cfq-iosched: don't idle if a deep seek queue is slow
[pandora-kernel.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private)
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct rb_node *active;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
93                         .count = 0, .min_vdisktime = 0, }
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         atomic_t ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* pending metadata requests */
134         int meta_pending;
135         /* number of requests that are on the dispatch list or inside driver */
136         int dispatched;
137
138         /* io prio of this group */
139         unsigned short ioprio, org_ioprio;
140         unsigned short ioprio_class, org_ioprio_class;
141
142         pid_t pid;
143
144         u32 seek_history;
145         sector_t last_request_pos;
146
147         struct cfq_rb_root *service_tree;
148         struct cfq_queue *new_cfqq;
149         struct cfq_group *cfqg;
150         struct cfq_group *orig_cfqg;
151         /* Number of sectors dispatched from queue in single dispatch round */
152         unsigned long nr_sectors;
153 };
154
155 /*
156  * First index in the service_trees.
157  * IDLE is handled separately, so it has negative index
158  */
159 enum wl_prio_t {
160         BE_WORKLOAD = 0,
161         RT_WORKLOAD = 1,
162         IDLE_WORKLOAD = 2,
163 };
164
165 /*
166  * Second index in the service_trees.
167  */
168 enum wl_type_t {
169         ASYNC_WORKLOAD = 0,
170         SYNC_NOIDLE_WORKLOAD = 1,
171         SYNC_WORKLOAD = 2
172 };
173
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176         /* group service_tree member */
177         struct rb_node rb_node;
178
179         /* group service_tree key */
180         u64 vdisktime;
181         unsigned int weight;
182         bool on_st;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /* Per group busy queus average. Useful for workload slice calc. */
188         unsigned int busy_queues_avg[2];
189         /*
190          * rr lists of queues with requests, onle rr for each priority class.
191          * Counts are embedded in the cfq_rb_root
192          */
193         struct cfq_rb_root service_trees[2][3];
194         struct cfq_rb_root service_tree_idle;
195
196         unsigned long saved_workload_slice;
197         enum wl_type_t saved_workload;
198         enum wl_prio_t saved_serving_prio;
199         struct blkio_group blkg;
200 #ifdef CONFIG_CFQ_GROUP_IOSCHED
201         struct hlist_node cfqd_node;
202         atomic_t ref;
203 #endif
204         /* number of requests that are on the dispatch list or inside driver */
205         int dispatched;
206 };
207
208 /*
209  * Per block device queue structure
210  */
211 struct cfq_data {
212         struct request_queue *queue;
213         /* Root service tree for cfq_groups */
214         struct cfq_rb_root grp_service_tree;
215         struct cfq_group root_group;
216
217         /*
218          * The priority currently being served
219          */
220         enum wl_prio_t serving_prio;
221         enum wl_type_t serving_type;
222         unsigned long workload_expires;
223         struct cfq_group *serving_group;
224         bool noidle_tree_requires_idle;
225
226         /*
227          * Each priority tree is sorted by next_request position.  These
228          * trees are used when determining if two or more queues are
229          * interleaving requests (see cfq_close_cooperator).
230          */
231         struct rb_root prio_trees[CFQ_PRIO_LISTS];
232
233         unsigned int busy_queues;
234
235         int rq_in_driver;
236         int rq_in_flight[2];
237
238         /*
239          * queue-depth detection
240          */
241         int rq_queued;
242         int hw_tag;
243         /*
244          * hw_tag can be
245          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
246          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
247          *  0 => no NCQ
248          */
249         int hw_tag_est_depth;
250         unsigned int hw_tag_samples;
251
252         /*
253          * idle window management
254          */
255         struct timer_list idle_slice_timer;
256         struct work_struct unplug_work;
257
258         struct cfq_queue *active_queue;
259         struct cfq_io_context *active_cic;
260
261         /*
262          * async queue for each priority case
263          */
264         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
265         struct cfq_queue *async_idle_cfqq;
266
267         sector_t last_position;
268
269         /*
270          * tunables, see top of file
271          */
272         unsigned int cfq_quantum;
273         unsigned int cfq_fifo_expire[2];
274         unsigned int cfq_back_penalty;
275         unsigned int cfq_back_max;
276         unsigned int cfq_slice[2];
277         unsigned int cfq_slice_async_rq;
278         unsigned int cfq_slice_idle;
279         unsigned int cfq_group_idle;
280         unsigned int cfq_latency;
281         unsigned int cfq_group_isolation;
282
283         unsigned int cic_index;
284         struct list_head cic_list;
285
286         /*
287          * Fallback dummy cfqq for extreme OOM conditions
288          */
289         struct cfq_queue oom_cfqq;
290
291         unsigned long last_delayed_sync;
292
293         /* List of cfq groups being managed on this device*/
294         struct hlist_head cfqg_list;
295         struct rcu_head rcu;
296 };
297
298 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
299
300 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
301                                             enum wl_prio_t prio,
302                                             enum wl_type_t type)
303 {
304         if (!cfqg)
305                 return NULL;
306
307         if (prio == IDLE_WORKLOAD)
308                 return &cfqg->service_tree_idle;
309
310         return &cfqg->service_trees[prio][type];
311 }
312
313 enum cfqq_state_flags {
314         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
315         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
316         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
317         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
318         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
319         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
320         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
321         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
322         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
323         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
324         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
325         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
326         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
327 };
328
329 #define CFQ_CFQQ_FNS(name)                                              \
330 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
331 {                                                                       \
332         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
333 }                                                                       \
334 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
335 {                                                                       \
336         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
337 }                                                                       \
338 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
339 {                                                                       \
340         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
341 }
342
343 CFQ_CFQQ_FNS(on_rr);
344 CFQ_CFQQ_FNS(wait_request);
345 CFQ_CFQQ_FNS(must_dispatch);
346 CFQ_CFQQ_FNS(must_alloc_slice);
347 CFQ_CFQQ_FNS(fifo_expire);
348 CFQ_CFQQ_FNS(idle_window);
349 CFQ_CFQQ_FNS(prio_changed);
350 CFQ_CFQQ_FNS(slice_new);
351 CFQ_CFQQ_FNS(sync);
352 CFQ_CFQQ_FNS(coop);
353 CFQ_CFQQ_FNS(split_coop);
354 CFQ_CFQQ_FNS(deep);
355 CFQ_CFQQ_FNS(wait_busy);
356 #undef CFQ_CFQQ_FNS
357
358 #ifdef CONFIG_CFQ_GROUP_IOSCHED
359 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
360         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
361                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
362                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
363
364 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
365         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
366                                 blkg_path(&(cfqg)->blkg), ##args);      \
367
368 #else
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
370         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
371 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
372 #endif
373 #define cfq_log(cfqd, fmt, args...)     \
374         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
375
376 /* Traverses through cfq group service trees */
377 #define for_each_cfqg_st(cfqg, i, j, st) \
378         for (i = 0; i <= IDLE_WORKLOAD; i++) \
379                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
380                         : &cfqg->service_tree_idle; \
381                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
382                         (i == IDLE_WORKLOAD && j == 0); \
383                         j++, st = i < IDLE_WORKLOAD ? \
384                         &cfqg->service_trees[i][j]: NULL) \
385
386
387 static inline bool iops_mode(struct cfq_data *cfqd)
388 {
389         /*
390          * If we are not idling on queues and it is a NCQ drive, parallel
391          * execution of requests is on and measuring time is not possible
392          * in most of the cases until and unless we drive shallower queue
393          * depths and that becomes a performance bottleneck. In such cases
394          * switch to start providing fairness in terms of number of IOs.
395          */
396         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
397                 return true;
398         else
399                 return false;
400 }
401
402 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
403 {
404         if (cfq_class_idle(cfqq))
405                 return IDLE_WORKLOAD;
406         if (cfq_class_rt(cfqq))
407                 return RT_WORKLOAD;
408         return BE_WORKLOAD;
409 }
410
411
412 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
413 {
414         if (!cfq_cfqq_sync(cfqq))
415                 return ASYNC_WORKLOAD;
416         if (!cfq_cfqq_idle_window(cfqq))
417                 return SYNC_NOIDLE_WORKLOAD;
418         return SYNC_WORKLOAD;
419 }
420
421 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
422                                         struct cfq_data *cfqd,
423                                         struct cfq_group *cfqg)
424 {
425         if (wl == IDLE_WORKLOAD)
426                 return cfqg->service_tree_idle.count;
427
428         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
429                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
430                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
431 }
432
433 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
434                                         struct cfq_group *cfqg)
435 {
436         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
437                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
438 }
439
440 static void cfq_dispatch_insert(struct request_queue *, struct request *);
441 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
442                                        struct io_context *, gfp_t);
443 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
444                                                 struct io_context *);
445
446 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
447                                             bool is_sync)
448 {
449         return cic->cfqq[is_sync];
450 }
451
452 static inline void cic_set_cfqq(struct cfq_io_context *cic,
453                                 struct cfq_queue *cfqq, bool is_sync)
454 {
455         cic->cfqq[is_sync] = cfqq;
456 }
457
458 #define CIC_DEAD_KEY    1ul
459 #define CIC_DEAD_INDEX_SHIFT    1
460
461 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
462 {
463         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
464 }
465
466 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
467 {
468         struct cfq_data *cfqd = cic->key;
469
470         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
471                 return NULL;
472
473         return cfqd;
474 }
475
476 /*
477  * We regard a request as SYNC, if it's either a read or has the SYNC bit
478  * set (in which case it could also be direct WRITE).
479  */
480 static inline bool cfq_bio_sync(struct bio *bio)
481 {
482         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
483 }
484
485 /*
486  * scheduler run of queue, if there are requests pending and no one in the
487  * driver that will restart queueing
488  */
489 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
490 {
491         if (cfqd->busy_queues) {
492                 cfq_log(cfqd, "schedule dispatch");
493                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
494         }
495 }
496
497 static int cfq_queue_empty(struct request_queue *q)
498 {
499         struct cfq_data *cfqd = q->elevator->elevator_data;
500
501         return !cfqd->rq_queued;
502 }
503
504 /*
505  * Scale schedule slice based on io priority. Use the sync time slice only
506  * if a queue is marked sync and has sync io queued. A sync queue with async
507  * io only, should not get full sync slice length.
