Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/cmarinas...
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181         unsigned int new_weight;
182         bool needs_update;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /*
188          * Per group busy queus average. Useful for workload slice calc. We
189          * create the array for each prio class but at run time it is used
190          * only for RT and BE class and slot for IDLE class remains unused.
191          * This is primarily done to avoid confusion and a gcc warning.
192          */
193         unsigned int busy_queues_avg[CFQ_PRIO_NR];
194         /*
195          * rr lists of queues with requests. We maintain service trees for
196          * RT and BE classes. These trees are subdivided in subclasses
197          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198          * class there is no subclassification and all the cfq queues go on
199          * a single tree service_tree_idle.
200          * Counts are embedded in the cfq_rb_root
201          */
202         struct cfq_rb_root service_trees[2][3];
203         struct cfq_rb_root service_tree_idle;
204
205         unsigned long saved_workload_slice;
206         enum wl_type_t saved_workload;
207         enum wl_prio_t saved_serving_prio;
208         struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210         struct hlist_node cfqd_node;
211         int ref;
212 #endif
213         /* number of requests that are on the dispatch list or inside driver */
214         int dispatched;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303         struct rcu_head rcu;
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 rq1;
670         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
671                 return rq2;
672         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
673                 return rq1;
674         else if ((rq2->cmd_flags & REQ_META) &&
675                  !(rq1->cmd_flags & REQ_META))
676                 return rq2;
677
678         s1 = blk_rq_pos(rq1);
679         s2 = blk_rq_pos(rq2);
680
681         /*
682          * by definition, 1KiB is 2 sectors
683          */
684         back_max = cfqd->cfq_back_max * 2;
685
686         /*
687          * Strict one way elevator _except_ in the case where we allow
688          * short backward seeks which are biased as twice the cost of a
689          * similar forward seek.
690          */
691         if (s1 >= last)
692                 d1 = s1 - last;
693         else if (s1 + back_max >= last)
694                 d1 = (last - s1) * cfqd->cfq_back_penalty;
695         else
696                 wrap |= CFQ_RQ1_WRAP;
697
698         if (s2 >= last)
699                 d2 = s2 - last;
700         else if (s2 + back_max >= last)
701                 d2 = (last - s2) * cfqd->cfq_back_penalty;
702         else
703                 wrap |= CFQ_RQ2_WRAP;
704
705         /* Found required data */
706
707         /*
708          * By doing switch() on the bit mask "wrap" we avoid having to
709          * check two variables for all permutations: --> faster!
710          */
711         switch (wrap) {
712         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
713                 if (d1 < d2)
714                         return rq1;
715                 else if (d2 < d1)
716                         return rq2;
717                 else {
718                         if (s1 >= s2)
719                                 return rq1;
720                         else
721                                 return rq2;
722                 }
723
724         case CFQ_RQ2_WRAP:
725                 return rq1;
726         case CFQ_RQ1_WRAP:
727                 return rq2;
728         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
729         default:
730                 /*
731                  * Since both rqs are wrapped,
732                  * start with the one that's further behind head
733                  * (--> only *one* back seek required),
734                  * since back seek takes more time than forward.
735                  */
736                 if (s1 <= s2)
737                         return rq1;
738                 else
739                         return rq2;
740         }
741 }
742
743 /*
744  * The below is leftmost cache rbtree addon
745  */
746 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
747 {
748         /* Service tree is empty */
749         if (!root->count)
750                 return NULL;
751
752         if (!root->left)
753                 root->left = rb_first(&root->rb);
754
755         if (root->left)
756                 return rb_entry(root->left, struct cfq_queue, rb_node);
757
758         return NULL;
759 }
760
761 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
762 {
763         if (!root->left)
764                 root->left = rb_first(&root->rb);
765
766         if (root->left)
767                 return rb_entry_cfqg(root->left);
768
769         return NULL;
770 }
771
772 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
773 {
774         rb_erase(n, root);
775         RB_CLEAR_NODE(n);
776 }
777
778 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
779 {
780         if (root->left == n)
781                 root->left = NULL;
782         rb_erase_init(n, &root->rb);
783         --root->count;
784 }
785
786 /*
787  * would be nice to take fifo expire time into account as well
788  */
789 static struct request *
790 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
791                   struct request *last)
792 {
793         struct rb_node *rbnext = rb_next(&last->rb_node);
794         struct rb_node *rbprev = rb_prev(&last->rb_node);
795         struct request *next = NULL, *prev = NULL;
796
797         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
798
799         if (rbprev)
800                 prev = rb_entry_rq(rbprev);
801
802         if (rbnext)
803                 next = rb_entry_rq(rbnext);
804         else {
805                 rbnext = rb_first(&cfqq->sort_list);
806                 if (rbnext && rbnext != &last->rb_node)
807                         next = rb_entry_rq(rbnext);
808         }
809
810         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
811 }
812
813 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
814                                       struct cfq_queue *cfqq)
815 {
816         /*
817          * just an approximation, should be ok.
818          */
819         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
820                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
821 }
822
823 static inline s64
824 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
825 {
826         return cfqg->vdisktime - st->min_vdisktime;
827 }
828
829 static void
830 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
831 {
832         struct rb_node **node = &st->rb.rb_node;
833         struct rb_node *parent = NULL;
834         struct cfq_group *__cfqg;
835         s64 key = cfqg_key(st, cfqg);
836         int left = 1;
837
838         while (*node != NULL) {
839                 parent = *node;
840                 __cfqg = rb_entry_cfqg(parent);
841
842                 if (key < cfqg_key(st, __cfqg))
843                         node = &parent->rb_left;
844                 else {
845                         node = &parent->rb_right;
846                         left = 0;
847                 }
848         }
849
850         if (left)
851                 st->left = &cfqg->rb_node;
852
853         rb_link_node(&cfqg->rb_node, parent, node);
854         rb_insert_color(&cfqg->rb_node, &st->rb);
855 }
856
857 static void
858 cfq_update_group_weight(struct cfq_group *cfqg)
859 {
860         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
861         if (cfqg->needs_update) {
862                 cfqg->weight = cfqg->new_weight;
863                 cfqg->needs_update = false;
864         }
865 }
866
867 static void
868 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
869 {
870         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
871
872         cfq_update_group_weight(cfqg);
873         __cfq_group_service_tree_add(st, cfqg);
874         st->total_weight += cfqg->weight;
875 }
876
877 static void
878 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
879 {
880         struct cfq_rb_root *st = &cfqd->grp_service_tree;
881         struct cfq_group *__cfqg;
882         struct rb_node *n;
883
884         cfqg->nr_cfqq++;
885         if (!RB_EMPTY_NODE(&cfqg->rb_node))
886                 return;
887
888         /*
889          * Currently put the group at the end. Later implement something
890          * so that groups get lesser vtime based on their weights, so that
891          * if group does not loose all if it was not continuously backlogged.
892          */
893         n = rb_last(&st->rb);
894         if (n) {
895                 __cfqg = rb_entry_cfqg(n);
896                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
897         } else
898                 cfqg->vdisktime = st->min_vdisktime;
899         cfq_group_service_tree_add(st, cfqg);
900 }
901
902 static void
903 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
904 {
905         st->total_weight -= cfqg->weight;
906         if (!RB_EMPTY_NODE(&cfqg->rb_node))
907                 cfq_rb_erase(&cfqg->rb_node, st);
908 }
909
910 static void
911 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
912 {
913         struct cfq_rb_root *st = &cfqd->grp_service_tree;
914
915         BUG_ON(cfqg->nr_cfqq < 1);
916         cfqg->nr_cfqq--;
917
918         /* If there are other cfq queues under this group, don't delete it */
919         if (cfqg->nr_cfqq)
920                 return;
921
922         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
923         cfq_group_service_tree_del(st, cfqg);
924         cfqg->saved_workload_slice = 0;
925         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
926 }
927
928 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
929                                                 unsigned int *unaccounted_time)
930 {
931         unsigned int slice_used;
932
933         /*
934          * Queue got expired before even a single request completed or
935          * got expired immediately after first request completion.
936          */
937         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
938                 /*
939                  * Also charge the seek time incurred to the group, otherwise
940                  * if there are mutiple queues in the group, each can dispatch
941                  * a single request on seeky media and cause lots of seek time
942                  * and group will never know it.
943                  */
944                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
945                                         1);
946         } else {
947                 slice_used = jiffies - cfqq->slice_start;
948                 if (slice_used > cfqq->allocated_slice) {
949                         *unaccounted_time = slice_used - cfqq->allocated_slice;
950                         slice_used = cfqq->allocated_slice;
951                 }
952                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
953                         *unaccounted_time += cfqq->slice_start -
954                                         cfqq->dispatch_start;
955         }
956
957         return slice_used;
958 }
959
960 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
961                                 struct cfq_queue *cfqq)
962 {
963         struct cfq_rb_root *st = &cfqd->grp_service_tree;
964         unsigned int used_sl, charge, unaccounted_sl = 0;
965         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
966                         - cfqg->service_tree_idle.count;
967
968         BUG_ON(nr_sync < 0);
969         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
970
971         if (iops_mode(cfqd))
972                 charge = cfqq->slice_dispatch;
973         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
974                 charge = cfqq->allocated_slice;
975
976         /* Can't update vdisktime while group is on service tree */
977         cfq_group_service_tree_del(st, cfqg);
978         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
979         /* If a new weight was requested, update now, off tree */
980         cfq_group_service_tree_add(st, cfqg);
981
982         /* This group is being expired. Save the context */
983         if (time_after(cfqd->workload_expires, jiffies)) {
984                 cfqg->saved_workload_slice = cfqd->workload_expires
985                                                 - jiffies;
986                 cfqg->saved_workload = cfqd->serving_type;
987                 cfqg->saved_serving_prio = cfqd->serving_prio;
988         } else
989                 cfqg->saved_workload_slice = 0;
990
991         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
992                                         st->min_vdisktime);
993         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
994                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
995                         iops_mode(cfqd), cfqq->nr_sectors);
996         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
997                                           unaccounted_sl);
998         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
999 }
1000
1001 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1002 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1003 {
1004         if (blkg)
1005                 return container_of(blkg, struct cfq_group, blkg);
1006         return NULL;
1007 }
1008
1009 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1010                                         unsigned int weight)
1011 {
1012         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1013         cfqg->new_weight = weight;
1014         cfqg->needs_update = true;
1015 }
1016
1017 static struct cfq_group * cfq_find_alloc_cfqg(struct cfq_data *cfqd,
1018                 struct blkio_cgroup *blkcg, int create)
1019 {
1020         struct cfq_group *cfqg = NULL;
1021         void *key = cfqd;
1022         int i, j;
1023         struct cfq_rb_root *st;
1024         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1025         unsigned int major, minor;
1026
1027         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1028         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1029                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030                 cfqg->blkg.dev = MKDEV(major, minor);
1031                 goto done;
1032         }
1033         if (cfqg || !create)
1034                 goto done;
1035
1036         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1037         if (!cfqg)
1038                 goto done;
1039
1040         for_each_cfqg_st(cfqg, i, j, st)
1041                 *st = CFQ_RB_ROOT;
1042         RB_CLEAR_NODE(&cfqg->rb_node);
1043
1044         /*
1045          * Take the initial reference that will be released on destroy
1046          * This can be thought of a joint reference by cgroup and
1047          * elevator which will be dropped by either elevator exit
1048          * or cgroup deletion path depending on who is exiting first.
1049          */
1050         cfqg->ref = 1;
1051
1052         /*
1053          * Add group onto cgroup list. It might happen that bdi->dev is
1054          * not initialized yet. Initialize this new group without major
1055          * and minor info and this info will be filled in once a new thread
1056          * comes for IO. See code above.
1057          */
1058         if (bdi->dev) {
1059                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1060                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1061                                         MKDEV(major, minor));
1062         } else
1063                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1064                                         0);
1065
1066         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1067
1068         /* Add group on cfqd list */
1069         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1070
1071 done:
1072         return cfqg;
1073 }
1074
1075 /*
1076  * Search for the cfq group current task belongs to. If create = 1, then also
1077  * create the cfq group if it does not exist. request_queue lock must be held.
