CFQ: move think time check variables to a separate struct
[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         /* number of requests that are on the dispatch list or inside driver */
133         int dispatched;
134
135         /* io prio of this group */
136         unsigned short ioprio, org_ioprio;
137         unsigned short ioprio_class;
138
139         pid_t pid;
140
141         u32 seek_history;
142         sector_t last_request_pos;
143
144         struct cfq_rb_root *service_tree;
145         struct cfq_queue *new_cfqq;
146         struct cfq_group *cfqg;
147         /* Number of sectors dispatched from queue in single dispatch round */
148         unsigned long nr_sectors;
149 };
150
151 /*
152  * First index in the service_trees.
153  * IDLE is handled separately, so it has negative index
154  */
155 enum wl_prio_t {
156         BE_WORKLOAD = 0,
157         RT_WORKLOAD = 1,
158         IDLE_WORKLOAD = 2,
159         CFQ_PRIO_NR,
160 };
161
162 /*
163  * Second index in the service_trees.
164  */
165 enum wl_type_t {
166         ASYNC_WORKLOAD = 0,
167         SYNC_NOIDLE_WORKLOAD = 1,
168         SYNC_WORKLOAD = 2
169 };
170
171 /* This is per cgroup per device grouping structure */
172 struct cfq_group {
173         /* group service_tree member */
174         struct rb_node rb_node;
175
176         /* group service_tree key */
177         u64 vdisktime;
178         unsigned int weight;
179         unsigned int new_weight;
180         bool needs_update;
181
182         /* number of cfqq currently on this group */
183         int nr_cfqq;
184
185         /*
186          * Per group busy queues average. Useful for workload slice calc. We
187          * create the array for each prio class but at run time it is used
188          * only for RT and BE class and slot for IDLE class remains unused.
189          * This is primarily done to avoid confusion and a gcc warning.
190          */
191         unsigned int busy_queues_avg[CFQ_PRIO_NR];
192         /*
193          * rr lists of queues with requests. We maintain service trees for
194          * RT and BE classes. These trees are subdivided in subclasses
195          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196          * class there is no subclassification and all the cfq queues go on
197          * a single tree service_tree_idle.
198          * Counts are embedded in the cfq_rb_root
199          */
200         struct cfq_rb_root service_trees[2][3];
201         struct cfq_rb_root service_tree_idle;
202
203         unsigned long saved_workload_slice;
204         enum wl_type_t saved_workload;
205         enum wl_prio_t saved_serving_prio;
206         struct blkio_group blkg;
207 #ifdef CONFIG_CFQ_GROUP_IOSCHED
208         struct hlist_node cfqd_node;
209         int ref;
210 #endif
211         /* number of requests that are on the dispatch list or inside driver */
212         int dispatched;
213 };
214
215 /*
216  * Per block device queue structure
217  */
218 struct cfq_data {
219         struct request_queue *queue;
220         /* Root service tree for cfq_groups */
221         struct cfq_rb_root grp_service_tree;
222         struct cfq_group root_group;
223
224         /*
225          * The priority currently being served
226          */
227         enum wl_prio_t serving_prio;
228         enum wl_type_t serving_type;
229         unsigned long workload_expires;
230         struct cfq_group *serving_group;
231
232         /*
233          * Each priority tree is sorted by next_request position.  These
234          * trees are used when determining if two or more queues are
235          * interleaving requests (see cfq_close_cooperator).
236          */
237         struct rb_root prio_trees[CFQ_PRIO_LISTS];
238
239         unsigned int busy_queues;
240         unsigned int busy_sync_queues;
241
242         int rq_in_driver;
243         int rq_in_flight[2];
244
245         /*
246          * queue-depth detection
247          */
248         int rq_queued;
249         int hw_tag;
250         /*
251          * hw_tag can be
252          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
253          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
254          *  0 => no NCQ
255          */
256         int hw_tag_est_depth;
257         unsigned int hw_tag_samples;
258
259         /*
260          * idle window management
261          */
262         struct timer_list idle_slice_timer;
263         struct work_struct unplug_work;
264
265         struct cfq_queue *active_queue;
266         struct cfq_io_context *active_cic;
267
268         /*
269          * async queue for each priority case
270          */
271         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
272         struct cfq_queue *async_idle_cfqq;
273
274         sector_t last_position;
275
276         /*
277          * tunables, see top of file
278          */
279         unsigned int cfq_quantum;
280         unsigned int cfq_fifo_expire[2];
281         unsigned int cfq_back_penalty;
282         unsigned int cfq_back_max;
283         unsigned int cfq_slice[2];
284         unsigned int cfq_slice_async_rq;
285         unsigned int cfq_slice_idle;
286         unsigned int cfq_group_idle;
287         unsigned int cfq_latency;
288
289         unsigned int cic_index;
290         struct list_head cic_list;
291
292         /*
293          * Fallback dummy cfqq for extreme OOM conditions
294          */
295         struct cfq_queue oom_cfqq;
296
297         unsigned long last_delayed_sync;
298
299         /* List of cfq groups being managed on this device*/
300         struct hlist_head cfqg_list;
301
302         /* Number of groups which are on blkcg->blkg_list */
303         unsigned int nr_blkcg_linked_grps;
304 };
305
306 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
307
308 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
309                                             enum wl_prio_t prio,
310                                             enum wl_type_t type)
311 {
312         if (!cfqg)
313                 return NULL;
314
315         if (prio == IDLE_WORKLOAD)
316                 return &cfqg->service_tree_idle;
317
318         return &cfqg->service_trees[prio][type];
319 }
320
321 enum cfqq_state_flags {
322         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
323         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
324         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
325         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
326         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
327         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
328         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
329         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
330         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
331         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
332         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
333         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
334         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
335 };
336
337 #define CFQ_CFQQ_FNS(name)                                              \
338 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
339 {                                                                       \
340         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
341 }                                                                       \
342 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
343 {                                                                       \
344         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
345 }                                                                       \
346 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
347 {                                                                       \
348         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
349 }
350
351 CFQ_CFQQ_FNS(on_rr);
352 CFQ_CFQQ_FNS(wait_request);
353 CFQ_CFQQ_FNS(must_dispatch);
354 CFQ_CFQQ_FNS(must_alloc_slice);
355 CFQ_CFQQ_FNS(fifo_expire);
356 CFQ_CFQQ_FNS(idle_window);
357 CFQ_CFQQ_FNS(prio_changed);
358 CFQ_CFQQ_FNS(slice_new);
359 CFQ_CFQQ_FNS(sync);
360 CFQ_CFQQ_FNS(coop);
361 CFQ_CFQQ_FNS(split_coop);
362 CFQ_CFQQ_FNS(deep);
363 CFQ_CFQQ_FNS(wait_busy);
364 #undef CFQ_CFQQ_FNS
365
366 #ifdef CONFIG_CFQ_GROUP_IOSCHED
367 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
368         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
369                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
370                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
371
372 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
373         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
374                                 blkg_path(&(cfqg)->blkg), ##args)       \
375
376 #else
377 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
378         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
380 #endif
381 #define cfq_log(cfqd, fmt, args...)     \
382         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
383
384 /* Traverses through cfq group service trees */
385 #define for_each_cfqg_st(cfqg, i, j, st) \
386         for (i = 0; i <= IDLE_WORKLOAD; i++) \
387                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
388                         : &cfqg->service_tree_idle; \
389                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
390                         (i == IDLE_WORKLOAD && j == 0); \
391                         j++, st = i < IDLE_WORKLOAD ? \
392                         &cfqg->service_trees[i][j]: NULL) \
393
394
395 static inline bool iops_mode(struct cfq_data *cfqd)
396 {
397         /*
398          * If we are not idling on queues and it is a NCQ drive, parallel
399          * execution of requests is on and measuring time is not possible
400          * in most of the cases until and unless we drive shallower queue
401          * depths and that becomes a performance bottleneck. In such cases
402          * switch to start providing fairness in terms of number of IOs.
403          */
404         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
405                 return true;
406         else
407                 return false;
408 }
409
410 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
411 {
412         if (cfq_class_idle(cfqq))
413                 return IDLE_WORKLOAD;
414         if (cfq_class_rt(cfqq))
415                 return RT_WORKLOAD;
416         return BE_WORKLOAD;
417 }
418
419
420 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
421 {
422         if (!cfq_cfqq_sync(cfqq))
423                 return ASYNC_WORKLOAD;
424         if (!cfq_cfqq_idle_window(cfqq))
425                 return SYNC_NOIDLE_WORKLOAD;
426         return SYNC_WORKLOAD;
427 }
428
429 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
430                                         struct cfq_data *cfqd,
431                                         struct cfq_group *cfqg)
432 {
433         if (wl == IDLE_WORKLOAD)
434                 return cfqg->service_tree_idle.count;
435
436         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
437                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
438                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
439 }
440
441 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
442                                         struct cfq_group *cfqg)
443 {
444         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
445                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
446 }
447
448 static void cfq_dispatch_insert(struct request_queue *, struct request *);
449 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
450                                        struct io_context *, gfp_t);
451 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
452                                                 struct io_context *);
453
454 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
455                                             bool is_sync)
456 {
457         return cic->cfqq[is_sync];
458 }
459
460 static inline void cic_set_cfqq(struct cfq_io_context *cic,
461                                 struct cfq_queue *cfqq, bool is_sync)
462 {
463         cic->cfqq[is_sync] = cfqq;
464 }
465
466 #define CIC_DEAD_KEY    1ul
467 #define CIC_DEAD_INDEX_SHIFT    1
468
469 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
470 {
471         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
472 }
473
474 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
475 {
476         struct cfq_data *cfqd = cic->key;
477
478         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
479                 return NULL;
480
481         return cfqd;
482 }
483
484 /*
485  * We regard a request as SYNC, if it's either a read or has the SYNC bit
486  * set (in which case it could also be direct WRITE).
487  */
488 static inline bool cfq_bio_sync(struct bio *bio)
489 {
490         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
491 }
492
493 /*
494  * scheduler run of queue, if there are requests pending and no one in the
495  * driver that will restart queueing
496  */
497 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
498 {
499         if (cfqd->busy_queues) {
500                 cfq_log(cfqd, "schedule dispatch");
501                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
502         }
503 }
504
505 /*
506  * Scale schedule slice based on io priority. Use the sync time slice only
507  * if a queue is marked sync and has sync io queued. A sync queue with async
508  * io only, should not get full sync slice length.
509  */
510 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
511                                  unsigned short prio)
512 {
513         const int base_slice = cfqd->cfq_slice[sync];
514
515         WARN_ON(prio >= IOPRIO_BE_NR);
516
517         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
518 }
519
520 static inline int
521 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
522 {
523         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
524 }
525
526 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
527 {
528         u64 d = delta << CFQ_SERVICE_SHIFT;
529
530         d = d * BLKIO_WEIGHT_DEFAULT;
531         do_div(d, cfqg->weight);
532         return d;
533 }
534
535 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
536 {
537         s64 delta = (s64)(vdisktime - min_vdisktime);
538         if (delta > 0)
539                 min_vdisktime = vdisktime;
540
541         return min_vdisktime;
542 }
543
544 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
545 {
546         s64 delta = (s64)(vdisktime - min_vdisktime);
547         if (delta < 0)
548                 min_vdisktime = vdisktime;
549
550         return min_vdisktime;
551 }
552
553 static void update_min_vdisktime(struct cfq_rb_root *st)
554 {
555         struct cfq_group *cfqg;
556
557         if (st->left) {
558                 cfqg = rb_entry_cfqg(st->left);
559                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
560                                                   cfqg->vdisktime);
561         }
562 }
563
564 /*
565  * get averaged number of queues of RT/BE priority.
566  * average is updated, with a formula that gives more weight to higher numbers,
567  * to quickly follows sudden increases and decrease slowly
568  */
569
570 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
571                                         struct cfq_group *cfqg, bool rt)
572 {
573         unsigned min_q, max_q;
574         unsigned mult  = cfq_hist_divisor - 1;
575         unsigned round = cfq_hist_divisor / 2;
576         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
577
578         min_q = min(cfqg->busy_queues_avg[rt], busy);
579         max_q = max(cfqg->busy_queues_avg[rt], busy);
580         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
581                 cfq_hist_divisor;
582         return cfqg->busy_queues_avg[rt];
583 }
584
585 static inline unsigned
586 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
587 {
588         struct cfq_rb_root *st = &cfqd->grp_service_tree;
589
590         return cfq_target_latency * cfqg->weight / st->total_weight;
591 }
592
593 static inline unsigned
594 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
595 {
596         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
597         if (cfqd->cfq_latency) {
598                 /*
599                  * interested queues (we consider only the ones with the same
600                  * priority class in the cfq group)
601                  */
602                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
603                                                 cfq_class_rt(cfqq));
604                 unsigned sync_slice = cfqd->cfq_slice[1];
605                 unsigned expect_latency = sync_slice * iq;
606                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
607
608                 if (expect_latency > group_slice) {
609                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
610                         /* scale low_slice according to IO priority
611                          * and sync vs async */
612                         unsigned low_slice =
613                                 min(slice, base_low_slice * slice / sync_slice);
614                         /* the adapted slice value is scaled to fit all iqs
615                          * into the target latency */
616                         slice = max(slice * group_slice / expect_latency,
617                                     low_slice);
618                 }
619         }
620         return slice;
621 }
622
623 static inline void
624 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
625 {
626         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
627
628         cfqq->slice_start = jiffies;
629         cfqq->slice_end = jiffies + slice;
630         cfqq->allocated_slice = slice;
631         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
632 }
633
634 /*
635  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
636  * isn't valid until the first request from the dispatch is activated
637  * and the slice time set.
638  */
639 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
640 {
641         if (cfq_cfqq_slice_new(cfqq))
642                 return false;
643         if (time_before(jiffies, cfqq->slice_end))
644                 return false;
645
646         return true;
647 }
648
649 /*
650  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
651  * We choose the request that is closest to the head right now. Distance
652  * behind the head is penalized and only allowed to a certain extent.