508  */
509 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
510                                  unsigned short prio)
511 {
512         const int base_slice = cfqd->cfq_slice[sync];
513
514         WARN_ON(prio >= IOPRIO_BE_NR);
515
516         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
517 }
518
519 static inline int
520 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
521 {
522         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
523 }
524
525 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
526 {
527         u64 d = delta << CFQ_SERVICE_SHIFT;
528
529         d = d * BLKIO_WEIGHT_DEFAULT;
530         do_div(d, cfqg->weight);
531         return d;
532 }
533
534 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
535 {
536         s64 delta = (s64)(vdisktime - min_vdisktime);
537         if (delta > 0)
538                 min_vdisktime = vdisktime;
539
540         return min_vdisktime;
541 }
542
543 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
544 {
545         s64 delta = (s64)(vdisktime - min_vdisktime);
546         if (delta < 0)
547                 min_vdisktime = vdisktime;
548
549         return min_vdisktime;
550 }
551
552 static void update_min_vdisktime(struct cfq_rb_root *st)
553 {
554         u64 vdisktime = st->min_vdisktime;
555         struct cfq_group *cfqg;
556
557         if (st->active) {
558                 cfqg = rb_entry_cfqg(st->active);
559                 vdisktime = cfqg->vdisktime;
560         }
561
562         if (st->left) {
563                 cfqg = rb_entry_cfqg(st->left);
564                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
565         }
566
567         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
568 }
569
570 /*
571  * get averaged number of queues of RT/BE priority.
572  * average is updated, with a formula that gives more weight to higher numbers,
573  * to quickly follows sudden increases and decrease slowly
574  */
575
576 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
577                                         struct cfq_group *cfqg, bool rt)
578 {
579         unsigned min_q, max_q;
580         unsigned mult  = cfq_hist_divisor - 1;
581         unsigned round = cfq_hist_divisor / 2;
582         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
583
584         min_q = min(cfqg->busy_queues_avg[rt], busy);
585         max_q = max(cfqg->busy_queues_avg[rt], busy);
586         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
587                 cfq_hist_divisor;
588         return cfqg->busy_queues_avg[rt];
589 }
590
591 static inline unsigned
592 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
593 {
594         struct cfq_rb_root *st = &cfqd->grp_service_tree;
595
596         return cfq_target_latency * cfqg->weight / st->total_weight;
597 }
598
599 static inline void
600 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
601 {
602         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
603         if (cfqd->cfq_latency) {
604                 /*
605                  * interested queues (we consider only the ones with the same
606                  * priority class in the cfq group)
607                  */
608                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
609                                                 cfq_class_rt(cfqq));
610                 unsigned sync_slice = cfqd->cfq_slice[1];
611                 unsigned expect_latency = sync_slice * iq;
612                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
613
614                 if (expect_latency > group_slice) {
615                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
616                         /* scale low_slice according to IO priority
617                          * and sync vs async */
618                         unsigned low_slice =
619                                 min(slice, base_low_slice * slice / sync_slice);
620                         /* the adapted slice value is scaled to fit all iqs
621                          * into the target latency */
622                         slice = max(slice * group_slice / expect_latency,
623                                     low_slice);
624                 }
625         }
626         cfqq->slice_start = jiffies;
627         cfqq->slice_end = jiffies + slice;
628         cfqq->allocated_slice = slice;
629         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
630 }
631
632 /*
633  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
634  * isn't valid until the first request from the dispatch is activated
635  * and the slice time set.
636  */
637 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
638 {
639         if (cfq_cfqq_slice_new(cfqq))
640                 return false;
641         if (time_before(jiffies, cfqq->slice_end))
642                 return false;
643
644         return true;
645 }
646
647 /*
648  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
649  * We choose the request that is closest to the head right now. Distance
650  * behind the head is penalized and only allowed to a certain extent.
651  */
652 static struct request *
653 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
654 {
655         sector_t s1, s2, d1 = 0, d2 = 0;
656         unsigned long back_max;
657 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
658 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
659         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
660
661         if (rq1 == NULL || rq1 == rq2)
662                 return rq2;
663         if (rq2 == NULL)
664                 return rq1;
665
666         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
667                 return rq1;
668         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
669                 return rq2;
670         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
671                 return rq1;
672         else if ((rq2->cmd_flags & REQ_META) &&
673                  !(rq1->cmd_flags & REQ_META))
674                 return rq2;
675
676         s1 = blk_rq_pos(rq1);
677         s2 = blk_rq_pos(rq2);
678
679         /*
680          * by definition, 1KiB is 2 sectors
681          */
682         back_max = cfqd->cfq_back_max * 2;
683
684         /*
685          * Strict one way elevator _except_ in the case where we allow
686          * short backward seeks which are biased as twice the cost of a
687          * similar forward seek.
688          */
689         if (s1 >= last)
690                 d1 = s1 - last;
691         else if (s1 + back_max >= last)
692                 d1 = (last - s1) * cfqd->cfq_back_penalty;
693         else
694                 wrap |= CFQ_RQ1_WRAP;
695
696         if (s2 >= last)
697                 d2 = s2 - last;
698         else if (s2 + back_max >= last)
699                 d2 = (last - s2) * cfqd->cfq_back_penalty;
700         else
701                 wrap |= CFQ_RQ2_WRAP;
702
703         /* Found required data */
704
705         /*
706          * By doing switch() on the bit mask "wrap" we avoid having to
707          * check two variables for all permutations: --> faster!
708          */
709         switch (wrap) {
710         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
711                 if (d1 < d2)
712                         return rq1;
713                 else if (d2 < d1)
714                         return rq2;
715                 else {
716                         if (s1 >= s2)
717                                 return rq1;
718                         else
719                                 return rq2;
720                 }
721
722         case CFQ_RQ2_WRAP:
723                 return rq1;
724         case CFQ_RQ1_WRAP:
725                 return rq2;
726         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727         default:
728                 /*
729                  * Since both rqs are wrapped,
730                  * start with the one that's further behind head
731                  * (--> only *one* back seek required),
732                  * since back seek takes more time than forward.
733                  */
734                 if (s1 <= s2)
735                         return rq1;
736                 else
737                         return rq2;
738         }
739 }
740
741 /*
742  * The below is leftmost cache rbtree addon
743  */
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
745 {
746         /* Service tree is empty */
747         if (!root->count)
748                 return NULL;
749
750         if (!root->left)
751                 root->left = rb_first(&root->rb);
752
753         if (root->left)
754                 return rb_entry(root->left, struct cfq_queue, rb_node);
755
756         return NULL;
757 }
758
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760 {
761         if (!root->left)
762                 root->left = rb_first(&root->rb);
763
764         if (root->left)
765                 return rb_entry_cfqg(root->left);
766
767         return NULL;
768 }
769
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
771 {
772         rb_erase(n, root);
773         RB_CLEAR_NODE(n);
774 }
775
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
777 {
778         if (root->left == n)
779                 root->left = NULL;
780         rb_erase_init(n, &root->rb);
781         --root->count;
782 }
783
784 /*
785  * would be nice to take fifo expire time into account as well
786  */
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789                   struct request *last)
790 {
791         struct rb_node *rbnext = rb_next(&last->rb_node);
792         struct rb_node *rbprev = rb_prev(&last->rb_node);
793         struct request *next = NULL, *prev = NULL;
794
795         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
796
797         if (rbprev)
798                 prev = rb_entry_rq(rbprev);
799
800         if (rbnext)
801                 next = rb_entry_rq(rbnext);
802         else {
803                 rbnext = rb_first(&cfqq->sort_list);
804                 if (rbnext && rbnext != &last->rb_node)
805                         next = rb_entry_rq(rbnext);
806         }
807
808         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
809 }
810
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812                                       struct cfq_queue *cfqq)
813 {
814         /*
815          * just an approximation, should be ok.
816          */
817         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
819 }
820
821 static inline s64
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
823 {
824         return cfqg->vdisktime - st->min_vdisktime;
825 }
826
827 static void
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
829 {
830         struct rb_node **node = &st->rb.rb_node;
831         struct rb_node *parent = NULL;
832         struct cfq_group *__cfqg;
833         s64 key = cfqg_key(st, cfqg);
834         int left = 1;
835
836         while (*node != NULL) {
837                 parent = *node;
838                 __cfqg = rb_entry_cfqg(parent);
839
840                 if (key < cfqg_key(st, __cfqg))
841                         node = &parent->rb_left;
842                 else {
843                         node = &parent->rb_right;
844                         left = 0;
845                 }
846         }
847
848         if (left)
849                 st->left = &cfqg->rb_node;
850
851         rb_link_node(&cfqg->rb_node, parent, node);
852         rb_insert_color(&cfqg->rb_node, &st->rb);
853 }
854
855 static void
856 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
857 {
858         struct cfq_rb_root *st = &cfqd->grp_service_tree;
859         struct cfq_group *__cfqg;
860         struct rb_node *n;
861
862         cfqg->nr_cfqq++;
863         if (cfqg->on_st)
864                 return;
865
866         /*
867          * Currently put the group at the end. Later implement something
868          * so that groups get lesser vtime based on their weights, so that
869          * if group does not loose all if it was not continously backlogged.
870          */
871         n = rb_last(&st->rb);
872         if (n) {
873                 __cfqg = rb_entry_cfqg(n);
874                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
875         } else
876                 cfqg->vdisktime = st->min_vdisktime;
877
878         __cfq_group_service_tree_add(st, cfqg);
879         cfqg->on_st = true;
880         st->total_weight += cfqg->weight;
881 }
882
883 static void
884 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
885 {
886         struct cfq_rb_root *st = &cfqd->grp_service_tree;
887
888         if (st->active == &cfqg->rb_node)
889                 st->active = NULL;
890
891         BUG_ON(cfqg->nr_cfqq < 1);
892         cfqg->nr_cfqq--;
893
894         /* If there are other cfq queues under this group, don't delete it */
895         if (cfqg->nr_cfqq)
896                 return;
897
898         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
899         cfqg->on_st = false;
900         st->total_weight -= cfqg->weight;
901         if (!RB_EMPTY_NODE(&cfqg->rb_node))
902                 cfq_rb_erase(&cfqg->rb_node, st);
903         cfqg->saved_workload_slice = 0;
904         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
905 }
906
907 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
908 {
909         unsigned int slice_used;
910
911         /*
912          * Queue got expired before even a single request completed or
913          * got expired immediately after first request completion.
914          */
915         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
916                 /*
917                  * Also charge the seek time incurred to the group, otherwise
918                  * if there are mutiple queues in the group, each can dispatch
919                  * a single request on seeky media and cause lots of seek time
920                  * and group will never know it.