1078  */
1079 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1080 {
1081         struct blkio_cgroup *blkcg;
1082         struct cfq_group *cfqg = NULL;
1083
1084         rcu_read_lock();
1085         blkcg = task_blkio_cgroup(current);
1086         cfqg = cfq_find_alloc_cfqg(cfqd, blkcg, create);
1087         if (!cfqg && create)
1088                 cfqg = &cfqd->root_group;
1089         rcu_read_unlock();
1090         return cfqg;
1091 }
1092
1093 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1094 {
1095         cfqg->ref++;
1096         return cfqg;
1097 }
1098
1099 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1100 {
1101         /* Currently, all async queues are mapped to root group */
1102         if (!cfq_cfqq_sync(cfqq))
1103                 cfqg = &cfqq->cfqd->root_group;
1104
1105         cfqq->cfqg = cfqg;
1106         /* cfqq reference on cfqg */
1107         cfqq->cfqg->ref++;
1108 }
1109
1110 static void cfq_put_cfqg(struct cfq_group *cfqg)
1111 {
1112         struct cfq_rb_root *st;
1113         int i, j;
1114
1115         BUG_ON(cfqg->ref <= 0);
1116         cfqg->ref--;
1117         if (cfqg->ref)
1118                 return;
1119         for_each_cfqg_st(cfqg, i, j, st)
1120                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1121         kfree(cfqg);
1122 }
1123
1124 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1125 {
1126         /* Something wrong if we are trying to remove same group twice */
1127         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1128
1129         hlist_del_init(&cfqg->cfqd_node);
1130
1131         /*
1132          * Put the reference taken at the time of creation so that when all
1133          * queues are gone, group can be destroyed.
1134          */
1135         cfq_put_cfqg(cfqg);
1136 }
1137
1138 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1139 {
1140         struct hlist_node *pos, *n;
1141         struct cfq_group *cfqg;
1142
1143         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1144                 /*
1145                  * If cgroup removal path got to blk_group first and removed
1146                  * it from cgroup list, then it will take care of destroying
1147                  * cfqg also.
1148                  */
1149                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1150                         cfq_destroy_cfqg(cfqd, cfqg);
1151         }
1152 }
1153
1154 /*
1155  * Blk cgroup controller notification saying that blkio_group object is being
1156  * delinked as associated cgroup object is going away. That also means that
1157  * no new IO will come in this group. So get rid of this group as soon as
1158  * any pending IO in the group is finished.
1159  *
1160  * This function is called under rcu_read_lock(). key is the rcu protected
1161  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1162  * read lock.
1163  *
1164  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1165  * it should not be NULL as even if elevator was exiting, cgroup deltion
1166  * path got to it first.
1167  */
1168 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1169 {
1170         unsigned long  flags;
1171         struct cfq_data *cfqd = key;
1172
1173         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1174         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1175         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1176 }
1177
1178 #else /* GROUP_IOSCHED */
1179 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1180 {
1181         return &cfqd->root_group;
1182 }
1183
1184 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1185 {
1186         return cfqg;
1187 }
1188
1189 static inline void
1190 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1191         cfqq->cfqg = cfqg;
1192 }
1193
1194 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1195 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1196
1197 #endif /* GROUP_IOSCHED */
1198
1199 /*
1200  * The cfqd->service_trees holds all pending cfq_queue's that have
1201  * requests waiting to be processed. It is sorted in the order that
1202  * we will service the queues.
1203  */
1204 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1205                                  bool add_front)
1206 {
1207         struct rb_node **p, *parent;
1208         struct cfq_queue *__cfqq;
1209         unsigned long rb_key;
1210         struct cfq_rb_root *service_tree;
1211         int left;
1212         int new_cfqq = 1;
1213         int group_changed = 0;
1214
1215         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1216                                                 cfqq_type(cfqq));
1217         if (cfq_class_idle(cfqq)) {
1218                 rb_key = CFQ_IDLE_DELAY;
1219                 parent = rb_last(&service_tree->rb);
1220                 if (parent && parent != &cfqq->rb_node) {
1221                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1222                         rb_key += __cfqq->rb_key;
1223                 } else
1224                         rb_key += jiffies;
1225         } else if (!add_front) {
1226                 /*
1227                  * Get our rb key offset. Subtract any residual slice
1228                  * value carried from last service. A negative resid
1229                  * count indicates slice overrun, and this should position
1230                  * the next service time further away in the tree.
1231                  */
1232                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1233                 rb_key -= cfqq->slice_resid;
1234                 cfqq->slice_resid = 0;
1235         } else {
1236                 rb_key = -HZ;
1237                 __cfqq = cfq_rb_first(service_tree);
1238                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1239         }
1240
1241         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1242                 new_cfqq = 0;
1243                 /*
1244                  * same position, nothing more to do
1245                  */
1246                 if (rb_key == cfqq->rb_key &&
1247                     cfqq->service_tree == service_tree)
1248                         return;
1249
1250                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1251                 cfqq->service_tree = NULL;
1252         }
1253
1254         left = 1;
1255         parent = NULL;
1256         cfqq->service_tree = service_tree;
1257         p = &service_tree->rb.rb_node;
1258         while (*p) {
1259                 struct rb_node **n;
1260
1261                 parent = *p;
1262                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1263
1264                 /*
1265                  * sort by key, that represents service time.
1266                  */
1267                 if (time_before(rb_key, __cfqq->rb_key))
1268                         n = &(*p)->rb_left;
1269                 else {
1270                         n = &(*p)->rb_right;
1271                         left = 0;
1272                 }
1273
1274                 p = n;
1275         }
1276
1277         if (left)
1278                 service_tree->left = &cfqq->rb_node;
1279
1280         cfqq->rb_key = rb_key;
1281         rb_link_node(&cfqq->rb_node, parent, p);
1282         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1283         service_tree->count++;
1284         if ((add_front || !new_cfqq) && !group_changed)
1285                 return;
1286         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1287 }
1288
1289 static struct cfq_queue *
1290 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1291                      sector_t sector, struct rb_node **ret_parent,
1292                      struct rb_node ***rb_link)
1293 {
1294         struct rb_node **p, *parent;
1295         struct cfq_queue *cfqq = NULL;
1296
1297         parent = NULL;
1298         p = &root->rb_node;
1299         while (*p) {
1300                 struct rb_node **n;
1301
1302                 parent = *p;
1303                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1304
1305                 /*
1306                  * Sort strictly based on sector.  Smallest to the left,
1307                  * largest to the right.
1308                  */
1309                 if (sector > blk_rq_pos(cfqq->next_rq))
1310                         n = &(*p)->rb_right;
1311                 else if (sector < blk_rq_pos(cfqq->next_rq))
1312                         n = &(*p)->rb_left;
1313                 else
1314                         break;
1315                 p = n;
1316                 cfqq = NULL;
1317         }
1318
1319         *ret_parent = parent;
1320         if (rb_link)
1321                 *rb_link = p;
1322         return cfqq;
1323 }
1324
1325 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1326 {
1327         struct rb_node **p, *parent;
1328         struct cfq_queue *__cfqq;
1329
1330         if (cfqq->p_root) {
1331                 rb_erase(&cfqq->p_node, cfqq->p_root);
1332                 cfqq->p_root = NULL;
1333         }
1334
1335         if (cfq_class_idle(cfqq))
1336                 return;
1337         if (!cfqq->next_rq)
1338                 return;
1339
1340         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1341         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1342                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1343         if (!__cfqq) {
1344                 rb_link_node(&cfqq->p_node, parent, p);
1345                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1346         } else
1347                 cfqq->p_root = NULL;
1348 }
1349
1350 /*
1351  * Update cfqq's position in the service tree.
1352  */
1353 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1354 {
1355         /*
1356          * Resorting requires the cfqq to be on the RR list already.
1357          */
1358         if (cfq_cfqq_on_rr(cfqq)) {
1359                 cfq_service_tree_add(cfqd, cfqq, 0);
1360                 cfq_prio_tree_add(cfqd, cfqq);
1361         }
1362 }
1363
1364 /*
1365  * add to busy list of queues for service, trying to be fair in ordering
1366  * the pending list according to last request service
1367  */
1368 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1369 {
1370         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1371         BUG_ON(cfq_cfqq_on_rr(cfqq));
1372         cfq_mark_cfqq_on_rr(cfqq);
1373         cfqd->busy_queues++;
1374         if (cfq_cfqq_sync(cfqq))
1375                 cfqd->busy_sync_queues++;
1376
1377         cfq_resort_rr_list(cfqd, cfqq);
1378 }
1379
1380 /*
1381  * Called when the cfqq no longer has requests pending, remove it from
1382  * the service tree.
1383  */
1384 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1385 {
1386         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1387         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1388         cfq_clear_cfqq_on_rr(cfqq);
1389
1390         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1391                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1392                 cfqq->service_tree = NULL;
1393         }
1394         if (cfqq->p_root) {
1395                 rb_erase(&cfqq->p_node, cfqq->p_root);
1396                 cfqq->p_root = NULL;
1397         }
1398
1399         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1400         BUG_ON(!cfqd->busy_queues);
1401         cfqd->busy_queues--;
1402         if (cfq_cfqq_sync(cfqq))
1403                 cfqd->busy_sync_queues--;
1404 }
1405
1406 /*
1407  * rb tree support functions
1408  */
1409 static void cfq_del_rq_rb(struct request *rq)
1410 {
1411         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1412         const int sync = rq_is_sync(rq);
1413
1414         BUG_ON(!cfqq->queued[sync]);
1415         cfqq->queued[sync]--;
1416
1417         elv_rb_del(&cfqq->sort_list, rq);
1418
1419         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1420                 /*
1421                  * Queue will be deleted from service tree when we actually
1422                  * expire it later. Right now just remove it from prio tree
1423                  * as it is empty.
1424                  */
1425                 if (cfqq->p_root) {
1426                         rb_erase(&cfqq->p_node, cfqq->p_root);
1427                         cfqq->p_root = NULL;
1428                 }
1429         }
1430 }
1431
1432 static void cfq_add_rq_rb(struct request *rq)
1433 {
1434         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1435         struct cfq_data *cfqd = cfqq->cfqd;
1436         struct request *__alias, *prev;
1437
1438         cfqq->queued[rq_is_sync(rq)]++;
1439
1440         /*
1441          * looks a little odd, but the first insert might return an alias.
1442          * if that happens, put the alias on the dispatch list
1443          */
1444         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1445                 cfq_dispatch_insert(cfqd->queue, __alias);
1446
1447         if (!cfq_cfqq_on_rr(cfqq))
1448                 cfq_add_cfqq_rr(cfqd, cfqq);
1449
1450         /*
1451          * check if this request is a better next-serve candidate
1452          */
1453         prev = cfqq->next_rq;
1454         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1455
1456         /*
1457          * adjust priority tree position, if ->next_rq changes
1458          */
1459         if (prev != cfqq->next_rq)
1460                 cfq_prio_tree_add(cfqd, cfqq);
1461
1462         BUG_ON(!cfqq->next_rq);
1463 }
1464
1465 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1466 {
1467         elv_rb_del(&cfqq->sort_list, rq);
1468         cfqq->queued[rq_is_sync(rq)]--;
1469         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1470                                         rq_data_dir(rq), rq_is_sync(rq));
1471         cfq_add_rq_rb(rq);
1472         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1473                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1474                         rq_is_sync(rq));
1475 }
1476
1477 static struct request *
1478 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1479 {
1480         struct task_struct *tsk = current;
1481         struct cfq_io_context *cic;
1482         struct cfq_queue *cfqq;
1483
1484         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1485         if (!cic)
1486                 return NULL;
1487
1488         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1489         if (cfqq) {
1490                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1491
1492                 return elv_rb_find(&cfqq->sort_list, sector);
1493         }
1494
1495         return NULL;
1496 }
1497
1498 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1499 {
1500         struct cfq_data *cfqd = q->elevator->elevator_data;
1501
1502         cfqd->rq_in_driver++;
1503         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1504                                                 cfqd->rq_in_driver);
1505
1506         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1507 }
1508
1509 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1510 {
1511         struct cfq_data *cfqd = q->elevator->elevator_data;
1512
1513         WARN_ON(!cfqd->rq_in_driver);
1514         cfqd->rq_in_driver--;
1515         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1516                                                 cfqd->rq_in_driver);
1517 }
1518
1519 static void cfq_remove_request(struct request *rq)
1520 {
1521         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1522
1523         if (cfqq->next_rq == rq)
1524                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1525
1526         list_del_init(&rq->queuelist);
1527         cfq_del_rq_rb(rq);
1528
1529         cfqq->cfqd->rq_queued--;
1530         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1531                                         rq_data_dir(rq), rq_is_sync(rq));
1532         if (rq->cmd_flags & REQ_META) {
1533                 WARN_ON(!cfqq->meta_pending);
1534                 cfqq->meta_pending--;
1535         }
1536 }
1537
1538 static int cfq_merge(struct request_queue *q, struct request **req,
1539                      struct bio *bio)
1540 {
1541         struct cfq_data *cfqd = q->elevator->elevator_data;
1542         struct request *__rq;
1543
1544         __rq = cfq_find_rq_fmerge(cfqd, bio);
1545         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1546                 *req = __rq;
1547                 return ELEVATOR_FRONT_MERGE;
1548         }
1549
1550         return ELEVATOR_NO_MERGE;
1551 }
1552
1553 static void cfq_merged_request(struct request_queue *q, struct request *req,
1554                                int type)
1555 {
1556         if (type == ELEVATOR_FRONT_MERGE) {
1557                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1558
1559                 cfq_reposition_rq_rb(cfqq, req);
1560         }
1561 }
1562
1563 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1564                                 struct bio *bio)
1565 {
1566         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1567                                         bio_data_dir(bio), cfq_bio_sync(bio));
1568 }
1569
1570 static void
1571 cfq_merged_requests(struct request_queue *q, struct request *rq,
1572                     struct request *next)
1573 {
1574         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1575         /*
1576          * reposition in fifo if next is older than rq
1577          */
1578         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1579             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1580                 list_move(&rq->queuelist, &next->queuelist);
1581                 rq_set_fifo_time(rq, rq_fifo_time(next));
1582         }
1583
1584         if (cfqq->next_rq == next)
1585                 cfqq->next_rq = rq;
1586         cfq_remove_request(next);
1587         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1588                                         rq_data_dir(next), rq_is_sync(next));
1589 }
1590
1591 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1592                            struct bio *bio)
1593 {
1594         struct cfq_data *cfqd = q->elevator->elevator_data;
1595         struct cfq_io_context *cic;
1596         struct cfq_queue *cfqq;
1597
1598         /*
1599          * Disallow merge of a sync bio into an async request.