653  */
654 static struct request *
655 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
656 {
657         sector_t s1, s2, d1 = 0, d2 = 0;
658         unsigned long back_max;
659 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
660 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
661         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
662
663         if (rq1 == NULL || rq1 == rq2)
664                 return rq2;
665         if (rq2 == NULL)
666                 return rq1;
667
668         if (rq_is_sync(rq1) != rq_is_sync(rq2))
669                 return rq_is_sync(rq1) ? rq1 : rq2;
670
671         s1 = blk_rq_pos(rq1);
672         s2 = blk_rq_pos(rq2);
673
674         /*
675          * by definition, 1KiB is 2 sectors
676          */
677         back_max = cfqd->cfq_back_max * 2;
678
679         /*
680          * Strict one way elevator _except_ in the case where we allow
681          * short backward seeks which are biased as twice the cost of a
682          * similar forward seek.
683          */
684         if (s1 >= last)
685                 d1 = s1 - last;
686         else if (s1 + back_max >= last)
687                 d1 = (last - s1) * cfqd->cfq_back_penalty;
688         else
689                 wrap |= CFQ_RQ1_WRAP;
690
691         if (s2 >= last)
692                 d2 = s2 - last;
693         else if (s2 + back_max >= last)
694                 d2 = (last - s2) * cfqd->cfq_back_penalty;
695         else
696                 wrap |= CFQ_RQ2_WRAP;
697
698         /* Found required data */
699
700         /*
701          * By doing switch() on the bit mask "wrap" we avoid having to
702          * check two variables for all permutations: --> faster!
703          */
704         switch (wrap) {
705         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
706                 if (d1 < d2)
707                         return rq1;
708                 else if (d2 < d1)
709                         return rq2;
710                 else {
711                         if (s1 >= s2)
712                                 return rq1;
713                         else
714                                 return rq2;
715                 }
716
717         case CFQ_RQ2_WRAP:
718                 return rq1;
719         case CFQ_RQ1_WRAP:
720                 return rq2;
721         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
722         default:
723                 /*
724                  * Since both rqs are wrapped,
725                  * start with the one that's further behind head
726                  * (--> only *one* back seek required),
727                  * since back seek takes more time than forward.
728                  */
729                 if (s1 <= s2)
730                         return rq1;
731                 else
732                         return rq2;
733         }
734 }
735
736 /*
737  * The below is leftmost cache rbtree addon
738  */
739 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
740 {
741         /* Service tree is empty */
742         if (!root->count)
743                 return NULL;
744
745         if (!root->left)
746                 root->left = rb_first(&root->rb);
747
748         if (root->left)
749                 return rb_entry(root->left, struct cfq_queue, rb_node);
750
751         return NULL;
752 }
753
754 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
755 {
756         if (!root->left)
757                 root->left = rb_first(&root->rb);
758
759         if (root->left)
760                 return rb_entry_cfqg(root->left);
761
762         return NULL;
763 }
764
765 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
766 {
767         rb_erase(n, root);
768         RB_CLEAR_NODE(n);
769 }
770
771 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
772 {
773         if (root->left == n)
774                 root->left = NULL;
775         rb_erase_init(n, &root->rb);
776         --root->count;
777 }
778
779 /*
780  * would be nice to take fifo expire time into account as well
781  */
782 static struct request *
783 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
784                   struct request *last)
785 {
786         struct rb_node *rbnext = rb_next(&last->rb_node);
787         struct rb_node *rbprev = rb_prev(&last->rb_node);
788         struct request *next = NULL, *prev = NULL;
789
790         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
791
792         if (rbprev)
793                 prev = rb_entry_rq(rbprev);
794
795         if (rbnext)
796                 next = rb_entry_rq(rbnext);
797         else {
798                 rbnext = rb_first(&cfqq->sort_list);
799                 if (rbnext && rbnext != &last->rb_node)
800                         next = rb_entry_rq(rbnext);
801         }
802
803         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
804 }
805
806 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
807                                       struct cfq_queue *cfqq)
808 {
809         /*
810          * just an approximation, should be ok.
811          */
812         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
813                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
814 }
815
816 static inline s64
817 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
818 {
819         return cfqg->vdisktime - st->min_vdisktime;
820 }
821
822 static void
823 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
824 {
825         struct rb_node **node = &st->rb.rb_node;
826         struct rb_node *parent = NULL;
827         struct cfq_group *__cfqg;
828         s64 key = cfqg_key(st, cfqg);
829         int left = 1;
830
831         while (*node != NULL) {
832                 parent = *node;
833                 __cfqg = rb_entry_cfqg(parent);
834
835                 if (key < cfqg_key(st, __cfqg))
836                         node = &parent->rb_left;
837                 else {
838                         node = &parent->rb_right;
839                         left = 0;
840                 }
841         }
842
843         if (left)
844                 st->left = &cfqg->rb_node;
845
846         rb_link_node(&cfqg->rb_node, parent, node);
847         rb_insert_color(&cfqg->rb_node, &st->rb);
848 }
849
850 static void
851 cfq_update_group_weight(struct cfq_group *cfqg)
852 {
853         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
854         if (cfqg->needs_update) {
855                 cfqg->weight = cfqg->new_weight;
856                 cfqg->needs_update = false;
857         }
858 }
859
860 static void
861 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
862 {
863         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
864
865         cfq_update_group_weight(cfqg);
866         __cfq_group_service_tree_add(st, cfqg);
867         st->total_weight += cfqg->weight;
868 }
869
870 static void
871 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
872 {
873         struct cfq_rb_root *st = &cfqd->grp_service_tree;
874         struct cfq_group *__cfqg;
875         struct rb_node *n;
876
877         cfqg->nr_cfqq++;
878         if (!RB_EMPTY_NODE(&cfqg->rb_node))
879                 return;
880
881         /*
882          * Currently put the group at the end. Later implement something
883          * so that groups get lesser vtime based on their weights, so that
884          * if group does not loose all if it was not continuously backlogged.
885          */
886         n = rb_last(&st->rb);
887         if (n) {
888                 __cfqg = rb_entry_cfqg(n);
889                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
890         } else
891                 cfqg->vdisktime = st->min_vdisktime;
892         cfq_group_service_tree_add(st, cfqg);
893 }
894
895 static void
896 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
897 {
898         st->total_weight -= cfqg->weight;
899         if (!RB_EMPTY_NODE(&cfqg->rb_node))
900                 cfq_rb_erase(&cfqg->rb_node, st);
901 }
902
903 static void
904 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
905 {
906         struct cfq_rb_root *st = &cfqd->grp_service_tree;
907
908         BUG_ON(cfqg->nr_cfqq < 1);
909         cfqg->nr_cfqq--;
910
911         /* If there are other cfq queues under this group, don't delete it */
912         if (cfqg->nr_cfqq)
913                 return;
914
915         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
916         cfq_group_service_tree_del(st, cfqg);
917         cfqg->saved_workload_slice = 0;
918         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
919 }
920
921 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
922                                                 unsigned int *unaccounted_time)
923 {
924         unsigned int slice_used;
925
926         /*
927          * Queue got expired before even a single request completed or
928          * got expired immediately after first request completion.
929          */
930         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
931                 /*
932                  * Also charge the seek time incurred to the group, otherwise
933                  * if there are mutiple queues in the group, each can dispatch
934                  * a single request on seeky media and cause lots of seek time
935                  * and group will never know it.
936                  */
937                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
938                                         1);
939         } else {
940                 slice_used = jiffies - cfqq->slice_start;
941                 if (slice_used > cfqq->allocated_slice) {
942                         *unaccounted_time = slice_used - cfqq->allocated_slice;
943                         slice_used = cfqq->allocated_slice;
944                 }
945                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
946                         *unaccounted_time += cfqq->slice_start -
947                                         cfqq->dispatch_start;
948         }
949
950         return slice_used;
951 }
952
953 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
954                                 struct cfq_queue *cfqq)
955 {
956         struct cfq_rb_root *st = &cfqd->grp_service_tree;
957         unsigned int used_sl, charge, unaccounted_sl = 0;
958         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
959                         - cfqg->service_tree_idle.count;
960
961         BUG_ON(nr_sync < 0);
962         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
963
964         if (iops_mode(cfqd))
965                 charge = cfqq->slice_dispatch;
966         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
967                 charge = cfqq->allocated_slice;
968
969         /* Can't update vdisktime while group is on service tree */
970         cfq_group_service_tree_del(st, cfqg);
971         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
972         /* If a new weight was requested, update now, off tree */
973         cfq_group_service_tree_add(st, cfqg);
974
975         /* This group is being expired. Save the context */
976         if (time_after(cfqd->workload_expires, jiffies)) {
977                 cfqg->saved_workload_slice = cfqd->workload_expires
978                                                 - jiffies;
979                 cfqg->saved_workload = cfqd->serving_type;
980                 cfqg->saved_serving_prio = cfqd->serving_prio;
981         } else
982                 cfqg->saved_workload_slice = 0;
983
984         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
985                                         st->min_vdisktime);
986         cfq_log_cfqq(cfqq->cfqd, cfqq,
987                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
988                      used_sl, cfqq->slice_dispatch, charge,
989                      iops_mode(cfqd), cfqq->nr_sectors);
990         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
991                                           unaccounted_sl);
992         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
993 }
994
995 #ifdef CONFIG_CFQ_GROUP_IOSCHED
996 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
997 {
998         if (blkg)
999                 return container_of(blkg, struct cfq_group, blkg);
1000         return NULL;
1001 }
1002
1003 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1004                                           unsigned int weight)
1005 {
1006         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1007         cfqg->new_weight = weight;
1008         cfqg->needs_update = true;
1009 }
1010
1011 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1012                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1013 {
1014         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1015         unsigned int major, minor;
1016
1017         /*
1018          * Add group onto cgroup list. It might happen that bdi->dev is
1019          * not initialized yet. Initialize this new group without major
1020          * and minor info and this info will be filled in once a new thread
1021          * comes for IO.
1022          */
1023         if (bdi->dev) {
1024                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1025                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1026                                         (void *)cfqd, MKDEV(major, minor));
1027         } else
1028                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1029                                         (void *)cfqd, 0);
1030
1031         cfqd->nr_blkcg_linked_grps++;
1032         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1033
1034         /* Add group on cfqd list */
1035         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1036 }
1037
1038 /*
1039  * Should be called from sleepable context. No request queue lock as per
1040  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1041  * from sleepable context.
1042  */
1043 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1044 {
1045         struct cfq_group *cfqg = NULL;
1046         int i, j, ret;
1047         struct cfq_rb_root *st;
1048
1049         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1050         if (!cfqg)
1051                 return NULL;
1052
1053         for_each_cfqg_st(cfqg, i, j, st)
1054                 *st = CFQ_RB_ROOT;
1055         RB_CLEAR_NODE(&cfqg->rb_node);
1056
1057         /*
1058          * Take the initial reference that will be released on destroy
1059          * This can be thought of a joint reference by cgroup and
1060          * elevator which will be dropped by either elevator exit
1061          * or cgroup deletion path depending on who is exiting first.
1062          */
1063         cfqg->ref = 1;
1064
1065         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1066         if (ret) {
1067                 kfree(cfqg);
1068                 return NULL;
1069         }
1070
1071         return cfqg;
1072 }
1073
1074 static struct cfq_group *
1075 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1076 {
1077         struct cfq_group *cfqg = NULL;
1078         void *key = cfqd;
1079         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1080         unsigned int major, minor;
1081
1082         /*
1083          * This is the common case when there are no blkio cgroups.
1084          * Avoid lookup in this case
1085          */
1086         if (blkcg == &blkio_root_cgroup)
1087                 cfqg = &cfqd->root_group;
1088         else
1089                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1090
1091         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1092                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1093                 cfqg->blkg.dev = MKDEV(major, minor);
1094         }
1095
1096         return cfqg;
1097 }
1098
1099 /*
1100  * Search for the cfq group current task belongs to. request_queue lock must
1101  * be held.
1102  */
1103 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1104 {
1105         struct blkio_cgroup *blkcg;
1106         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1107         struct request_queue *q = cfqd->queue;
1108
1109         rcu_read_lock();
1110         blkcg = task_blkio_cgroup(current);
1111         cfqg = cfq_find_cfqg(cfqd, blkcg);
1112         if (cfqg) {
1113                 rcu_read_unlock();
1114                 return cfqg;
1115         }
1116
1117         /*
1118          * Need to allocate a group. Allocation of group also needs allocation
1119          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1120          * we need to drop rcu lock and queue_lock before we call alloc.
1121          *
1122          * Not taking any queue reference here and assuming that queue is
1123          * around by the time we return. CFQ queue allocation code does
1124          * the same. It might be racy though.
1125          */
1126
1127         rcu_read_unlock();
1128         spin_unlock_irq(q->queue_lock);
1129
1130         cfqg = cfq_alloc_cfqg(cfqd);
1131
1132         spin_lock_irq(q->queue_lock);
1133
1134         rcu_read_lock();
1135         blkcg = task_blkio_cgroup(current);
1136
1137         /*
1138          * If some other thread already allocated the group while we were
1139          * not holding queue lock, free up the group
1140          */
1141         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1142
1143         if (__cfqg) {
1144                 kfree(cfqg);
1145                 rcu_read_unlock();
1146                 return __cfqg;
1147         }
1148
1149         if (!cfqg)
1150                 cfqg = &cfqd->root_group;
1151
1152         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1153         rcu_read_unlock();
1154         return cfqg;
1155 }
1156
1157 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1158 {
1159         cfqg->ref++;
1160         return cfqg;
1161 }
1162
1163 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1164 {
1165         /* Currently, all async queues are mapped to root group */
1166         if (!cfq_cfqq_sync(cfqq))
1167                 cfqg = &cfqq->cfqd->root_group;
1168
1169         cfqq->cfqg = cfqg;
1170         /* cfqq reference on cfqg */
1171         cfqq->cfqg->ref++;
1172 }
1173
1174 static void cfq_put_cfqg(struct cfq_group *cfqg)
1175 {
1176         struct cfq_rb_root *st;
1177         int i, j;
1178
1179         BUG_ON(cfqg->ref <= 0);
1180         cfqg->ref--;
1181         if (cfqg->ref)
1182                 return;
1183         for_each_cfqg_st(cfqg, i, j, st)
1184                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1185         free_percpu(cfqg->blkg.stats_cpu);
1186         kfree(cfqg);
1187 }
1188
1189 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1190 {
1191         /* Something wrong if we are trying to remove same group twice */
1192         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1193
1194         hlist_del_init(&cfqg->cfqd_node);
1195
1196         /*
1197          * Put the reference taken at the time of creation so that when all
1198          * queues are gone, group can be destroyed.