921                  */
922                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
923                                         1);
924         } else {
925                 slice_used = jiffies - cfqq->slice_start;
926                 if (slice_used > cfqq->allocated_slice)
927                         slice_used = cfqq->allocated_slice;
928         }
929
930         return slice_used;
931 }
932
933 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
934                                 struct cfq_queue *cfqq)
935 {
936         struct cfq_rb_root *st = &cfqd->grp_service_tree;
937         unsigned int used_sl, charge;
938         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
939                         - cfqg->service_tree_idle.count;
940
941         BUG_ON(nr_sync < 0);
942         used_sl = charge = cfq_cfqq_slice_usage(cfqq);
943
944         if (iops_mode(cfqd))
945                 charge = cfqq->slice_dispatch;
946         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
947                 charge = cfqq->allocated_slice;
948
949         /* Can't update vdisktime while group is on service tree */
950         cfq_rb_erase(&cfqg->rb_node, st);
951         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
952         __cfq_group_service_tree_add(st, cfqg);
953
954         /* This group is being expired. Save the context */
955         if (time_after(cfqd->workload_expires, jiffies)) {
956                 cfqg->saved_workload_slice = cfqd->workload_expires
957                                                 - jiffies;
958                 cfqg->saved_workload = cfqd->serving_type;
959                 cfqg->saved_serving_prio = cfqd->serving_prio;
960         } else
961                 cfqg->saved_workload_slice = 0;
962
963         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
964                                         st->min_vdisktime);
965         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
966                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
967                         iops_mode(cfqd), cfqq->nr_sectors);
968         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
969         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
970 }
971
972 #ifdef CONFIG_CFQ_GROUP_IOSCHED
973 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
974 {
975         if (blkg)
976                 return container_of(blkg, struct cfq_group, blkg);
977         return NULL;
978 }
979
980 void
981 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
982 {
983         cfqg_of_blkg(blkg)->weight = weight;
984 }
985
986 static struct cfq_group *
987 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
988 {
989         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
990         struct cfq_group *cfqg = NULL;
991         void *key = cfqd;
992         int i, j;
993         struct cfq_rb_root *st;
994         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
995         unsigned int major, minor;
996
997         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
998         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
999                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1000                 cfqg->blkg.dev = MKDEV(major, minor);
1001                 goto done;
1002         }
1003         if (cfqg || !create)
1004                 goto done;
1005
1006         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1007         if (!cfqg)
1008                 goto done;
1009
1010         for_each_cfqg_st(cfqg, i, j, st)
1011                 *st = CFQ_RB_ROOT;
1012         RB_CLEAR_NODE(&cfqg->rb_node);
1013
1014         /*
1015          * Take the initial reference that will be released on destroy
1016          * This can be thought of a joint reference by cgroup and
1017          * elevator which will be dropped by either elevator exit
1018          * or cgroup deletion path depending on who is exiting first.
1019          */
1020         atomic_set(&cfqg->ref, 1);
1021
1022         /*
1023          * Add group onto cgroup list. It might happen that bdi->dev is
1024          * not initiliazed yet. Initialize this new group without major
1025          * and minor info and this info will be filled in once a new thread
1026          * comes for IO. See code above.
1027          */
1028         if (bdi->dev) {
1029                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1031                                         MKDEV(major, minor));
1032         } else
1033                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1034                                         0);
1035
1036         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1037
1038         /* Add group on cfqd list */
1039         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1040
1041 done:
1042         return cfqg;
1043 }
1044
1045 /*
1046  * Search for the cfq group current task belongs to. If create = 1, then also
1047  * create the cfq group if it does not exist. request_queue lock must be held.
1048  */
1049 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1050 {
1051         struct cgroup *cgroup;
1052         struct cfq_group *cfqg = NULL;
1053
1054         rcu_read_lock();
1055         cgroup = task_cgroup(current, blkio_subsys_id);
1056         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1057         if (!cfqg && create)
1058                 cfqg = &cfqd->root_group;
1059         rcu_read_unlock();
1060         return cfqg;
1061 }
1062
1063 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1064 {
1065         atomic_inc(&cfqg->ref);
1066         return cfqg;
1067 }
1068
1069 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1070 {
1071         /* Currently, all async queues are mapped to root group */
1072         if (!cfq_cfqq_sync(cfqq))
1073                 cfqg = &cfqq->cfqd->root_group;
1074
1075         cfqq->cfqg = cfqg;
1076         /* cfqq reference on cfqg */
1077         atomic_inc(&cfqq->cfqg->ref);
1078 }
1079
1080 static void cfq_put_cfqg(struct cfq_group *cfqg)
1081 {
1082         struct cfq_rb_root *st;
1083         int i, j;
1084
1085         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1086         if (!atomic_dec_and_test(&cfqg->ref))
1087                 return;
1088         for_each_cfqg_st(cfqg, i, j, st)
1089                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1090         kfree(cfqg);
1091 }
1092
1093 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1094 {
1095         /* Something wrong if we are trying to remove same group twice */
1096         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1097
1098         hlist_del_init(&cfqg->cfqd_node);
1099
1100         /*
1101          * Put the reference taken at the time of creation so that when all
1102          * queues are gone, group can be destroyed.
1103          */
1104         cfq_put_cfqg(cfqg);
1105 }
1106
1107 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1108 {
1109         struct hlist_node *pos, *n;
1110         struct cfq_group *cfqg;
1111
1112         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1113                 /*
1114                  * If cgroup removal path got to blk_group first and removed
1115                  * it from cgroup list, then it will take care of destroying
1116                  * cfqg also.
1117                  */
1118                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1119                         cfq_destroy_cfqg(cfqd, cfqg);
1120         }
1121 }
1122
1123 /*
1124  * Blk cgroup controller notification saying that blkio_group object is being
1125  * delinked as associated cgroup object is going away. That also means that
1126  * no new IO will come in this group. So get rid of this group as soon as
1127  * any pending IO in the group is finished.
1128  *
1129  * This function is called under rcu_read_lock(). key is the rcu protected
1130  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1131  * read lock.
1132  *
1133  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1134  * it should not be NULL as even if elevator was exiting, cgroup deltion
1135  * path got to it first.
1136  */
1137 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1138 {
1139         unsigned long  flags;
1140         struct cfq_data *cfqd = key;
1141
1142         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1143         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1144         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1145 }
1146
1147 #else /* GROUP_IOSCHED */
1148 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1149 {
1150         return &cfqd->root_group;
1151 }
1152
1153 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1154 {
1155         return cfqg;
1156 }
1157
1158 static inline void
1159 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1160         cfqq->cfqg = cfqg;
1161 }
1162
1163 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1164 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1165
1166 #endif /* GROUP_IOSCHED */
1167
1168 /*
1169  * The cfqd->service_trees holds all pending cfq_queue's that have
1170  * requests waiting to be processed. It is sorted in the order that
1171  * we will service the queues.
1172  */
1173 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1174                                  bool add_front)
1175 {
1176         struct rb_node **p, *parent;
1177         struct cfq_queue *__cfqq;
1178         unsigned long rb_key;
1179         struct cfq_rb_root *service_tree;
1180         int left;
1181         int new_cfqq = 1;
1182         int group_changed = 0;
1183
1184 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1185         if (!cfqd->cfq_group_isolation
1186             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1187             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1188                 /* Move this cfq to root group */
1189                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1190                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1191                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1192                 cfqq->orig_cfqg = cfqq->cfqg;
1193                 cfqq->cfqg = &cfqd->root_group;
1194                 atomic_inc(&cfqd->root_group.ref);
1195                 group_changed = 1;
1196         } else if (!cfqd->cfq_group_isolation
1197                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1198                 /* cfqq is sequential now needs to go to its original group */
1199                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1200                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1201                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1202                 cfq_put_cfqg(cfqq->cfqg);
1203                 cfqq->cfqg = cfqq->orig_cfqg;
1204                 cfqq->orig_cfqg = NULL;
1205                 group_changed = 1;
1206                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1207         }
1208 #endif
1209
1210         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1211                                                 cfqq_type(cfqq));
1212         if (cfq_class_idle(cfqq)) {
1213                 rb_key = CFQ_IDLE_DELAY;
1214                 parent = rb_last(&service_tree->rb);
1215                 if (parent && parent != &cfqq->rb_node) {
1216                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1217                         rb_key += __cfqq->rb_key;
1218                 } else
1219                         rb_key += jiffies;
1220         } else if (!add_front) {
1221                 /*
1222                  * Get our rb key offset. Subtract any residual slice
1223                  * value carried from last service. A negative resid
1224                  * count indicates slice overrun, and this should position
1225                  * the next service time further away in the tree.
1226                  */
1227                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1228                 rb_key -= cfqq->slice_resid;
1229                 cfqq->slice_resid = 0;
1230         } else {
1231                 rb_key = -HZ;
1232                 __cfqq = cfq_rb_first(service_tree);
1233                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1234         }
1235
1236         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1237                 new_cfqq = 0;
1238                 /*
1239                  * same position, nothing more to do
1240                  */
1241                 if (rb_key == cfqq->rb_key &&
1242                     cfqq->service_tree == service_tree)
1243                         return;
1244
1245                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1246                 cfqq->service_tree = NULL;
1247         }
1248
1249         left = 1;
1250         parent = NULL;
1251         cfqq->service_tree = service_tree;
1252         p = &service_tree->rb.rb_node;
1253         while (*p) {
1254                 struct rb_node **n;
1255
1256                 parent = *p;
1257                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1258
1259                 /*
1260                  * sort by key, that represents service time.
1261                  */
1262                 if (time_before(rb_key, __cfqq->rb_key))
1263                         n = &(*p)->rb_left;
1264                 else {
1265                         n = &(*p)->rb_right;
1266                         left = 0;
1267                 }
1268
1269                 p = n;
1270         }
1271
1272         if (left)
1273                 service_tree->left = &cfqq->rb_node;
1274
1275         cfqq->rb_key = rb_key;
1276         rb_link_node(&cfqq->rb_node, parent, p);
1277         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1278         service_tree->count++;
1279         if ((add_front || !new_cfqq) && !group_changed)
1280                 return;
1281         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1282 }
1283
1284 static struct cfq_queue *
1285 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1286                      sector_t sector, struct rb_node **ret_parent,
1287                      struct rb_node ***rb_link)
1288 {
1289         struct rb_node **p, *parent;
1290         struct cfq_queue *cfqq = NULL;
1291
1292         parent = NULL;
1293         p = &root->rb_node;
1294         while (*p) {
1295                 struct rb_node **n;
1296
1297                 parent = *p;
1298                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1299
1300                 /*
1301                  * Sort strictly based on sector.  Smallest to the left,
1302                  * largest to the right.