1600          */
1601         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1602                 return false;
1603
1604         /*
1605          * Lookup the cfqq that this bio will be queued with. Allow
1606          * merge only if rq is queued there.
1607          */
1608         cic = cfq_cic_lookup(cfqd, current->io_context);
1609         if (!cic)
1610                 return false;
1611
1612         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1613         return cfqq == RQ_CFQQ(rq);
1614 }
1615
1616 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1617 {
1618         del_timer(&cfqd->idle_slice_timer);
1619         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1620 }
1621
1622 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1623                                    struct cfq_queue *cfqq)
1624 {
1625         if (cfqq) {
1626                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1627                                 cfqd->serving_prio, cfqd->serving_type);
1628                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1629                 cfqq->slice_start = 0;
1630                 cfqq->dispatch_start = jiffies;
1631                 cfqq->allocated_slice = 0;
1632                 cfqq->slice_end = 0;
1633                 cfqq->slice_dispatch = 0;
1634                 cfqq->nr_sectors = 0;
1635
1636                 cfq_clear_cfqq_wait_request(cfqq);
1637                 cfq_clear_cfqq_must_dispatch(cfqq);
1638                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1639                 cfq_clear_cfqq_fifo_expire(cfqq);
1640                 cfq_mark_cfqq_slice_new(cfqq);
1641
1642                 cfq_del_timer(cfqd, cfqq);
1643         }
1644
1645         cfqd->active_queue = cfqq;
1646 }
1647
1648 /*
1649  * current cfqq expired its slice (or was too idle), select new one
1650  */
1651 static void
1652 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1653                     bool timed_out)
1654 {
1655         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1656
1657         if (cfq_cfqq_wait_request(cfqq))
1658                 cfq_del_timer(cfqd, cfqq);
1659
1660         cfq_clear_cfqq_wait_request(cfqq);
1661         cfq_clear_cfqq_wait_busy(cfqq);
1662
1663         /*
1664          * If this cfqq is shared between multiple processes, check to
1665          * make sure that those processes are still issuing I/Os within
1666          * the mean seek distance.  If not, it may be time to break the
1667          * queues apart again.
1668          */
1669         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1670                 cfq_mark_cfqq_split_coop(cfqq);
1671
1672         /*
1673          * store what was left of this slice, if the queue idled/timed out
1674          */
1675         if (timed_out) {
1676                 if (cfq_cfqq_slice_new(cfqq))
1677                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1678                 else
1679                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1680                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1681         }
1682
1683         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1684
1685         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1686                 cfq_del_cfqq_rr(cfqd, cfqq);
1687
1688         cfq_resort_rr_list(cfqd, cfqq);
1689
1690         if (cfqq == cfqd->active_queue)
1691                 cfqd->active_queue = NULL;
1692
1693         if (cfqd->active_cic) {
1694                 put_io_context(cfqd->active_cic->ioc);
1695                 cfqd->active_cic = NULL;
1696         }
1697 }
1698
1699 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1700 {
1701         struct cfq_queue *cfqq = cfqd->active_queue;
1702
1703         if (cfqq)
1704                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1705 }
1706
1707 /*
1708  * Get next queue for service. Unless we have a queue preemption,
1709  * we'll simply select the first cfqq in the service tree.
1710  */
1711 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1712 {
1713         struct cfq_rb_root *service_tree =
1714                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1715                                         cfqd->serving_type);
1716
1717         if (!cfqd->rq_queued)
1718                 return NULL;
1719
1720         /* There is nothing to dispatch */
1721         if (!service_tree)
1722                 return NULL;
1723         if (RB_EMPTY_ROOT(&service_tree->rb))
1724                 return NULL;
1725         return cfq_rb_first(service_tree);
1726 }
1727
1728 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1729 {
1730         struct cfq_group *cfqg;
1731         struct cfq_queue *cfqq;
1732         int i, j;
1733         struct cfq_rb_root *st;
1734
1735         if (!cfqd->rq_queued)
1736                 return NULL;
1737
1738         cfqg = cfq_get_next_cfqg(cfqd);
1739         if (!cfqg)
1740                 return NULL;
1741
1742         for_each_cfqg_st(cfqg, i, j, st)
1743                 if ((cfqq = cfq_rb_first(st)) != NULL)
1744                         return cfqq;
1745         return NULL;
1746 }
1747
1748 /*
1749  * Get and set a new active queue for service.
1750  */
1751 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1752                                               struct cfq_queue *cfqq)
1753 {
1754         if (!cfqq)
1755                 cfqq = cfq_get_next_queue(cfqd);
1756
1757         __cfq_set_active_queue(cfqd, cfqq);
1758         return cfqq;
1759 }
1760
1761 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1762                                           struct request *rq)
1763 {
1764         if (blk_rq_pos(rq) >= cfqd->last_position)
1765                 return blk_rq_pos(rq) - cfqd->last_position;
1766         else
1767                 return cfqd->last_position - blk_rq_pos(rq);
1768 }
1769
1770 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1771                                struct request *rq)
1772 {
1773         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1774 }
1775
1776 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1777                                     struct cfq_queue *cur_cfqq)
1778 {
1779         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1780         struct rb_node *parent, *node;
1781         struct cfq_queue *__cfqq;
1782         sector_t sector = cfqd->last_position;
1783
1784         if (RB_EMPTY_ROOT(root))
1785                 return NULL;
1786
1787         /*
1788          * First, if we find a request starting at the end of the last
1789          * request, choose it.
1790          */
1791         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1792         if (__cfqq)
1793                 return __cfqq;
1794
1795         /*
1796          * If the exact sector wasn't found, the parent of the NULL leaf
1797          * will contain the closest sector.
1798          */
1799         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1800         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1801                 return __cfqq;
1802
1803         if (blk_rq_pos(__cfqq->next_rq) < sector)
1804                 node = rb_next(&__cfqq->p_node);
1805         else
1806                 node = rb_prev(&__cfqq->p_node);
1807         if (!node)
1808                 return NULL;
1809
1810         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1811         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1812                 return __cfqq;
1813
1814         return NULL;
1815 }
1816
1817 /*
1818  * cfqd - obvious
1819  * cur_cfqq - passed in so that we don't decide that the current queue is
1820  *            closely cooperating with itself.
1821  *
1822  * So, basically we're assuming that that cur_cfqq has dispatched at least
1823  * one request, and that cfqd->last_position reflects a position on the disk
1824  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1825  * assumption.
1826  */
1827 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1828                                               struct cfq_queue *cur_cfqq)
1829 {
1830         struct cfq_queue *cfqq;
1831
1832         if (cfq_class_idle(cur_cfqq))
1833                 return NULL;
1834         if (!cfq_cfqq_sync(cur_cfqq))
1835                 return NULL;
1836         if (CFQQ_SEEKY(cur_cfqq))
1837                 return NULL;
1838
1839         /*
1840          * Don't search priority tree if it's the only queue in the group.
1841          */
1842         if (cur_cfqq->cfqg->nr_cfqq == 1)
1843                 return NULL;
1844
1845         /*
1846          * We should notice if some of the queues are cooperating, eg
1847          * working closely on the same area of the disk. In that case,
1848          * we can group them together and don't waste time idling.
1849          */
1850         cfqq = cfqq_close(cfqd, cur_cfqq);
1851         if (!cfqq)
1852                 return NULL;
1853
1854         /* If new queue belongs to different cfq_group, don't choose it */
1855         if (cur_cfqq->cfqg != cfqq->cfqg)
1856                 return NULL;
1857
1858         /*
1859          * It only makes sense to merge sync queues.
1860          */
1861         if (!cfq_cfqq_sync(cfqq))
1862                 return NULL;
1863         if (CFQQ_SEEKY(cfqq))
1864                 return NULL;
1865
1866         /*
1867          * Do not merge queues of different priority classes
1868          */
1869         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1870                 return NULL;
1871
1872         return cfqq;
1873 }
1874
1875 /*
1876  * Determine whether we should enforce idle window for this queue.
1877  */
1878
1879 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1880 {
1881         enum wl_prio_t prio = cfqq_prio(cfqq);
1882         struct cfq_rb_root *service_tree = cfqq->service_tree;
1883
1884         BUG_ON(!service_tree);
1885         BUG_ON(!service_tree->count);
1886
1887         if (!cfqd->cfq_slice_idle)
1888                 return false;
1889
1890         /* We never do for idle class queues. */
1891         if (prio == IDLE_WORKLOAD)
1892                 return false;
1893
1894         /* We do for queues that were marked with idle window flag. */
1895         if (cfq_cfqq_idle_window(cfqq) &&
1896            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1897                 return true;
1898
1899         /*
1900          * Otherwise, we do only if they are the last ones
1901          * in their service tree.
1902          */
1903         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1904                 return true;
1905         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1906                         service_tree->count);
1907         return false;
1908 }
1909
1910 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1911 {
1912         struct cfq_queue *cfqq = cfqd->active_queue;
1913         struct cfq_io_context *cic;
1914         unsigned long sl, group_idle = 0;
1915
1916         /*
1917          * SSD device without seek penalty, disable idling. But only do so
1918          * for devices that support queuing, otherwise we still have a problem
1919          * with sync vs async workloads.
1920          */
1921         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1922                 return;
1923
1924         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1925         WARN_ON(cfq_cfqq_slice_new(cfqq));
1926
1927         /*
1928          * idle is disabled, either manually or by past process history
1929          */
1930         if (!cfq_should_idle(cfqd, cfqq)) {
1931                 /* no queue idling. Check for group idling */
1932                 if (cfqd->cfq_group_idle)
1933                         group_idle = cfqd->cfq_group_idle;
1934                 else
1935                         return;
1936         }
1937
1938         /*
1939          * still active requests from this queue, don't idle
1940          */
1941         if (cfqq->dispatched)
1942                 return;
1943
1944         /*
1945          * task has exited, don't wait
1946          */
1947         cic = cfqd->active_cic;
1948         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1949                 return;
1950
1951         /*
1952          * If our average think time is larger than the remaining time
1953          * slice, then don't idle. This avoids overrunning the allotted
1954          * time slice.
1955          */
1956         if (sample_valid(cic->ttime_samples) &&
1957             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1958                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1959                                 cic->ttime_mean);
1960                 return;
1961         }
1962
1963         /* There are other queues in the group, don't do group idle */
1964         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1965                 return;
1966
1967         cfq_mark_cfqq_wait_request(cfqq);
1968
1969         if (group_idle)
1970                 sl = cfqd->cfq_group_idle;
1971         else
1972                 sl = cfqd->cfq_slice_idle;
1973
1974         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1975         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1976         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1977                         group_idle ? 1 : 0);
1978 }
1979
1980 /*
1981  * Move request from internal lists to the request queue dispatch list.