1199          */
1200         cfq_put_cfqg(cfqg);
1201 }
1202
1203 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1204 {
1205         struct hlist_node *pos, *n;
1206         struct cfq_group *cfqg;
1207
1208         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1209                 /*
1210                  * If cgroup removal path got to blk_group first and removed
1211                  * it from cgroup list, then it will take care of destroying
1212                  * cfqg also.
1213                  */
1214                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1215                         cfq_destroy_cfqg(cfqd, cfqg);
1216         }
1217 }
1218
1219 /*
1220  * Blk cgroup controller notification saying that blkio_group object is being
1221  * delinked as associated cgroup object is going away. That also means that
1222  * no new IO will come in this group. So get rid of this group as soon as
1223  * any pending IO in the group is finished.
1224  *
1225  * This function is called under rcu_read_lock(). key is the rcu protected
1226  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1227  * read lock.
1228  *
1229  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1230  * it should not be NULL as even if elevator was exiting, cgroup deltion
1231  * path got to it first.
1232  */
1233 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1234 {
1235         unsigned long  flags;
1236         struct cfq_data *cfqd = key;
1237
1238         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1239         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1240         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1241 }
1242
1243 #else /* GROUP_IOSCHED */
1244 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1245 {
1246         return &cfqd->root_group;
1247 }
1248
1249 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1250 {
1251         return cfqg;
1252 }
1253
1254 static inline void
1255 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1256         cfqq->cfqg = cfqg;
1257 }
1258
1259 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1260 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1261
1262 #endif /* GROUP_IOSCHED */
1263
1264 /*
1265  * The cfqd->service_trees holds all pending cfq_queue's that have
1266  * requests waiting to be processed. It is sorted in the order that
1267  * we will service the queues.
1268  */
1269 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1270                                  bool add_front)
1271 {
1272         struct rb_node **p, *parent;
1273         struct cfq_queue *__cfqq;
1274         unsigned long rb_key;
1275         struct cfq_rb_root *service_tree;
1276         int left;
1277         int new_cfqq = 1;
1278
1279         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1280                                                 cfqq_type(cfqq));
1281         if (cfq_class_idle(cfqq)) {
1282                 rb_key = CFQ_IDLE_DELAY;
1283                 parent = rb_last(&service_tree->rb);
1284                 if (parent && parent != &cfqq->rb_node) {
1285                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1286                         rb_key += __cfqq->rb_key;
1287                 } else
1288                         rb_key += jiffies;
1289         } else if (!add_front) {
1290                 /*
1291                  * Get our rb key offset. Subtract any residual slice
1292                  * value carried from last service. A negative resid
1293                  * count indicates slice overrun, and this should position
1294                  * the next service time further away in the tree.
1295                  */
1296                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1297                 rb_key -= cfqq->slice_resid;
1298                 cfqq->slice_resid = 0;
1299         } else {
1300                 rb_key = -HZ;
1301                 __cfqq = cfq_rb_first(service_tree);
1302                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1303         }
1304
1305         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1306                 new_cfqq = 0;
1307                 /*
1308                  * same position, nothing more to do
1309                  */
1310                 if (rb_key == cfqq->rb_key &&
1311                     cfqq->service_tree == service_tree)
1312                         return;
1313
1314                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1315                 cfqq->service_tree = NULL;
1316         }
1317
1318         left = 1;
1319         parent = NULL;
1320         cfqq->service_tree = service_tree;
1321         p = &service_tree->rb.rb_node;
1322         while (*p) {
1323                 struct rb_node **n;
1324
1325                 parent = *p;
1326                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1327
1328                 /*
1329                  * sort by key, that represents service time.
1330                  */
1331                 if (time_before(rb_key, __cfqq->rb_key))
1332                         n = &(*p)->rb_left;
1333                 else {
1334                         n = &(*p)->rb_right;
1335                         left = 0;
1336                 }
1337
1338                 p = n;
1339         }
1340
1341         if (left)
1342                 service_tree->left = &cfqq->rb_node;
1343
1344         cfqq->rb_key = rb_key;
1345         rb_link_node(&cfqq->rb_node, parent, p);
1346         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1347         service_tree->count++;
1348         if (add_front || !new_cfqq)
1349                 return;
1350         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1351 }
1352
1353 static struct cfq_queue *
1354 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1355                      sector_t sector, struct rb_node **ret_parent,
1356                      struct rb_node ***rb_link)
1357 {
1358         struct rb_node **p, *parent;
1359         struct cfq_queue *cfqq = NULL;
1360
1361         parent = NULL;
1362         p = &root->rb_node;
1363         while (*p) {
1364                 struct rb_node **n;
1365
1366                 parent = *p;
1367                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1368
1369                 /*
1370                  * Sort strictly based on sector.  Smallest to the left,
1371                  * largest to the right.
1372                  */
1373                 if (sector > blk_rq_pos(cfqq->next_rq))
1374                         n = &(*p)->rb_right;
1375                 else if (sector < blk_rq_pos(cfqq->next_rq))
1376                         n = &(*p)->rb_left;
1377                 else
1378                         break;
1379                 p = n;
1380                 cfqq = NULL;
1381         }
1382
1383         *ret_parent = parent;
1384         if (rb_link)
1385                 *rb_link = p;
1386         return cfqq;
1387 }
1388
1389 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1390 {
1391         struct rb_node **p, *parent;
1392         struct cfq_queue *__cfqq;
1393
1394         if (cfqq->p_root) {
1395                 rb_erase(&cfqq->p_node, cfqq->p_root);
1396                 cfqq->p_root = NULL;
1397         }
1398
1399         if (cfq_class_idle(cfqq))
1400                 return;
1401         if (!cfqq->next_rq)
1402                 return;
1403
1404         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1405         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1406                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1407         if (!__cfqq) {
1408                 rb_link_node(&cfqq->p_node, parent, p);
1409                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1410         } else
1411                 cfqq->p_root = NULL;
1412 }
1413
1414 /*
1415  * Update cfqq's position in the service tree.
1416  */
1417 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1418 {
1419         /*
1420          * Resorting requires the cfqq to be on the RR list already.
1421          */
1422         if (cfq_cfqq_on_rr(cfqq)) {
1423                 cfq_service_tree_add(cfqd, cfqq, 0);
1424                 cfq_prio_tree_add(cfqd, cfqq);
1425         }
1426 }
1427
1428 /*
1429  * add to busy list of queues for service, trying to be fair in ordering
1430  * the pending list according to last request service
1431  */
1432 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1433 {
1434         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1435         BUG_ON(cfq_cfqq_on_rr(cfqq));
1436         cfq_mark_cfqq_on_rr(cfqq);
1437         cfqd->busy_queues++;
1438         if (cfq_cfqq_sync(cfqq))
1439                 cfqd->busy_sync_queues++;
1440
1441         cfq_resort_rr_list(cfqd, cfqq);
1442 }
1443
1444 /*
1445  * Called when the cfqq no longer has requests pending, remove it from
1446  * the service tree.
1447  */
1448 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1449 {
1450         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1451         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1452         cfq_clear_cfqq_on_rr(cfqq);
1453
1454         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1455                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1456                 cfqq->service_tree = NULL;
1457         }
1458         if (cfqq->p_root) {
1459                 rb_erase(&cfqq->p_node, cfqq->p_root);
1460                 cfqq->p_root = NULL;
1461         }
1462
1463         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1464         BUG_ON(!cfqd->busy_queues);
1465         cfqd->busy_queues--;
1466         if (cfq_cfqq_sync(cfqq))
1467                 cfqd->busy_sync_queues--;
1468 }
1469
1470 /*
1471  * rb tree support functions
1472  */
1473 static void cfq_del_rq_rb(struct request *rq)
1474 {
1475         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1476         const int sync = rq_is_sync(rq);
1477
1478         BUG_ON(!cfqq->queued[sync]);
1479         cfqq->queued[sync]--;
1480
1481         elv_rb_del(&cfqq->sort_list, rq);
1482
1483         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1484                 /*
1485                  * Queue will be deleted from service tree when we actually
1486                  * expire it later. Right now just remove it from prio tree
1487                  * as it is empty.
1488                  */
1489                 if (cfqq->p_root) {
1490                         rb_erase(&cfqq->p_node, cfqq->p_root);
1491                         cfqq->p_root = NULL;
1492                 }
1493         }
1494 }
1495
1496 static void cfq_add_rq_rb(struct request *rq)
1497 {
1498         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1499         struct cfq_data *cfqd = cfqq->cfqd;
1500         struct request *prev;
1501
1502         cfqq->queued[rq_is_sync(rq)]++;
1503
1504         elv_rb_add(&cfqq->sort_list, rq);
1505
1506         if (!cfq_cfqq_on_rr(cfqq))
1507                 cfq_add_cfqq_rr(cfqd, cfqq);
1508
1509         /*
1510          * check if this request is a better next-serve candidate
1511          */
1512         prev = cfqq->next_rq;
1513         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1514
1515         /*
1516          * adjust priority tree position, if ->next_rq changes
1517          */
1518         if (prev != cfqq->next_rq)
1519                 cfq_prio_tree_add(cfqd, cfqq);
1520
1521         BUG_ON(!cfqq->next_rq);
1522 }
1523
1524 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1525 {
1526         elv_rb_del(&cfqq->sort_list, rq);
1527         cfqq->queued[rq_is_sync(rq)]--;
1528         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1529                                         rq_data_dir(rq), rq_is_sync(rq));
1530         cfq_add_rq_rb(rq);
1531         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1532                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1533                         rq_is_sync(rq));
1534 }
1535
1536 static struct request *
1537 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1538 {
1539         struct task_struct *tsk = current;
1540         struct cfq_io_context *cic;
1541         struct cfq_queue *cfqq;
1542
1543         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1544         if (!cic)
1545                 return NULL;
1546
1547         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1548         if (cfqq) {
1549                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1550
1551                 return elv_rb_find(&cfqq->sort_list, sector);
1552         }
1553
1554         return NULL;
1555 }
1556
1557 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1558 {
1559         struct cfq_data *cfqd = q->elevator->elevator_data;
1560
1561         cfqd->rq_in_driver++;
1562         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1563                                                 cfqd->rq_in_driver);
1564
1565         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1566 }
1567
1568 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1569 {
1570         struct cfq_data *cfqd = q->elevator->elevator_data;
1571
1572         WARN_ON(!cfqd->rq_in_driver);
1573         cfqd->rq_in_driver--;
1574         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1575                                                 cfqd->rq_in_driver);
1576 }
1577
1578 static void cfq_remove_request(struct request *rq)
1579 {
1580         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1581
1582         if (cfqq->next_rq == rq)
1583                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1584
1585         list_del_init(&rq->queuelist);
1586         cfq_del_rq_rb(rq);
1587
1588         cfqq->cfqd->rq_queued--;
1589         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1590                                         rq_data_dir(rq), rq_is_sync(rq));
1591 }
1592
1593 static int cfq_merge(struct request_queue *q, struct request **req,
1594                      struct bio *bio)
1595 {
1596         struct cfq_data *cfqd = q->elevator->elevator_data;
1597         struct request *__rq;
1598
1599         __rq = cfq_find_rq_fmerge(cfqd, bio);
1600         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1601                 *req = __rq;
1602                 return ELEVATOR_FRONT_MERGE;
1603         }
1604
1605         return ELEVATOR_NO_MERGE;
1606 }
1607
1608 static void cfq_merged_request(struct request_queue *q, struct request *req,
1609                                int type)
1610 {
1611         if (type == ELEVATOR_FRONT_MERGE) {
1612                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1613
1614                 cfq_reposition_rq_rb(cfqq, req);
1615         }
1616 }
1617
1618 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1619                                 struct bio *bio)
1620 {
1621         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1622                                         bio_data_dir(bio), cfq_bio_sync(bio));
1623 }
1624
1625 static void
1626 cfq_merged_requests(struct request_queue *q, struct request *rq,
1627                     struct request *next)
1628 {
1629         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1630         /*
1631          * reposition in fifo if next is older than rq
1632          */
1633         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1634             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1635                 list_move(&rq->queuelist, &next->queuelist);
1636                 rq_set_fifo_time(rq, rq_fifo_time(next));
1637         }
1638
1639         if (cfqq->next_rq == next)
1640                 cfqq->next_rq = rq;
1641         cfq_remove_request(next);
1642         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1643                                         rq_data_dir(next), rq_is_sync(next));
1644 }
1645
1646 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1647                            struct bio *bio)
1648 {
1649         struct cfq_data *cfqd = q->elevator->elevator_data;
1650         struct cfq_io_context *cic;
1651         struct cfq_queue *cfqq;
1652
1653         /*
1654          * Disallow merge of a sync bio into an async request.
1655          */
1656         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1657                 return false;
1658
1659         /*
1660          * Lookup the cfqq that this bio will be queued with. Allow
1661          * merge only if rq is queued there.
1662          */
1663         cic = cfq_cic_lookup(cfqd, current->io_context);
1664         if (!cic)
1665                 return false;
1666
1667         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1668         return cfqq == RQ_CFQQ(rq);
1669 }
1670
1671 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1672 {
1673         del_timer(&cfqd->idle_slice_timer);
1674         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1675 }
1676
1677 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1678                                    struct cfq_queue *cfqq)
1679 {
1680         if (cfqq) {
1681                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1682                                 cfqd->serving_prio, cfqd->serving_type);
1683                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1684                 cfqq->slice_start = 0;
1685                 cfqq->dispatch_start = jiffies;
1686                 cfqq->allocated_slice = 0;
1687                 cfqq->slice_end = 0;
1688                 cfqq->slice_dispatch = 0;
1689                 cfqq->nr_sectors = 0;
1690
1691                 cfq_clear_cfqq_wait_request(cfqq);
1692                 cfq_clear_cfqq_must_dispatch(cfqq);
1693                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1694                 cfq_clear_cfqq_fifo_expire(cfqq);
1695                 cfq_mark_cfqq_slice_new(cfqq);
1696
1697                 cfq_del_timer(cfqd, cfqq);
1698         }
1699
1700         cfqd->active_queue = cfqq;
1701 }
1702
1703 /*
1704  * current cfqq expired its slice (or was too idle), select new one
1705  */
1706 static void
1707 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1708                     bool timed_out)
1709 {
1710         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1711
1712         if (cfq_cfqq_wait_request(cfqq))
1713                 cfq_del_timer(cfqd, cfqq);
1714
1715         cfq_clear_cfqq_wait_request(cfqq);
1716         cfq_clear_cfqq_wait_busy(cfqq);
1717
1718         /*
1719          * If this cfqq is shared between multiple processes, check to
1720          * make sure that those processes are still issuing I/Os within
1721          * the mean seek distance.  If not, it may be time to break the
1722          * queues apart again.
