiting within our mean seek
2342          * distance, let it run.  The expire code will check for close
2343          * cooperators and put the close queue at the front of the service
2344          * tree.  If possible, merge the expiring queue with the new cfqq.
2345          */
2346         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2347         if (new_cfqq) {
2348                 if (!cfqq->new_cfqq)
2349                         cfq_setup_merge(cfqq, new_cfqq);
2350                 goto expire;
2351         }
2352
2353         /*
2354          * No requests pending. If the active queue still has requests in
2355          * flight or is idling for a new request, allow either of these
2356          * conditions to happen (or time out) before selecting a new queue.
2357          */
2358         if (timer_pending(&cfqd->idle_slice_timer)) {
2359                 cfqq = NULL;
2360                 goto keep_queue;
2361         }
2362
2363         /*
2364          * This is a deep seek queue, but the device is much faster than
2365          * the queue can deliver, don't idle
2366          **/
2367         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2368             (cfq_cfqq_slice_new(cfqq) ||
2369             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2370                 cfq_clear_cfqq_deep(cfqq);
2371                 cfq_clear_cfqq_idle_window(cfqq);
2372         }
2373
2374         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2375                 cfqq = NULL;
2376                 goto keep_queue;
2377         }
2378
2379         /*
2380          * If group idle is enabled and there are requests dispatched from
2381          * this group, wait for requests to complete.
2382          */
2383 check_group_idle:
2384         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2385             && cfqq->cfqg->dispatched) {
2386                 cfqq = NULL;
2387                 goto keep_queue;
2388         }
2389
2390 expire:
2391         cfq_slice_expired(cfqd, 0);
2392 new_queue:
2393         /*
2394          * Current queue expired. Check if we have to switch to a new
2395          * service tree
2396          */
2397         if (!new_cfqq)
2398                 cfq_choose_cfqg(cfqd);
2399
2400         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2401 keep_queue:
2402         return cfqq;
2403 }
2404
2405 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2406 {
2407         int dispatched = 0;
2408
2409         while (cfqq->next_rq) {
2410                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2411                 dispatched++;
2412         }
2413
2414         BUG_ON(!list_empty(&cfqq->fifo));
2415
2416         /* By default cfqq is not expired if it is empty. Do it explicitly */
2417         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2418         return dispatched;
2419 }
2420
2421 /*
2422  * Drain our current requests. Used for barriers and when switching
2423  * io schedulers on-the-fly.
2424  */
2425 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2426 {
2427         struct cfq_queue *cfqq;
2428         int dispatched = 0;
2429
2430         /* Expire the timeslice of the current active queue first */
2431         cfq_slice_expired(cfqd, 0);
2432         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2433                 __cfq_set_active_queue(cfqd, cfqq);
2434                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2435         }
2436
2437         BUG_ON(cfqd->busy_queues);
2438
2439         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2440         return dispatched;
2441 }
2442
2443 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2444         struct cfq_queue *cfqq)
2445 {
2446         /* the queue hasn't finished any request, can't estimate */
2447         if (cfq_cfqq_slice_new(cfqq))
2448                 return true;
2449         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2450                 cfqq->slice_end))
2451                 return true;
2452
2453         return false;
2454 }
2455
2456 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2457 {
2458         unsigned int max_dispatch;
2459
2460         /*
2461          * Drain async requests before we start sync IO
2462          */
2463         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2464                 return false;
2465
2466         /*
2467          * If this is an async queue and we have sync IO in flight, let it wait
2468          */
2469         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2470                 return false;
2471
2472         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2473         if (cfq_class_idle(cfqq))
2474                 max_dispatch = 1;
2475
2476         /*
2477          * Does this cfqq already have too much IO in flight?
2478          */
2479         if (cfqq->dispatched >= max_dispatch) {
2480                 bool promote_sync = false;
2481                 /*
2482                  * idle queue must always only have a single IO in flight
2483                  */
2484                 if (cfq_class_idle(cfqq))
2485                         return false;
2486
2487                 /*
2488                  * If there is only one sync queue
2489                  * we can ignore async queue here and give the sync
2490                  * queue no dispatch limit. The reason is a sync queue can
2491                  * preempt async queue, limiting the sync queue doesn't make
2492                  * sense. This is useful for aiostress test.
2493                  */
2494                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2495                         promote_sync = true;
2496
2497                 /*
2498                  * We have other queues, don't allow more IO from this one
2499                  */
2500                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2501                                 !promote_sync)
2502                         return false;
2503
2504                 /*
2505                  * Sole queue user, no limit
2506                  */
2507                 if (cfqd->busy_queues == 1 || promote_sync)
2508                         max_dispatch = -1;
2509                 else
2510                         /*
2511                          * Normally we start throttling cfqq when cfq_quantum/2
2512                          * requests have been dispatched. But we can drive
2513                          * deeper queue depths at the beginning of slice
2514                          * subjected to upper limit of cfq_quantum.
2515                          * */
2516                         max_dispatch = cfqd->cfq_quantum;
2517         }
2518
2519         /*
2520          * Async queues must wait a bit before being allowed dispatch.
2521          * We also ramp up the dispatch depth gradually for async IO,
2522          * based on the last sync IO we serviced
2523          */
2524         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2525                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2526                 unsigned int depth;
2527
2528                 depth = last_sync / cfqd->cfq_slice[1];
2529                 if (!depth && !cfqq->dispatched)
2530                         depth = 1;
2531                 if (depth < max_dispatch)
2532                         max_dispatch = depth;
2533         }
2534
2535         /*
2536          * If we're below the current max, allow a dispatch
2537          */
2538         return cfqq->dispatched < max_dispatch;
2539 }
2540
2541 /*
2542  * Dispatch a request from cfqq, moving them to the request queue
2543  * dispatch list.
2544  */
2545 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2546 {
2547         struct request *rq;
2548
2549         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2550
2551         if (!cfq_may_dispatch(cfqd, cfqq))
2552                 return false;
2553
2554         /*
2555          * follow expired path, else get first next available
2556          */
2557         rq = cfq_check_fifo(cfqq);
2558         if (!rq)
2559                 rq = cfqq->next_rq;
2560
2561         /*
2562          * insert request into driver dispatch list
2563          */
2564         cfq_dispatch_insert(cfqd->queue, rq);
2565
2566         if (!cfqd->active_cic) {
2567                 struct cfq_io_context *cic = RQ_CIC(rq);
2568
2569                 atomic_long_inc(&cic->ioc->refcount);
2570                 cfqd->active_cic = cic;
2571         }
2572
2573         return true;
2574 }
2575
2576 /*
2577  * Find the cfqq that we need to service and move a request from that to the
2578  * dispatch list
2579  */
2580 static int cfq_dispatch_requests(struct request_queue *q, int force)
2581 {
2582         struct cfq_data *cfqd = q->elevator->elevator_data;
2583         struct cfq_queue *cfqq;
2584
2585         if (!cfqd->busy_queues)
2586                 return 0;
2587
2588         if (unlikely(force))
2589                 return cfq_forced_dispatch(cfqd);
2590
2591         cfqq = cfq_select_queue(cfqd);
2592         if (!cfqq)
2593                 return 0;
2594
2595         /*
2596          * Dispatch a request from this cfqq, if it is allowed
2597          */
2598         if (!cfq_dispatch_request(cfqd, cfqq))
2599                 return 0;
2600
2601         cfqq->slice_dispatch++;
2602         cfq_clear_cfqq_must_dispatch(cfqq);
2603
2604         /*
2605          * expire an async queue immediately if it has used up its slice. idle
2606          * queue always expire after 1 dispatch round.