1982  */
1983 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1984 {
1985         struct cfq_data *cfqd = q->elevator->elevator_data;
1986         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1987
1988         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1989
1990         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1991         cfq_remove_request(rq);
1992         cfqq->dispatched++;
1993         (RQ_CFQG(rq))->dispatched++;
1994         elv_dispatch_sort(q, rq);
1995
1996         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1997         cfqq->nr_sectors += blk_rq_sectors(rq);
1998         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1999                                         rq_data_dir(rq), rq_is_sync(rq));
2000 }
2001
2002 /*
2003  * return expired entry, or NULL to just start from scratch in rbtree
2004  */
2005 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2006 {
2007         struct request *rq = NULL;
2008
2009         if (cfq_cfqq_fifo_expire(cfqq))
2010                 return NULL;
2011
2012         cfq_mark_cfqq_fifo_expire(cfqq);
2013
2014         if (list_empty(&cfqq->fifo))
2015                 return NULL;
2016
2017         rq = rq_entry_fifo(cfqq->fifo.next);
2018         if (time_before(jiffies, rq_fifo_time(rq)))
2019                 rq = NULL;
2020
2021         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2022         return rq;
2023 }
2024
2025 static inline int
2026 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2027 {
2028         const int base_rq = cfqd->cfq_slice_async_rq;
2029
2030         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2031
2032         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2033 }
2034
2035 /*
2036  * Must be called with the queue_lock held.
2037  */
2038 static int cfqq_process_refs(struct cfq_queue *cfqq)
2039 {
2040         int process_refs, io_refs;
2041
2042         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2043         process_refs = cfqq->ref - io_refs;
2044         BUG_ON(process_refs < 0);
2045         return process_refs;
2046 }
2047
2048 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2049 {
2050         int process_refs, new_process_refs;
2051         struct cfq_queue *__cfqq;
2052
2053         /*
2054          * If there are no process references on the new_cfqq, then it is
2055          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2056          * chain may have dropped their last reference (not just their
2057          * last process reference).
2058          */
2059         if (!cfqq_process_refs(new_cfqq))
2060                 return;
2061
2062         /* Avoid a circular list and skip interim queue merges */
2063         while ((__cfqq = new_cfqq->new_cfqq)) {
2064                 if (__cfqq == cfqq)
2065                         return;
2066                 new_cfqq = __cfqq;
2067         }
2068
2069         process_refs = cfqq_process_refs(cfqq);
2070         new_process_refs = cfqq_process_refs(new_cfqq);
2071         /*
2072          * If the process for the cfqq has gone away, there is no
2073          * sense in merging the queues.
2074          */
2075         if (process_refs == 0 || new_process_refs == 0)
2076                 return;
2077
2078         /*
2079          * Merge in the direction of the lesser amount of work.
2080          */
2081         if (new_process_refs >= process_refs) {
2082                 cfqq->new_cfqq = new_cfqq;
2083                 new_cfqq->ref += process_refs;
2084         } else {
2085                 new_cfqq->new_cfqq = cfqq;
2086                 cfqq->ref += new_process_refs;
2087         }
2088 }
2089
2090 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2091                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2092 {
2093         struct cfq_queue *queue;
2094         int i;
2095         bool key_valid = false;
2096         unsigned long lowest_key = 0;
2097         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2098
2099         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2100                 /* select the one with lowest rb_key */
2101                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2102                 if (queue &&
2103                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2104                         lowest_key = queue->rb_key;
2105                         cur_best = i;
2106                         key_valid = true;
2107                 }
2108         }
2109
2110         return cur_best;
2111 }
2112
2113 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2114 {
2115         unsigned slice;
2116         unsigned count;
2117         struct cfq_rb_root *st;
2118         unsigned group_slice;
2119         enum wl_prio_t original_prio = cfqd->serving_prio;
2120
2121         /* Choose next priority. RT > BE > IDLE */
2122         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2123                 cfqd->serving_prio = RT_WORKLOAD;
2124         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2125                 cfqd->serving_prio = BE_WORKLOAD;
2126         else {
2127                 cfqd->serving_prio = IDLE_WORKLOAD;
2128                 cfqd->workload_expires = jiffies + 1;
2129                 return;
2130         }
2131
2132         if (original_prio != cfqd->serving_prio)
2133                 goto new_workload;
2134
2135         /*
2136          * For RT and BE, we have to choose also the type
2137          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2138          * expiration time
2139          */
2140         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2141         count = st->count;
2142
2143         /*
2144          * check workload expiration, and that we still have other queues ready
2145          */
2146         if (count && !time_after(jiffies, cfqd->workload_expires))
2147                 return;
2148
2149 new_workload:
2150         /* otherwise select new workload type */
2151         cfqd->serving_type =
2152                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2153         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2154         count = st->count;
2155
2156         /*
2157          * the workload slice is computed as a fraction of target latency
2158          * proportional to the number of queues in that workload, over
2159          * all the queues in the same priority class
2160          */
2161         group_slice = cfq_group_slice(cfqd, cfqg);
2162
2163         slice = group_slice * count /
2164                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2165                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2166
2167         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2168                 unsigned int tmp;
2169
2170                 /*
2171                  * Async queues are currently system wide. Just taking
2172                  * proportion of queues with-in same group will lead to higher
2173                  * async ratio system wide as generally root group is going
2174                  * to have higher weight. A more accurate thing would be to
2175                  * calculate system wide asnc/sync ratio.
2176                  */
2177                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2178                 tmp = tmp/cfqd->busy_queues;
2179                 slice = min_t(unsigned, slice, tmp);
2180
2181                 /* async workload slice is scaled down according to
2182                  * the sync/async slice ratio. */
2183                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2184         } else
2185                 /* sync workload slice is at least 2 * cfq_slice_idle */
2186                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2187
2188         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2189         cfq_log(cfqd, "workload slice:%d", slice);
2190         cfqd->workload_expires = jiffies + slice;
2191 }
2192
2193 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2194 {
2195         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2196         struct cfq_group *cfqg;
2197
2198         if (RB_EMPTY_ROOT(&st->rb))
2199                 return NULL;
2200         cfqg = cfq_rb_first_group(st);
2201         update_min_vdisktime(st);
2202         return cfqg;
2203 }
2204
2205 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2206 {
2207         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2208
2209         cfqd->serving_group = cfqg;
2210
2211         /* Restore the workload type data */
2212         if (cfqg->saved_workload_slice) {
2213                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2214                 cfqd->serving_type = cfqg->saved_workload;
2215                 cfqd->serving_prio = cfqg->saved_serving_prio;
2216         } else
2217                 cfqd->workload_expires = jiffies - 1;
2218
2219         choose_service_tree(cfqd, cfqg);
2220 }
2221
2222 /*
2223  * Select a queue for service. If we have a current active queue,
2224  * check whether to continue servicing it, or retrieve and set a new one.
2225  */
2226 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2227 {
2228         struct cfq_queue *cfqq, *new_cfqq = NULL;
2229
2230         cfqq = cfqd->active_queue;
2231         if (!cfqq)
2232                 goto new_queue;
2233
2234         if (!cfqd->rq_queued)
2235                 return NULL;
2236
2237         /*
2238          * We were waiting for group to get backlogged. Expire the queue
2239          */
2240         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2241                 goto expire;
2242
2243         /*
2244          * The active queue has run out of time, expire it and select new.
2245          */
2246         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2247                 /*
2248                  * If slice had not expired at the completion of last request
2249                  * we might not have turned on wait_busy flag. Don't expire
2250                  * the queue yet. Allow the group to get backlogged.
2251                  *
2252                  * The very fact that we have used the slice, that means we
2253                  * have been idling all along on this queue and it should be
2254                  * ok to wait for this request to complete.
2255                  */
2256                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2257                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2258                         cfqq = NULL;
2259                         goto keep_queue;
2260                 } else
2261                         goto check_group_idle;
2262         }
2263
2264         /*
2265          * The active queue has requests and isn't expired, allow it to
2266          * dispatch.
2267          */
2268         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2269                 goto keep_queue;
2270
2271         /*
2272          * If another queue has a request waiting within our mean seek
2273          * distance, let it run.  The expire code will check for close
2274          * cooperators and put the close queue at the front of the service
2275          * tree.  If possible, merge the expiring queue with the new cfqq.
2276          */
2277         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2278         if (new_cfqq) {
2279                 if (!cfqq->new_cfqq)
2280                         cfq_setup_merge(cfqq, new_cfqq);
2281                 goto expire;
2282         }
2283
2284         /*
2285          * No requests pending. If the active queue still has requests in
2286          * flight or is idling for a new request, allow either of these
2287          * conditions to happen (or time out) before selecting a new queue.
2288          */
2289         if (timer_pending(&cfqd->idle_slice_timer)) {
2290                 cfqq = NULL;
2291                 goto keep_queue;
2292         }
2293
2294         /*
2295          * This is a deep seek queue, but the device is much faster than
2296          * the queue can deliver, don't idle
2297          **/
2298         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2299             (cfq_cfqq_slice_new(cfqq) ||
2300             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2301                 cfq_clear_cfqq_deep(cfqq);
2302                 cfq_clear_cfqq_idle_window(cfqq);
2303         }
2304
2305         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2306                 cfqq = NULL;
2307                 goto keep_queue;
2308         }
2309
2310         /*
2311          * If group idle is enabled and there are requests dispatched from
2312          * this group, wait for requests to complete.
2313          */
2314 check_group_idle:
2315         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2316             && cfqq->cfqg->dispatched) {
2317                 cfqq = NULL;
2318                 goto keep_queue;
2319         }
2320
2321 expire:
2322         cfq_slice_expired(cfqd, 0);
2323 new_queue:
2324         /*
2325          * Current queue expired. Check if we have to switch to a new
2326          * service tree
2327          */
2328         if (!new_cfqq)
2329                 cfq_choose_cfqg(cfqd);
2330
2331         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2332 keep_queue:
2333         return cfqq;
2334 }
2335
2336 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2337 {
2338         int dispatched = 0;
2339
2340         while (cfqq->next_rq) {
2341                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2342                 dispatched++;
2343         }
2344
2345         BUG_ON(!list_empty(&cfqq->fifo));
2346
2347         /* By default cfqq is not expired if it is empty. Do it explicitly */
2348         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2349         return dispatched;
2350 }
2351
2352 /*
2353  * Drain our current requests. Used for barriers and when switching
2354  * io schedulers on-the-fly.
2355  */
2356 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2357 {
2358         struct cfq_queue *cfqq;
2359         int dispatched = 0;
2360
2361         /* Expire the timeslice of the current active queue first */
2362         cfq_slice_expired(cfqd, 0);
2363         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2364                 __cfq_set_active_queue(cfqd, cfqq);
2365                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2366         }
2367
2368         BUG_ON(cfqd->busy_queues);
2369
2370         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2371         return dispatched;
2372 }
2373
2374 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2375         struct cfq_queue *cfqq)
2376 {
2377         /* the queue hasn't finished any request, can't estimate */
2378         if (cfq_cfqq_slice_new(cfqq))
2379                 return true;
2380         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2381                 cfqq->slice_end))
2382                 return true;
2383
2384         return false;
2385 }
2386
2387 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2388 {
2389         unsigned int max_dispatch;
2390
2391         /*
2392          * Drain async requests before we start sync IO
2393          */
2394         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2395                 return false;
2396
2397         /*
2398          * If this is an async queue and we have sync IO in flight, let it wait
2399          */
2400         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2401                 return false;
2402
2403         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2404         if (cfq_class_idle(cfqq))
2405                 max_dispatch = 1;
2406
2407         /*
2408          * Does this cfqq already have too much IO in flight?
2409          */
2410         if (cfqq->dispatched >= max_dispatch) {
2411                 bool promote_sync = false;
2412                 /*
2413                  * idle queue must always only have a single IO in flight
2414                  */
2415                 if (cfq_class_idle(cfqq))
2416                         return false;
2417
2418                 /*
2419                  * If there is only one sync queue
2420                  * we can ignore async queue here and give the sync
2421                  * queue no dispatch limit. The reason is a sync queue can
2422                  * preempt async queue, limiting the sync queue doesn't make
2423                  * sense. This is useful for aiostress test.
2424                  */
2425                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2426                         promote_sync = true;
2427
2428                 /*
2429                  * We have other queues, don't allow more IO from this one
2430                  */
2431                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2432                                 !promote_sync)
2433                         return false;
2434
2435                 /*
2436                  * Sole queue user, no limit
2437                  */
2438                 if (cfqd->busy_queues == 1 || promote_sync)
2439                         max_dispatch = -1;
2440                 else
2441                         /*
2442                          * Normally we start throttling cfqq when cfq_quantum/2
2443                          * requests have been dispatched. But we can drive
2444                          * deeper queue depths at the beginning of slice
2445                          * subjected to upper limit of cfq_quantum.
2446                          * */
2447                         max_dispatch = cfqd->cfq_quantum;
2448         }
2449
2450         /*
2451          * Async queues must wait a bit before being allowed dispatch.
2452          * We also ramp up the dispatch depth gradually for async IO,
2453          * based on the last sync IO we serviced
2454          */
2455         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2456                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2457                 unsigned int depth;
2458
2459                 depth = last_sync / cfqd->cfq_slice[1];
2460                 if (!depth && !cfqq->dispatched)
2461                         depth = 1;
2462                 if (depth < max_dispatch)
2463                         max_dispatch = depth;
2464         }
2465
2466         /*
2467          * If we're below the current max, allow a dispatch
2468          */
2469         return cfqq->dispatched < max_dispatch;
2470 }
2471
2472 /*
2473  * Dispatch a request from cfqq, moving them to the request queue
2474  * dispatch list.