1303                  */
1304                 if (sector > blk_rq_pos(cfqq->next_rq))
1305                         n = &(*p)->rb_right;
1306                 else if (sector < blk_rq_pos(cfqq->next_rq))
1307                         n = &(*p)->rb_left;
1308                 else
1309                         break;
1310                 p = n;
1311                 cfqq = NULL;
1312         }
1313
1314         *ret_parent = parent;
1315         if (rb_link)
1316                 *rb_link = p;
1317         return cfqq;
1318 }
1319
1320 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1321 {
1322         struct rb_node **p, *parent;
1323         struct cfq_queue *__cfqq;
1324
1325         if (cfqq->p_root) {
1326                 rb_erase(&cfqq->p_node, cfqq->p_root);
1327                 cfqq->p_root = NULL;
1328         }
1329
1330         if (cfq_class_idle(cfqq))
1331                 return;
1332         if (!cfqq->next_rq)
1333                 return;
1334
1335         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1336         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1337                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1338         if (!__cfqq) {
1339                 rb_link_node(&cfqq->p_node, parent, p);
1340                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1341         } else
1342                 cfqq->p_root = NULL;
1343 }
1344
1345 /*
1346  * Update cfqq's position in the service tree.
1347  */
1348 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1349 {
1350         /*
1351          * Resorting requires the cfqq to be on the RR list already.
1352          */
1353         if (cfq_cfqq_on_rr(cfqq)) {
1354                 cfq_service_tree_add(cfqd, cfqq, 0);
1355                 cfq_prio_tree_add(cfqd, cfqq);
1356         }
1357 }
1358
1359 /*
1360  * add to busy list of queues for service, trying to be fair in ordering
1361  * the pending list according to last request service
1362  */
1363 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1364 {
1365         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1366         BUG_ON(cfq_cfqq_on_rr(cfqq));
1367         cfq_mark_cfqq_on_rr(cfqq);
1368         cfqd->busy_queues++;
1369
1370         cfq_resort_rr_list(cfqd, cfqq);
1371 }
1372
1373 /*
1374  * Called when the cfqq no longer has requests pending, remove it from
1375  * the service tree.
1376  */
1377 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1378 {
1379         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1380         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1381         cfq_clear_cfqq_on_rr(cfqq);
1382
1383         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1384                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1385                 cfqq->service_tree = NULL;
1386         }
1387         if (cfqq->p_root) {
1388                 rb_erase(&cfqq->p_node, cfqq->p_root);
1389                 cfqq->p_root = NULL;
1390         }
1391
1392         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1393         BUG_ON(!cfqd->busy_queues);
1394         cfqd->busy_queues--;
1395 }
1396
1397 /*
1398  * rb tree support functions
1399  */
1400 static void cfq_del_rq_rb(struct request *rq)
1401 {
1402         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1403         const int sync = rq_is_sync(rq);
1404
1405         BUG_ON(!cfqq->queued[sync]);
1406         cfqq->queued[sync]--;
1407
1408         elv_rb_del(&cfqq->sort_list, rq);
1409
1410         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1411                 /*
1412                  * Queue will be deleted from service tree when we actually
1413                  * expire it later. Right now just remove it from prio tree
1414                  * as it is empty.
1415                  */
1416                 if (cfqq->p_root) {
1417                         rb_erase(&cfqq->p_node, cfqq->p_root);
1418                         cfqq->p_root = NULL;
1419                 }
1420         }
1421 }
1422
1423 static void cfq_add_rq_rb(struct request *rq)
1424 {
1425         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1426         struct cfq_data *cfqd = cfqq->cfqd;
1427         struct request *__alias, *prev;
1428
1429         cfqq->queued[rq_is_sync(rq)]++;
1430
1431         /*
1432          * looks a little odd, but the first insert might return an alias.
1433          * if that happens, put the alias on the dispatch list
1434          */
1435         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1436                 cfq_dispatch_insert(cfqd->queue, __alias);
1437
1438         if (!cfq_cfqq_on_rr(cfqq))
1439                 cfq_add_cfqq_rr(cfqd, cfqq);
1440
1441         /*
1442          * check if this request is a better next-serve candidate
1443          */
1444         prev = cfqq->next_rq;
1445         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1446
1447         /*
1448          * adjust priority tree position, if ->next_rq changes
1449          */
1450         if (prev != cfqq->next_rq)
1451                 cfq_prio_tree_add(cfqd, cfqq);
1452
1453         BUG_ON(!cfqq->next_rq);
1454 }
1455
1456 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1457 {
1458         elv_rb_del(&cfqq->sort_list, rq);
1459         cfqq->queued[rq_is_sync(rq)]--;
1460         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1461                                         rq_data_dir(rq), rq_is_sync(rq));
1462         cfq_add_rq_rb(rq);
1463         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1464                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1465                         rq_is_sync(rq));
1466 }
1467
1468 static struct request *
1469 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1470 {
1471         struct task_struct *tsk = current;
1472         struct cfq_io_context *cic;
1473         struct cfq_queue *cfqq;
1474
1475         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1476         if (!cic)
1477                 return NULL;
1478
1479         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1480         if (cfqq) {
1481                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1482
1483                 return elv_rb_find(&cfqq->sort_list, sector);
1484         }
1485
1486         return NULL;
1487 }
1488
1489 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1490 {
1491         struct cfq_data *cfqd = q->elevator->elevator_data;
1492
1493         cfqd->rq_in_driver++;
1494         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1495                                                 cfqd->rq_in_driver);
1496
1497         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1498 }
1499
1500 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1501 {
1502         struct cfq_data *cfqd = q->elevator->elevator_data;
1503
1504         WARN_ON(!cfqd->rq_in_driver);
1505         cfqd->rq_in_driver--;
1506         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1507                                                 cfqd->rq_in_driver);
1508 }
1509
1510 static void cfq_remove_request(struct request *rq)
1511 {
1512         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1513
1514         if (cfqq->next_rq == rq)
1515                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1516
1517         list_del_init(&rq->queuelist);
1518         cfq_del_rq_rb(rq);
1519
1520         cfqq->cfqd->rq_queued--;
1521         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1522                                         rq_data_dir(rq), rq_is_sync(rq));
1523         if (rq->cmd_flags & REQ_META) {
1524                 WARN_ON(!cfqq->meta_pending);
1525                 cfqq->meta_pending--;
1526         }
1527 }
1528
1529 static int cfq_merge(struct request_queue *q, struct request **req,
1530                      struct bio *bio)
1531 {
1532         struct cfq_data *cfqd = q->elevator->elevator_data;
1533         struct request *__rq;
1534
1535         __rq = cfq_find_rq_fmerge(cfqd, bio);
1536         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1537                 *req = __rq;
1538                 return ELEVATOR_FRONT_MERGE;
1539         }
1540
1541         return ELEVATOR_NO_MERGE;
1542 }
1543
1544 static void cfq_merged_request(struct request_queue *q, struct request *req,
1545                                int type)
1546 {
1547         if (type == ELEVATOR_FRONT_MERGE) {
1548                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1549
1550                 cfq_reposition_rq_rb(cfqq, req);
1551         }
1552 }
1553
1554 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1555                                 struct bio *bio)
1556 {
1557         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1558                                         bio_data_dir(bio), cfq_bio_sync(bio));
1559 }
1560
1561 static void
1562 cfq_merged_requests(struct request_queue *q, struct request *rq,
1563                     struct request *next)
1564 {
1565         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1566         /*
1567          * reposition in fifo if next is older than rq
1568          */
1569         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1570             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1571                 list_move(&rq->queuelist, &next->queuelist);
1572                 rq_set_fifo_time(rq, rq_fifo_time(next));
1573         }
1574
1575         if (cfqq->next_rq == next)
1576                 cfqq->next_rq = rq;
1577         cfq_remove_request(next);
1578         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1579                                         rq_data_dir(next), rq_is_sync(next));
1580 }
1581
1582 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1583                            struct bio *bio)
1584 {
1585         struct cfq_data *cfqd = q->elevator->elevator_data;
1586         struct cfq_io_context *cic;
1587         struct cfq_queue *cfqq;
1588
1589         /*
1590          * Disallow merge of a sync bio into an async request.
1591          */
1592         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1593                 return false;
1594
1595         /*
1596          * Lookup the cfqq that this bio will be queued with. Allow
1597          * merge only if rq is queued there.
1598          */
1599         cic = cfq_cic_lookup(cfqd, current->io_context);
1600         if (!cic)
1601                 return false;
1602
1603         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1604         return cfqq == RQ_CFQQ(rq);
1605 }
1606
1607 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1608 {
1609         del_timer(&cfqd->idle_slice_timer);
1610         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1611 }
1612
1613 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1614                                    struct cfq_queue *cfqq)
1615 {
1616         if (cfqq) {
1617                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1618                                 cfqd->serving_prio, cfqd->serving_type);
1619                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1620                 cfqq->slice_start = 0;
1621                 cfqq->dispatch_start = jiffies;
1622                 cfqq->allocated_slice = 0;
1623                 cfqq->slice_end = 0;
1624                 cfqq->slice_dispatch = 0;
1625                 cfqq->nr_sectors = 0;
1626
1627                 cfq_clear_cfqq_wait_request(cfqq);
1628                 cfq_clear_cfqq_must_dispatch(cfqq);
1629                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1630                 cfq_clear_cfqq_fifo_expire(cfqq);
1631                 cfq_mark_cfqq_slice_new(cfqq);
1632
1633                 cfq_del_timer(cfqd, cfqq);
1634         }
1635
1636         cfqd->active_queue = cfqq;
1637 }
1638
1639 /*
1640  * current cfqq expired its slice (or was too idle), select new one
1641  */
1642 static void
1643 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1644                     bool timed_out)
1645 {
1646         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1647
1648         if (cfq_cfqq_wait_request(cfqq))
1649                 cfq_del_timer(cfqd, cfqq);
1650
1651         cfq_clear_cfqq_wait_request(cfqq);
1652         cfq_clear_cfqq_wait_busy(cfqq);
1653
1654         /*
1655          * If this cfqq is shared between multiple processes, check to
1656          * make sure that those processes are still issuing I/Os within
1657          * the mean seek distance.  If not, it may be time to break the
1658          * queues apart again.
1659          */
1660         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1661                 cfq_mark_cfqq_split_coop(cfqq);
1662
1663         /*
1664          * store what was left of this slice, if the queue idled/timed out
1665          */
1666         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1667                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1668                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1669         }
1670
1671         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1672
1673         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1674                 cfq_del_cfqq_rr(cfqd, cfqq);
1675
1676         cfq_resort_rr_list(cfqd, cfqq);
1677
1678         if (cfqq == cfqd->active_queue)
1679                 cfqd->active_queue = NULL;
1680
1681         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1682                 cfqd->grp_service_tree.active = NULL;
1683
1684         if (cfqd->active_cic) {
1685                 put_io_context(cfqd->active_cic->ioc);
1686                 cfqd->active_cic = NULL;
1687         }
1688 }
1689
1690 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1691 {
1692         struct cfq_queue *cfqq = cfqd->active_queue;
1693
1694         if (cfqq)
1695                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1696 }
1697
1698 /*
1699  * Get next queue for service. Unless we have a queue preemption,
1700  * we'll simply select the first cfqq in the service tree.