2607          */
2608         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2609             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2610             cfq_class_idle(cfqq))) {
2611                 cfqq->slice_end = jiffies + 1;
2612                 cfq_slice_expired(cfqd, 0);
2613         }
2614
2615         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2616         return 1;
2617 }
2618
2619 /*
2620  * task holds one reference to the queue, dropped when task exits. each rq
2621  * in-flight on this queue also holds a reference, dropped when rq is freed.
2622  *
2623  * Each cfq queue took a reference on the parent group. Drop it now.
2624  * queue lock must be held here.
2625  */
2626 static void cfq_put_queue(struct cfq_queue *cfqq)
2627 {
2628         struct cfq_data *cfqd = cfqq->cfqd;
2629         struct cfq_group *cfqg;
2630
2631         BUG_ON(cfqq->ref <= 0);
2632
2633         cfqq->ref--;
2634         if (cfqq->ref)
2635                 return;
2636
2637         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2638         BUG_ON(rb_first(&cfqq->sort_list));
2639         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2640         cfqg = cfqq->cfqg;
2641
2642         if (unlikely(cfqd->active_queue == cfqq)) {
2643                 __cfq_slice_expired(cfqd, cfqq, 0);
2644                 cfq_schedule_dispatch(cfqd);
2645         }
2646
2647         BUG_ON(cfq_cfqq_on_rr(cfqq));
2648         kmem_cache_free(cfq_pool, cfqq);
2649         cfq_put_cfqg(cfqg);
2650 }
2651
2652 /*
2653  * Call func for each cic attached to this ioc.
2654  */
2655 static void
2656 call_for_each_cic(struct io_context *ioc,
2657                   void (*func)(struct io_context *, struct cfq_io_context *))
2658 {
2659         struct cfq_io_context *cic;
2660         struct hlist_node *n;
2661
2662         rcu_read_lock();
2663
2664         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2665                 func(ioc, cic);
2666
2667         rcu_read_unlock();
2668 }
2669
2670 static void cfq_cic_free_rcu(struct rcu_head *head)
2671 {
2672         struct cfq_io_context *cic;
2673
2674         cic = container_of(head, struct cfq_io_context, rcu_head);
2675
2676         kmem_cache_free(cfq_ioc_pool, cic);
2677         elv_ioc_count_dec(cfq_ioc_count);
2678
2679         if (ioc_gone) {
2680                 /*
2681                  * CFQ scheduler is exiting, grab exit lock and check
2682                  * the pending io context count. If it hits zero,
2683                  * complete ioc_gone and set it back to NULL
2684                  */
2685                 spin_lock(&ioc_gone_lock);
2686                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2687                         complete(ioc_gone);
2688                         ioc_gone = NULL;
2689                 }
2690                 spin_unlock(&ioc_gone_lock);
2691         }
2692 }
2693
2694 static void cfq_cic_free(struct cfq_io_context *cic)
2695 {
2696         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2697 }
2698
2699 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2700 {
2701         unsigned long flags;
2702         unsigned long dead_key = (unsigned long) cic->key;
2703
2704         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2705
2706         spin_lock_irqsave(&ioc->lock, flags);
2707         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2708         hlist_del_rcu(&cic->cic_list);
2709         spin_unlock_irqrestore(&ioc->lock, flags);
2710
2711         cfq_cic_free(cic);
2712 }
2713
2714 /*
2715  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2716  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2717  * and ->trim() which is called with the task lock held
2718  */
2719 static void cfq_free_io_context(struct io_context *ioc)
2720 {
2721         /*
2722          * ioc->refcount is zero here, or we are called from elv_unregister(),
2723          * so no more cic's are allowed to be linked into this ioc.  So it
2724          * should be ok to iterate over the known list, we will see all cic's
2725          * since no new ones are added.
2726          */
2727         call_for_each_cic(ioc, cic_free_func);
2728 }
2729
2730 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2731 {
2732         struct cfq_queue *__cfqq, *next;
2733
2734         /*
2735          * If this queue was scheduled to merge with another queue, be
2736          * sure to drop the reference taken on that queue (and others in
2737          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2738          */
2739         __cfqq = cfqq->new_cfqq;
2740         while (__cfqq) {
2741                 if (__cfqq == cfqq) {
2742                         WARN(1, "cfqq->new_cfqq loop detected\n");
2743                         break;
2744                 }
2745                 next = __cfqq->new_cfqq;
2746                 cfq_put_queue(__cfqq);
2747                 __cfqq = next;
2748         }
2749 }
2750
2751 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2752 {
2753         if (unlikely(cfqq == cfqd->active_queue)) {
2754                 __cfq_slice_expired(cfqd, cfqq, 0);
2755                 cfq_schedule_dispatch(cfqd);
2756         }
2757
2758         cfq_put_cooperator(cfqq);
2759
2760         cfq_put_queue(cfqq);
2761 }
2762
2763 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2764                                          struct cfq_io_context *cic)
2765 {
2766         struct io_context *ioc = cic->ioc;
2767
2768         list_del_init(&cic->queue_list);
2769
2770         /*
2771          * Make sure dead mark is seen for dead queues
2772          */
2773         smp_wmb();
2774         cic->key = cfqd_dead_key(cfqd);
2775
2776         if (ioc->ioc_data == cic)
2777                 rcu_assign_pointer(ioc->ioc_data, NULL);
2778
2779         if (cic->cfqq[BLK_RW_ASYNC]) {
2780                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2781                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2782         }
2783
2784         if (cic->cfqq[BLK_RW_SYNC]) {
2785                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2786                 cic->cfqq[BLK_RW_SYNC] = NULL;
2787         }
2788 }
2789
2790 static void cfq_exit_single_io_context(struct io_context *ioc,
2791                                        struct cfq_io_context *cic)
2792 {
2793         struct cfq_data *cfqd = cic_to_cfqd(cic);
2794
2795         if (cfqd) {
2796                 struct request_queue *q = cfqd->queue;
2797                 unsigned long flags;
2798
2799                 spin_lock_irqsave(q->queue_lock, flags);
2800
2801                 /*
2802                  * Ensure we get a fresh copy of the ->key to prevent
2803                  * race between exiting task and queue
2804                  */
2805                 smp_read_barrier_depends();
2806                 if (cic->key == cfqd)
2807                         __cfq_exit_single_io_context(cfqd, cic);
2808
2809                 spin_unlock_irqrestore(q->queue_lock, flags);
2810         }
2811 }
2812
2813 /*
2814  * The process that ioc belongs to has exited, we need to clean up
2815  * and put the internal structures we have that belongs to that process.