2475  */
2476 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2477 {
2478         struct request *rq;
2479
2480         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2481
2482         if (!cfq_may_dispatch(cfqd, cfqq))
2483                 return false;
2484
2485         /*
2486          * follow expired path, else get first next available
2487          */
2488         rq = cfq_check_fifo(cfqq);
2489         if (!rq)
2490                 rq = cfqq->next_rq;
2491
2492         /*
2493          * insert request into driver dispatch list
2494          */
2495         cfq_dispatch_insert(cfqd->queue, rq);
2496
2497         if (!cfqd->active_cic) {
2498                 struct cfq_io_context *cic = RQ_CIC(rq);
2499
2500                 atomic_long_inc(&cic->ioc->refcount);
2501                 cfqd->active_cic = cic;
2502         }
2503
2504         return true;
2505 }
2506
2507 /*
2508  * Find the cfqq that we need to service and move a request from that to the
2509  * dispatch list
2510  */
2511 static int cfq_dispatch_requests(struct request_queue *q, int force)
2512 {
2513         struct cfq_data *cfqd = q->elevator->elevator_data;
2514         struct cfq_queue *cfqq;
2515
2516         if (!cfqd->busy_queues)
2517                 return 0;
2518
2519         if (unlikely(force))
2520                 return cfq_forced_dispatch(cfqd);
2521
2522         cfqq = cfq_select_queue(cfqd);
2523         if (!cfqq)
2524                 return 0;
2525
2526         /*
2527          * Dispatch a request from this cfqq, if it is allowed
2528          */
2529         if (!cfq_dispatch_request(cfqd, cfqq))
2530                 return 0;
2531
2532         cfqq->slice_dispatch++;
2533         cfq_clear_cfqq_must_dispatch(cfqq);
2534
2535         /*
2536          * expire an async queue immediately if it has used up its slice. idle
2537          * queue always expire after 1 dispatch round.
2538          */
2539         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2540             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2541             cfq_class_idle(cfqq))) {
2542                 cfqq->slice_end = jiffies + 1;
2543                 cfq_slice_expired(cfqd, 0);
2544         }
2545
2546         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2547         return 1;
2548 }
2549
2550 /*
2551  * task holds one reference to the queue, dropped when task exits. each rq
2552  * in-flight on this queue also holds a reference, dropped when rq is freed.
2553  *
2554  * Each cfq queue took a reference on the parent group. Drop it now.
2555  * queue lock must be held here.
2556  */
2557 static void cfq_put_queue(struct cfq_queue *cfqq)
2558 {
2559         struct cfq_data *cfqd = cfqq->cfqd;
2560         struct cfq_group *cfqg;
2561
2562         BUG_ON(cfqq->ref <= 0);
2563
2564         cfqq->ref--;
2565         if (cfqq->ref)
2566                 return;
2567
2568         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2569         BUG_ON(rb_first(&cfqq->sort_list));
2570         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2571         cfqg = cfqq->cfqg;
2572
2573         if (unlikely(cfqd->active_queue == cfqq)) {
2574                 __cfq_slice_expired(cfqd, cfqq, 0);
2575                 cfq_schedule_dispatch(cfqd);
2576         }
2577
2578         BUG_ON(cfq_cfqq_on_rr(cfqq));
2579         kmem_cache_free(cfq_pool, cfqq);
2580         cfq_put_cfqg(cfqg);
2581 }
2582
2583 /*
2584  * Call func for each cic attached to this ioc.
2585  */
2586 static void
2587 call_for_each_cic(struct io_context *ioc,
2588                   void (*func)(struct io_context *, struct cfq_io_context *))
2589 {
2590         struct cfq_io_context *cic;
2591         struct hlist_node *n;
2592
2593         rcu_read_lock();
2594
2595         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2596                 func(ioc, cic);
2597
2598         rcu_read_unlock();
2599 }
2600
2601 static void cfq_cic_free_rcu(struct rcu_head *head)
2602 {
2603         struct cfq_io_context *cic;
2604
2605         cic = container_of(head, struct cfq_io_context, rcu_head);
2606
2607         kmem_cache_free(cfq_ioc_pool, cic);
2608         elv_ioc_count_dec(cfq_ioc_count);
2609
2610         if (ioc_gone) {
2611                 /*
2612                  * CFQ scheduler is exiting, grab exit lock and check
2613                  * the pending io context count. If it hits zero,
2614                  * complete ioc_gone and set it back to NULL
2615                  */
2616                 spin_lock(&ioc_gone_lock);
2617                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2618                         complete(ioc_gone);
2619                         ioc_gone = NULL;
2620                 }
2621                 spin_unlock(&ioc_gone_lock);
2622         }
2623 }
2624
2625 static void cfq_cic_free(struct cfq_io_context *cic)
2626 {
2627         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2628 }
2629
2630 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2631 {
2632         unsigned long flags;
2633         unsigned long dead_key = (unsigned long) cic->key;
2634
2635         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2636
2637         spin_lock_irqsave(&ioc->lock, flags);
2638         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2639         hlist_del_rcu(&cic->cic_list);
2640         spin_unlock_irqrestore(&ioc->lock, flags);
2641
2642         cfq_cic_free(cic);
2643 }
2644
2645 /*
2646  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2647  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2648  * and ->trim() which is called with the task lock held
2649  */
2650 static void cfq_free_io_context(struct io_context *ioc)
2651 {
2652         /*
2653          * ioc->refcount is zero here, or we are called from elv_unregister(),
2654          * so no more cic's are allowed to be linked into this ioc.  So it
2655          * should be ok to iterate over the known list, we will see all cic's
2656          * since no new ones are added.
2657          */
2658         call_for_each_cic(ioc, cic_free_func);
2659 }
2660
2661 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2662 {
2663         struct cfq_queue *__cfqq, *next;
2664
2665         /*
2666          * If this queue was scheduled to merge with another queue, be
2667          * sure to drop the reference taken on that queue (and others in
2668          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2669          */
2670         __cfqq = cfqq->new_cfqq;
2671         while (__cfqq) {
2672                 if (__cfqq == cfqq) {
2673                         WARN(1, "cfqq->new_cfqq loop detected\n");
2674                         break;
2675                 }
2676                 next = __cfqq->new_cfqq;
2677                 cfq_put_queue(__cfqq);
2678                 __cfqq = next;
2679         }
2680 }
2681
2682 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2683 {
2684         if (unlikely(cfqq == cfqd->active_queue)) {
2685                 __cfq_slice_expired(cfqd, cfqq, 0);
2686                 cfq_schedule_dispatch(cfqd);
2687         }
2688
2689         cfq_put_cooperator(cfqq);
2690
2691         cfq_put_queue(cfqq);
2692 }
2693
2694 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2695                                          struct cfq_io_context *cic)
2696 {
2697         struct io_context *ioc = cic->ioc;
2698
2699         list_del_init(&cic->queue_list);
2700
2701         /*
2702          * Make sure dead mark is seen for dead queues
2703          */
2704         smp_wmb();
2705         cic->key = cfqd_dead_key(cfqd);
2706
2707         if (ioc->ioc_data == cic)
2708                 rcu_assign_pointer(ioc->ioc_data, NULL);
2709
2710         if (cic->cfqq[BLK_RW_ASYNC]) {
2711                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2712                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2713         }
2714
2715         if (cic->cfqq[BLK_RW_SYNC]) {
2716                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2717                 cic->cfqq[BLK_RW_SYNC] = NULL;
2718         }
2719 }
2720
2721 static void cfq_exit_single_io_context(struct io_context *ioc,
2722                                        struct cfq_io_context *cic)
2723 {
2724         struct cfq_data *cfqd = cic_to_cfqd(cic);
2725
2726         if (cfqd) {
2727                 struct request_queue *q = cfqd->queue;
2728                 unsigned long flags;
2729
2730                 spin_lock_irqsave(q->queue_lock, flags);
2731
2732                 /*
2733                  * Ensure we get a fresh copy of the ->key to prevent
2734                  * race between exiting task and queue
2735                  */
2736                 smp_read_barrier_depends();
2737                 if (cic->key == cfqd)
2738                         __cfq_exit_single_io_context(cfqd, cic);
2739
2740                 spin_unlock_irqrestore(q->queue_lock, flags);
2741         }
2742 }
2743
2744 /*
2745  * The process that ioc belongs to has exited, we need to clean up
2746  * and put the internal structures we have that belongs to that process.
2747  */
2748 static void cfq_exit_io_context(struct io_context *ioc)
2749 {
2750         call_for_each_cic(ioc, cfq_exit_single_io_context);
2751 }
2752
2753 static struct cfq_io_context *
2754 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2755 {
2756         struct cfq_io_context *cic;
2757
2758         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2759                                                         cfqd->queue->node);
2760         if (cic) {
2761                 cic->last_end_request = jiffies;
2762                 INIT_LIST_HEAD(&cic->queue_list);
2763                 INIT_HLIST_NODE(&cic->cic_list);
2764                 cic->dtor = cfq_free_io_context;
2765                 cic->exit = cfq_exit_io_context;
2766                 elv_ioc_count_inc(cfq_ioc_count);
2767         }
2768
2769         return cic;
2770 }
2771
2772 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2773 {
2774         struct task_struct *tsk = current;
2775         int ioprio_class;
2776
2777         if (!cfq_cfqq_prio_changed(cfqq))
2778                 return;
2779
2780         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2781         switch (ioprio_class) {
2782         default:
2783                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2784         case IOPRIO_CLASS_NONE:
2785                 /*
2786                  * no prio set, inherit CPU scheduling settings
2787                  */
2788                 cfqq->ioprio = task_nice_ioprio(tsk);
2789                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2790                 break;
2791         case IOPRIO_CLASS_RT:
2792                 cfqq->ioprio = task_ioprio(ioc);
2793                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2794                 break;
2795         case IOPRIO_CLASS_BE:
2796                 cfqq->ioprio = task_ioprio(ioc);
2797                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2798                 break;
2799         case IOPRIO_CLASS_IDLE:
2800                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2801                 cfqq->ioprio = 7;
2802                 cfq_clear_cfqq_idle_window(cfqq);
2803                 break;
2804         }
2805
2806         /*
2807          * keep track of original prio settings in case we have to temporarily
2808          * elevate the priority of this queue
2809          */
2810         cfqq->org_ioprio = cfqq->ioprio;
2811         cfqq->org_ioprio_class = cfqq->ioprio_class;
2812         cfq_clear_cfqq_prio_changed(cfqq);
2813 }
2814
2815 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2816 {
2817         struct cfq_data *cfqd = cic_to_cfqd(cic);
2818         struct cfq_queue *cfqq;
2819         unsigned long flags;
2820
2821         if (unlikely(!cfqd))
2822                 return;
2823
2824         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2825
2826         cfqq = cic->cfqq[BLK_RW_ASYNC];
2827         if (cfqq) {
2828                 struct cfq_queue *new_cfqq;
2829                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2830                                                 GFP_ATOMIC);
2831                 if (new_cfqq) {
2832                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2833                         cfq_put_queue(cfqq);
2834                 }
2835         }
2836
2837         cfqq = cic->cfqq[BLK_RW_SYNC];
2838         if (cfqq)
2839                 cfq_mark_cfqq_prio_changed(cfqq);
2840
2841         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2842 }
2843
2844 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2845 {
2846         call_for_each_cic(ioc, changed_ioprio);
2847         ioc->ioprio_changed = 0;
2848 }
2849
2850 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2851                           pid_t pid, bool is_sync)
2852 {
2853         RB_CLEAR_NODE(&cfqq->rb_node);
2854         RB_CLEAR_NODE(&cfqq->p_node);
2855         INIT_LIST_HEAD(&cfqq->fifo);
2856
2857         cfqq->ref = 0;
2858         cfqq->cfqd = cfqd;
2859
2860         cfq_mark_cfqq_prio_changed(cfqq);
2861
2862         if (is_sync) {
2863                 if (!cfq_class_idle(cfqq))
2864                         cfq_mark_cfqq_idle_window(cfqq);
2865                 cfq_mark_cfqq_sync(cfqq);
2866         }
2867         cfqq->pid = pid;
2868 }
2869
2870 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2871 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2872 {
2873         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2874         struct cfq_data *cfqd = cic_to_cfqd(cic);
2875         unsigned long flags;
2876         struct request_queue *q;
2877
2878         if (unlikely(!cfqd))
2879                 return;
2880
2881         q = cfqd->queue;
2882
2883         spin_lock_irqsave(q->queue_lock, flags);
2884
2885         if (sync_cfqq) {
2886                 /*
2887                  * Drop reference to sync queue. A new sync queue will be
2888                  * assigned in new group upon arrival of a fresh request.