1723          */
1724         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1725                 cfq_mark_cfqq_split_coop(cfqq);
1726
1727         /*
1728          * store what was left of this slice, if the queue idled/timed out
1729          */
1730         if (timed_out) {
1731                 if (cfq_cfqq_slice_new(cfqq))
1732                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1733                 else
1734                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1735                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1736         }
1737
1738         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1739
1740         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1741                 cfq_del_cfqq_rr(cfqd, cfqq);
1742
1743         cfq_resort_rr_list(cfqd, cfqq);
1744
1745         if (cfqq == cfqd->active_queue)
1746                 cfqd->active_queue = NULL;
1747
1748         if (cfqd->active_cic) {
1749                 put_io_context(cfqd->active_cic->ioc);
1750                 cfqd->active_cic = NULL;
1751         }
1752 }
1753
1754 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1755 {
1756         struct cfq_queue *cfqq = cfqd->active_queue;
1757
1758         if (cfqq)
1759                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1760 }
1761
1762 /*
1763  * Get next queue for service. Unless we have a queue preemption,
1764  * we'll simply select the first cfqq in the service tree.
1765  */
1766 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1767 {
1768         struct cfq_rb_root *service_tree =
1769                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1770                                         cfqd->serving_type);
1771
1772         if (!cfqd->rq_queued)
1773                 return NULL;
1774
1775         /* There is nothing to dispatch */
1776         if (!service_tree)
1777                 return NULL;
1778         if (RB_EMPTY_ROOT(&service_tree->rb))
1779                 return NULL;
1780         return cfq_rb_first(service_tree);
1781 }
1782
1783 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1784 {
1785         struct cfq_group *cfqg;
1786         struct cfq_queue *cfqq;
1787         int i, j;
1788         struct cfq_rb_root *st;
1789
1790         if (!cfqd->rq_queued)
1791                 return NULL;
1792
1793         cfqg = cfq_get_next_cfqg(cfqd);
1794         if (!cfqg)
1795                 return NULL;
1796
1797         for_each_cfqg_st(cfqg, i, j, st)
1798                 if ((cfqq = cfq_rb_first(st)) != NULL)
1799                         return cfqq;
1800         return NULL;
1801 }
1802
1803 /*
1804  * Get and set a new active queue for service.
1805  */
1806 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1807                                               struct cfq_queue *cfqq)
1808 {
1809         if (!cfqq)
1810                 cfqq = cfq_get_next_queue(cfqd);
1811
1812         __cfq_set_active_queue(cfqd, cfqq);
1813         return cfqq;
1814 }
1815
1816 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1817                                           struct request *rq)
1818 {
1819         if (blk_rq_pos(rq) >= cfqd->last_position)
1820                 return blk_rq_pos(rq) - cfqd->last_position;
1821         else
1822                 return cfqd->last_position - blk_rq_pos(rq);
1823 }
1824
1825 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1826                                struct request *rq)
1827 {
1828         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1829 }
1830
1831 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1832                                     struct cfq_queue *cur_cfqq)
1833 {
1834         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1835         struct rb_node *parent, *node;
1836         struct cfq_queue *__cfqq;
1837         sector_t sector = cfqd->last_position;
1838
1839         if (RB_EMPTY_ROOT(root))
1840                 return NULL;
1841
1842         /*
1843          * First, if we find a request starting at the end of the last
1844          * request, choose it.
1845          */
1846         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1847         if (__cfqq)
1848                 return __cfqq;
1849
1850         /*
1851          * If the exact sector wasn't found, the parent of the NULL leaf
1852          * will contain the closest sector.
1853          */
1854         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1855         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1856                 return __cfqq;
1857
1858         if (blk_rq_pos(__cfqq->next_rq) < sector)
1859                 node = rb_next(&__cfqq->p_node);
1860         else
1861                 node = rb_prev(&__cfqq->p_node);
1862         if (!node)
1863                 return NULL;
1864
1865         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1866         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1867                 return __cfqq;
1868
1869         return NULL;
1870 }
1871
1872 /*
1873  * cfqd - obvious
1874  * cur_cfqq - passed in so that we don't decide that the current queue is
1875  *            closely cooperating with itself.
1876  *
1877  * So, basically we're assuming that that cur_cfqq has dispatched at least
1878  * one request, and that cfqd->last_position reflects a position on the disk
1879  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1880  * assumption.
1881  */
1882 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1883                                               struct cfq_queue *cur_cfqq)
1884 {
1885         struct cfq_queue *cfqq;
1886
1887         if (cfq_class_idle(cur_cfqq))
1888                 return NULL;
1889         if (!cfq_cfqq_sync(cur_cfqq))
1890                 return NULL;
1891         if (CFQQ_SEEKY(cur_cfqq))
1892                 return NULL;
1893
1894         /*
1895          * Don't search priority tree if it's the only queue in the group.
1896          */
1897         if (cur_cfqq->cfqg->nr_cfqq == 1)
1898                 return NULL;
1899
1900         /*
1901          * We should notice if some of the queues are cooperating, eg
1902          * working closely on the same area of the disk. In that case,
1903          * we can group them together and don't waste time idling.
1904          */
1905         cfqq = cfqq_close(cfqd, cur_cfqq);
1906         if (!cfqq)
1907                 return NULL;
1908
1909         /* If new queue belongs to different cfq_group, don't choose it */
1910         if (cur_cfqq->cfqg != cfqq->cfqg)
1911                 return NULL;
1912
1913         /*
1914          * It only makes sense to merge sync queues.
1915          */
1916         if (!cfq_cfqq_sync(cfqq))
1917                 return NULL;
1918         if (CFQQ_SEEKY(cfqq))
1919                 return NULL;
1920
1921         /*
1922          * Do not merge queues of different priority classes
1923          */
1924         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1925                 return NULL;
1926
1927         return cfqq;
1928 }
1929
1930 /*
1931  * Determine whether we should enforce idle window for this queue.
1932  */
1933
1934 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1935 {
1936         enum wl_prio_t prio = cfqq_prio(cfqq);
1937         struct cfq_rb_root *service_tree = cfqq->service_tree;
1938
1939         BUG_ON(!service_tree);
1940         BUG_ON(!service_tree->count);
1941
1942         if (!cfqd->cfq_slice_idle)
1943                 return false;
1944
1945         /* We never do for idle class queues. */
1946         if (prio == IDLE_WORKLOAD)
1947                 return false;
1948
1949         /* We do for queues that were marked with idle window flag. */
1950         if (cfq_cfqq_idle_window(cfqq) &&
1951            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1952                 return true;
1953
1954         /*
1955          * Otherwise, we do only if they are the last ones
1956          * in their service tree.
1957          */
1958         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1959                 return true;
1960         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1961                         service_tree->count);
1962         return false;
1963 }
1964
1965 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1966 {
1967         struct cfq_queue *cfqq = cfqd->active_queue;
1968         struct cfq_io_context *cic;
1969         unsigned long sl, group_idle = 0;
1970
1971         /*
1972          * SSD device without seek penalty, disable idling. But only do so
1973          * for devices that support queuing, otherwise we still have a problem
1974          * with sync vs async workloads.
1975          */
1976         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1977                 return;
1978
1979         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1980         WARN_ON(cfq_cfqq_slice_new(cfqq));
1981
1982         /*
1983          * idle is disabled, either manually or by past process history
1984          */
1985         if (!cfq_should_idle(cfqd, cfqq)) {
1986                 /* no queue idling. Check for group idling */
1987                 if (cfqd->cfq_group_idle)
1988                         group_idle = cfqd->cfq_group_idle;
1989                 else
1990                         return;
1991         }
1992
1993         /*
1994          * still active requests from this queue, don't idle
1995          */
1996         if (cfqq->dispatched)
1997                 return;
1998
1999         /*
2000          * task has exited, don't wait
2001          */
2002         cic = cfqd->active_cic;
2003         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2004                 return;
2005
2006         /*
2007          * If our average think time is larger than the remaining time
2008          * slice, then don't idle. This avoids overrunning the allotted
2009          * time slice.
2010          */
2011         if (sample_valid(cic->ttime.ttime_samples) &&
2012             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2013                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2014                              cic->ttime.ttime_mean);
2015                 return;
2016         }
2017
2018         /* There are other queues in the group, don't do group idle */
2019         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2020                 return;
2021
2022         cfq_mark_cfqq_wait_request(cfqq);
2023
2024         if (group_idle)
2025                 sl = cfqd->cfq_group_idle;
2026         else
2027                 sl = cfqd->cfq_slice_idle;
2028
2029         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2030         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2031         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2032                         group_idle ? 1 : 0);
2033 }
2034
2035 /*
2036  * Move request from internal lists to the request queue dispatch list.
2037  */
2038 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2039 {
2040         struct cfq_data *cfqd = q->elevator->elevator_data;
2041         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2042
2043         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2044
2045         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2046         cfq_remove_request(rq);
2047         cfqq->dispatched++;
2048         (RQ_CFQG(rq))->dispatched++;
2049         elv_dispatch_sort(q, rq);
2050
2051         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2052         cfqq->nr_sectors += blk_rq_sectors(rq);
2053         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2054                                         rq_data_dir(rq), rq_is_sync(rq));
2055 }
2056
2057 /*
2058  * return expired entry, or NULL to just start from scratch in rbtree
2059  */
2060 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2061 {
2062         struct request *rq = NULL;
2063
2064         if (cfq_cfqq_fifo_expire(cfqq))
2065                 return NULL;
2066
2067         cfq_mark_cfqq_fifo_expire(cfqq);
2068
2069         if (list_empty(&cfqq->fifo))
2070                 return NULL;
2071
2072         rq = rq_entry_fifo(cfqq->fifo.next);
2073         if (time_before(jiffies, rq_fifo_time(rq)))
2074                 rq = NULL;
2075
2076         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2077         return rq;
2078 }
2079
2080 static inline int
2081 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2082 {
2083         const int base_rq = cfqd->cfq_slice_async_rq;
2084
2085         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2086
2087         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2088 }
2089
2090 /*
2091  * Must be called with the queue_lock held.
2092  */
2093 static int cfqq_process_refs(struct cfq_queue *cfqq)
2094 {
2095         int process_refs, io_refs;
2096
2097         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2098         process_refs = cfqq->ref - io_refs;
2099         BUG_ON(process_refs < 0);
2100         return process_refs;
2101 }
2102
2103 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2104 {
2105         int process_refs, new_process_refs;
2106         struct cfq_queue *__cfqq;
2107
2108         /*
2109          * If there are no process references on the new_cfqq, then it is
2110          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2111          * chain may have dropped their last reference (not just their
2112          * last process reference).
2113          */
2114         if (!cfqq_process_refs(new_cfqq))
2115                 return;
2116
2117         /* Avoid a circular list and skip interim queue merges */
2118         while ((__cfqq = new_cfqq->new_cfqq)) {
2119                 if (__cfqq == cfqq)
2120                         return;
2121                 new_cfqq = __cfqq;
2122         }
2123
2124         process_refs = cfqq_process_refs(cfqq);
2125         new_process_refs = cfqq_process_refs(new_cfqq);
2126         /*
2127          * If the process for the cfqq has gone away, there is no
2128          * sense in merging the queues.
2129          */
2130         if (process_refs == 0 || new_process_refs == 0)
2131                 return;
2132
2133         /*
2134          * Merge in the direction of the lesser amount of work.
2135          */
2136         if (new_process_refs >= process_refs) {
2137                 cfqq->new_cfqq = new_cfqq;
2138                 new_cfqq->ref += process_refs;
2139         } else {
2140                 new_cfqq->new_cfqq = cfqq;
2141                 cfqq->ref += new_process_refs;
2142         }
2143 }
2144
2145 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2146                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2147 {
2148         struct cfq_queue *queue;
2149         int i;
2150         bool key_valid = false;
2151         unsigned long lowest_key = 0;
2152         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2153
2154         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2155                 /* select the one with lowest rb_key */
2156                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2157                 if (queue &&
2158                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2159                         lowest_key = queue->rb_key;
2160                         cur_best = i;
2161                         key_valid = true;
2162                 }
2163         }
2164
2165         return cur_best;
2166 }
2167
2168 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2169 {
2170         unsigned slice;
2171         unsigned count;
2172         struct cfq_rb_root *st;
2173         unsigned group_slice;
2174         enum wl_prio_t original_prio = cfqd->serving_prio;
2175
2176         /* Choose next priority. RT > BE > IDLE */
2177         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2178                 cfqd->serving_prio = RT_WORKLOAD;
2179         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2180                 cfqd->serving_prio = BE_WORKLOAD;
2181         else {
2182                 cfqd->serving_prio = IDLE_WORKLOAD;
2183                 cfqd->workload_expires = jiffies + 1;
2184                 return;
2185         }
2186
2187         if (original_prio != cfqd->serving_prio)
2188                 goto new_workload;
2189
2190         /*
2191          * For RT and BE, we have to choose also the type
2192          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2193          * expiration time
2194          */
2195         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2196         count = st->count;
2197
2198         /*
2199          * check workload expiration, and that we still have other queues ready
2200          */
2201         if (count && !time_after(jiffies, cfqd->workload_expires))
2202                 return;
2203
2204 new_workload:
2205         /* otherwise select new workload type */
2206         cfqd->serving_type =
2207                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2208         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2209         count = st->count;
2210
2211         /*
2212          * the workload slice is computed as a fraction of target latency
2213          * proportional to the number of queues in that workload, over
2214          * all the queues in the same priority class
2215          */
2216         group_slice = cfq_group_slice(cfqd, cfqg);
2217
2218         slice = group_slice * count /
2219                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2220                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2221
2222         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2223                 unsigned int tmp;
2224
2225                 /*
2226                  * Async queues are currently system wide. Just taking
2227                  * proportion of queues with-in same group will lead to higher
2228                  * async ratio system wide as generally root group is going
2229                  * to have higher weight. A more accurate thing would be to
2230                  * calculate system wide asnc/sync ratio.