2816  */
2817 static void cfq_exit_io_context(struct io_context *ioc)
2818 {
2819         call_for_each_cic(ioc, cfq_exit_single_io_context);
2820 }
2821
2822 static struct cfq_io_context *
2823 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2824 {
2825         struct cfq_io_context *cic;
2826
2827         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2828                                                         cfqd->queue->node);
2829         if (cic) {
2830                 cic->last_end_request = jiffies;
2831                 INIT_LIST_HEAD(&cic->queue_list);
2832                 INIT_HLIST_NODE(&cic->cic_list);
2833                 cic->dtor = cfq_free_io_context;
2834                 cic->exit = cfq_exit_io_context;
2835                 elv_ioc_count_inc(cfq_ioc_count);
2836         }
2837
2838         return cic;
2839 }
2840
2841 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2842 {
2843         struct task_struct *tsk = current;
2844         int ioprio_class;
2845
2846         if (!cfq_cfqq_prio_changed(cfqq))
2847                 return;
2848
2849         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2850         switch (ioprio_class) {
2851         default:
2852                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2853         case IOPRIO_CLASS_NONE:
2854                 /*
2855                  * no prio set, inherit CPU scheduling settings
2856                  */
2857                 cfqq->ioprio = task_nice_ioprio(tsk);
2858                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2859                 break;
2860         case IOPRIO_CLASS_RT:
2861                 cfqq->ioprio = task_ioprio(ioc);
2862                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2863                 break;
2864         case IOPRIO_CLASS_BE:
2865                 cfqq->ioprio = task_ioprio(ioc);
2866                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2867                 break;
2868         case IOPRIO_CLASS_IDLE:
2869                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2870                 cfqq->ioprio = 7;
2871                 cfq_clear_cfqq_idle_window(cfqq);
2872                 break;
2873         }
2874
2875         /*
2876          * keep track of original prio settings in case we have to temporarily
2877          * elevate the priority of this queue
2878          */
2879         cfqq->org_ioprio = cfqq->ioprio;
2880         cfqq->org_ioprio_class = cfqq->ioprio_class;
2881         cfq_clear_cfqq_prio_changed(cfqq);
2882 }
2883
2884 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2885 {
2886         struct cfq_data *cfqd = cic_to_cfqd(cic);
2887         struct cfq_queue *cfqq;
2888         unsigned long flags;
2889
2890         if (unlikely(!cfqd))
2891                 return;
2892
2893         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2894
2895         cfqq = cic->cfqq[BLK_RW_ASYNC];
2896         if (cfqq) {
2897                 struct cfq_queue *new_cfqq;
2898                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2899                                                 GFP_ATOMIC);
2900                 if (new_cfqq) {
2901                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2902                         cfq_put_queue(cfqq);
2903                 }
2904         }
2905
2906         cfqq = cic->cfqq[BLK_RW_SYNC];
2907         if (cfqq)
2908                 cfq_mark_cfqq_prio_changed(cfqq);
2909
2910         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2911 }
2912
2913 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2914 {
2915         call_for_each_cic(ioc, changed_ioprio);
2916         ioc->ioprio_changed = 0;
2917 }
2918
2919 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2920                           pid_t pid, bool is_sync)
2921 {
2922         RB_CLEAR_NODE(&cfqq->rb_node);
2923         RB_CLEAR_NODE(&cfqq->p_node);
2924         INIT_LIST_HEAD(&cfqq->fifo);
2925
2926         cfqq->ref = 0;
2927         cfqq->cfqd = cfqd;
2928
2929         cfq_mark_cfqq_prio_changed(cfqq);
2930
2931         if (is_sync) {
2932                 if (!cfq_class_idle(cfqq))
2933                         cfq_mark_cfqq_idle_window(cfqq);
2934                 cfq_mark_cfqq_sync(cfqq);
2935         }
2936         cfqq->pid = pid;
2937 }
2938
2939 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2940 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2941 {
2942         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2943         struct cfq_data *cfqd = cic_to_cfqd(cic);
2944         unsigned long flags;
2945         struct request_queue *q;
2946
2947         if (unlikely(!cfqd))
2948                 return;
2949
2950         q = cfqd->queue;
2951
2952         spin_lock_irqsave(q->queue_lock, flags);
2953
2954         if (sync_cfqq) {
2955                 /*
2956                  * Drop reference to sync queue. A new sync queue will be
2957                  * assigned in new group upon arrival of a fresh request.
2958                  */
2959                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2960                 cic_set_cfqq(cic, NULL, 1);
2961                 cfq_put_queue(sync_cfqq);
2962         }
2963
2964         spin_unlock_irqrestore(q->queue_lock, flags);
2965 }
2966
2967 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2968 {
2969         call_for_each_cic(ioc, changed_cgroup);
2970         ioc->cgroup_changed = 0;
2971 }
2972 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2973
2974 static struct cfq_queue *
2975 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2976                      struct io_context *ioc, gfp_t gfp_mask)
2977 {
2978         struct cfq_queue *cfqq, *new_cfqq = NULL;
2979         struct cfq_io_context *cic;
2980         struct cfq_group *cfqg;
2981
2982 retry:
2983         cfqg = cfq_get_cfqg(cfqd);
2984         cic = cfq_cic_lookup(cfqd, ioc);
2985         /* cic always exists here */
2986         cfqq = cic_to_cfqq(cic, is_sync);
2987
2988         /*
2989          * Always try a new alloc if we fell back to the OOM cfqq
2990          * originally, since it should just be a temporary situation.
2991          */
2992         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2993                 cfqq = NULL;
2994                 if (new_cfqq) {
2995                         cfqq = new_cfqq;
2996                         new_cfqq = NULL;
2997                 } else if (gfp_mask & __GFP_WAIT) {
2998                         spin_unlock_irq(cfqd->queue->queue_lock);
2999                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3000                                         gfp_mask | __GFP_ZERO,
3001                                         cfqd->queue->node);
3002                         spin_lock_irq(cfqd->queue->queue_lock);
3003                         if (new_cfqq)
3004                                 goto retry;
3005                 } else {
3006                         cfqq = kmem_cache_alloc_node(cfq_pool,
3007                                         gfp_mask | __GFP_ZERO,
3008                                         cfqd->queue->node);
3009                 }
3010
3011                 if (cfqq) {
3012                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3013                         cfq_init_prio_data(cfqq, ioc);
3014                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3015                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3016                 } else
3017                         cfqq = &cfqd->oom_cfqq;
3018         }
3019
3020         if (new_cfqq)
3021                 kmem_cache_free(cfq_pool, new_cfqq);
3022
3023         return cfqq;
3024 }
3025
3026 static struct cfq_queue **
3027 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3028 {
3029         switch (ioprio_class) {
3030         case IOPRIO_CLASS_RT:
3031                 return &cfqd->async_cfqq[0][ioprio];
3032         case IOPRIO_CLASS_BE:
3033                 return &cfqd->async_cfqq[1][ioprio];
3034         case IOPRIO_CLASS_IDLE:
3035                 return &cfqd->async_idle_cfqq;
3036         default:
3037                 BUG();
3038         }
3039 }
3040
3041 static struct cfq_queue *
3042 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3043               gfp_t gfp_mask)
3044 {
3045         const int ioprio = task_ioprio(ioc);
3046         const int ioprio_class = task_ioprio_class(ioc);
3047         struct cfq_queue **async_cfqq = NULL;
3048         struct cfq_queue *cfqq = NULL;
3049
3050         if (!is_sync) {
3051                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3052                 cfqq = *async_cfqq;
3053         }
3054
3055         if (!cfqq)
3056                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3057
3058         /*
3059          * pin the queue now that it's allocated, scheduler exit will prune it
3060          */
3061         if (!is_sync && !(*async_cfqq)) {
3062                 cfqq->ref++;
3063                 *async_cfqq = cfqq;
3064         }
3065
3066         cfqq->ref++;
3067         return cfqq;
3068 }
3069
3070 /*
3071  * We drop cfq io contexts lazily, so we may find a dead one.