2889                  */
2890                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2891                 cic_set_cfqq(cic, NULL, 1);
2892                 cfq_put_queue(sync_cfqq);
2893         }
2894
2895         spin_unlock_irqrestore(q->queue_lock, flags);
2896 }
2897
2898 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2899 {
2900         call_for_each_cic(ioc, changed_cgroup);
2901         ioc->cgroup_changed = 0;
2902 }
2903 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2904
2905 static struct cfq_queue *
2906 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2907                      struct io_context *ioc, gfp_t gfp_mask)
2908 {
2909         struct cfq_queue *cfqq, *new_cfqq = NULL;
2910         struct cfq_io_context *cic;
2911         struct cfq_group *cfqg;
2912
2913 retry:
2914         cfqg = cfq_get_cfqg(cfqd, 1);
2915         cic = cfq_cic_lookup(cfqd, ioc);
2916         /* cic always exists here */
2917         cfqq = cic_to_cfqq(cic, is_sync);
2918
2919         /*
2920          * Always try a new alloc if we fell back to the OOM cfqq
2921          * originally, since it should just be a temporary situation.
2922          */
2923         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2924                 cfqq = NULL;
2925                 if (new_cfqq) {
2926                         cfqq = new_cfqq;
2927                         new_cfqq = NULL;
2928                 } else if (gfp_mask & __GFP_WAIT) {
2929                         spin_unlock_irq(cfqd->queue->queue_lock);
2930                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2931                                         gfp_mask | __GFP_ZERO,
2932                                         cfqd->queue->node);
2933                         spin_lock_irq(cfqd->queue->queue_lock);
2934                         if (new_cfqq)
2935                                 goto retry;
2936                 } else {
2937                         cfqq = kmem_cache_alloc_node(cfq_pool,
2938                                         gfp_mask | __GFP_ZERO,
2939                                         cfqd->queue->node);
2940                 }
2941
2942                 if (cfqq) {
2943                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2944                         cfq_init_prio_data(cfqq, ioc);
2945                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2946                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2947                 } else
2948                         cfqq = &cfqd->oom_cfqq;
2949         }
2950
2951         if (new_cfqq)
2952                 kmem_cache_free(cfq_pool, new_cfqq);
2953
2954         return cfqq;
2955 }
2956
2957 static struct cfq_queue **
2958 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2959 {
2960         switch (ioprio_class) {
2961         case IOPRIO_CLASS_RT:
2962                 return &cfqd->async_cfqq[0][ioprio];
2963         case IOPRIO_CLASS_BE:
2964                 return &cfqd->async_cfqq[1][ioprio];
2965         case IOPRIO_CLASS_IDLE:
2966                 return &cfqd->async_idle_cfqq;
2967         default:
2968                 BUG();
2969         }
2970 }
2971
2972 static struct cfq_queue *
2973 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2974               gfp_t gfp_mask)
2975 {
2976         const int ioprio = task_ioprio(ioc);
2977         const int ioprio_class = task_ioprio_class(ioc);
2978         struct cfq_queue **async_cfqq = NULL;
2979         struct cfq_queue *cfqq = NULL;
2980
2981         if (!is_sync) {
2982                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2983                 cfqq = *async_cfqq;
2984         }
2985
2986         if (!cfqq)
2987                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2988
2989         /*
2990          * pin the queue now that it's allocated, scheduler exit will prune it
2991          */
2992         if (!is_sync && !(*async_cfqq)) {
2993                 cfqq->ref++;
2994                 *async_cfqq = cfqq;
2995         }
2996
2997         cfqq->ref++;
2998         return cfqq;
2999 }
3000
3001 /*
3002  * We drop cfq io contexts lazily, so we may find a dead one.
3003  */
3004 static void
3005 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3006                   struct cfq_io_context *cic)
3007 {
3008         unsigned long flags;
3009
3010         WARN_ON(!list_empty(&cic->queue_list));
3011         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3012
3013         spin_lock_irqsave(&ioc->lock, flags);
3014
3015         BUG_ON(ioc->ioc_data == cic);
3016
3017         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3018         hlist_del_rcu(&cic->cic_list);
3019         spin_unlock_irqrestore(&ioc->lock, flags);
3020
3021         cfq_cic_free(cic);
3022 }
3023
3024 static struct cfq_io_context *
3025 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3026 {
3027         struct cfq_io_context *cic;
3028         unsigned long flags;
3029
3030         if (unlikely(!ioc))
3031                 return NULL;
3032
3033         rcu_read_lock();
3034
3035         /*
3036          * we maintain a last-hit cache, to avoid browsing over the tree
3037          */
3038         cic = rcu_dereference(ioc->ioc_data);
3039         if (cic && cic->key == cfqd) {
3040                 rcu_read_unlock();
3041                 return cic;
3042         }
3043
3044         do {
3045                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3046                 rcu_read_unlock();
3047                 if (!cic)
3048                         break;
3049                 if (unlikely(cic->key != cfqd)) {
3050                         cfq_drop_dead_cic(cfqd, ioc, cic);
3051                         rcu_read_lock();
3052                         continue;
3053                 }
3054
3055                 spin_lock_irqsave(&ioc->lock, flags);
3056                 rcu_assign_pointer(ioc->ioc_data, cic);
3057                 spin_unlock_irqrestore(&ioc->lock, flags);
3058                 break;
3059         } while (1);
3060
3061         return cic;
3062 }
3063
3064 /*
3065  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3066  * the process specific cfq io context when entered from the block layer.
3067  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3068  */
3069 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3070                         struct cfq_io_context *cic, gfp_t gfp_mask)
3071 {
3072         unsigned long flags;
3073         int ret;
3074
3075         ret = radix_tree_preload(gfp_mask);
3076         if (!ret) {
3077                 cic->ioc = ioc;
3078                 cic->key = cfqd;
3079
3080                 spin_lock_irqsave(&ioc->lock, flags);
3081                 ret = radix_tree_insert(&ioc->radix_root,
3082                                                 cfqd->cic_index, cic);
3083                 if (!ret)
3084                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3085                 spin_unlock_irqrestore(&ioc->lock, flags);
3086
3087                 radix_tree_preload_end();
3088
3089                 if (!ret) {
3090                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3091                         list_add(&cic->queue_list, &cfqd->cic_list);
3092                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3093                 }
3094         }
3095
3096         if (ret)
3097                 printk(KERN_ERR "cfq: cic link failed!\n");
3098
3099         return ret;
3100 }
3101
3102 /*
3103  * Setup general io context and cfq io context. There can be several cfq
3104  * io contexts per general io context, if this process is doing io to more
3105  * than one device managed by cfq.
3106  */
3107 static struct cfq_io_context *
3108 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3109 {
3110         struct io_context *ioc = NULL;
3111         struct cfq_io_context *cic;
3112
3113         might_sleep_if(gfp_mask & __GFP_WAIT);
3114
3115         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3116         if (!ioc)
3117                 return NULL;
3118
3119         cic = cfq_cic_lookup(cfqd, ioc);
3120         if (cic)
3121                 goto out;
3122
3123         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3124         if (cic == NULL)
3125                 goto err;
3126
3127         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3128                 goto err_free;
3129
3130 out:
3131         smp_read_barrier_depends();
3132         if (unlikely(ioc->ioprio_changed))
3133                 cfq_ioc_set_ioprio(ioc);
3134
3135 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3136         if (unlikely(ioc->cgroup_changed))
3137                 cfq_ioc_set_cgroup(ioc);
3138 #endif
3139         return cic;
3140 err_free:
3141         cfq_cic_free(cic);
3142 err:
3143         put_io_context(ioc);
3144         return NULL;
3145 }
3146
3147 static void
3148 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3149 {
3150         unsigned long elapsed = jiffies - cic->last_end_request;
3151         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3152
3153         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3154         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3155         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3156 }
3157
3158 static void
3159 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3160                        struct request *rq)
3161 {
3162         sector_t sdist = 0;
3163         sector_t n_sec = blk_rq_sectors(rq);
3164         if (cfqq->last_request_pos) {
3165                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3166                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3167                 else
3168                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3169         }
3170
3171         cfqq->seek_history <<= 1;
3172         if (blk_queue_nonrot(cfqd->queue))
3173                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3174         else
3175                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3176 }
3177
3178 /*
3179  * Disable idle window if the process thinks too long or seeks so much that
3180  * it doesn't matter
3181  */
3182 static void
3183 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3184                        struct cfq_io_context *cic)
3185 {
3186         int old_idle, enable_idle;
3187
3188         /*
3189          * Don't idle for async or idle io prio class
3190          */
3191         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3192                 return;
3193
3194         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3195
3196         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3197                 cfq_mark_cfqq_deep(cfqq);
3198
3199         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3200                 enable_idle = 0;
3201         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3202             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3203                 enable_idle = 0;
3204         else if (sample_valid(cic->ttime_samples)) {
3205                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3206                         enable_idle = 0;
3207                 else
3208                         enable_idle = 1;
3209         }
3210
3211         if (old_idle != enable_idle) {
3212                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3213                 if (enable_idle)
3214                         cfq_mark_cfqq_idle_window(cfqq);
3215                 else
3216                         cfq_clear_cfqq_idle_window(cfqq);
3217         }
3218 }
3219
3220 /*
3221  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3222  * no or if we aren't sure, a 1 will cause a preempt.
3223  */
3224 static bool
3225 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3226                    struct request *rq)
3227 {
3228         struct cfq_queue *cfqq;
3229
3230         cfqq = cfqd->active_queue;
3231         if (!cfqq)
3232                 return false;
3233
3234         if (cfq_class_idle(new_cfqq))
3235                 return false;
3236
3237         if (cfq_class_idle(cfqq))
3238                 return true;
3239
3240         /*
3241          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3242          */
3243         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3244                 return false;
3245
3246         /*
3247          * if the new request is sync, but the currently running queue is
3248          * not, let the sync request have priority.
3249          */
3250         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3251                 return true;
3252
3253         if (new_cfqq->cfqg != cfqq->cfqg)
3254                 return false;
3255
3256         if (cfq_slice_used(cfqq))
3257                 return true;
3258
3259         /* Allow preemption only if we are idling on sync-noidle tree */
3260         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3261             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3262             new_cfqq->service_tree->count == 2 &&
3263             RB_EMPTY_ROOT(&cfqq->sort_list))
3264                 return true;
3265
3266         /*
3267          * So both queues are sync. Let the new request get disk time if
3268          * it's a metadata request and the current queue is doing regular IO.
3269          */
3270         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3271                 return true;
3272
3273         /*
3274          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3275          */
3276         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3277                 return true;
3278
3279         /* An idle queue should not be idle now for some reason */
3280         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3281                 return true;
3282
3283         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3284                 return false;
3285
3286         /*
3287          * if this request is as-good as one we would expect from the
3288          * current cfqq, let it preempt
3289          */
3290         if (cfq_rq_close(cfqd, cfqq, rq))
3291                 return true;
3292
3293         return false;
3294 }
3295
3296 /*
3297  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3298  * let it have half of its nominal slice.
3299  */
3300 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3301 {
3302         struct cfq_queue *old_cfqq = cfqd->active_queue;
3303
3304         cfq_log_cfqq(cfqd, cfqq, "preempt");
3305         cfq_slice_expired(cfqd, 1);
3306
3307         /*
3308          * workload type is changed, don't save slice, otherwise preempt
3309          * doesn't happen
3310          */
3311         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3312                 cfqq->cfqg->saved_workload_slice = 0;
3313
3314         /*
3315          * Put the new queue at the front of the of the current list,
3316          * so we know that it will be selected next.
3317          */
3318         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3319
3320         cfq_service_tree_add(cfqd, cfqq, 1);
3321
3322         cfqq->slice_end = 0;
3323         cfq_mark_cfqq_slice_new(cfqq);
3324 }
3325
3326 /*
3327  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3328  * something we should do about it
3329  */
3330 static void
3331 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3332                 struct request *rq)
3333 {
3334         struct cfq_io_context *cic = RQ_CIC(rq);
3335
3336         cfqd->rq_queued++;
3337         if (rq->cmd_flags & REQ_META)
3338                 cfqq->meta_pending++;
3339
3340         cfq_update_io_thinktime(cfqd, cic);
3341         cfq_update_io_seektime(cfqd, cfqq, rq);
3342         cfq_update_idle_window(cfqd, cfqq, cic);
3343
3344         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3345
3346         if (cfqq == cfqd->active_queue) {
3347                 /*
3348                  * Remember that we saw a request from this process, but
3349                  * don't start queuing just yet. Otherwise we risk seeing lots
3350                  * of tiny requests, because we disrupt the normal plugging
3351                  * and merging. If the request is already larger than a single
3352                  * page, let it rip immediately. For that case we assume that
3353                  * merging is already done. Ditto for a busy system that
3354                  * has other work pending, don't risk delaying until the
3355                  * idle timer unplug to continue working.