1701  */
1702 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1703 {
1704         struct cfq_rb_root *service_tree =
1705                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1706                                         cfqd->serving_type);
1707
1708         if (!cfqd->rq_queued)
1709                 return NULL;
1710
1711         /* There is nothing to dispatch */
1712         if (!service_tree)
1713                 return NULL;
1714         if (RB_EMPTY_ROOT(&service_tree->rb))
1715                 return NULL;
1716         return cfq_rb_first(service_tree);
1717 }
1718
1719 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1720 {
1721         struct cfq_group *cfqg;
1722         struct cfq_queue *cfqq;
1723         int i, j;
1724         struct cfq_rb_root *st;
1725
1726         if (!cfqd->rq_queued)
1727                 return NULL;
1728
1729         cfqg = cfq_get_next_cfqg(cfqd);
1730         if (!cfqg)
1731                 return NULL;
1732
1733         for_each_cfqg_st(cfqg, i, j, st)
1734                 if ((cfqq = cfq_rb_first(st)) != NULL)
1735                         return cfqq;
1736         return NULL;
1737 }
1738
1739 /*
1740  * Get and set a new active queue for service.
1741  */
1742 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1743                                               struct cfq_queue *cfqq)
1744 {
1745         if (!cfqq)
1746                 cfqq = cfq_get_next_queue(cfqd);
1747
1748         __cfq_set_active_queue(cfqd, cfqq);
1749         return cfqq;
1750 }
1751
1752 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1753                                           struct request *rq)
1754 {
1755         if (blk_rq_pos(rq) >= cfqd->last_position)
1756                 return blk_rq_pos(rq) - cfqd->last_position;
1757         else
1758                 return cfqd->last_position - blk_rq_pos(rq);
1759 }
1760
1761 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1762                                struct request *rq)
1763 {
1764         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1765 }
1766
1767 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1768                                     struct cfq_queue *cur_cfqq)
1769 {
1770         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1771         struct rb_node *parent, *node;
1772         struct cfq_queue *__cfqq;
1773         sector_t sector = cfqd->last_position;
1774
1775         if (RB_EMPTY_ROOT(root))
1776                 return NULL;
1777
1778         /*
1779          * First, if we find a request starting at the end of the last
1780          * request, choose it.
1781          */
1782         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1783         if (__cfqq)
1784                 return __cfqq;
1785
1786         /*
1787          * If the exact sector wasn't found, the parent of the NULL leaf
1788          * will contain the closest sector.
1789          */
1790         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1791         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1792                 return __cfqq;
1793
1794         if (blk_rq_pos(__cfqq->next_rq) < sector)
1795                 node = rb_next(&__cfqq->p_node);
1796         else
1797                 node = rb_prev(&__cfqq->p_node);
1798         if (!node)
1799                 return NULL;
1800
1801         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1802         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1803                 return __cfqq;
1804
1805         return NULL;
1806 }
1807
1808 /*
1809  * cfqd - obvious
1810  * cur_cfqq - passed in so that we don't decide that the current queue is
1811  *            closely cooperating with itself.
1812  *
1813  * So, basically we're assuming that that cur_cfqq has dispatched at least
1814  * one request, and that cfqd->last_position reflects a position on the disk
1815  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1816  * assumption.
1817  */
1818 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1819                                               struct cfq_queue *cur_cfqq)
1820 {
1821         struct cfq_queue *cfqq;
1822
1823         if (cfq_class_idle(cur_cfqq))
1824                 return NULL;
1825         if (!cfq_cfqq_sync(cur_cfqq))
1826                 return NULL;
1827         if (CFQQ_SEEKY(cur_cfqq))
1828                 return NULL;
1829
1830         /*
1831          * Don't search priority tree if it's the only queue in the group.
1832          */
1833         if (cur_cfqq->cfqg->nr_cfqq == 1)
1834                 return NULL;
1835
1836         /*
1837          * We should notice if some of the queues are cooperating, eg
1838          * working closely on the same area of the disk. In that case,
1839          * we can group them together and don't waste time idling.
1840          */
1841         cfqq = cfqq_close(cfqd, cur_cfqq);
1842         if (!cfqq)
1843                 return NULL;
1844
1845         /* If new queue belongs to different cfq_group, don't choose it */
1846         if (cur_cfqq->cfqg != cfqq->cfqg)
1847                 return NULL;
1848
1849         /*
1850          * It only makes sense to merge sync queues.
1851          */
1852         if (!cfq_cfqq_sync(cfqq))
1853                 return NULL;
1854         if (CFQQ_SEEKY(cfqq))
1855                 return NULL;
1856
1857         /*
1858          * Do not merge queues of different priority classes
1859          */
1860         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1861                 return NULL;
1862
1863         return cfqq;
1864 }
1865
1866 /*
1867  * Determine whether we should enforce idle window for this queue.
1868  */
1869
1870 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1871 {
1872         enum wl_prio_t prio = cfqq_prio(cfqq);
1873         struct cfq_rb_root *service_tree = cfqq->service_tree;
1874
1875         BUG_ON(!service_tree);
1876         BUG_ON(!service_tree->count);
1877
1878         if (!cfqd->cfq_slice_idle)
1879                 return false;
1880
1881         /* We never do for idle class queues. */
1882         if (prio == IDLE_WORKLOAD)
1883                 return false;
1884
1885         /* We do for queues that were marked with idle window flag. */
1886         if (cfq_cfqq_idle_window(cfqq) &&
1887            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1888                 return true;
1889
1890         /*
1891          * Otherwise, we do only if they are the last ones
1892          * in their service tree.
1893          */
1894         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1895                 return true;
1896         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1897                         service_tree->count);
1898         return false;
1899 }
1900
1901 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1902 {
1903         struct cfq_queue *cfqq = cfqd->active_queue;
1904         struct cfq_io_context *cic;
1905         unsigned long sl, group_idle = 0;
1906
1907         /*
1908          * SSD device without seek penalty, disable idling. But only do so
1909          * for devices that support queuing, otherwise we still have a problem
1910          * with sync vs async workloads.
1911          */
1912         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1913                 return;
1914
1915         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1916         WARN_ON(cfq_cfqq_slice_new(cfqq));
1917
1918         /*
1919          * idle is disabled, either manually or by past process history
1920          */
1921         if (!cfq_should_idle(cfqd, cfqq)) {
1922                 /* no queue idling. Check for group idling */
1923                 if (cfqd->cfq_group_idle)
1924                         group_idle = cfqd->cfq_group_idle;
1925                 else
1926                         return;
1927         }
1928
1929         /*
1930          * still active requests from this queue, don't idle
1931          */
1932         if (cfqq->dispatched)
1933                 return;
1934
1935         /*
1936          * task has exited, don't wait
1937          */
1938         cic = cfqd->active_cic;
1939         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1940                 return;
1941
1942         /*
1943          * If our average think time is larger than the remaining time
1944          * slice, then don't idle. This avoids overrunning the allotted
1945          * time slice.
1946          */
1947         if (sample_valid(cic->ttime_samples) &&
1948             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1949                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1950                                 cic->ttime_mean);
1951                 return;
1952         }
1953
1954         /* There are other queues in the group, don't do group idle */
1955         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1956                 return;
1957
1958         cfq_mark_cfqq_wait_request(cfqq);
1959
1960         if (group_idle)
1961                 sl = cfqd->cfq_group_idle;
1962         else
1963                 sl = cfqd->cfq_slice_idle;
1964
1965         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1966         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1967         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1968                         group_idle ? 1 : 0);
1969 }
1970
1971 /*
1972  * Move request from internal lists to the request queue dispatch list.
1973  */
1974 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1975 {
1976         struct cfq_data *cfqd = q->elevator->elevator_data;
1977         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1978
1979         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1980
1981         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1982         cfq_remove_request(rq);
1983         cfqq->dispatched++;
1984         (RQ_CFQG(rq))->dispatched++;
1985         elv_dispatch_sort(q, rq);
1986
1987         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1988         cfqq->nr_sectors += blk_rq_sectors(rq);
1989         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1990                                         rq_data_dir(rq), rq_is_sync(rq));
1991 }
1992
1993 /*
1994  * return expired entry, or NULL to just start from scratch in rbtree
1995  */
1996 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1997 {
1998         struct request *rq = NULL;
1999
2000         if (cfq_cfqq_fifo_expire(cfqq))
2001                 return NULL;
2002
2003         cfq_mark_cfqq_fifo_expire(cfqq);
2004
2005         if (list_empty(&cfqq->fifo))
2006                 return NULL;
2007
2008         rq = rq_entry_fifo(cfqq->fifo.next);
2009         if (time_before(jiffies, rq_fifo_time(rq)))
2010                 rq = NULL;
2011
2012         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2013         return rq;
2014 }
2015
2016 static inline int
2017 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2018 {
2019         const int base_rq = cfqd->cfq_slice_async_rq;
2020
2021         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2022
2023         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2024 }
2025
2026 /*
2027  * Must be called with the queue_lock held.
2028  */
2029 static int cfqq_process_refs(struct cfq_queue *cfqq)
2030 {
2031         int process_refs, io_refs;
2032
2033         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2034         process_refs = atomic_read(&cfqq->ref) - io_refs;
2035         BUG_ON(process_refs < 0);
2036         return process_refs;
2037 }
2038
2039 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2040 {
2041         int process_refs, new_process_refs;
2042         struct cfq_queue *__cfqq;
2043
2044         /*
2045          * If there are no process references on the new_cfqq, then it is
2046          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2047          * chain may have dropped their last reference (not just their
2048          * last process reference).
2049          */
2050         if (!cfqq_process_refs(new_cfqq))
2051                 return;
2052
2053         /* Avoid a circular list and skip interim queue merges */
2054         while ((__cfqq = new_cfqq->new_cfqq)) {
2055                 if (__cfqq == cfqq)
2056                         return;
2057                 new_cfqq = __cfqq;
2058         }
2059
2060         process_refs = cfqq_process_refs(cfqq);
2061         new_process_refs = cfqq_process_refs(new_cfqq);
2062         /*
2063          * If the process for the cfqq has gone away, there is no
2064          * sense in merging the queues.
2065          */
2066         if (process_refs == 0 || new_process_refs == 0)
2067                 return;
2068
2069         /*
2070          * Merge in the direction of the lesser amount of work.