3072  */
3073 static void
3074 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3075                   struct cfq_io_context *cic)
3076 {
3077         unsigned long flags;
3078
3079         WARN_ON(!list_empty(&cic->queue_list));
3080         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3081
3082         spin_lock_irqsave(&ioc->lock, flags);
3083
3084         BUG_ON(ioc->ioc_data == cic);
3085
3086         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3087         hlist_del_rcu(&cic->cic_list);
3088         spin_unlock_irqrestore(&ioc->lock, flags);
3089
3090         cfq_cic_free(cic);
3091 }
3092
3093 static struct cfq_io_context *
3094 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3095 {
3096         struct cfq_io_context *cic;
3097         unsigned long flags;
3098
3099         if (unlikely(!ioc))
3100                 return NULL;
3101
3102         rcu_read_lock();
3103
3104         /*
3105          * we maintain a last-hit cache, to avoid browsing over the tree
3106          */
3107         cic = rcu_dereference(ioc->ioc_data);
3108         if (cic && cic->key == cfqd) {
3109                 rcu_read_unlock();
3110                 return cic;
3111         }
3112
3113         do {
3114                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3115                 rcu_read_unlock();
3116                 if (!cic)
3117                         break;
3118                 if (unlikely(cic->key != cfqd)) {
3119                         cfq_drop_dead_cic(cfqd, ioc, cic);
3120                         rcu_read_lock();
3121                         continue;
3122                 }
3123
3124                 spin_lock_irqsave(&ioc->lock, flags);
3125                 rcu_assign_pointer(ioc->ioc_data, cic);
3126                 spin_unlock_irqrestore(&ioc->lock, flags);
3127                 break;
3128         } while (1);
3129
3130         return cic;
3131 }
3132
3133 /*
3134  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3135  * the process specific cfq io context when entered from the block layer.
3136  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3137  */
3138 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3139                         struct cfq_io_context *cic, gfp_t gfp_mask)
3140 {
3141         unsigned long flags;
3142         int ret;
3143
3144         ret = radix_tree_preload(gfp_mask);
3145         if (!ret) {
3146                 cic->ioc = ioc;
3147                 cic->key = cfqd;
3148
3149                 spin_lock_irqsave(&ioc->lock, flags);
3150                 ret = radix_tree_insert(&ioc->radix_root,
3151                                                 cfqd->cic_index, cic);
3152                 if (!ret)
3153                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3154                 spin_unlock_irqrestore(&ioc->lock, flags);
3155
3156                 radix_tree_preload_end();
3157
3158                 if (!ret) {
3159                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3160                         list_add(&cic->queue_list, &cfqd->cic_list);
3161                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3162                 }
3163         }
3164
3165         if (ret)
3166                 printk(KERN_ERR "cfq: cic link failed!\n");
3167
3168         return ret;
3169 }
3170
3171 /*
3172  * Setup general io context and cfq io context. There can be several cfq
3173  * io contexts per general io context, if this process is doing io to more
3174  * than one device managed by cfq.
3175  */
3176 static struct cfq_io_context *
3177 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3178 {
3179         struct io_context *ioc = NULL;
3180         struct cfq_io_context *cic;
3181
3182         might_sleep_if(gfp_mask & __GFP_WAIT);
3183
3184         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3185         if (!ioc)
3186                 return NULL;
3187
3188         cic = cfq_cic_lookup(cfqd, ioc);
3189         if (cic)
3190                 goto out;
3191
3192         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3193         if (cic == NULL)
3194                 goto err;
3195
3196         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3197                 goto err_free;
3198
3199 out:
3200         smp_read_barrier_depends();
3201         if (unlikely(ioc->ioprio_changed))
3202                 cfq_ioc_set_ioprio(ioc);
3203
3204 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3205         if (unlikely(ioc->cgroup_changed))
3206                 cfq_ioc_set_cgroup(ioc);
3207 #endif
3208         return cic;
3209 err_free:
3210         cfq_cic_free(cic);
3211 err:
3212         put_io_context(ioc);
3213         return NULL;
3214 }
3215
3216 static void
3217 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3218 {
3219         unsigned long elapsed = jiffies - cic->last_end_request;
3220         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3221
3222         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3223         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3224         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3225 }
3226
3227 static void
3228 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3229                        struct request *rq)
3230 {
3231         sector_t sdist = 0;
3232         sector_t n_sec = blk_rq_sectors(rq);
3233         if (cfqq->last_request_pos) {
3234                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3235                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3236                 else
3237                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3238         }
3239
3240         cfqq->seek_history <<= 1;
3241         if (blk_queue_nonrot(cfqd->queue))
3242                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3243         else
3244                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3245 }
3246
3247 /*
3248  * Disable idle window if the process thinks too long or seeks so much that
3249  * it doesn't matter
3250  */
3251 static void
3252 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3253                        struct cfq_io_context *cic)
3254 {
3255         int old_idle, enable_idle;
3256
3257         /*
3258          * Don't idle for async or idle io prio class
3259          */
3260         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3261                 return;
3262
3263         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3264
3265         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3266                 cfq_mark_cfqq_deep(cfqq);
3267
3268         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3269                 enable_idle = 0;
3270         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3271             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3272                 enable_idle = 0;
3273         else if (sample_valid(cic->ttime_samples)) {
3274                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3275                         enable_idle = 0;
3276                 else
3277                         enable_idle = 1;
3278         }
3279
3280         if (old_idle != enable_idle) {
3281                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3282                 if (enable_idle)
3283                         cfq_mark_cfqq_idle_window(cfqq);
3284                 else
3285                         cfq_clear_cfqq_idle_window(cfqq);
3286         }
3287 }
3288
3289 /*
3290  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3291  * no or if we aren't sure, a 1 will cause a preempt.
3292  */
3293 static bool
3294 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3295                    struct request *rq)
3296 {
3297         struct cfq_queue *cfqq;
3298
3299         cfqq = cfqd->active_queue;
3300         if (!cfqq)
3301                 return false;
3302
3303         if (cfq_class_idle(new_cfqq))
3304                 return false;
3305
3306         if (cfq_class_idle(cfqq))
3307                 return true;
3308
3309         /*
3310          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3311          */
3312         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3313                 return false;
3314
3315         /*
3316          * if the new request is sync, but the currently running queue is
3317          * not, let the sync request have priority.
3318          */
3319         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3320                 return true;
3321
3322         if (new_cfqq->cfqg != cfqq->cfqg)
3323                 return false;
3324
3325         if (cfq_slice_used(cfqq))
3326                 return true;
3327
3328         /* Allow preemption only if we are idling on sync-noidle tree */
3329         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3330             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3331             new_cfqq->service_tree->count == 2 &&
3332             RB_EMPTY_ROOT(&cfqq->sort_list))
3333                 return true;
3334
3335         /*
3336          * So both queues are sync. Let the new request get disk time if
3337          * it's a metadata request and the current queue is doing regular IO.