2231                  */
2232                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2233                 tmp = tmp/cfqd->busy_queues;
2234                 slice = min_t(unsigned, slice, tmp);
2235
2236                 /* async workload slice is scaled down according to
2237                  * the sync/async slice ratio. */
2238                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2239         } else
2240                 /* sync workload slice is at least 2 * cfq_slice_idle */
2241                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2242
2243         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2244         cfq_log(cfqd, "workload slice:%d", slice);
2245         cfqd->workload_expires = jiffies + slice;
2246 }
2247
2248 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2249 {
2250         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2251         struct cfq_group *cfqg;
2252
2253         if (RB_EMPTY_ROOT(&st->rb))
2254                 return NULL;
2255         cfqg = cfq_rb_first_group(st);
2256         update_min_vdisktime(st);
2257         return cfqg;
2258 }
2259
2260 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2261 {
2262         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2263
2264         cfqd->serving_group = cfqg;
2265
2266         /* Restore the workload type data */
2267         if (cfqg->saved_workload_slice) {
2268                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2269                 cfqd->serving_type = cfqg->saved_workload;
2270                 cfqd->serving_prio = cfqg->saved_serving_prio;
2271         } else
2272                 cfqd->workload_expires = jiffies - 1;
2273
2274         choose_service_tree(cfqd, cfqg);
2275 }
2276
2277 /*
2278  * Select a queue for service. If we have a current active queue,
2279  * check whether to continue servicing it, or retrieve and set a new one.
2280  */
2281 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2282 {
2283         struct cfq_queue *cfqq, *new_cfqq = NULL;
2284
2285         cfqq = cfqd->active_queue;
2286         if (!cfqq)
2287                 goto new_queue;
2288
2289         if (!cfqd->rq_queued)
2290                 return NULL;
2291
2292         /*
2293          * We were waiting for group to get backlogged. Expire the queue
2294          */
2295         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2296                 goto expire;
2297
2298         /*
2299          * The active queue has run out of time, expire it and select new.
2300          */
2301         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2302                 /*
2303                  * If slice had not expired at the completion of last request
2304                  * we might not have turned on wait_busy flag. Don't expire
2305                  * the queue yet. Allow the group to get backlogged.
2306                  *
2307                  * The very fact that we have used the slice, that means we
2308                  * have been idling all along on this queue and it should be
2309                  * ok to wait for this request to complete.
2310                  */
2311                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2312                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2313                         cfqq = NULL;
2314                         goto keep_queue;
2315                 } else
2316                         goto check_group_idle;
2317         }
2318
2319         /*
2320          * The active queue has requests and isn't expired, allow it to
2321          * dispatch.
2322          */
2323         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2324                 goto keep_queue;
2325
2326         /*
2327          * If another queue has a request waiting within our mean seek
2328          * distance, let it run.  The expire code will check for close
2329          * cooperators and put the close queue at the front of the service
2330          * tree.  If possible, merge the expiring queue with the new cfqq.
2331          */
2332         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2333         if (new_cfqq) {
2334                 if (!cfqq->new_cfqq)
2335                         cfq_setup_merge(cfqq, new_cfqq);
2336                 goto expire;
2337         }
2338
2339         /*
2340          * No requests pending. If the active queue still has requests in
2341          * flight or is idling for a new request, allow either of these
2342          * conditions to happen (or time out) before selecting a new queue.
2343          */
2344         if (timer_pending(&cfqd->idle_slice_timer)) {
2345                 cfqq = NULL;
2346                 goto keep_queue;
2347         }
2348
2349         /*
2350          * This is a deep seek queue, but the device is much faster than
2351          * the queue can deliver, don't idle
2352          **/
2353         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2354             (cfq_cfqq_slice_new(cfqq) ||
2355             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2356                 cfq_clear_cfqq_deep(cfqq);
2357                 cfq_clear_cfqq_idle_window(cfqq);
2358         }
2359
2360         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2361                 cfqq = NULL;
2362                 goto keep_queue;
2363         }
2364
2365         /*
2366          * If group idle is enabled and there are requests dispatched from
2367          * this group, wait for requests to complete.
2368          */
2369 check_group_idle:
2370         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2371             && cfqq->cfqg->dispatched) {
2372                 cfqq = NULL;
2373                 goto keep_queue;
2374         }
2375
2376 expire:
2377         cfq_slice_expired(cfqd, 0);
2378 new_queue:
2379         /*
2380          * Current queue expired. Check if we have to switch to a new
2381          * service tree
2382          */
2383         if (!new_cfqq)
2384                 cfq_choose_cfqg(cfqd);
2385
2386         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2387 keep_queue:
2388         return cfqq;
2389 }
2390
2391 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2392 {
2393         int dispatched = 0;
2394
2395         while (cfqq->next_rq) {
2396                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2397                 dispatched++;
2398         }
2399
2400         BUG_ON(!list_empty(&cfqq->fifo));
2401
2402         /* By default cfqq is not expired if it is empty. Do it explicitly */
2403         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2404         return dispatched;
2405 }
2406
2407 /*
2408  * Drain our current requests. Used for barriers and when switching
2409  * io schedulers on-the-fly.
2410  */
2411 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2412 {
2413         struct cfq_queue *cfqq;
2414         int dispatched = 0;
2415
2416         /* Expire the timeslice of the current active queue first */
2417         cfq_slice_expired(cfqd, 0);
2418         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2419                 __cfq_set_active_queue(cfqd, cfqq);
2420                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2421         }
2422
2423         BUG_ON(cfqd->busy_queues);
2424
2425         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2426         return dispatched;
2427 }
2428
2429 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2430         struct cfq_queue *cfqq)
2431 {
2432         /* the queue hasn't finished any request, can't estimate */
2433         if (cfq_cfqq_slice_new(cfqq))
2434                 return true;
2435         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2436                 cfqq->slice_end))
2437                 return true;
2438
2439         return false;
2440 }
2441
2442 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2443 {
2444         unsigned int max_dispatch;
2445
2446         /*
2447          * Drain async requests before we start sync IO
2448          */
2449         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2450                 return false;
2451
2452         /*
2453          * If this is an async queue and we have sync IO in flight, let it wait
2454          */
2455         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2456                 return false;
2457
2458         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2459         if (cfq_class_idle(cfqq))
2460                 max_dispatch = 1;
2461
2462         /*
2463          * Does this cfqq already have too much IO in flight?
2464          */
2465         if (cfqq->dispatched >= max_dispatch) {
2466                 bool promote_sync = false;
2467                 /*
2468                  * idle queue must always only have a single IO in flight
2469                  */
2470                 if (cfq_class_idle(cfqq))
2471                         return false;
2472
2473                 /*
2474                  * If there is only one sync queue
2475                  * we can ignore async queue here and give the sync
2476                  * queue no dispatch limit. The reason is a sync queue can
2477                  * preempt async queue, limiting the sync queue doesn't make
2478                  * sense. This is useful for aiostress test.
2479                  */
2480                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2481                         promote_sync = true;
2482
2483                 /*
2484                  * We have other queues, don't allow more IO from this one
2485                  */
2486                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2487                                 !promote_sync)
2488                         return false;
2489
2490                 /*
2491                  * Sole queue user, no limit
2492                  */
2493                 if (cfqd->busy_queues == 1 || promote_sync)
2494                         max_dispatch = -1;
2495                 else
2496                         /*
2497                          * Normally we start throttling cfqq when cfq_quantum/2
2498                          * requests have been dispatched. But we can drive
2499                          * deeper queue depths at the beginning of slice
2500                          * subjected to upper limit of cfq_quantum.
2501                          * */
2502                         max_dispatch = cfqd->cfq_quantum;
2503         }
2504
2505         /*
2506          * Async queues must wait a bit before being allowed dispatch.
2507          * We also ramp up the dispatch depth gradually for async IO,
2508          * based on the last sync IO we serviced
2509          */
2510         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2511                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2512                 unsigned int depth;
2513
2514                 depth = last_sync / cfqd->cfq_slice[1];
2515                 if (!depth && !cfqq->dispatched)
2516                         depth = 1;
2517                 if (depth < max_dispatch)
2518                         max_dispatch = depth;
2519         }
2520
2521         /*
2522          * If we're below the current max, allow a dispatch
2523          */
2524         return cfqq->dispatched < max_dispatch;
2525 }
2526
2527 /*
2528  * Dispatch a request from cfqq, moving them to the request queue
2529  * dispatch list.
2530  */
2531 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2532 {
2533         struct request *rq;
2534
2535         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2536
2537         if (!cfq_may_dispatch(cfqd, cfqq))
2538                 return false;
2539
2540         /*
2541          * follow expired path, else get first next available
2542          */
2543         rq = cfq_check_fifo(cfqq);
2544         if (!rq)
2545                 rq = cfqq->next_rq;
2546
2547         /*
2548          * insert request into driver dispatch list
2549          */
2550         cfq_dispatch_insert(cfqd->queue, rq);
2551
2552         if (!cfqd->active_cic) {
2553                 struct cfq_io_context *cic = RQ_CIC(rq);
2554
2555                 atomic_long_inc(&cic->ioc->refcount);
2556                 cfqd->active_cic = cic;
2557         }
2558
2559         return true;
2560 }
2561
2562 /*
2563  * Find the cfqq that we need to service and move a request from that to the
2564  * dispatch list
2565  */
2566 static int cfq_dispatch_requests(struct request_queue *q, int force)
2567 {
2568         struct cfq_data *cfqd = q->elevator->elevator_data;
2569         struct cfq_queue *cfqq;
2570
2571         if (!cfqd->busy_queues)
2572                 return 0;
2573
2574         if (unlikely(force))
2575                 return cfq_forced_dispatch(cfqd);
2576
2577         cfqq = cfq_select_queue(cfqd);
2578         if (!cfqq)
2579                 return 0;
2580
2581         /*
2582          * Dispatch a request from this cfqq, if it is allowed
2583          */
2584         if (!cfq_dispatch_request(cfqd, cfqq))
2585                 return 0;
2586
2587         cfqq->slice_dispatch++;
2588         cfq_clear_cfqq_must_dispatch(cfqq);
2589
2590         /*
2591          * expire an async queue immediately if it has used up its slice. idle
2592          * queue always expire after 1 dispatch round.
2593          */
2594         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2595             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2596             cfq_class_idle(cfqq))) {
2597                 cfqq->slice_end = jiffies + 1;
2598                 cfq_slice_expired(cfqd, 0);
2599         }
2600
2601         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2602         return 1;
2603 }
2604
2605 /*
2606  * task holds one reference to the queue, dropped when task exits. each rq
2607  * in-flight on this queue also holds a reference, dropped when rq is freed.
2608  *
2609  * Each cfq queue took a reference on the parent group. Drop it now.
2610  * queue lock must be held here.
2611  */
2612 static void cfq_put_queue(struct cfq_queue *cfqq)
2613 {
2614         struct cfq_data *cfqd = cfqq->cfqd;
2615         struct cfq_group *cfqg;
2616
2617         BUG_ON(cfqq->ref <= 0);
2618
2619         cfqq->ref--;
2620         if (cfqq->ref)
2621                 return;
2622
2623         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2624         BUG_ON(rb_first(&cfqq->sort_list));
2625         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2626         cfqg = cfqq->cfqg;
2627
2628         if (unlikely(cfqd->active_queue == cfqq)) {
2629                 __cfq_slice_expired(cfqd, cfqq, 0);
2630                 cfq_schedule_dispatch(cfqd);
2631         }
2632
2633         BUG_ON(cfq_cfqq_on_rr(cfqq));
2634         kmem_cache_free(cfq_pool, cfqq);
2635         cfq_put_cfqg(cfqg);
2636 }
2637
2638 /*
2639  * Call func for each cic attached to this ioc.
2640  */
2641 static void
2642 call_for_each_cic(struct io_context *ioc,
2643                   void (*func)(struct io_context *, struct cfq_io_context *))
2644 {
2645         struct cfq_io_context *cic;
2646         struct hlist_node *n;
2647
2648         rcu_read_lock();
2649
2650         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2651                 func(ioc, cic);
2652
2653         rcu_read_unlock();
2654 }
2655
2656 static void cfq_cic_free_rcu(struct rcu_head *head)
2657 {
2658         struct cfq_io_context *cic;
2659
2660         cic = container_of(head, struct cfq_io_context, rcu_head);
2661
2662         kmem_cache_free(cfq_ioc_pool, cic);
2663         elv_ioc_count_dec(cfq_ioc_count);
2664
2665         if (ioc_gone) {
2666                 /*
2667                  * CFQ scheduler is exiting, grab exit lock and check
2668                  * the pending io context count. If it hits zero,
2669                  * complete ioc_gone and set it back to NULL
2670                  */
2671                 spin_lock(&ioc_gone_lock);
2672                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2673                         complete(ioc_gone);
2674                         ioc_gone = NULL;
2675                 }
2676                 spin_unlock(&ioc_gone_lock);
2677         }
2678 }
2679
2680 static void cfq_cic_free(struct cfq_io_context *cic)
2681 {
2682         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2683 }
2684
2685 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2686 {
2687         unsigned long flags;
2688         unsigned long dead_key = (unsigned long) cic->key;
2689
2690         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2691
2692         spin_lock_irqsave(&ioc->lock, flags);
2693         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2694         hlist_del_rcu(&cic->cic_list);
2695         spin_unlock_irqrestore(&ioc->lock, flags);
2696
2697         cfq_cic_free(cic);
2698 }
2699
2700 /*
2701  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2702  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2703  * and ->trim() which is called with the task lock held
2704  */
2705 static void cfq_free_io_context(struct io_context *ioc)
2706 {
2707         /*
2708          * ioc->refcount is zero here, or we are called from elv_unregister(),
2709          * so no more cic's are allowed to be linked into this ioc.  So it
2710          * should be ok to iterate over the known list, we will see all cic's
2711          * since no new ones are added.