2071          */
2072         if (new_process_refs >= process_refs) {
2073                 cfqq->new_cfqq = new_cfqq;
2074                 atomic_add(process_refs, &new_cfqq->ref);
2075         } else {
2076                 new_cfqq->new_cfqq = cfqq;
2077                 atomic_add(new_process_refs, &cfqq->ref);
2078         }
2079 }
2080
2081 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2082                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2083 {
2084         struct cfq_queue *queue;
2085         int i;
2086         bool key_valid = false;
2087         unsigned long lowest_key = 0;
2088         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2089
2090         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2091                 /* select the one with lowest rb_key */
2092                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2093                 if (queue &&
2094                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2095                         lowest_key = queue->rb_key;
2096                         cur_best = i;
2097                         key_valid = true;
2098                 }
2099         }
2100
2101         return cur_best;
2102 }
2103
2104 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2105 {
2106         unsigned slice;
2107         unsigned count;
2108         struct cfq_rb_root *st;
2109         unsigned group_slice;
2110
2111         if (!cfqg) {
2112                 cfqd->serving_prio = IDLE_WORKLOAD;
2113                 cfqd->workload_expires = jiffies + 1;
2114                 return;
2115         }
2116
2117         /* Choose next priority. RT > BE > IDLE */
2118         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2119                 cfqd->serving_prio = RT_WORKLOAD;
2120         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2121                 cfqd->serving_prio = BE_WORKLOAD;
2122         else {
2123                 cfqd->serving_prio = IDLE_WORKLOAD;
2124                 cfqd->workload_expires = jiffies + 1;
2125                 return;
2126         }
2127
2128         /*
2129          * For RT and BE, we have to choose also the type
2130          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2131          * expiration time
2132          */
2133         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2134         count = st->count;
2135
2136         /*
2137          * check workload expiration, and that we still have other queues ready
2138          */
2139         if (count && !time_after(jiffies, cfqd->workload_expires))
2140                 return;
2141
2142         /* otherwise select new workload type */
2143         cfqd->serving_type =
2144                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2145         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2146         count = st->count;
2147
2148         /*
2149          * the workload slice is computed as a fraction of target latency
2150          * proportional to the number of queues in that workload, over
2151          * all the queues in the same priority class
2152          */
2153         group_slice = cfq_group_slice(cfqd, cfqg);
2154
2155         slice = group_slice * count /
2156                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2157                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2158
2159         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2160                 unsigned int tmp;
2161
2162                 /*
2163                  * Async queues are currently system wide. Just taking
2164                  * proportion of queues with-in same group will lead to higher
2165                  * async ratio system wide as generally root group is going
2166                  * to have higher weight. A more accurate thing would be to
2167                  * calculate system wide asnc/sync ratio.
2168                  */
2169                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2170                 tmp = tmp/cfqd->busy_queues;
2171                 slice = min_t(unsigned, slice, tmp);
2172
2173                 /* async workload slice is scaled down according to
2174                  * the sync/async slice ratio. */
2175                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2176         } else
2177                 /* sync workload slice is at least 2 * cfq_slice_idle */
2178                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2179
2180         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2181         cfq_log(cfqd, "workload slice:%d", slice);
2182         cfqd->workload_expires = jiffies + slice;
2183         cfqd->noidle_tree_requires_idle = false;
2184 }
2185
2186 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2187 {
2188         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2189         struct cfq_group *cfqg;
2190
2191         if (RB_EMPTY_ROOT(&st->rb))
2192                 return NULL;
2193         cfqg = cfq_rb_first_group(st);
2194         st->active = &cfqg->rb_node;
2195         update_min_vdisktime(st);
2196         return cfqg;
2197 }
2198
2199 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2200 {
2201         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2202
2203         cfqd->serving_group = cfqg;
2204
2205         /* Restore the workload type data */
2206         if (cfqg->saved_workload_slice) {
2207                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2208                 cfqd->serving_type = cfqg->saved_workload;
2209                 cfqd->serving_prio = cfqg->saved_serving_prio;
2210         } else
2211                 cfqd->workload_expires = jiffies - 1;
2212
2213         choose_service_tree(cfqd, cfqg);
2214 }
2215
2216 /*
2217  * Select a queue for service. If we have a current active queue,
2218  * check whether to continue servicing it, or retrieve and set a new one.
2219  */
2220 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2221 {
2222         struct cfq_queue *cfqq, *new_cfqq = NULL;
2223
2224         cfqq = cfqd->active_queue;
2225         if (!cfqq)
2226                 goto new_queue;
2227
2228         if (!cfqd->rq_queued)
2229                 return NULL;
2230
2231         /*
2232          * We were waiting for group to get backlogged. Expire the queue
2233          */
2234         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2235                 goto expire;
2236
2237         /*
2238          * The active queue has run out of time, expire it and select new.
2239          */
2240         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2241                 /*
2242                  * If slice had not expired at the completion of last request
2243                  * we might not have turned on wait_busy flag. Don't expire
2244                  * the queue yet. Allow the group to get backlogged.
2245                  *
2246                  * The very fact that we have used the slice, that means we
2247                  * have been idling all along on this queue and it should be
2248                  * ok to wait for this request to complete.
2249                  */
2250                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2251                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2252                         cfqq = NULL;
2253                         goto keep_queue;
2254                 } else
2255                         goto check_group_idle;
2256         }
2257
2258         /*
2259          * The active queue has requests and isn't expired, allow it to
2260          * dispatch.
2261          */
2262         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2263                 goto keep_queue;
2264
2265         /*
2266          * If another queue has a request waiting within our mean seek
2267          * distance, let it run.  The expire code will check for close
2268          * cooperators and put the close queue at the front of the service
2269          * tree.  If possible, merge the expiring queue with the new cfqq.
2270          */
2271         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2272         if (new_cfqq) {
2273                 if (!cfqq->new_cfqq)
2274                         cfq_setup_merge(cfqq, new_cfqq);
2275                 goto expire;
2276         }
2277
2278         /*
2279          * No requests pending. If the active queue still has requests in
2280          * flight or is idling for a new request, allow either of these
2281          * conditions to happen (or time out) before selecting a new queue.
2282          */
2283         if (timer_pending(&cfqd->idle_slice_timer)) {
2284                 cfqq = NULL;
2285                 goto keep_queue;
2286         }
2287
2288         /*
2289          * This is a deep seek queue, but the device is much faster than
2290          * the queue can deliver, don't idle
2291          **/
2292         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2293             (cfq_cfqq_slice_new(cfqq) ||
2294             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2295                 cfq_clear_cfqq_deep(cfqq);
2296                 cfq_clear_cfqq_idle_window(cfqq);
2297         }
2298
2299         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2300                 cfqq = NULL;
2301                 goto keep_queue;
2302         }
2303
2304         /*
2305          * If group idle is enabled and there are requests dispatched from
2306          * this group, wait for requests to complete.
2307          */
2308 check_group_idle:
2309         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2310             && cfqq->cfqg->dispatched) {
2311                 cfqq = NULL;
2312                 goto keep_queue;
2313         }
2314
2315 expire:
2316         cfq_slice_expired(cfqd, 0);
2317 new_queue:
2318         /*
2319          * Current queue expired. Check if we have to switch to a new
2320          * service tree
2321          */
2322         if (!new_cfqq)
2323                 cfq_choose_cfqg(cfqd);
2324
2325         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2326 keep_queue:
2327         return cfqq;
2328 }
2329
2330 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2331 {
2332         int dispatched = 0;
2333
2334         while (cfqq->next_rq) {
2335                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2336                 dispatched++;
2337         }
2338
2339         BUG_ON(!list_empty(&cfqq->fifo));
2340
2341         /* By default cfqq is not expired if it is empty. Do it explicitly */
2342         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2343         return dispatched;
2344 }
2345
2346 /*
2347  * Drain our current requests. Used for barriers and when switching
2348  * io schedulers on-the-fly.
2349  */
2350 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2351 {
2352         struct cfq_queue *cfqq;
2353         int dispatched = 0;
2354
2355         /* Expire the timeslice of the current active queue first */
2356         cfq_slice_expired(cfqd, 0);
2357         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2358                 __cfq_set_active_queue(cfqd, cfqq);
2359                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2360         }
2361
2362         BUG_ON(cfqd->busy_queues);
2363
2364         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2365         return dispatched;
2366 }
2367
2368 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2369         struct cfq_queue *cfqq)
2370 {
2371         /* the queue hasn't finished any request, can't estimate */
2372         if (cfq_cfqq_slice_new(cfqq))
2373                 return true;
2374         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2375                 cfqq->slice_end))
2376                 return true;
2377
2378         return false;
2379 }
2380
2381 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2382 {
2383         unsigned int max_dispatch;
2384
2385         /*
2386          * Drain async requests before we start sync IO
2387          */
2388         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2389                 return false;
2390
2391         /*
2392          * If this is an async queue and we have sync IO in flight, let it wait
2393          */
2394         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2395                 return false;
2396
2397         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2398         if (cfq_class_idle(cfqq))
2399                 max_dispatch = 1;
2400
2401         /*
2402          * Does this cfqq already have too much IO in flight?
2403          */
2404         if (cfqq->dispatched >= max_dispatch) {
2405                 /*
2406                  * idle queue must always only have a single IO in flight
2407                  */
2408                 if (cfq_class_idle(cfqq))
2409                         return false;
2410
2411                 /*
2412                  * We have other queues, don't allow more IO from this one
2413                  */
2414                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2415                         return false;
2416
2417                 /*
2418                  * Sole queue user, no limit
2419                  */
2420                 if (cfqd->busy_queues == 1)
2421                         max_dispatch = -1;
2422                 else
2423                         /*
2424                          * Normally we start throttling cfqq when cfq_quantum/2
2425                          * requests have been dispatched. But we can drive
2426                          * deeper queue depths at the beginning of slice
2427                          * subjected to upper limit of cfq_quantum.
2428                          * */
2429                         max_dispatch = cfqd->cfq_quantum;
2430         }
2431
2432         /*
2433          * Async queues must wait a bit before being allowed dispatch.
2434          * We also ramp up the dispatch depth gradually for async IO,
2435          * based on the last sync IO we serviced
2436          */
2437         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2438                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2439                 unsigned int depth;
2440
2441                 depth = last_sync / cfqd->cfq_slice[1];
2442                 if (!depth && !cfqq->dispatched)
2443                         depth = 1;
2444                 if (depth < max_dispatch)
2445                         max_dispatch = depth;
2446         }
2447
2448         /*
2449          * If we're below the current max, allow a dispatch
2450          */
2451         return cfqq->dispatched < max_dispatch;
2452 }
2453
2454 /*
2455  * Dispatch a request from cfqq, moving them to the request queue
2456  * dispatch list.