3338          */
3339         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3340                 return true;
3341
3342         /*
3343          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3344          */
3345         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3346                 return true;
3347
3348         /* An idle queue should not be idle now for some reason */
3349         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3350                 return true;
3351
3352         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3353                 return false;
3354
3355         /*
3356          * if this request is as-good as one we would expect from the
3357          * current cfqq, let it preempt
3358          */
3359         if (cfq_rq_close(cfqd, cfqq, rq))
3360                 return true;
3361
3362         return false;
3363 }
3364
3365 /*
3366  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3367  * let it have half of its nominal slice.
3368  */
3369 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3370 {
3371         struct cfq_queue *old_cfqq = cfqd->active_queue;
3372
3373         cfq_log_cfqq(cfqd, cfqq, "preempt");
3374         cfq_slice_expired(cfqd, 1);
3375
3376         /*
3377          * workload type is changed, don't save slice, otherwise preempt
3378          * doesn't happen
3379          */
3380         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3381                 cfqq->cfqg->saved_workload_slice = 0;
3382
3383         /*
3384          * Put the new queue at the front of the of the current list,
3385          * so we know that it will be selected next.
3386          */
3387         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3388
3389         cfq_service_tree_add(cfqd, cfqq, 1);
3390
3391         cfqq->slice_end = 0;
3392         cfq_mark_cfqq_slice_new(cfqq);
3393 }
3394
3395 /*
3396  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3397  * something we should do about it
3398  */
3399 static void
3400 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3401                 struct request *rq)
3402 {
3403         struct cfq_io_context *cic = RQ_CIC(rq);
3404
3405         cfqd->rq_queued++;
3406         if (rq->cmd_flags & REQ_META)
3407                 cfqq->meta_pending++;
3408
3409         cfq_update_io_thinktime(cfqd, cic);
3410         cfq_update_io_seektime(cfqd, cfqq, rq);
3411         cfq_update_idle_window(cfqd, cfqq, cic);
3412
3413         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3414
3415         if (cfqq == cfqd->active_queue) {
3416                 /*
3417                  * Remember that we saw a request from this process, but
3418                  * don't start queuing just yet. Otherwise we risk seeing lots
3419                  * of tiny requests, because we disrupt the normal plugging
3420                  * and merging. If the request is already larger than a single
3421                  * page, let it rip immediately. For that case we assume that
3422                  * merging is already done. Ditto for a busy system that
3423                  * has other work pending, don't risk delaying until the
3424                  * idle timer unplug to continue working.
3425                  */
3426                 if (cfq_cfqq_wait_request(cfqq)) {
3427                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3428                             cfqd->busy_queues > 1) {
3429                                 cfq_del_timer(cfqd, cfqq);
3430                                 cfq_clear_cfqq_wait_request(cfqq);
3431                                 __blk_run_queue(cfqd->queue);
3432                         } else {
3433                                 cfq_blkiocg_update_idle_time_stats(
3434                                                 &cfqq->cfqg->blkg);
3435                                 cfq_mark_cfqq_must_dispatch(cfqq);
3436                         }
3437                 }
3438         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3439                 /*
3440                  * not the active queue - expire current slice if it is
3441                  * idle and has expired it's mean thinktime or this new queue
3442                  * has some old slice time left and is of higher priority or
3443                  * this new queue is RT and the current one is BE
3444                  */
3445                 cfq_preempt_queue(cfqd, cfqq);
3446                 __blk_run_queue(cfqd->queue);
3447         }
3448 }
3449
3450 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3451 {
3452         struct cfq_data *cfqd = q->elevator->elevator_data;
3453         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3454
3455         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3456         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3457
3458         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3459         list_add_tail(&rq->queuelist, &cfqq->fifo);
3460         cfq_add_rq_rb(rq);
3461         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3462                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3463                         rq_is_sync(rq));
3464         cfq_rq_enqueued(cfqd, cfqq, rq);
3465 }
3466
3467 /*
3468  * Update hw_tag based on peak queue depth over 50 samples under
3469  * sufficient load.
3470  */
3471 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3472 {
3473         struct cfq_queue *cfqq = cfqd->active_queue;
3474
3475         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3476                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3477
3478         if (cfqd->hw_tag == 1)
3479                 return;
3480
3481         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3482             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3483                 return;
3484
3485         /*
3486          * If active queue hasn't enough requests and can idle, cfq might not
3487          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3488          * case
3489          */
3490         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3491             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3492             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3493                 return;
3494
3495         if (cfqd->hw_tag_samples++ < 50)
3496                 return;
3497
3498         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3499                 cfqd->hw_tag = 1;
3500         else
3501                 cfqd->hw_tag = 0;
3502 }
3503
3504 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3505 {
3506         struct cfq_io_context *cic = cfqd->active_cic;
3507
3508         /* If the queue already has requests, don't wait */
3509         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3510                 return false;
3511
3512         /* If there are other queues in the group, don't wait */
3513         if (cfqq->cfqg->nr_cfqq > 1)
3514                 return false;
3515
3516         if (cfq_slice_used(cfqq))
3517                 return true;
3518
3519         /* if slice left is less than think time, wait busy */
3520         if (cic && sample_valid(cic->ttime_samples)
3521             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3522                 return true;
3523
3524         /*
3525          * If think times is less than a jiffy than ttime_mean=0 and above
3526          * will not be true. It might happen that slice has not expired yet
3527          * but will expire soon (4-5 ns) during select_queue(). To cover the
3528          * case where think time is less than a jiffy, mark the queue wait
3529          * busy if only 1 jiffy is left in the slice.
3530          */
3531         if (cfqq->slice_end - jiffies == 1)
3532                 return true;
3533
3534         return false;
3535 }
3536
3537 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3538 {
3539         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3540         struct cfq_data *cfqd = cfqq->cfqd;
3541         const int sync = rq_is_sync(rq);
3542         unsigned long now;
3543
3544         now = jiffies;
3545         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3546                      !!(rq->cmd_flags & REQ_NOIDLE));
3547
3548         cfq_update_hw_tag(cfqd);
3549
3550         WARN_ON(!cfqd->rq_in_driver);
3551         WARN_ON(!cfqq->dispatched);
3552         cfqd->rq_in_driver--;
3553         cfqq->dispatched--;
3554         (RQ_CFQG(rq))->dispatched--;
3555         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3556                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3557                         rq_data_dir(rq), rq_is_sync(rq));
3558
3559         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3560
3561         if (sync) {
3562                 RQ_CIC(rq)->last_end_request = now;
3563                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3564                         cfqd->last_delayed_sync = now;
3565         }
3566
3567         /*
3568          * If this is the active queue, check if it needs to be expired,
3569          * or if we want to idle in case it has no pending requests.
3570          */
3571         if (cfqd->active_queue == cfqq) {
3572                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3573
3574                 if (cfq_cfqq_slice_new(cfqq)) {
3575                         cfq_set_prio_slice(cfqd, cfqq);
3576                         cfq_clear_cfqq_slice_new(cfqq);
3577                 }
3578
3579                 /*
3580                  * Should we wait for next request to come in before we expire
3581                  * the queue.