2712          */
2713         call_for_each_cic(ioc, cic_free_func);
2714 }
2715
2716 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2717 {
2718         struct cfq_queue *__cfqq, *next;
2719
2720         /*
2721          * If this queue was scheduled to merge with another queue, be
2722          * sure to drop the reference taken on that queue (and others in
2723          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2724          */
2725         __cfqq = cfqq->new_cfqq;
2726         while (__cfqq) {
2727                 if (__cfqq == cfqq) {
2728                         WARN(1, "cfqq->new_cfqq loop detected\n");
2729                         break;
2730                 }
2731                 next = __cfqq->new_cfqq;
2732                 cfq_put_queue(__cfqq);
2733                 __cfqq = next;
2734         }
2735 }
2736
2737 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2738 {
2739         if (unlikely(cfqq == cfqd->active_queue)) {
2740                 __cfq_slice_expired(cfqd, cfqq, 0);
2741                 cfq_schedule_dispatch(cfqd);
2742         }
2743
2744         cfq_put_cooperator(cfqq);
2745
2746         cfq_put_queue(cfqq);
2747 }
2748
2749 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2750                                          struct cfq_io_context *cic)
2751 {
2752         struct io_context *ioc = cic->ioc;
2753
2754         list_del_init(&cic->queue_list);
2755
2756         /*
2757          * Make sure dead mark is seen for dead queues
2758          */
2759         smp_wmb();
2760         cic->key = cfqd_dead_key(cfqd);
2761
2762         rcu_read_lock();
2763         if (rcu_dereference(ioc->ioc_data) == cic) {
2764                 rcu_read_unlock();
2765                 spin_lock(&ioc->lock);
2766                 rcu_assign_pointer(ioc->ioc_data, NULL);
2767                 spin_unlock(&ioc->lock);
2768         } else
2769                 rcu_read_unlock();
2770
2771         if (cic->cfqq[BLK_RW_ASYNC]) {
2772                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2773                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2774         }
2775
2776         if (cic->cfqq[BLK_RW_SYNC]) {
2777                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2778                 cic->cfqq[BLK_RW_SYNC] = NULL;
2779         }
2780 }
2781
2782 static void cfq_exit_single_io_context(struct io_context *ioc,
2783                                        struct cfq_io_context *cic)
2784 {
2785         struct cfq_data *cfqd = cic_to_cfqd(cic);
2786
2787         if (cfqd) {
2788                 struct request_queue *q = cfqd->queue;
2789                 unsigned long flags;
2790
2791                 spin_lock_irqsave(q->queue_lock, flags);
2792
2793                 /*
2794                  * Ensure we get a fresh copy of the ->key to prevent
2795                  * race between exiting task and queue
2796                  */
2797                 smp_read_barrier_depends();
2798                 if (cic->key == cfqd)
2799                         __cfq_exit_single_io_context(cfqd, cic);
2800
2801                 spin_unlock_irqrestore(q->queue_lock, flags);
2802         }
2803 }
2804
2805 /*
2806  * The process that ioc belongs to has exited, we need to clean up
2807  * and put the internal structures we have that belongs to that process.
2808  */
2809 static void cfq_exit_io_context(struct io_context *ioc)
2810 {
2811         call_for_each_cic(ioc, cfq_exit_single_io_context);
2812 }
2813
2814 static struct cfq_io_context *
2815 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2816 {
2817         struct cfq_io_context *cic;
2818
2819         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2820                                                         cfqd->queue->node);
2821         if (cic) {
2822                 cic->ttime.last_end_request = jiffies;
2823                 INIT_LIST_HEAD(&cic->queue_list);
2824                 INIT_HLIST_NODE(&cic->cic_list);
2825                 cic->dtor = cfq_free_io_context;
2826                 cic->exit = cfq_exit_io_context;
2827                 elv_ioc_count_inc(cfq_ioc_count);
2828         }
2829
2830         return cic;
2831 }
2832
2833 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2834 {
2835         struct task_struct *tsk = current;
2836         int ioprio_class;
2837
2838         if (!cfq_cfqq_prio_changed(cfqq))
2839                 return;
2840
2841         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2842         switch (ioprio_class) {
2843         default:
2844                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2845         case IOPRIO_CLASS_NONE:
2846                 /*
2847                  * no prio set, inherit CPU scheduling settings
2848                  */
2849                 cfqq->ioprio = task_nice_ioprio(tsk);
2850                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2851                 break;
2852         case IOPRIO_CLASS_RT:
2853                 cfqq->ioprio = task_ioprio(ioc);
2854                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2855                 break;
2856         case IOPRIO_CLASS_BE:
2857                 cfqq->ioprio = task_ioprio(ioc);
2858                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2859                 break;
2860         case IOPRIO_CLASS_IDLE:
2861                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2862                 cfqq->ioprio = 7;
2863                 cfq_clear_cfqq_idle_window(cfqq);
2864                 break;
2865         }
2866
2867         /*
2868          * keep track of original prio settings in case we have to temporarily
2869          * elevate the priority of this queue
2870          */
2871         cfqq->org_ioprio = cfqq->ioprio;
2872         cfq_clear_cfqq_prio_changed(cfqq);
2873 }
2874
2875 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2876 {
2877         struct cfq_data *cfqd = cic_to_cfqd(cic);
2878         struct cfq_queue *cfqq;
2879         unsigned long flags;
2880
2881         if (unlikely(!cfqd))
2882                 return;
2883
2884         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2885
2886         cfqq = cic->cfqq[BLK_RW_ASYNC];
2887         if (cfqq) {
2888                 struct cfq_queue *new_cfqq;
2889                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2890                                                 GFP_ATOMIC);
2891                 if (new_cfqq) {
2892                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2893                         cfq_put_queue(cfqq);
2894                 }
2895         }
2896
2897         cfqq = cic->cfqq[BLK_RW_SYNC];
2898         if (cfqq)
2899                 cfq_mark_cfqq_prio_changed(cfqq);
2900
2901         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2902 }
2903
2904 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2905 {
2906         call_for_each_cic(ioc, changed_ioprio);
2907         ioc->ioprio_changed = 0;
2908 }
2909
2910 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2911                           pid_t pid, bool is_sync)
2912 {
2913         RB_CLEAR_NODE(&cfqq->rb_node);
2914         RB_CLEAR_NODE(&cfqq->p_node);
2915         INIT_LIST_HEAD(&cfqq->fifo);
2916
2917         cfqq->ref = 0;
2918         cfqq->cfqd = cfqd;
2919
2920         cfq_mark_cfqq_prio_changed(cfqq);
2921
2922         if (is_sync) {
2923                 if (!cfq_class_idle(cfqq))
2924                         cfq_mark_cfqq_idle_window(cfqq);
2925                 cfq_mark_cfqq_sync(cfqq);
2926         }
2927         cfqq->pid = pid;
2928 }
2929
2930 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2931 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2932 {
2933         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2934         struct cfq_data *cfqd = cic_to_cfqd(cic);
2935         unsigned long flags;
2936         struct request_queue *q;
2937
2938         if (unlikely(!cfqd))
2939                 return;
2940
2941         q = cfqd->queue;
2942
2943         spin_lock_irqsave(q->queue_lock, flags);
2944
2945         if (sync_cfqq) {
2946                 /*
2947                  * Drop reference to sync queue. A new sync queue will be
2948                  * assigned in new group upon arrival of a fresh request.
2949                  */
2950                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2951                 cic_set_cfqq(cic, NULL, 1);
2952                 cfq_put_queue(sync_cfqq);
2953         }
2954
2955         spin_unlock_irqrestore(q->queue_lock, flags);
2956 }
2957
2958 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2959 {
2960         call_for_each_cic(ioc, changed_cgroup);
2961         ioc->cgroup_changed = 0;
2962 }
2963 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2964
2965 static struct cfq_queue *
2966 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2967                      struct io_context *ioc, gfp_t gfp_mask)
2968 {
2969         struct cfq_queue *cfqq, *new_cfqq = NULL;
2970         struct cfq_io_context *cic;
2971         struct cfq_group *cfqg;
2972
2973 retry:
2974         cfqg = cfq_get_cfqg(cfqd);
2975         cic = cfq_cic_lookup(cfqd, ioc);
2976         /* cic always exists here */
2977         cfqq = cic_to_cfqq(cic, is_sync);
2978
2979         /*
2980          * Always try a new alloc if we fell back to the OOM cfqq
2981          * originally, since it should just be a temporary situation.
2982          */
2983         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2984                 cfqq = NULL;
2985                 if (new_cfqq) {
2986                         cfqq = new_cfqq;
2987                         new_cfqq = NULL;
2988                 } else if (gfp_mask & __GFP_WAIT) {
2989                         spin_unlock_irq(cfqd->queue->queue_lock);
2990                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2991                                         gfp_mask | __GFP_ZERO,
2992                                         cfqd->queue->node);
2993                         spin_lock_irq(cfqd->queue->queue_lock);
2994                         if (new_cfqq)
2995                                 goto retry;
2996                 } else {
2997                         cfqq = kmem_cache_alloc_node(cfq_pool,
2998                                         gfp_mask | __GFP_ZERO,
2999                                         cfqd->queue->node);
3000                 }
3001
3002                 if (cfqq) {
3003                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3004                         cfq_init_prio_data(cfqq, ioc);
3005                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3006                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3007                 } else
3008                         cfqq = &cfqd->oom_cfqq;
3009         }
3010
3011         if (new_cfqq)
3012                 kmem_cache_free(cfq_pool, new_cfqq);
3013
3014         return cfqq;
3015 }
3016
3017 static struct cfq_queue **
3018 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3019 {
3020         switch (ioprio_class) {
3021         case IOPRIO_CLASS_RT:
3022                 return &cfqd->async_cfqq[0][ioprio];
3023         case IOPRIO_CLASS_BE:
3024                 return &cfqd->async_cfqq[1][ioprio];
3025         case IOPRIO_CLASS_IDLE:
3026                 return &cfqd->async_idle_cfqq;
3027         default:
3028                 BUG();
3029         }
3030 }
3031
3032 static struct cfq_queue *
3033 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3034               gfp_t gfp_mask)
3035 {
3036         const int ioprio = task_ioprio(ioc);
3037         const int ioprio_class = task_ioprio_class(ioc);
3038         struct cfq_queue **async_cfqq = NULL;
3039         struct cfq_queue *cfqq = NULL;
3040
3041         if (!is_sync) {
3042                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3043                 cfqq = *async_cfqq;
3044         }
3045
3046         if (!cfqq)
3047                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3048
3049         /*
3050          * pin the queue now that it's allocated, scheduler exit will prune it
3051          */
3052         if (!is_sync && !(*async_cfqq)) {
3053                 cfqq->ref++;
3054                 *async_cfqq = cfqq;
3055         }
3056
3057         cfqq->ref++;
3058         return cfqq;
3059 }
3060
3061 /*
3062  * We drop cfq io contexts lazily, so we may find a dead one.
3063  */
3064 static void
3065 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3066                   struct cfq_io_context *cic)
3067 {
3068         unsigned long flags;
3069
3070         WARN_ON(!list_empty(&cic->queue_list));
3071         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3072
3073         spin_lock_irqsave(&ioc->lock, flags);
3074
3075         BUG_ON(rcu_dereference_check(ioc->ioc_data,
3076                 lockdep_is_held(&ioc->lock)) == cic);
3077
3078         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3079         hlist_del_rcu(&cic->cic_list);
3080         spin_unlock_irqrestore(&ioc->lock, flags);
3081
3082         cfq_cic_free(cic);
3083 }
3084
3085 static struct cfq_io_context *
3086 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3087 {
3088         struct cfq_io_context *cic;
3089         unsigned long flags;
3090
3091         if (unlikely(!ioc))
3092                 return NULL;
3093
3094         rcu_read_lock();
3095
3096         /*
3097          * we maintain a last-hit cache, to avoid browsing over the tree
3098          */
3099         cic = rcu_dereference(ioc->ioc_data);
3100         if (cic && cic->key == cfqd) {
3101                 rcu_read_unlock();
3102                 return cic;
3103         }
3104
3105         do {
3106                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3107                 rcu_read_unlock();
3108                 if (!cic)
3109                         break;
3110                 if (unlikely(cic->key != cfqd)) {
3111                         cfq_drop_dead_cic(cfqd, ioc, cic);
3112                         rcu_read_lock();
3113                         continue;
3114                 }
3115
3116                 spin_lock_irqsave(&ioc->lock, flags);
3117                 rcu_assign_pointer(ioc->ioc_data, cic);
3118                 spin_unlock_irqrestore(&ioc->lock, flags);
3119                 break;
3120         } while (1);
3121
3122         return cic;
3123 }
3124
3125 /*
3126  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3127  * the process specific cfq io context when entered from the block layer.
3128  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3129  */
3130 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3131                         struct cfq_io_context *cic, gfp_t gfp_mask)
3132 {
3133         unsigned long flags;
3134         int ret;
3135
3136         ret = radix_tree_preload(gfp_mask);
3137         if (!ret) {
3138                 cic->ioc = ioc;
3139                 cic->key = cfqd;
3140
3141                 spin_lock_irqsave(&ioc->lock, flags);
3142                 ret = radix_tree_insert(&ioc->radix_root,
3143                                                 cfqd->cic_index, cic);
3144                 if (!ret)
3145                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3146                 spin_unlock_irqrestore(&ioc->lock, flags);
3147
3148                 radix_tree_preload_end();
3149
3150                 if (!ret) {
3151                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3152                         list_add(&cic->queue_list, &cfqd->cic_list);
3153                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3154                 }
3155         }
3156
3157         if (ret)
3158                 printk(KERN_ERR "cfq: cic link failed!\n");
3159
3160         return ret;
3161 }
3162
3163 /*
3164  * Setup general io context and cfq io context. There can be several cfq
3165  * io contexts per general io context, if this process is doing io to more
3166  * than one device managed by cfq.