3356                  */
3357                 if (cfq_cfqq_wait_request(cfqq)) {
3358                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3359                             cfqd->busy_queues > 1) {
3360                                 cfq_del_timer(cfqd, cfqq);
3361                                 cfq_clear_cfqq_wait_request(cfqq);
3362                                 __blk_run_queue(cfqd->queue);
3363                         } else {
3364                                 cfq_blkiocg_update_idle_time_stats(
3365                                                 &cfqq->cfqg->blkg);
3366                                 cfq_mark_cfqq_must_dispatch(cfqq);
3367                         }
3368                 }
3369         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3370                 /*
3371                  * not the active queue - expire current slice if it is
3372                  * idle and has expired it's mean thinktime or this new queue
3373                  * has some old slice time left and is of higher priority or
3374                  * this new queue is RT and the current one is BE
3375                  */
3376                 cfq_preempt_queue(cfqd, cfqq);
3377                 __blk_run_queue(cfqd->queue);
3378         }
3379 }
3380
3381 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3382 {
3383         struct cfq_data *cfqd = q->elevator->elevator_data;
3384         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3385
3386         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3387         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3388
3389         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3390         list_add_tail(&rq->queuelist, &cfqq->fifo);
3391         cfq_add_rq_rb(rq);
3392         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3393                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3394                         rq_is_sync(rq));
3395         cfq_rq_enqueued(cfqd, cfqq, rq);
3396 }
3397
3398 /*
3399  * Update hw_tag based on peak queue depth over 50 samples under
3400  * sufficient load.
3401  */
3402 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3403 {
3404         struct cfq_queue *cfqq = cfqd->active_queue;
3405
3406         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3407                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3408
3409         if (cfqd->hw_tag == 1)
3410                 return;
3411
3412         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3413             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3414                 return;
3415
3416         /*
3417          * If active queue hasn't enough requests and can idle, cfq might not
3418          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3419          * case
3420          */
3421         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3422             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3423             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3424                 return;
3425
3426         if (cfqd->hw_tag_samples++ < 50)
3427                 return;
3428
3429         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3430                 cfqd->hw_tag = 1;
3431         else
3432                 cfqd->hw_tag = 0;
3433 }
3434
3435 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3436 {
3437         struct cfq_io_context *cic = cfqd->active_cic;
3438
3439         /* If the queue already has requests, don't wait */
3440         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3441                 return false;
3442
3443         /* If there are other queues in the group, don't wait */
3444         if (cfqq->cfqg->nr_cfqq > 1)
3445                 return false;
3446
3447         if (cfq_slice_used(cfqq))
3448                 return true;
3449
3450         /* if slice left is less than think time, wait busy */
3451         if (cic && sample_valid(cic->ttime_samples)
3452             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3453                 return true;
3454
3455         /*
3456          * If think times is less than a jiffy than ttime_mean=0 and above
3457          * will not be true. It might happen that slice has not expired yet
3458          * but will expire soon (4-5 ns) during select_queue(). To cover the
3459          * case where think time is less than a jiffy, mark the queue wait
3460          * busy if only 1 jiffy is left in the slice.
3461          */
3462         if (cfqq->slice_end - jiffies == 1)
3463                 return true;
3464
3465         return false;
3466 }
3467
3468 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3469 {
3470         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3471         struct cfq_data *cfqd = cfqq->cfqd;
3472         const int sync = rq_is_sync(rq);
3473         unsigned long now;
3474
3475         now = jiffies;
3476         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3477                      !!(rq->cmd_flags & REQ_NOIDLE));
3478
3479         cfq_update_hw_tag(cfqd);
3480
3481         WARN_ON(!cfqd->rq_in_driver);
3482         WARN_ON(!cfqq->dispatched);
3483         cfqd->rq_in_driver--;
3484         cfqq->dispatched--;
3485         (RQ_CFQG(rq))->dispatched--;
3486         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3487                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3488                         rq_data_dir(rq), rq_is_sync(rq));
3489
3490         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3491
3492         if (sync) {
3493                 RQ_CIC(rq)->last_end_request = now;
3494                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3495                         cfqd->last_delayed_sync = now;
3496         }
3497
3498         /*
3499          * If this is the active queue, check if it needs to be expired,
3500          * or if we want to idle in case it has no pending requests.
3501          */
3502         if (cfqd->active_queue == cfqq) {
3503                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3504
3505                 if (cfq_cfqq_slice_new(cfqq)) {
3506                         cfq_set_prio_slice(cfqd, cfqq);
3507                         cfq_clear_cfqq_slice_new(cfqq);
3508                 }
3509
3510                 /*
3511                  * Should we wait for next request to come in before we expire
3512                  * the queue.
3513                  */
3514                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3515                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3516                         if (!cfqd->cfq_slice_idle)
3517                                 extend_sl = cfqd->cfq_group_idle;
3518                         cfqq->slice_end = jiffies + extend_sl;
3519                         cfq_mark_cfqq_wait_busy(cfqq);
3520                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3521                 }
3522
3523                 /*
3524                  * Idling is not enabled on:
3525                  * - expired queues
3526                  * - idle-priority queues
3527                  * - async queues
3528                  * - queues with still some requests queued
3529                  * - when there is a close cooperator
3530                  */
3531                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3532                         cfq_slice_expired(cfqd, 1);
3533                 else if (sync && cfqq_empty &&
3534                          !cfq_close_cooperator(cfqd, cfqq)) {
3535                         cfq_arm_slice_timer(cfqd);
3536                 }
3537         }
3538
3539         if (!cfqd->rq_in_driver)
3540                 cfq_schedule_dispatch(cfqd);
3541 }
3542
3543 /*
3544  * we temporarily boost lower priority queues if they are holding fs exclusive
3545  * resources. they are boosted to normal prio (CLASS_BE/4)
3546  */
3547 static void cfq_prio_boost(struct cfq_queue *cfqq)
3548 {
3549         if (has_fs_excl()) {
3550                 /*
3551                  * boost idle prio on transactions that would lock out other
3552                  * users of the filesystem
3553                  */
3554                 if (cfq_class_idle(cfqq))
3555                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3556                 if (cfqq->ioprio > IOPRIO_NORM)
3557                         cfqq->ioprio = IOPRIO_NORM;
3558         } else {
3559                 /*
3560                  * unboost the queue (if needed)
3561                  */
3562                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3563                 cfqq->ioprio = cfqq->org_ioprio;
3564         }
3565 }
3566
3567 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3568 {
3569         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3570                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3571                 return ELV_MQUEUE_MUST;
3572         }
3573
3574         return ELV_MQUEUE_MAY;
3575 }
3576
3577 static int cfq_may_queue(struct request_queue *q, int rw)
3578 {
3579         struct cfq_data *cfqd = q->elevator->elevator_data;
3580         struct task_struct *tsk = current;
3581         struct cfq_io_context *cic;
3582         struct cfq_queue *cfqq;
3583
3584         /*
3585          * don't force setup of a queue from here, as a call to may_queue
3586          * does not necessarily imply that a request actually will be queued.
3587          * so just lookup a possibly existing queue, or return 'may queue'
3588          * if that fails
3589          */
3590         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3591         if (!cic)
3592                 return ELV_MQUEUE_MAY;
3593
3594         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3595         if (cfqq) {
3596                 cfq_init_prio_data(cfqq, cic->ioc);
3597                 cfq_prio_boost(cfqq);
3598
3599                 return __cfq_may_queue(cfqq);
3600         }
3601
3602         return ELV_MQUEUE_MAY;
3603 }
3604
3605 /*
3606  * queue lock held here
3607  */
3608 static void cfq_put_request(struct request *rq)
3609 {
3610         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3611
3612         if (cfqq) {
3613                 const int rw = rq_data_dir(rq);
3614
3615                 BUG_ON(!cfqq->allocated[rw]);
3616                 cfqq->allocated[rw]--;
3617
3618                 put_io_context(RQ_CIC(rq)->ioc);
3619
3620                 rq->elevator_private[0] = NULL;
3621                 rq->elevator_private[1] = NULL;
3622
3623                 /* Put down rq reference on cfqg */
3624                 cfq_put_cfqg(RQ_CFQG(rq));
3625                 rq->elevator_private[2] = NULL;
3626
3627                 cfq_put_queue(cfqq);
3628         }
3629 }
3630
3631 static struct cfq_queue *
3632 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3633                 struct cfq_queue *cfqq)
3634 {
3635         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3636         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3637         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3638         cfq_put_queue(cfqq);
3639         return cic_to_cfqq(cic, 1);
3640 }
3641
3642 /*
3643  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3644  * was the last process referring to said cfqq.
3645  */
3646 static struct cfq_queue *
3647 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3648 {
3649         if (cfqq_process_refs(cfqq) == 1) {
3650                 cfqq->pid = current->pid;
3651                 cfq_clear_cfqq_coop(cfqq);
3652                 cfq_clear_cfqq_split_coop(cfqq);
3653                 return cfqq;
3654         }
3655
3656         cic_set_cfqq(cic, NULL, 1);
3657
3658         cfq_put_cooperator(cfqq);
3659
3660         cfq_put_queue(cfqq);
3661         return NULL;
3662 }
3663 /*
3664  * Allocate cfq data structures associated with this request.
3665  */
3666 static int
3667 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3668 {
3669         struct cfq_data *cfqd = q->elevator->elevator_data;
3670         struct cfq_io_context *cic;
3671         const int rw = rq_data_dir(rq);
3672         const bool is_sync = rq_is_sync(rq);
3673         struct cfq_queue *cfqq;
3674         unsigned long flags;
3675
3676         might_sleep_if(gfp_mask & __GFP_WAIT);
3677
3678         cic = cfq_get_io_context(cfqd, gfp_mask);
3679
3680         spin_lock_irqsave(q->queue_lock, flags);
3681
3682         if (!cic)
3683                 goto queue_fail;
3684
3685 new_queue:
3686         cfqq = cic_to_cfqq(cic, is_sync);
3687         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3688                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3689                 cic_set_cfqq(cic, cfqq, is_sync);
3690         } else {
3691                 /*
3692                  * If the queue was seeky for too long, break it apart.
3693                  */
3694                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3695                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3696                         cfqq = split_cfqq(cic, cfqq);
3697                         if (!cfqq)
3698                                 goto new_queue;
3699                 }
3700
3701                 /*
3702                  * Check to see if this queue is scheduled to merge with
3703                  * another, closely cooperating queue.  The merging of
3704                  * queues happens here as it must be done in process context.
3705                  * The reference on new_cfqq was taken in merge_cfqqs.