3582                  */
3583                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3584                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3585                         if (!cfqd->cfq_slice_idle)
3586                                 extend_sl = cfqd->cfq_group_idle;
3587                         cfqq->slice_end = jiffies + extend_sl;
3588                         cfq_mark_cfqq_wait_busy(cfqq);
3589                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3590                 }
3591
3592                 /*
3593                  * Idling is not enabled on:
3594                  * - expired queues
3595                  * - idle-priority queues
3596                  * - async queues
3597                  * - queues with still some requests queued
3598                  * - when there is a close cooperator
3599                  */
3600                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3601                         cfq_slice_expired(cfqd, 1);
3602                 else if (sync && cfqq_empty &&
3603                          !cfq_close_cooperator(cfqd, cfqq)) {
3604                         cfq_arm_slice_timer(cfqd);
3605                 }
3606         }
3607
3608         if (!cfqd->rq_in_driver)
3609                 cfq_schedule_dispatch(cfqd);
3610 }
3611
3612 /*
3613  * we temporarily boost lower priority queues if they are holding fs exclusive
3614  * resources. they are boosted to normal prio (CLASS_BE/4)
3615  */
3616 static void cfq_prio_boost(struct cfq_queue *cfqq)
3617 {
3618         if (has_fs_excl()) {
3619                 /*
3620                  * boost idle prio on transactions that would lock out other
3621                  * users of the filesystem
3622                  */
3623                 if (cfq_class_idle(cfqq))
3624                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3625                 if (cfqq->ioprio > IOPRIO_NORM)
3626                         cfqq->ioprio = IOPRIO_NORM;
3627         } else {
3628                 /*
3629                  * unboost the queue (if needed)
3630                  */
3631                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3632                 cfqq->ioprio = cfqq->org_ioprio;
3633         }
3634 }
3635
3636 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3637 {
3638         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3639                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3640                 return ELV_MQUEUE_MUST;
3641         }
3642
3643         return ELV_MQUEUE_MAY;
3644 }
3645
3646 static int cfq_may_queue(struct request_queue *q, int rw)
3647 {
3648         struct cfq_data *cfqd = q->elevator->elevator_data;
3649         struct task_struct *tsk = current;
3650         struct cfq_io_context *cic;
3651         struct cfq_queue *cfqq;
3652
3653         /*
3654          * don't force setup of a queue from here, as a call to may_queue
3655          * does not necessarily imply that a request actually will be queued.
3656          * so just lookup a possibly existing queue, or return 'may queue'
3657          * if that fails
3658          */
3659         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3660         if (!cic)
3661                 return ELV_MQUEUE_MAY;
3662
3663         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3664         if (cfqq) {
3665                 cfq_init_prio_data(cfqq, cic->ioc);
3666                 cfq_prio_boost(cfqq);
3667
3668                 return __cfq_may_queue(cfqq);
3669         }
3670
3671         return ELV_MQUEUE_MAY;
3672 }
3673
3674 /*
3675  * queue lock held here
3676  */
3677 static void cfq_put_request(struct request *rq)
3678 {
3679         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3680
3681         if (cfqq) {
3682                 const int rw = rq_data_dir(rq);
3683
3684                 BUG_ON(!cfqq->allocated[rw]);
3685                 cfqq->allocated[rw]--;
3686
3687                 put_io_context(RQ_CIC(rq)->ioc);
3688
3689                 rq->elevator_private[0] = NULL;
3690                 rq->elevator_private[1] = NULL;
3691
3692                 /* Put down rq reference on cfqg */
3693                 cfq_put_cfqg(RQ_CFQG(rq));
3694                 rq->elevator_private[2] = NULL;
3695
3696                 cfq_put_queue(cfqq);
3697         }
3698 }
3699
3700 static struct cfq_queue *
3701 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3702                 struct cfq_queue *cfqq)
3703 {
3704         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3705         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3706         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3707         cfq_put_queue(cfqq);
3708         return cic_to_cfqq(cic, 1);
3709 }
3710
3711 /*
3712  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3713  * was the last process referring to said cfqq.
3714  */
3715 static struct cfq_queue *
3716 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3717 {
3718         if (cfqq_process_refs(cfqq) == 1) {
3719                 cfqq->pid = current->pid;
3720                 cfq_clear_cfqq_coop(cfqq);
3721                 cfq_clear_cfqq_split_coop(cfqq);
3722                 return cfqq;
3723         }
3724
3725         cic_set_cfqq(cic, NULL, 1);
3726
3727         cfq_put_cooperator(cfqq);
3728
3729         cfq_put_queue(cfqq);
3730         return NULL;
3731 }
3732 /*
3733  * Allocate cfq data structures associated with this request.
3734  */
3735 static int
3736 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3737 {
3738         struct cfq_data *cfqd = q->elevator->elevator_data;
3739         struct cfq_io_context *cic;
3740         const int rw = rq_data_dir(rq);
3741         const bool is_sync = rq_is_sync(rq);
3742         struct cfq_queue *cfqq;
3743         unsigned long flags;
3744
3745         might_sleep_if(gfp_mask & __GFP_WAIT);
3746
3747         cic = cfq_get_io_context(cfqd, gfp_mask);
3748
3749         spin_lock_irqsave(q->queue_lock, flags);
3750
3751         if (!cic)
3752                 goto queue_fail;
3753
3754 new_queue:
3755         cfqq = cic_to_cfqq(cic, is_sync);
3756         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3757                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3758                 cic_set_cfqq(cic, cfqq, is_sync);
3759         } else {
3760                 /*
3761                  * If the queue was seeky for too long, break it apart.
3762                  */
3763                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3764                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3765                         cfqq = split_cfqq(cic, cfqq);
3766                         if (!cfqq)
3767                                 goto new_queue;
3768                 }
3769
3770                 /*
3771                  * Check to see if this queue is scheduled to merge with
3772                  * another, closely cooperating queue.  The merging of
3773                  * queues happens here as it must be done in process context.
3774                  * The reference on new_cfqq was taken in merge_cfqqs.