3167  */
3168 static struct cfq_io_context *
3169 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3170 {
3171         struct io_context *ioc = NULL;
3172         struct cfq_io_context *cic;
3173
3174         might_sleep_if(gfp_mask & __GFP_WAIT);
3175
3176         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3177         if (!ioc)
3178                 return NULL;
3179
3180         cic = cfq_cic_lookup(cfqd, ioc);
3181         if (cic)
3182                 goto out;
3183
3184         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3185         if (cic == NULL)
3186                 goto err;
3187
3188         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3189                 goto err_free;
3190
3191 out:
3192         smp_read_barrier_depends();
3193         if (unlikely(ioc->ioprio_changed))
3194                 cfq_ioc_set_ioprio(ioc);
3195
3196 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3197         if (unlikely(ioc->cgroup_changed))
3198                 cfq_ioc_set_cgroup(ioc);
3199 #endif
3200         return cic;
3201 err_free:
3202         cfq_cic_free(cic);
3203 err:
3204         put_io_context(ioc);
3205         return NULL;
3206 }
3207
3208 static void
3209 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3210 {
3211         unsigned long elapsed = jiffies - ttime->last_end_request;
3212         elapsed = min(elapsed, 2UL * slice_idle);
3213
3214         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3215         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3216         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3217 }
3218
3219 static void
3220 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3221         struct cfq_io_context *cic)
3222 {
3223         if (cfq_cfqq_sync(cfqq))
3224                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
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.ttime_samples)) {
3274                 if (cic->ttime.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          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3337          */
3338         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3339                 return true;
3340
3341         /* An idle queue should not be idle now for some reason */
3342         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3343                 return true;
3344
3345         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3346                 return false;
3347
3348         /*
3349          * if this request is as-good as one we would expect from the
3350          * current cfqq, let it preempt
3351          */
3352         if (cfq_rq_close(cfqd, cfqq, rq))
3353                 return true;
3354
3355         return false;
3356 }
3357
3358 /*
3359  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3360  * let it have half of its nominal slice.
3361  */
3362 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3363 {
3364         struct cfq_queue *old_cfqq = cfqd->active_queue;
3365
3366         cfq_log_cfqq(cfqd, cfqq, "preempt");
3367         cfq_slice_expired(cfqd, 1);
3368
3369         /*
3370          * workload type is changed, don't save slice, otherwise preempt
3371          * doesn't happen
3372          */
3373         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3374                 cfqq->cfqg->saved_workload_slice = 0;
3375
3376         /*
3377          * Put the new queue at the front of the of the current list,
3378          * so we know that it will be selected next.
3379          */
3380         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3381
3382         cfq_service_tree_add(cfqd, cfqq, 1);
3383
3384         cfqq->slice_end = 0;
3385         cfq_mark_cfqq_slice_new(cfqq);
3386 }
3387
3388 /*
3389  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3390  * something we should do about it
3391  */
3392 static void
3393 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3394                 struct request *rq)
3395 {
3396         struct cfq_io_context *cic = RQ_CIC(rq);
3397
3398         cfqd->rq_queued++;
3399
3400         cfq_update_io_thinktime(cfqd, cfqq, cic);
3401         cfq_update_io_seektime(cfqd, cfqq, rq);
3402         cfq_update_idle_window(cfqd, cfqq, cic);
3403
3404         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3405
3406         if (cfqq == cfqd->active_queue) {
3407                 /*
3408                  * Remember that we saw a request from this process, but
3409                  * don't start queuing just yet. Otherwise we risk seeing lots
3410                  * of tiny requests, because we disrupt the normal plugging
3411                  * and merging. If the request is already larger than a single
3412                  * page, let it rip immediately. For that case we assume that
3413                  * merging is already done. Ditto for a busy system that
3414                  * has other work pending, don't risk delaying until the
3415                  * idle timer unplug to continue working.
3416                  */
3417                 if (cfq_cfqq_wait_request(cfqq)) {
3418                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3419                             cfqd->busy_queues > 1) {
3420                                 cfq_del_timer(cfqd, cfqq);
3421                                 cfq_clear_cfqq_wait_request(cfqq);
3422                                 __blk_run_queue(cfqd->queue);
3423                         } else {
3424                                 cfq_blkiocg_update_idle_time_stats(
3425                                                 &cfqq->cfqg->blkg);
3426                                 cfq_mark_cfqq_must_dispatch(cfqq);
3427                         }
3428                 }
3429         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3430                 /*
3431                  * not the active queue - expire current slice if it is
3432                  * idle and has expired it's mean thinktime or this new queue
3433                  * has some old slice time left and is of higher priority or
3434                  * this new queue is RT and the current one is BE
3435                  */
3436                 cfq_preempt_queue(cfqd, cfqq);
3437                 __blk_run_queue(cfqd->queue);
3438         }
3439 }
3440
3441 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3442 {
3443         struct cfq_data *cfqd = q->elevator->elevator_data;
3444         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3445
3446         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3447         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3448
3449         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3450         list_add_tail(&rq->queuelist, &cfqq->fifo);
3451         cfq_add_rq_rb(rq);
3452         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3453                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3454                         rq_is_sync(rq));
3455         cfq_rq_enqueued(cfqd, cfqq, rq);
3456 }
3457
3458 /*
3459  * Update hw_tag based on peak queue depth over 50 samples under
3460  * sufficient load.
3461  */
3462 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3463 {
3464         struct cfq_queue *cfqq = cfqd->active_queue;
3465
3466         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3467                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3468
3469         if (cfqd->hw_tag == 1)
3470                 return;
3471
3472         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3473             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3474                 return;
3475
3476         /*
3477          * If active queue hasn't enough requests and can idle, cfq might not
3478          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3479          * case
3480          */
3481         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3482             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3483             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3484                 return;
3485
3486         if (cfqd->hw_tag_samples++ < 50)
3487                 return;
3488
3489         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3490                 cfqd->hw_tag = 1;
3491         else
3492                 cfqd->hw_tag = 0;
3493 }
3494
3495 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3496 {
3497         struct cfq_io_context *cic = cfqd->active_cic;
3498
3499         /* If the queue already has requests, don't wait */
3500         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3501                 return false;
3502
3503         /* If there are other queues in the group, don't wait */
3504         if (cfqq->cfqg->nr_cfqq > 1)
3505                 return false;
3506
3507         if (cfq_slice_used(cfqq))
3508                 return true;
3509
3510         /* if slice left is less than think time, wait busy */
3511         if (cic && sample_valid(cic->ttime.ttime_samples)
3512             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3513                 return true;
3514
3515         /*
3516          * If think times is less than a jiffy than ttime_mean=0 and above
3517          * will not be true. It might happen that slice has not expired yet
3518          * but will expire soon (4-5 ns) during select_queue(). To cover the
3519          * case where think time is less than a jiffy, mark the queue wait
3520          * busy if only 1 jiffy is left in the slice.
3521          */
3522         if (cfqq->slice_end - jiffies == 1)
3523                 return true;
3524
3525         return false;
3526 }
3527
3528 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3529 {
3530         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3531         struct cfq_data *cfqd = cfqq->cfqd;
3532         const int sync = rq_is_sync(rq);
3533         unsigned long now;
3534
3535         now = jiffies;
3536         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3537                      !!(rq->cmd_flags & REQ_NOIDLE));
3538
3539         cfq_update_hw_tag(cfqd);
3540
3541         WARN_ON(!cfqd->rq_in_driver);
3542         WARN_ON(!cfqq->dispatched);
3543         cfqd->rq_in_driver--;
3544         cfqq->dispatched--;
3545         (RQ_CFQG(rq))->dispatched--;
3546         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3547                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3548                         rq_data_dir(rq), rq_is_sync(rq));
3549
3550         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3551
3552         if (sync) {
3553                 RQ_CIC(rq)->ttime.last_end_request = now;
3554                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3555                         cfqd->last_delayed_sync = now;
3556         }
3557
3558         /*
3559          * If this is the active queue, check if it needs to be expired,
3560          * or if we want to idle in case it has no pending requests.
3561          */
3562         if (cfqd->active_queue == cfqq) {
3563                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3564
3565                 if (cfq_cfqq_slice_new(cfqq)) {
3566                         cfq_set_prio_slice(cfqd, cfqq);
3567                         cfq_clear_cfqq_slice_new(cfqq);
3568                 }
3569
3570                 /*
3571                  * Should we wait for next request to come in before we expire
3572                  * the queue.
3573                  */
3574                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3575                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3576                         if (!cfqd->cfq_slice_idle)
3577                                 extend_sl = cfqd->cfq_group_idle;
3578                         cfqq->slice_end = jiffies + extend_sl;
3579                         cfq_mark_cfqq_wait_busy(cfqq);
3580                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3581                 }
3582
3583                 /*
3584                  * Idling is not enabled on:
3585                  * - expired queues
3586                  * - idle-priority queues
3587                  * - async queues
3588                  * - queues with still some requests queued
3589                  * - when there is a close cooperator
3590                  */
3591                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3592                         cfq_slice_expired(cfqd, 1);
3593                 else if (sync && cfqq_empty &&
3594                          !cfq_close_cooperator(cfqd, cfqq)) {
3595                         cfq_arm_slice_timer(cfqd);
3596                 }
3597         }
3598
3599         if (!cfqd->rq_in_driver)
3600                 cfq_schedule_dispatch(cfqd);
3601 }
3602
3603 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3604 {
3605         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3606                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3607                 return ELV_MQUEUE_MUST;
3608         }
3609
3610         return ELV_MQUEUE_MAY;
3611 }
3612
3613 static int cfq_may_queue(struct request_queue *q, int rw)
3614 {
3615         struct cfq_data *cfqd = q->elevator->elevator_data;
3616         struct task_struct *tsk = current;
3617         struct cfq_io_context *cic;
3618         struct cfq_queue *cfqq;
3619
3620         /*
3621          * don't force setup of a queue from here, as a call to may_queue
3622          * does not necessarily imply that a request actually will be queued.
3623          * so just lookup a possibly existing queue, or return 'may queue'
3624          * if that fails
3625          */
3626         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3627         if (!cic)
3628                 return ELV_MQUEUE_MAY;
3629
3630         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3631         if (cfqq) {
3632                 cfq_init_prio_data(cfqq, cic->ioc);
3633
3634                 return __cfq_may_queue(cfqq);
3635         }
3636
3637         return ELV_MQUEUE_MAY;
3638 }
3639
3640 /*
3641  * queue lock held here
3642  */
3643 static void cfq_put_request(struct request *rq)
3644 {
3645         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3646
3647         if (cfqq) {
3648                 const int rw = rq_data_dir(rq);
3649
3650                 BUG_ON(!cfqq->allocated[rw]);
3651                 cfqq->allocated[rw]--;
3652
3653                 put_io_context(RQ_CIC(rq)->ioc);
3654
3655                 rq->elevator_private[0] = NULL;
3656                 rq->elevator_private[1] = NULL;
3657
3658                 /* Put down rq reference on cfqg */
3659                 cfq_put_cfqg(RQ_CFQG(rq));
3660                 rq->elevator_private[2] = NULL;
3661
3662                 cfq_put_queue(cfqq);
3663         }
3664 }
3665
3666 static struct cfq_queue *
3667 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3668                 struct cfq_queue *cfqq)
3669 {
3670         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3671         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3672         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3673         cfq_put_queue(cfqq);
3674         return cic_to_cfqq(cic, 1);
3675 }
3676
3677 /*
3678  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3679  * was the last process referring to said cfqq.
3680  */
3681 static struct cfq_queue *
3682 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3683 {
3684         if (cfqq_process_refs(cfqq) == 1) {
3685                 cfqq->pid = current->pid;
3686                 cfq_clear_cfqq_coop(cfqq);
3687                 cfq_clear_cfqq_split_coop(cfqq);
3688                 return cfqq;
3689         }
3690
3691         cic_set_cfqq(cic, NULL, 1);
3692
3693         cfq_put_cooperator(cfqq);
3694
3695         cfq_put_queue(cfqq);
3696         return NULL;
3697 }
3698 /*
3699  * Allocate cfq data structures associated with this request.
3700  */
3701 static int
3702 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3703 {
3704         struct cfq_data *cfqd = q->elevator->elevator_data;
3705         struct cfq_io_context *cic;
3706         const int rw = rq_data_dir(rq);
3707         const bool is_sync = rq_is_sync(rq);
3708         struct cfq_queue *cfqq;
3709         unsigned long flags;
3710
3711         might_sleep_if(gfp_mask & __GFP_WAIT);
3712
3713         cic = cfq_get_io_context(cfqd, gfp_mask);
3714
3715         spin_lock_irqsave(q->queue_lock, flags);
3716
3717         if (!cic)
3718                 goto queue_fail;
3719
3720 new_queue:
3721         cfqq = cic_to_cfqq(cic, is_sync);
3722         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3723                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3724                 cic_set_cfqq(cic, cfqq, is_sync);
3725         } else {
3726                 /*
3727                  * If the queue was seeky for too long, break it apart.
3728                  */
3729                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3730                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3731                         cfqq = split_cfqq(cic, cfqq);
3732                         if (!cfqq)
3733                                 goto new_queue;
3734                 }
3735
3736                 /*
3737                  * Check to see if this queue is scheduled to merge with
3738                  * another, closely cooperating queue.  The merging of
3739                  * queues happens here as it must be done in process context.
3740                  * The reference on new_cfqq was taken in merge_cfqqs.