2457  */
2458 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2459 {
2460         struct request *rq;
2461
2462         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2463
2464         if (!cfq_may_dispatch(cfqd, cfqq))
2465                 return false;
2466
2467         /*
2468          * follow expired path, else get first next available
2469          */
2470         rq = cfq_check_fifo(cfqq);
2471         if (!rq)
2472                 rq = cfqq->next_rq;
2473
2474         /*
2475          * insert request into driver dispatch list
2476          */
2477         cfq_dispatch_insert(cfqd->queue, rq);
2478
2479         if (!cfqd->active_cic) {
2480                 struct cfq_io_context *cic = RQ_CIC(rq);
2481
2482                 atomic_long_inc(&cic->ioc->refcount);
2483                 cfqd->active_cic = cic;
2484         }
2485
2486         return true;
2487 }
2488
2489 /*
2490  * Find the cfqq that we need to service and move a request from that to the
2491  * dispatch list
2492  */
2493 static int cfq_dispatch_requests(struct request_queue *q, int force)
2494 {
2495         struct cfq_data *cfqd = q->elevator->elevator_data;
2496         struct cfq_queue *cfqq;
2497
2498         if (!cfqd->busy_queues)
2499                 return 0;
2500
2501         if (unlikely(force))
2502                 return cfq_forced_dispatch(cfqd);
2503
2504         cfqq = cfq_select_queue(cfqd);
2505         if (!cfqq)
2506                 return 0;
2507
2508         /*
2509          * Dispatch a request from this cfqq, if it is allowed
2510          */
2511         if (!cfq_dispatch_request(cfqd, cfqq))
2512                 return 0;
2513
2514         cfqq->slice_dispatch++;
2515         cfq_clear_cfqq_must_dispatch(cfqq);
2516
2517         /*
2518          * expire an async queue immediately if it has used up its slice. idle
2519          * queue always expire after 1 dispatch round.
2520          */
2521         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2522             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2523             cfq_class_idle(cfqq))) {
2524                 cfqq->slice_end = jiffies + 1;
2525                 cfq_slice_expired(cfqd, 0);
2526         }
2527
2528         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2529         return 1;
2530 }
2531
2532 /*
2533  * task holds one reference to the queue, dropped when task exits. each rq
2534  * in-flight on this queue also holds a reference, dropped when rq is freed.
2535  *
2536  * Each cfq queue took a reference on the parent group. Drop it now.
2537  * queue lock must be held here.
2538  */
2539 static void cfq_put_queue(struct cfq_queue *cfqq)
2540 {
2541         struct cfq_data *cfqd = cfqq->cfqd;
2542         struct cfq_group *cfqg, *orig_cfqg;
2543
2544         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2545
2546         if (!atomic_dec_and_test(&cfqq->ref))
2547                 return;
2548
2549         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2550         BUG_ON(rb_first(&cfqq->sort_list));
2551         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2552         cfqg = cfqq->cfqg;
2553         orig_cfqg = cfqq->orig_cfqg;
2554
2555         if (unlikely(cfqd->active_queue == cfqq)) {
2556                 __cfq_slice_expired(cfqd, cfqq, 0);
2557                 cfq_schedule_dispatch(cfqd);
2558         }
2559
2560         BUG_ON(cfq_cfqq_on_rr(cfqq));
2561         kmem_cache_free(cfq_pool, cfqq);
2562         cfq_put_cfqg(cfqg);
2563         if (orig_cfqg)
2564                 cfq_put_cfqg(orig_cfqg);
2565 }
2566
2567 /*
2568  * Must always be called with the rcu_read_lock() held
2569  */
2570 static void
2571 __call_for_each_cic(struct io_context *ioc,
2572                     void (*func)(struct io_context *, struct cfq_io_context *))
2573 {
2574         struct cfq_io_context *cic;
2575         struct hlist_node *n;
2576
2577         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2578                 func(ioc, cic);
2579 }
2580
2581 /*
2582  * Call func for each cic attached to this ioc.
2583  */
2584 static void
2585 call_for_each_cic(struct io_context *ioc,
2586                   void (*func)(struct io_context *, struct cfq_io_context *))
2587 {
2588         rcu_read_lock();
2589         __call_for_each_cic(ioc, func);
2590         rcu_read_unlock();
2591 }
2592
2593 static void cfq_cic_free_rcu(struct rcu_head *head)
2594 {
2595         struct cfq_io_context *cic;
2596
2597         cic = container_of(head, struct cfq_io_context, rcu_head);
2598
2599         kmem_cache_free(cfq_ioc_pool, cic);
2600         elv_ioc_count_dec(cfq_ioc_count);
2601
2602         if (ioc_gone) {
2603                 /*
2604                  * CFQ scheduler is exiting, grab exit lock and check
2605                  * the pending io context count. If it hits zero,
2606                  * complete ioc_gone and set it back to NULL
2607                  */
2608                 spin_lock(&ioc_gone_lock);
2609                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2610                         complete(ioc_gone);
2611                         ioc_gone = NULL;
2612                 }
2613                 spin_unlock(&ioc_gone_lock);
2614         }
2615 }
2616
2617 static void cfq_cic_free(struct cfq_io_context *cic)
2618 {
2619         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2620 }
2621
2622 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2623 {
2624         unsigned long flags;
2625         unsigned long dead_key = (unsigned long) cic->key;
2626
2627         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2628
2629         spin_lock_irqsave(&ioc->lock, flags);
2630         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2631         hlist_del_rcu(&cic->cic_list);
2632         spin_unlock_irqrestore(&ioc->lock, flags);
2633
2634         cfq_cic_free(cic);
2635 }
2636
2637 /*
2638  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2639  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2640  * and ->trim() which is called with the task lock held
2641  */
2642 static void cfq_free_io_context(struct io_context *ioc)
2643 {
2644         /*
2645          * ioc->refcount is zero here, or we are called from elv_unregister(),
2646          * so no more cic's are allowed to be linked into this ioc.  So it
2647          * should be ok to iterate over the known list, we will see all cic's
2648          * since no new ones are added.
2649          */
2650         __call_for_each_cic(ioc, cic_free_func);
2651 }
2652
2653 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2654 {
2655         struct cfq_queue *__cfqq, *next;
2656
2657         /*
2658          * If this queue was scheduled to merge with another queue, be
2659          * sure to drop the reference taken on that queue (and others in
2660          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2661          */
2662         __cfqq = cfqq->new_cfqq;
2663         while (__cfqq) {
2664                 if (__cfqq == cfqq) {
2665                         WARN(1, "cfqq->new_cfqq loop detected\n");
2666                         break;
2667                 }
2668                 next = __cfqq->new_cfqq;
2669                 cfq_put_queue(__cfqq);
2670                 __cfqq = next;
2671         }
2672 }
2673
2674 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2675 {
2676         if (unlikely(cfqq == cfqd->active_queue)) {
2677                 __cfq_slice_expired(cfqd, cfqq, 0);
2678                 cfq_schedule_dispatch(cfqd);
2679         }
2680
2681         cfq_put_cooperator(cfqq);
2682
2683         cfq_put_queue(cfqq);
2684 }
2685
2686 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2687                                          struct cfq_io_context *cic)
2688 {
2689         struct io_context *ioc = cic->ioc;
2690
2691         list_del_init(&cic->queue_list);
2692
2693         /*
2694          * Make sure dead mark is seen for dead queues
2695          */
2696         smp_wmb();
2697         cic->key = cfqd_dead_key(cfqd);
2698
2699         if (ioc->ioc_data == cic)
2700                 rcu_assign_pointer(ioc->ioc_data, NULL);
2701
2702         if (cic->cfqq[BLK_RW_ASYNC]) {
2703                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2704                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2705         }
2706
2707         if (cic->cfqq[BLK_RW_SYNC]) {
2708                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2709                 cic->cfqq[BLK_RW_SYNC] = NULL;
2710         }
2711 }
2712
2713 static void cfq_exit_single_io_context(struct io_context *ioc,
2714                                        struct cfq_io_context *cic)
2715 {
2716         struct cfq_data *cfqd = cic_to_cfqd(cic);
2717
2718         if (cfqd) {
2719                 struct request_queue *q = cfqd->queue;
2720                 unsigned long flags;
2721
2722                 spin_lock_irqsave(q->queue_lock, flags);
2723
2724                 /*
2725                  * Ensure we get a fresh copy of the ->key to prevent
2726                  * race between exiting task and queue
2727                  */
2728                 smp_read_barrier_depends();
2729                 if (cic->key == cfqd)
2730                         __cfq_exit_single_io_context(cfqd, cic);
2731
2732                 spin_unlock_irqrestore(q->queue_lock, flags);
2733         }
2734 }
2735
2736 /*
2737  * The process that ioc belongs to has exited, we need to clean up
2738  * and put the internal structures we have that belongs to that process.