3775                  */
3776                 if (cfqq->new_cfqq)
3777                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3778         }
3779
3780         cfqq->allocated[rw]++;
3781
3782         cfqq->ref++;
3783         rq->elevator_private[0] = cic;
3784         rq->elevator_private[1] = cfqq;
3785         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3786         spin_unlock_irqrestore(q->queue_lock, flags);
3787         return 0;
3788
3789 queue_fail:
3790         if (cic)
3791                 put_io_context(cic->ioc);
3792
3793         cfq_schedule_dispatch(cfqd);
3794         spin_unlock_irqrestore(q->queue_lock, flags);
3795         cfq_log(cfqd, "set_request fail");
3796         return 1;
3797 }
3798
3799 static void cfq_kick_queue(struct work_struct *work)
3800 {
3801         struct cfq_data *cfqd =
3802                 container_of(work, struct cfq_data, unplug_work);
3803         struct request_queue *q = cfqd->queue;
3804
3805         spin_lock_irq(q->queue_lock);
3806         __blk_run_queue(cfqd->queue);
3807         spin_unlock_irq(q->queue_lock);
3808 }
3809
3810 /*
3811  * Timer running if the active_queue is currently idling inside its time slice
3812  */
3813 static void cfq_idle_slice_timer(unsigned long data)
3814 {
3815         struct cfq_data *cfqd = (struct cfq_data *) data;
3816         struct cfq_queue *cfqq;
3817         unsigned long flags;
3818         int timed_out = 1;
3819
3820         cfq_log(cfqd, "idle timer fired");
3821
3822         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3823
3824         cfqq = cfqd->active_queue;
3825         if (cfqq) {
3826                 timed_out = 0;
3827
3828                 /*
3829                  * We saw a request before the queue expired, let it through
3830                  */
3831                 if (cfq_cfqq_must_dispatch(cfqq))
3832                         goto out_kick;
3833
3834                 /*
3835                  * expired
3836                  */
3837                 if (cfq_slice_used(cfqq))
3838                         goto expire;
3839
3840                 /*
3841                  * only expire and reinvoke request handler, if there are
3842                  * other queues with pending requests
3843                  */
3844                 if (!cfqd->busy_queues)
3845                         goto out_cont;
3846
3847                 /*
3848                  * not expired and it has a request pending, let it dispatch
3849                  */
3850                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3851                         goto out_kick;
3852
3853                 /*
3854                  * Queue depth flag is reset only when the idle didn't succeed
3855                  */
3856                 cfq_clear_cfqq_deep(cfqq);
3857         }
3858 expire:
3859         cfq_slice_expired(cfqd, timed_out);
3860 out_kick:
3861         cfq_schedule_dispatch(cfqd);
3862 out_cont:
3863         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3864 }
3865
3866 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3867 {
3868         del_timer_sync(&cfqd->idle_slice_timer);
3869         cancel_work_sync(&cfqd->unplug_work);
3870 }
3871
3872 static void cfq_put_async_queues(struct cfq_data *cfqd)
3873 {
3874         int i;
3875
3876         for (i = 0; i < IOPRIO_BE_NR; i++) {
3877                 if (cfqd->async_cfqq[0][i])
3878                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3879                 if (cfqd->async_cfqq[1][i])
3880                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3881         }
3882
3883         if (cfqd->async_idle_cfqq)
3884                 cfq_put_queue(cfqd->async_idle_cfqq);
3885 }
3886
3887 static void cfq_exit_queue(struct elevator_queue *e)
3888 {
3889         struct cfq_data *cfqd = e->elevator_data;
3890         struct request_queue *q = cfqd->queue;
3891         bool wait = false;
3892
3893         cfq_shutdown_timer_wq(cfqd);
3894
3895         spin_lock_irq(q->queue_lock);
3896
3897         if (cfqd->active_queue)
3898                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3899
3900         while (!list_empty(&cfqd->cic_list)) {
3901                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3902                                                         struct cfq_io_context,
3903                                                         queue_list);
3904
3905                 __cfq_exit_single_io_context(cfqd, cic);
3906         }
3907
3908         cfq_put_async_queues(cfqd);
3909         cfq_release_cfq_groups(cfqd);
3910
3911         /*
3912          * If there are groups which we could not unlink from blkcg list,
3913          * wait for a rcu period for them to be freed.
3914          */
3915         if (cfqd->nr_blkcg_linked_grps)
3916                 wait = true;
3917
3918         spin_unlock_irq(q->queue_lock);
3919
3920         cfq_shutdown_timer_wq(cfqd);
3921
3922         spin_lock(&cic_index_lock);
3923         ida_remove(&cic_index_ida, cfqd->cic_index);
3924         spin_unlock(&cic_index_lock);
3925
3926         /*
3927          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3928          * Do this wait only if there are other unlinked groups out
3929          * there. This can happen if cgroup deletion path claimed the
3930          * responsibility of cleaning up a group before queue cleanup code
3931          * get to the group.
3932          *
3933          * Do not call synchronize_rcu() unconditionally as there are drivers
3934          * which create/delete request queue hundreds of times during scan/boot
3935          * and synchronize_rcu() can take significant time and slow down boot.
3936          */
3937         if (wait)
3938                 synchronize_rcu();
3939
3940 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3941         /* Free up per cpu stats for root group */
3942         free_percpu(cfqd->root_group.blkg.stats_cpu);
3943 #endif
3944         kfree(cfqd);
3945 }
3946
3947 static int cfq_alloc_cic_index(void)
3948 {
3949         int index, error;
3950
3951         do {
3952                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3953                         return -ENOMEM;
3954
3955                 spin_lock(&cic_index_lock);
3956                 error = ida_get_new(&cic_index_ida, &index);
3957                 spin_unlock(&cic_index_lock);
3958                 if (error && error != -EAGAIN)
3959                         return error;
3960         } while (error);
3961
3962         return index;
3963 }
3964
3965 static void *cfq_init_queue(struct request_queue *q)
3966 {
3967         struct cfq_data *cfqd;
3968         int i, j;
3969         struct cfq_group *cfqg;
3970         struct cfq_rb_root *st;
3971
3972         i = cfq_alloc_cic_index();
3973         if (i < 0)
3974                 return NULL;
3975
3976         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3977         if (!cfqd)
3978                 return NULL;
3979
3980         /*
3981          * Don't need take queue_lock in the routine, since we are
3982          * initializing the ioscheduler, and nobody is using cfqd
3983          */
3984         cfqd->cic_index = i;
3985
3986         /* Init root service tree */
3987         cfqd->grp_service_tree = CFQ_RB_ROOT;
3988
3989         /* Init root group */
3990         cfqg = &cfqd->root_group;
3991         for_each_cfqg_st(cfqg, i, j, st)
3992                 *st = CFQ_RB_ROOT;
3993         RB_CLEAR_NODE(&cfqg->rb_node);
3994
3995         /* Give preference to root group over other groups */
3996         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3997
3998 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3999         /*
4000          * Set root group reference to 2. One reference will be dropped when
4001          * all groups on cfqd->cfqg_list are being deleted during queue exit.
4002          * Other reference will remain there as we don't want to delete this
4003          * group as it is statically allocated and gets destroyed when
4004          * throtl_data goes away.
4005          */
4006         cfqg->ref = 2;
4007
4008         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4009                 kfree(cfqg);
4010                 kfree(cfqd);
4011                 return NULL;
4012         }
4013
4014         rcu_read_lock();
4015
4016         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4017                                         (void *)cfqd, 0);
4018         rcu_read_unlock();
4019         cfqd->nr_blkcg_linked_grps++;
4020
4021         /* Add group on cfqd->cfqg_list */
4022         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4023 #endif
4024         /*
4025          * Not strictly needed (since RB_ROOT just clears the node and we
4026          * zeroed cfqd on alloc), but better be safe in case someone decides
4027          * to add magic to the rb code
4028          */
4029         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4030                 cfqd->prio_trees[i] = RB_ROOT;
4031
4032         /*
4033          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4034          * Grab a permanent reference to it, so that the normal code flow
4035          * will not attempt to free it.
4036          */
4037         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4038         cfqd->oom_cfqq.ref++;
4039         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4040
4041         INIT_LIST_HEAD(&cfqd->cic_list);
4042
4043         cfqd->queue = q;
4044
4045         init_timer(&cfqd->idle_slice_timer);
4046         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4047         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4048
4049         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4050
4051         cfqd->cfq_quantum = cfq_quantum;
4052         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4053         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4054         cfqd->cfq_back_max = cfq_back_max;
4055         cfqd->cfq_back_penalty = cfq_back_penalty;
4056         cfqd->cfq_slice[0] = cfq_slice_async;
4057         cfqd->cfq_slice[1] = cfq_slice_sync;
4058         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4059         cfqd->cfq_slice_idle = cfq_slice_idle;
4060         cfqd->cfq_group_idle = cfq_group_idle;
4061         cfqd->cfq_latency = 1;
4062         cfqd->hw_tag = -1;
4063         /*
4064          * we optimistically start assuming sync ops weren't delayed in last
4065          * second, in order to have larger depth for async operations.