3741                  */
3742                 if (cfqq->new_cfqq)
3743                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3744         }
3745
3746         cfqq->allocated[rw]++;
3747
3748         cfqq->ref++;
3749         rq->elevator_private[0] = cic;
3750         rq->elevator_private[1] = cfqq;
3751         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3752         spin_unlock_irqrestore(q->queue_lock, flags);
3753         return 0;
3754
3755 queue_fail:
3756         cfq_schedule_dispatch(cfqd);
3757         spin_unlock_irqrestore(q->queue_lock, flags);
3758         cfq_log(cfqd, "set_request fail");
3759         return 1;
3760 }
3761
3762 static void cfq_kick_queue(struct work_struct *work)
3763 {
3764         struct cfq_data *cfqd =
3765                 container_of(work, struct cfq_data, unplug_work);
3766         struct request_queue *q = cfqd->queue;
3767
3768         spin_lock_irq(q->queue_lock);
3769         __blk_run_queue(cfqd->queue);
3770         spin_unlock_irq(q->queue_lock);
3771 }
3772
3773 /*
3774  * Timer running if the active_queue is currently idling inside its time slice
3775  */
3776 static void cfq_idle_slice_timer(unsigned long data)
3777 {
3778         struct cfq_data *cfqd = (struct cfq_data *) data;
3779         struct cfq_queue *cfqq;
3780         unsigned long flags;
3781         int timed_out = 1;
3782
3783         cfq_log(cfqd, "idle timer fired");
3784
3785         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3786
3787         cfqq = cfqd->active_queue;
3788         if (cfqq) {
3789                 timed_out = 0;
3790
3791                 /*
3792                  * We saw a request before the queue expired, let it through
3793                  */
3794                 if (cfq_cfqq_must_dispatch(cfqq))
3795                         goto out_kick;
3796
3797                 /*
3798                  * expired
3799                  */
3800                 if (cfq_slice_used(cfqq))
3801                         goto expire;
3802
3803                 /*
3804                  * only expire and reinvoke request handler, if there are
3805                  * other queues with pending requests
3806                  */
3807                 if (!cfqd->busy_queues)
3808                         goto out_cont;
3809
3810                 /*
3811                  * not expired and it has a request pending, let it dispatch
3812                  */
3813                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3814                         goto out_kick;
3815
3816                 /*
3817                  * Queue depth flag is reset only when the idle didn't succeed
3818                  */
3819                 cfq_clear_cfqq_deep(cfqq);
3820         }
3821 expire:
3822         cfq_slice_expired(cfqd, timed_out);
3823 out_kick:
3824         cfq_schedule_dispatch(cfqd);
3825 out_cont:
3826         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3827 }
3828
3829 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3830 {
3831         del_timer_sync(&cfqd->idle_slice_timer);
3832         cancel_work_sync(&cfqd->unplug_work);
3833 }
3834
3835 static void cfq_put_async_queues(struct cfq_data *cfqd)
3836 {
3837         int i;
3838
3839         for (i = 0; i < IOPRIO_BE_NR; i++) {
3840                 if (cfqd->async_cfqq[0][i])
3841                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3842                 if (cfqd->async_cfqq[1][i])
3843                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3844         }
3845
3846         if (cfqd->async_idle_cfqq)
3847                 cfq_put_queue(cfqd->async_idle_cfqq);
3848 }
3849
3850 static void cfq_exit_queue(struct elevator_queue *e)
3851 {
3852         struct cfq_data *cfqd = e->elevator_data;
3853         struct request_queue *q = cfqd->queue;
3854         bool wait = false;
3855
3856         cfq_shutdown_timer_wq(cfqd);
3857
3858         spin_lock_irq(q->queue_lock);
3859
3860         if (cfqd->active_queue)
3861                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3862
3863         while (!list_empty(&cfqd->cic_list)) {
3864                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3865                                                         struct cfq_io_context,
3866                                                         queue_list);
3867
3868                 __cfq_exit_single_io_context(cfqd, cic);
3869         }
3870
3871         cfq_put_async_queues(cfqd);
3872         cfq_release_cfq_groups(cfqd);
3873
3874         /*
3875          * If there are groups which we could not unlink from blkcg list,
3876          * wait for a rcu period for them to be freed.
3877          */
3878         if (cfqd->nr_blkcg_linked_grps)
3879                 wait = true;
3880
3881         spin_unlock_irq(q->queue_lock);
3882
3883         cfq_shutdown_timer_wq(cfqd);
3884
3885         spin_lock(&cic_index_lock);
3886         ida_remove(&cic_index_ida, cfqd->cic_index);
3887         spin_unlock(&cic_index_lock);
3888
3889         /*
3890          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3891          * Do this wait only if there are other unlinked groups out
3892          * there. This can happen if cgroup deletion path claimed the
3893          * responsibility of cleaning up a group before queue cleanup code
3894          * get to the group.
3895          *
3896          * Do not call synchronize_rcu() unconditionally as there are drivers
3897          * which create/delete request queue hundreds of times during scan/boot
3898          * and synchronize_rcu() can take significant time and slow down boot.
3899          */
3900         if (wait)
3901                 synchronize_rcu();
3902
3903 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3904         /* Free up per cpu stats for root group */
3905         free_percpu(cfqd->root_group.blkg.stats_cpu);
3906 #endif
3907         kfree(cfqd);
3908 }
3909
3910 static int cfq_alloc_cic_index(void)
3911 {
3912         int index, error;
3913
3914         do {
3915                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3916                         return -ENOMEM;
3917
3918                 spin_lock(&cic_index_lock);
3919                 error = ida_get_new(&cic_index_ida, &index);
3920                 spin_unlock(&cic_index_lock);
3921                 if (error && error != -EAGAIN)
3922                         return error;
3923         } while (error);
3924
3925         return index;
3926 }
3927
3928 static void *cfq_init_queue(struct request_queue *q)
3929 {
3930         struct cfq_data *cfqd;
3931         int i, j;
3932         struct cfq_group *cfqg;
3933         struct cfq_rb_root *st;
3934
3935         i = cfq_alloc_cic_index();
3936         if (i < 0)
3937                 return NULL;
3938
3939         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3940         if (!cfqd) {
3941                 spin_lock(&cic_index_lock);
3942                 ida_remove(&cic_index_ida, i);
3943                 spin_unlock(&cic_index_lock);
3944                 return NULL;
3945         }
3946
3947         /*
3948          * Don't need take queue_lock in the routine, since we are
3949          * initializing the ioscheduler, and nobody is using cfqd
3950          */
3951         cfqd->cic_index = i;
3952
3953         /* Init root service tree */
3954         cfqd->grp_service_tree = CFQ_RB_ROOT;
3955
3956         /* Init root group */
3957         cfqg = &cfqd->root_group;
3958         for_each_cfqg_st(cfqg, i, j, st)
3959                 *st = CFQ_RB_ROOT;
3960         RB_CLEAR_NODE(&cfqg->rb_node);
3961
3962         /* Give preference to root group over other groups */
3963         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3964
3965 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3966         /*
3967          * Set root group reference to 2. One reference will be dropped when
3968          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3969          * Other reference will remain there as we don't want to delete this
3970          * group as it is statically allocated and gets destroyed when
3971          * throtl_data goes away.
3972          */
3973         cfqg->ref = 2;
3974
3975         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3976                 kfree(cfqg);
3977                 kfree(cfqd);
3978                 return NULL;
3979         }
3980
3981         rcu_read_lock();
3982
3983         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3984                                         (void *)cfqd, 0);
3985         rcu_read_unlock();
3986         cfqd->nr_blkcg_linked_grps++;
3987
3988         /* Add group on cfqd->cfqg_list */
3989         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3990 #endif
3991         /*
3992          * Not strictly needed (since RB_ROOT just clears the node and we
3993          * zeroed cfqd on alloc), but better be safe in case someone decides
3994          * to add magic to the rb code
3995          */
3996         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3997                 cfqd->prio_trees[i] = RB_ROOT;
3998
3999         /*
4000          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4001          * Grab a permanent reference to it, so that the normal code flow
4002          * will not attempt to free it.
4003          */
4004         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4005         cfqd->oom_cfqq.ref++;
4006         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4007
4008         INIT_LIST_HEAD(&cfqd->cic_list);
4009
4010         cfqd->queue = q;
4011
4012         init_timer(&cfqd->idle_slice_timer);
4013         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4014         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4015
4016         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4017
4018         cfqd->cfq_quantum = cfq_quantum;
4019         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4020         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4021         cfqd->cfq_back_max = cfq_back_max;
4022         cfqd->cfq_back_penalty = cfq_back_penalty;
4023         cfqd->cfq_slice[0] = cfq_slice_async;
4024         cfqd->cfq_slice[1] = cfq_slice_sync;
4025         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4026         cfqd->cfq_slice_idle = cfq_slice_idle;
4027         cfqd->cfq_group_idle = cfq_group_idle;
4028         cfqd->cfq_latency = 1;
4029         cfqd->hw_tag = -1;
4030         /*
4031          * we optimistically start assuming sync ops weren't delayed in last
4032          * second, in order to have larger depth for async operations.
4033          */
4034         cfqd->last_delayed_sync = jiffies - HZ;
4035         return cfqd;
4036 }
4037
4038 static void cfq_slab_kill(void)
4039 {
4040         /*
4041          * Caller already ensured that pending RCU callbacks are completed,
4042          * so we should have no busy allocations at this point.
4043          */
4044         if (cfq_pool)
4045                 kmem_cache_destroy(cfq_pool);
4046         if (cfq_ioc_pool)
4047                 kmem_cache_destroy(cfq_ioc_pool);
4048 }
4049
4050 static int __init cfq_slab_setup(void)
4051 {
4052         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4053         if (!cfq_pool)
4054                 goto fail;
4055
4056         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4057         if (!cfq_ioc_pool)
4058                 goto fail;
4059
4060         return 0;
4061 fail:
4062         cfq_slab_kill();
4063         return -ENOMEM;
4064 }
4065
4066 /*
4067  * sysfs parts below -->
4068  */
4069 static ssize_t
4070 cfq_var_show(unsigned int var, char *page)
4071 {
4072         return sprintf(page, "%d\n", var);
4073 }
4074
4075 static ssize_t
4076 cfq_var_store(unsigned int *var, const char *page, size_t count)
4077 {
4078         char *p = (char *) page;
4079
4080         *var = simple_strtoul(p, &p, 10);
4081         return count;
4082 }
4083
4084 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4085 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4086 {                                                                       \
4087         struct cfq_data *cfqd = e->elevator_data;                       \
4088         unsigned int __data = __VAR;                                    \
4089         if (__CONV)                                                     \
4090                 __data = jiffies_to_msecs(__data);                      \
4091         return cfq_var_show(__data, (page));                            \
4092 }
4093 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4094 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4095 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4096 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4097 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4098 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4099 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4100 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4101 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4102 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4103 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4104 #undef SHOW_FUNCTION
4105
4106 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4107 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4108 {                                                                       \
4109         struct cfq_data *cfqd = e->elevator_data;                       \
4110         unsigned int __data;                                            \
4111         int ret = cfq_var_store(&__data, (page), count);                \
4112         if (__data < (MIN))                                             \
4113                 __data = (MIN);                                         \
4114         else if (__data > (MAX))                                        \
4115                 __data = (MAX);                                         \
4116         if (__CONV)                                                     \
4117                 *(__PTR) = msecs_to_jiffies(__data);                    \
4118         else                                                            \
4119                 *(__PTR) = __data;                                      \
4120         return ret;                                                     \
4121 }
4122 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4123 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4124                 UINT_MAX, 1);
4125 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4126                 UINT_MAX, 1);
4127 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4128 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4129                 UINT_MAX, 0);
4130 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4131 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4132 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4133 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4134 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4135                 UINT_MAX, 0);
4136 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4137 #undef STORE_FUNCTION
4138
4139 #define CFQ_ATTR(name) \
4140         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4141
4142 static struct elv_fs_entry cfq_attrs[] = {
4143         CFQ_ATTR(quantum),
4144         CFQ_ATTR(fifo_expire_sync),
4145         CFQ_ATTR(fifo_expire_async),
4146         CFQ_ATTR(back_seek_max),
4147         CFQ_ATTR(back_seek_penalty),
4148         CFQ_ATTR(slice_sync),
4149         CFQ_ATTR(slice_async),
4150         CFQ_ATTR(slice_async_rq),
4151         CFQ_ATTR(slice_idle),
4152         CFQ_ATTR(group_idle),
4153         CFQ_ATTR(low_latency),
4154         __ATTR_NULL
4155 };
4156
4157 static struct elevator_type iosched_cfq = {
4158         .ops = {
4159                 .elevator_merge_fn =            cfq_merge,
4160                 .elevator_merged_fn =           cfq_merged_request,
4161                 .elevator_merge_req_fn =        cfq_merged_requests,
4162                 .elevator_allow_merge_fn =      cfq_allow_merge,
4163                 .elevator_bio_merged_fn =       cfq_bio_merged,
4164                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4165                 .elevator_add_req_fn =          cfq_insert_request,
4166                 .elevator_activate_req_fn =     cfq_activate_request,
4167                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4168                 .elevator_completed_req_fn =    cfq_completed_request,
4169                 .elevator_former_req_fn =       elv_rb_former_request,
4170                 .elevator_latter_req_fn =       elv_rb_latter_request,
4171                 .elevator_set_req_fn =          cfq_set_request,
4172                 .elevator_put_req_fn =          cfq_put_request,
4173                 .elevator_may_queue_fn =        cfq_may_queue,
4174                 .elevator_init_fn =             cfq_init_queue,
4175                 .elevator_exit_fn =             cfq_exit_queue,
4176                 .trim =                         cfq_free_io_context,
4177         },
4178         .elevator_attrs =       cfq_attrs,
4179         .elevator_name =        "cfq",
4180         .elevator_owner =       THIS_MODULE,
4181 };
4182
4183 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4184 static struct blkio_policy_type blkio_policy_cfq = {
4185         .ops = {
4186                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4187                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4188         },
4189         .plid = BLKIO_POLICY_PROP,
4190 };
4191 #else
4192 static struct blkio_policy_type blkio_policy_cfq;
4193 #endif
4194
4195 static int __init cfq_init(void)
4196 {
4197         /*
4198          * could be 0 on HZ < 1000 setups
4199          */
4200         if (!cfq_slice_async)
4201                 cfq_slice_async = 1;
4202         if (!cfq_slice_idle)
4203                 cfq_slice_idle = 1;
4204
4205 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4206         if (!cfq_group_idle)
4207                 cfq_group_idle = 1;
4208 #else
4209                 cfq_group_idle = 0;
4210 #endif
4211         if (cfq_slab_setup())
4212                 return -ENOMEM;
4213
4214         elv_register(&iosched_cfq);
4215         blkio_policy_register(&blkio_policy_cfq);
4216
4217         return 0;
4218 }
4219
4220 static void __exit cfq_exit(void)
4221 {
4222         DECLARE_COMPLETION_ONSTACK(all_gone);
4223         blkio_policy_unregister(&blkio_policy_cfq);
4224         elv_unregister(&iosched_cfq);
4225         ioc_gone = &all_gone;
4226         /* ioc_gone's update must be visible before reading ioc_count */
4227         smp_wmb();
4228
4229         /*
4230          * this also protects us from entering cfq_slab_kill() with
4231          * pending RCU callbacks
4232          */
4233         if (elv_ioc_count_read(cfq_ioc_count))
4234                 wait_for_completion(&all_gone);
4235         ida_destroy(&cic_index_ida);
4236         cfq_slab_kill();
4237 }
4238
4239 module_init(cfq_init);
4240 module_exit(cfq_exit);
4241
4242 MODULE_AUTHOR("Jens Axboe");
4243 MODULE_LICENSE("GPL");
4244 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");