3706                  */
3707                 if (cfqq->new_cfqq)
3708                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3709         }
3710
3711         cfqq->allocated[rw]++;
3712
3713         cfqq->ref++;
3714         rq->elevator_private[0] = cic;
3715         rq->elevator_private[1] = cfqq;
3716         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3717         spin_unlock_irqrestore(q->queue_lock, flags);
3718         return 0;
3719
3720 queue_fail:
3721         if (cic)
3722                 put_io_context(cic->ioc);
3723
3724         cfq_schedule_dispatch(cfqd);
3725         spin_unlock_irqrestore(q->queue_lock, flags);
3726         cfq_log(cfqd, "set_request fail");
3727         return 1;
3728 }
3729
3730 static void cfq_kick_queue(struct work_struct *work)
3731 {
3732         struct cfq_data *cfqd =
3733                 container_of(work, struct cfq_data, unplug_work);
3734         struct request_queue *q = cfqd->queue;
3735
3736         spin_lock_irq(q->queue_lock);
3737         __blk_run_queue(cfqd->queue);
3738         spin_unlock_irq(q->queue_lock);
3739 }
3740
3741 /*
3742  * Timer running if the active_queue is currently idling inside its time slice
3743  */
3744 static void cfq_idle_slice_timer(unsigned long data)
3745 {
3746         struct cfq_data *cfqd = (struct cfq_data *) data;
3747         struct cfq_queue *cfqq;
3748         unsigned long flags;
3749         int timed_out = 1;
3750
3751         cfq_log(cfqd, "idle timer fired");
3752
3753         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3754
3755         cfqq = cfqd->active_queue;
3756         if (cfqq) {
3757                 timed_out = 0;
3758
3759                 /*
3760                  * We saw a request before the queue expired, let it through
3761                  */
3762                 if (cfq_cfqq_must_dispatch(cfqq))
3763                         goto out_kick;
3764
3765                 /*
3766                  * expired
3767                  */
3768                 if (cfq_slice_used(cfqq))
3769                         goto expire;
3770
3771                 /*
3772                  * only expire and reinvoke request handler, if there are
3773                  * other queues with pending requests
3774                  */
3775                 if (!cfqd->busy_queues)
3776                         goto out_cont;
3777
3778                 /*
3779                  * not expired and it has a request pending, let it dispatch
3780                  */
3781                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3782                         goto out_kick;
3783
3784                 /*
3785                  * Queue depth flag is reset only when the idle didn't succeed
3786                  */
3787                 cfq_clear_cfqq_deep(cfqq);
3788         }
3789 expire:
3790         cfq_slice_expired(cfqd, timed_out);
3791 out_kick:
3792         cfq_schedule_dispatch(cfqd);
3793 out_cont:
3794         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3795 }
3796
3797 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3798 {
3799         del_timer_sync(&cfqd->idle_slice_timer);
3800         cancel_work_sync(&cfqd->unplug_work);
3801 }
3802
3803 static void cfq_put_async_queues(struct cfq_data *cfqd)
3804 {
3805         int i;
3806
3807         for (i = 0; i < IOPRIO_BE_NR; i++) {
3808                 if (cfqd->async_cfqq[0][i])
3809                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3810                 if (cfqd->async_cfqq[1][i])
3811                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3812         }
3813
3814         if (cfqd->async_idle_cfqq)
3815                 cfq_put_queue(cfqd->async_idle_cfqq);
3816 }
3817
3818 static void cfq_cfqd_free(struct rcu_head *head)
3819 {
3820         kfree(container_of(head, struct cfq_data, rcu));
3821 }
3822
3823 static void cfq_exit_queue(struct elevator_queue *e)
3824 {
3825         struct cfq_data *cfqd = e->elevator_data;
3826         struct request_queue *q = cfqd->queue;
3827
3828         cfq_shutdown_timer_wq(cfqd);
3829
3830         spin_lock_irq(q->queue_lock);
3831
3832         if (cfqd->active_queue)
3833                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3834
3835         while (!list_empty(&cfqd->cic_list)) {
3836                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3837                                                         struct cfq_io_context,
3838                                                         queue_list);
3839
3840                 __cfq_exit_single_io_context(cfqd, cic);
3841         }
3842
3843         cfq_put_async_queues(cfqd);
3844         cfq_release_cfq_groups(cfqd);
3845         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3846
3847         spin_unlock_irq(q->queue_lock);
3848
3849         cfq_shutdown_timer_wq(cfqd);
3850
3851         spin_lock(&cic_index_lock);
3852         ida_remove(&cic_index_ida, cfqd->cic_index);
3853         spin_unlock(&cic_index_lock);
3854
3855         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3856         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3857 }
3858
3859 static int cfq_alloc_cic_index(void)
3860 {
3861         int index, error;
3862
3863         do {
3864                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3865                         return -ENOMEM;
3866
3867                 spin_lock(&cic_index_lock);
3868                 error = ida_get_new(&cic_index_ida, &index);
3869                 spin_unlock(&cic_index_lock);
3870                 if (error && error != -EAGAIN)
3871                         return error;
3872         } while (error);
3873
3874         return index;
3875 }
3876
3877 static void *cfq_init_queue(struct request_queue *q)
3878 {
3879         struct cfq_data *cfqd;
3880         int i, j;
3881         struct cfq_group *cfqg;
3882         struct cfq_rb_root *st;
3883
3884         i = cfq_alloc_cic_index();
3885         if (i < 0)
3886                 return NULL;
3887
3888         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3889         if (!cfqd)
3890                 return NULL;
3891
3892         /*
3893          * Don't need take queue_lock in the routine, since we are
3894          * initializing the ioscheduler, and nobody is using cfqd
3895          */
3896         cfqd->cic_index = i;
3897
3898         /* Init root service tree */
3899         cfqd->grp_service_tree = CFQ_RB_ROOT;
3900
3901         /* Init root group */
3902         cfqg = &cfqd->root_group;
3903         for_each_cfqg_st(cfqg, i, j, st)
3904                 *st = CFQ_RB_ROOT;
3905         RB_CLEAR_NODE(&cfqg->rb_node);
3906
3907         /* Give preference to root group over other groups */
3908         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3909
3910 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3911         /*
3912          * Take a reference to root group which we never drop. This is just
3913          * to make sure that cfq_put_cfqg() does not try to kfree root group
3914          */
3915         cfqg->ref = 1;
3916         rcu_read_lock();
3917         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3918                                         (void *)cfqd, 0);
3919         rcu_read_unlock();
3920 #endif
3921         /*
3922          * Not strictly needed (since RB_ROOT just clears the node and we
3923          * zeroed cfqd on alloc), but better be safe in case someone decides
3924          * to add magic to the rb code
3925          */
3926         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3927                 cfqd->prio_trees[i] = RB_ROOT;
3928
3929         /*
3930          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3931          * Grab a permanent reference to it, so that the normal code flow
3932          * will not attempt to free it.
3933          */
3934         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3935         cfqd->oom_cfqq.ref++;
3936         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3937
3938         INIT_LIST_HEAD(&cfqd->cic_list);
3939
3940         cfqd->queue = q;
3941
3942         init_timer(&cfqd->idle_slice_timer);
3943         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3944         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3945
3946         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3947
3948         cfqd->cfq_quantum = cfq_quantum;
3949         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3950         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3951         cfqd->cfq_back_max = cfq_back_max;
3952         cfqd->cfq_back_penalty = cfq_back_penalty;
3953         cfqd->cfq_slice[0] = cfq_slice_async;
3954         cfqd->cfq_slice[1] = cfq_slice_sync;
3955         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3956         cfqd->cfq_slice_idle = cfq_slice_idle;
3957         cfqd->cfq_group_idle = cfq_group_idle;
3958         cfqd->cfq_latency = 1;
3959         cfqd->hw_tag = -1;
3960         /*
3961          * we optimistically start assuming sync ops weren't delayed in last
3962          * second, in order to have larger depth for async operations.
3963          */
3964         cfqd->last_delayed_sync = jiffies - HZ;
3965         return cfqd;
3966 }
3967
3968 static void cfq_slab_kill(void)
3969 {
3970         /*
3971          * Caller already ensured that pending RCU callbacks are completed,
3972          * so we should have no busy allocations at this point.
3973          */
3974         if (cfq_pool)
3975                 kmem_cache_destroy(cfq_pool);
3976         if (cfq_ioc_pool)
3977                 kmem_cache_destroy(cfq_ioc_pool);
3978 }
3979
3980 static int __init cfq_slab_setup(void)
3981 {
3982         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3983         if (!cfq_pool)
3984                 goto fail;
3985
3986         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3987         if (!cfq_ioc_pool)
3988                 goto fail;
3989
3990         return 0;
3991 fail:
3992         cfq_slab_kill();
3993         return -ENOMEM;
3994 }
3995
3996 /*
3997  * sysfs parts below -->
3998  */
3999 static ssize_t
4000 cfq_var_show(unsigned int var, char *page)
4001 {
4002         return sprintf(page, "%d\n", var);
4003 }
4004
4005 static ssize_t
4006 cfq_var_store(unsigned int *var, const char *page, size_t count)
4007 {
4008         char *p = (char *) page;
4009
4010         *var = simple_strtoul(p, &p, 10);
4011         return count;
4012 }
4013
4014 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4015 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4016 {                                                                       \
4017         struct cfq_data *cfqd = e->elevator_data;                       \
4018         unsigned int __data = __VAR;                                    \
4019         if (__CONV)                                                     \
4020                 __data = jiffies_to_msecs(__data);                      \
4021         return cfq_var_show(__data, (page));                            \
4022 }
4023 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4024 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4025 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4026 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4027 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4028 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4029 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4030 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4031 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4032 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4033 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4034 #undef SHOW_FUNCTION
4035
4036 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4037 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4038 {                                                                       \
4039         struct cfq_data *cfqd = e->elevator_data;                       \
4040         unsigned int __data;                                            \
4041         int ret = cfq_var_store(&__data, (page), count);                \
4042         if (__data < (MIN))                                             \
4043                 __data = (MIN);                                         \
4044         else if (__data > (MAX))                                        \
4045                 __data = (MAX);                                         \
4046         if (__CONV)                                                     \
4047                 *(__PTR) = msecs_to_jiffies(__data);                    \
4048         else                                                            \
4049                 *(__PTR) = __data;                                      \
4050         return ret;                                                     \
4051 }
4052 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4053 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4054                 UINT_MAX, 1);
4055 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4056                 UINT_MAX, 1);
4057 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4058 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4059                 UINT_MAX, 0);
4060 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4061 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4062 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4063 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4064 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4065                 UINT_MAX, 0);
4066 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4067 #undef STORE_FUNCTION
4068
4069 #define CFQ_ATTR(name) \
4070         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4071
4072 static struct elv_fs_entry cfq_attrs[] = {
4073         CFQ_ATTR(quantum),
4074         CFQ_ATTR(fifo_expire_sync),
4075         CFQ_ATTR(fifo_expire_async),
4076         CFQ_ATTR(back_seek_max),
4077         CFQ_ATTR(back_seek_penalty),
4078         CFQ_ATTR(slice_sync),
4079         CFQ_ATTR(slice_async),
4080         CFQ_ATTR(slice_async_rq),
4081         CFQ_ATTR(slice_idle),
4082         CFQ_ATTR(group_idle),
4083         CFQ_ATTR(low_latency),
4084         __ATTR_NULL
4085 };
4086
4087 static struct elevator_type iosched_cfq = {
4088         .ops = {
4089                 .elevator_merge_fn =            cfq_merge,
4090                 .elevator_merged_fn =           cfq_merged_request,
4091                 .elevator_merge_req_fn =        cfq_merged_requests,
4092                 .elevator_allow_merge_fn =      cfq_allow_merge,
4093                 .elevator_bio_merged_fn =       cfq_bio_merged,
4094                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4095                 .elevator_add_req_fn =          cfq_insert_request,
4096                 .elevator_activate_req_fn =     cfq_activate_request,
4097                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4098                 .elevator_completed_req_fn =    cfq_completed_request,
4099                 .elevator_former_req_fn =       elv_rb_former_request,
4100                 .elevator_latter_req_fn =       elv_rb_latter_request,
4101                 .elevator_set_req_fn =          cfq_set_request,
4102                 .elevator_put_req_fn =          cfq_put_request,
4103                 .elevator_may_queue_fn =        cfq_may_queue,
4104                 .elevator_init_fn =             cfq_init_queue,
4105                 .elevator_exit_fn =             cfq_exit_queue,
4106                 .trim =                         cfq_free_io_context,
4107         },
4108         .elevator_attrs =       cfq_attrs,
4109         .elevator_name =        "cfq",
4110         .elevator_owner =       THIS_MODULE,
4111 };
4112
4113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4114 static struct blkio_policy_type blkio_policy_cfq = {
4115         .ops = {
4116                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4117                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4118         },
4119         .plid = BLKIO_POLICY_PROP,
4120 };
4121 #else
4122 static struct blkio_policy_type blkio_policy_cfq;
4123 #endif
4124
4125 static int __init cfq_init(void)
4126 {
4127         /*
4128          * could be 0 on HZ < 1000 setups
4129          */
4130         if (!cfq_slice_async)
4131                 cfq_slice_async = 1;
4132         if (!cfq_slice_idle)
4133                 cfq_slice_idle = 1;
4134
4135 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4136         if (!cfq_group_idle)
4137                 cfq_group_idle = 1;
4138 #else
4139                 cfq_group_idle = 0;
4140 #endif
4141         if (cfq_slab_setup())
4142                 return -ENOMEM;
4143
4144         elv_register(&iosched_cfq);
4145         blkio_policy_register(&blkio_policy_cfq);
4146
4147         return 0;
4148 }
4149
4150 static void __exit cfq_exit(void)
4151 {
4152         DECLARE_COMPLETION_ONSTACK(all_gone);
4153         blkio_policy_unregister(&blkio_policy_cfq);
4154         elv_unregister(&iosched_cfq);
4155         ioc_gone = &all_gone;
4156         /* ioc_gone's update must be visible before reading ioc_count */
4157         smp_wmb();
4158
4159         /*
4160          * this also protects us from entering cfq_slab_kill() with
4161          * pending RCU callbacks
4162          */
4163         if (elv_ioc_count_read(cfq_ioc_count))
4164                 wait_for_completion(&all_gone);
4165         ida_destroy(&cic_index_ida);
4166         cfq_slab_kill();
4167 }
4168
4169 module_init(cfq_init);
4170 module_exit(cfq_exit);
4171
4172 MODULE_AUTHOR("Jens Axboe");
4173 MODULE_LICENSE("GPL");
4174 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");