4066          */
4067         cfqd->last_delayed_sync = jiffies - HZ;
4068         return cfqd;
4069 }
4070
4071 static void cfq_slab_kill(void)
4072 {
4073         /*
4074          * Caller already ensured that pending RCU callbacks are completed,
4075          * so we should have no busy allocations at this point.
4076          */
4077         if (cfq_pool)
4078                 kmem_cache_destroy(cfq_pool);
4079         if (cfq_ioc_pool)
4080                 kmem_cache_destroy(cfq_ioc_pool);
4081 }
4082
4083 static int __init cfq_slab_setup(void)
4084 {
4085         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4086         if (!cfq_pool)
4087                 goto fail;
4088
4089         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4090         if (!cfq_ioc_pool)
4091                 goto fail;
4092
4093         return 0;
4094 fail:
4095         cfq_slab_kill();
4096         return -ENOMEM;
4097 }
4098
4099 /*
4100  * sysfs parts below -->
4101  */
4102 static ssize_t
4103 cfq_var_show(unsigned int var, char *page)
4104 {
4105         return sprintf(page, "%d\n", var);
4106 }
4107
4108 static ssize_t
4109 cfq_var_store(unsigned int *var, const char *page, size_t count)
4110 {
4111         char *p = (char *) page;
4112
4113         *var = simple_strtoul(p, &p, 10);
4114         return count;
4115 }
4116
4117 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4118 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4119 {                                                                       \
4120         struct cfq_data *cfqd = e->elevator_data;                       \
4121         unsigned int __data = __VAR;                                    \
4122         if (__CONV)                                                     \
4123                 __data = jiffies_to_msecs(__data);                      \
4124         return cfq_var_show(__data, (page));                            \
4125 }
4126 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4127 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4128 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4129 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4130 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4131 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4132 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4133 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4134 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4135 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4136 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4137 #undef SHOW_FUNCTION
4138
4139 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4140 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4141 {                                                                       \
4142         struct cfq_data *cfqd = e->elevator_data;                       \
4143         unsigned int __data;                                            \
4144         int ret = cfq_var_store(&__data, (page), count);                \
4145         if (__data < (MIN))                                             \
4146                 __data = (MIN);                                         \
4147         else if (__data > (MAX))                                        \
4148                 __data = (MAX);                                         \
4149         if (__CONV)                                                     \
4150                 *(__PTR) = msecs_to_jiffies(__data);                    \
4151         else                                                            \
4152                 *(__PTR) = __data;                                      \
4153         return ret;                                                     \
4154 }
4155 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4156 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4157                 UINT_MAX, 1);
4158 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4159                 UINT_MAX, 1);
4160 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4161 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4162                 UINT_MAX, 0);
4163 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4164 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4165 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4166 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4167 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4168                 UINT_MAX, 0);
4169 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4170 #undef STORE_FUNCTION
4171
4172 #define CFQ_ATTR(name) \
4173         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4174
4175 static struct elv_fs_entry cfq_attrs[] = {
4176         CFQ_ATTR(quantum),
4177         CFQ_ATTR(fifo_expire_sync),
4178         CFQ_ATTR(fifo_expire_async),
4179         CFQ_ATTR(back_seek_max),
4180         CFQ_ATTR(back_seek_penalty),
4181         CFQ_ATTR(slice_sync),
4182         CFQ_ATTR(slice_async),
4183         CFQ_ATTR(slice_async_rq),
4184         CFQ_ATTR(slice_idle),
4185         CFQ_ATTR(group_idle),
4186         CFQ_ATTR(low_latency),
4187         __ATTR_NULL
4188 };
4189
4190 static struct elevator_type iosched_cfq = {
4191         .ops = {
4192                 .elevator_merge_fn =            cfq_merge,
4193                 .elevator_merged_fn =           cfq_merged_request,
4194                 .elevator_merge_req_fn =        cfq_merged_requests,
4195                 .elevator_allow_merge_fn =      cfq_allow_merge,
4196                 .elevator_bio_merged_fn =       cfq_bio_merged,
4197                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4198                 .elevator_add_req_fn =          cfq_insert_request,
4199                 .elevator_activate_req_fn =     cfq_activate_request,
4200                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4201                 .elevator_completed_req_fn =    cfq_completed_request,
4202                 .elevator_former_req_fn =       elv_rb_former_request,
4203                 .elevator_latter_req_fn =       elv_rb_latter_request,
4204                 .elevator_set_req_fn =          cfq_set_request,
4205                 .elevator_put_req_fn =          cfq_put_request,
4206                 .elevator_may_queue_fn =        cfq_may_queue,
4207                 .elevator_init_fn =             cfq_init_queue,
4208                 .elevator_exit_fn =             cfq_exit_queue,
4209                 .trim =                         cfq_free_io_context,
4210         },
4211         .elevator_attrs =       cfq_attrs,
4212         .elevator_name =        "cfq",
4213         .elevator_owner =       THIS_MODULE,
4214 };
4215
4216 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4217 static struct blkio_policy_type blkio_policy_cfq = {
4218         .ops = {
4219                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4220                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4221         },
4222         .plid = BLKIO_POLICY_PROP,
4223 };
4224 #else
4225 static struct blkio_policy_type blkio_policy_cfq;
4226 #endif
4227
4228 static int __init cfq_init(void)
4229 {
4230         /*
4231          * could be 0 on HZ < 1000 setups
4232          */
4233         if (!cfq_slice_async)
4234                 cfq_slice_async = 1;
4235         if (!cfq_slice_idle)
4236                 cfq_slice_idle = 1;
4237
4238 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4239         if (!cfq_group_idle)
4240                 cfq_group_idle = 1;
4241 #else
4242                 cfq_group_idle = 0;
4243 #endif
4244         if (cfq_slab_setup())
4245                 return -ENOMEM;
4246
4247         elv_register(&iosched_cfq);
4248         blkio_policy_register(&blkio_policy_cfq);
4249
4250         return 0;
4251 }
4252
4253 static void __exit cfq_exit(void)
4254 {
4255         DECLARE_COMPLETION_ONSTACK(all_gone);
4256         blkio_policy_unregister(&blkio_policy_cfq);
4257         elv_unregister(&iosched_cfq);
4258         ioc_gone = &all_gone;
4259         /* ioc_gone's update must be visible before reading ioc_count */
4260         smp_wmb();
4261
4262         /*
4263          * this also protects us from entering cfq_slab_kill() with
4264          * pending RCU callbacks
4265          */
4266         if (elv_ioc_count_read(cfq_ioc_count))
4267                 wait_for_completion(&all_gone);
4268         ida_destroy(&cic_index_ida);
4269         cfq_slab_kill();
4270 }
4271
4272 module_init(cfq_init);
4273 module_exit(cfq_exit);
4274
4275 MODULE_AUTHOR("Jens Axboe");
4276 MODULE_LICENSE("GPL");
4277 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");