2739  */
2740 static void cfq_exit_io_context(struct io_context *ioc)
2741 {
2742         call_for_each_cic(ioc, cfq_exit_single_io_context);
2743 }
2744
2745 static struct cfq_io_context *
2746 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2747 {
2748         struct cfq_io_context *cic;
2749
2750         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2751                                                         cfqd->queue->node);
2752         if (cic) {
2753                 cic->last_end_request = jiffies;
2754                 INIT_LIST_HEAD(&cic->queue_list);
2755                 INIT_HLIST_NODE(&cic->cic_list);
2756                 cic->dtor = cfq_free_io_context;
2757                 cic->exit = cfq_exit_io_context;
2758                 elv_ioc_count_inc(cfq_ioc_count);
2759         }
2760
2761         return cic;
2762 }
2763
2764 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2765 {
2766         struct task_struct *tsk = current;
2767         int ioprio_class;
2768
2769         if (!cfq_cfqq_prio_changed(cfqq))
2770                 return;
2771
2772         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2773         switch (ioprio_class) {
2774         default:
2775                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2776         case IOPRIO_CLASS_NONE:
2777                 /*
2778                  * no prio set, inherit CPU scheduling settings
2779                  */
2780                 cfqq->ioprio = task_nice_ioprio(tsk);
2781                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2782                 break;
2783         case IOPRIO_CLASS_RT:
2784                 cfqq->ioprio = task_ioprio(ioc);
2785                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2786                 break;
2787         case IOPRIO_CLASS_BE:
2788                 cfqq->ioprio = task_ioprio(ioc);
2789                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2790                 break;
2791         case IOPRIO_CLASS_IDLE:
2792                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2793                 cfqq->ioprio = 7;
2794                 cfq_clear_cfqq_idle_window(cfqq);
2795                 break;
2796         }
2797
2798         /*
2799          * keep track of original prio settings in case we have to temporarily
2800          * elevate the priority of this queue
2801          */
2802         cfqq->org_ioprio = cfqq->ioprio;
2803         cfqq->org_ioprio_class = cfqq->ioprio_class;
2804         cfq_clear_cfqq_prio_changed(cfqq);
2805 }
2806
2807 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2808 {
2809         struct cfq_data *cfqd = cic_to_cfqd(cic);
2810         struct cfq_queue *cfqq;
2811         unsigned long flags;
2812
2813         if (unlikely(!cfqd))
2814                 return;
2815
2816         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2817
2818         cfqq = cic->cfqq[BLK_RW_ASYNC];
2819         if (cfqq) {
2820                 struct cfq_queue *new_cfqq;
2821                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2822                                                 GFP_ATOMIC);
2823                 if (new_cfqq) {
2824                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2825                         cfq_put_queue(cfqq);
2826                 }
2827         }
2828
2829         cfqq = cic->cfqq[BLK_RW_SYNC];
2830         if (cfqq)
2831                 cfq_mark_cfqq_prio_changed(cfqq);
2832
2833         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2834 }
2835
2836 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2837 {
2838         call_for_each_cic(ioc, changed_ioprio);
2839         ioc->ioprio_changed = 0;
2840 }
2841
2842 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2843                           pid_t pid, bool is_sync)
2844 {
2845         RB_CLEAR_NODE(&cfqq->rb_node);
2846         RB_CLEAR_NODE(&cfqq->p_node);
2847         INIT_LIST_HEAD(&cfqq->fifo);
2848
2849         atomic_set(&cfqq->ref, 0);
2850         cfqq->cfqd = cfqd;
2851
2852         cfq_mark_cfqq_prio_changed(cfqq);
2853
2854         if (is_sync) {
2855                 if (!cfq_class_idle(cfqq))
2856                         cfq_mark_cfqq_idle_window(cfqq);
2857                 cfq_mark_cfqq_sync(cfqq);
2858         }
2859         cfqq->pid = pid;
2860 }
2861
2862 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2863 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2864 {
2865         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2866         struct cfq_data *cfqd = cic_to_cfqd(cic);
2867         unsigned long flags;
2868         struct request_queue *q;
2869
2870         if (unlikely(!cfqd))
2871                 return;
2872
2873         q = cfqd->queue;
2874
2875         spin_lock_irqsave(q->queue_lock, flags);
2876
2877         if (sync_cfqq) {
2878                 /*
2879                  * Drop reference to sync queue. A new sync queue will be
2880                  * assigned in new group upon arrival of a fresh request.
2881                  */
2882                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2883                 cic_set_cfqq(cic, NULL, 1);
2884                 cfq_put_queue(sync_cfqq);
2885         }
2886
2887         spin_unlock_irqrestore(q->queue_lock, flags);
2888 }
2889
2890 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2891 {
2892         call_for_each_cic(ioc, changed_cgroup);
2893         ioc->cgroup_changed = 0;
2894 }
2895 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2896
2897 static struct cfq_queue *
2898 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2899                      struct io_context *ioc, gfp_t gfp_mask)
2900 {
2901         struct cfq_queue *cfqq, *new_cfqq = NULL;
2902         struct cfq_io_context *cic;
2903         struct cfq_group *cfqg;
2904
2905 retry:
2906         cfqg = cfq_get_cfqg(cfqd, 1);
2907         cic = cfq_cic_lookup(cfqd, ioc);
2908         /* cic always exists here */
2909         cfqq = cic_to_cfqq(cic, is_sync);
2910
2911         /*
2912          * Always try a new alloc if we fell back to the OOM cfqq
2913          * originally, since it should just be a temporary situation.
2914          */
2915         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2916                 cfqq = NULL;
2917                 if (new_cfqq) {
2918                         cfqq = new_cfqq;
2919                         new_cfqq = NULL;
2920                 } else if (gfp_mask & __GFP_WAIT) {
2921                         spin_unlock_irq(cfqd->queue->queue_lock);
2922                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2923                                         gfp_mask | __GFP_ZERO,
2924                                         cfqd->queue->node);
2925                         spin_lock_irq(cfqd->queue->queue_lock);
2926                         if (new_cfqq)
2927                                 goto retry;
2928                 } else {
2929                         cfqq = kmem_cache_alloc_node(cfq_pool,
2930                                         gfp_mask | __GFP_ZERO,
2931                                         cfqd->queue->node);
2932                 }
2933
2934                 if (cfqq) {
2935                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2936                         cfq_init_prio_data(cfqq, ioc);
2937                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2938                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2939                 } else
2940                         cfqq = &cfqd->oom_cfqq;
2941         }
2942
2943         if (new_cfqq)
2944                 kmem_cache_free(cfq_pool, new_cfqq);
2945
2946         return cfqq;
2947 }
2948
2949 static struct cfq_queue **
2950 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2951 {
2952         switch (ioprio_class) {
2953         case IOPRIO_CLASS_RT:
2954                 return &cfqd->async_cfqq[0][ioprio];
2955         case IOPRIO_CLASS_BE:
2956                 return &cfqd->async_cfqq[1][ioprio];
2957         case IOPRIO_CLASS_IDLE:
2958                 return &cfqd->async_idle_cfqq;
2959         default:
2960                 BUG();
2961         }
2962 }
2963
2964 static struct cfq_queue *
2965 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2966               gfp_t gfp_mask)
2967 {
2968         const int ioprio = task_ioprio(ioc);
2969         const int ioprio_class = task_ioprio_class(ioc);
2970         struct cfq_queue **async_cfqq = NULL;
2971         struct cfq_queue *cfqq = NULL;
2972
2973         if (!is_sync) {
2974                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2975                 cfqq = *async_cfqq;
2976         }
2977
2978         if (!cfqq)
2979                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2980
2981         /*
2982          * pin the queue now that it's allocated, scheduler exit will prune it
2983          */
2984         if (!is_sync && !(*async_cfqq)) {
2985                 atomic_inc(&cfqq->ref);
2986                 *async_cfqq = cfqq;
2987         }
2988
2989         atomic_inc(&cfqq->ref);
2990         return cfqq;
2991 }
2992
2993 /*
2994  * We drop cfq io contexts lazily, so we may find a dead one.
2995  */
2996 static void
2997 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2998                   struct cfq_io_context *cic)
2999 {
3000         unsigned long flags;
3001
3002         WARN_ON(!list_empty(&cic->queue_list));
3003         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3004
3005         spin_lock_irqsave(&ioc->lock, flags);
3006
3007         BUG_ON(ioc->ioc_data == cic);
3008
3009         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3010         hlist_del_rcu(&cic->cic_list);
3011         spin_unlock_irqrestore(&ioc->lock, flags);
3012
3013         cfq_cic_free(cic);
3014 }
3015
3016 static struct cfq_io_context *
3017 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3018 {
3019         struct cfq_io_context *cic;
3020         unsigned long flags;
3021
3022         if (unlikely(!ioc))
3023                 return NULL;
3024
3025         rcu_read_lock();
3026
3027         /*
3028          * we maintain a last-hit cache, to avoid browsing over the tree
3029          */
3030         cic = rcu_dereference(ioc->ioc_data);
3031         if (cic && cic->key == cfqd) {
3032                 rcu_read_unlock();
3033                 return cic;
3034         }
3035
3036         do {
3037                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3038                 rcu_read_unlock();
3039                 if (!cic)
3040                         break;
3041                 if (unlikely(cic->key != cfqd)) {
3042                         cfq_drop_dead_cic(cfqd, ioc, cic);
3043                         rcu_read_lock();
3044                         continue;
3045                 }
3046
3047                 spin_lock_irqsave(&ioc->lock, flags);
3048                 rcu_assign_pointer(ioc->ioc_data, cic);
3049                 spin_unlock_irqrestore(&ioc->lock, flags);
3050                 break;
3051         } while (1);
3052
3053         return cic;
3054 }
3055
3056 /*
3057  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3058  * the process specific cfq io context when entered from the block layer.
3059  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3060  */
3061 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3062                         struct cfq_io_context *cic, gfp_t gfp_mask)
3063 {
3064         unsigned long flags;
3065         int ret;
3066
3067         ret = radix_tree_preload(gfp_mask);
3068         if (!ret) {
3069                 cic->ioc = ioc;
3070                 cic->key = cfqd;
3071
3072                 spin_lock_irqsave(&ioc->lock, flags);
3073                 ret = radix_tree_insert(&ioc->radix_root,
3074                                                 cfqd->cic_index, cic);
3075                 if (!ret)
3076                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3077                 spin_unlock_irqrestore(&ioc->lock, flags);
3078
3079                 radix_tree_preload_end();
3080
3081                 if (!ret) {
3082                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3083                         list_add(&cic->queue_list, &cfqd->cic_list);
3084                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3085                 }
3086         }
3087
3088         if (ret)
3089                 printk(KERN_ERR "cfq: cic link failed!\n");
3090
3091         return ret;
3092 }
3093
3094 /*
3095  * Setup general io context and cfq io context. There can be several cfq
3096  * io contexts per general io context, if this process is doing io to more
3097  * than one device managed by cfq.
3098  */
3099 static struct cfq_io_context *
3100 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3101 {
3102         struct io_context *ioc = NULL;
3103         struct cfq_io_context *cic;
3104
3105         might_sleep_if(gfp_mask & __GFP_WAIT);
3106
3107         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3108         if (!ioc)
3109                 return NULL;
3110
3111         cic = cfq_cic_lookup(cfqd, ioc);
3112         if (cic)
3113                 goto out;
3114
3115         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3116         if (cic == NULL)
3117                 goto err;
3118
3119         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3120                 goto err_free;
3121
3122 out:
3123         smp_read_barrier_depends();
3124         if (unlikely(ioc->ioprio_changed))
3125                 cfq_ioc_set_ioprio(ioc);
3126
3127 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3128         if (unlikely(ioc->cgroup_changed))
3129                 cfq_ioc_set_cgroup(ioc);
3130 #endif
3131         return cic;
3132 err_free:
3133         cfq_cic_free(cic);
3134 err:
3135         put_io_context(ioc);
3136         return NULL;
3137 }
3138
3139 static void
3140 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3141 {
3142         unsigned long elapsed = jiffies - cic->last_end_request;
3143         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3144
3145         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3146         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3147         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3148 }
3149
3150 static void
3151 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3152                        struct request *rq)
3153 {
3154         sector_t sdist = 0;
3155         sector_t n_sec = blk_rq_sectors(rq);
3156         if (cfqq->last_request_pos) {
3157                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3158                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3159                 else
3160                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3161         }
3162
3163         cfqq->seek_history <<= 1;
3164         if (blk_queue_nonrot(cfqd->queue))
3165                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3166         else
3167                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3168 }
3169
3170 /*
3171  * Disable idle window if the process thinks too long or seeks so much that
3172  * it doesn't matter
3173  */
3174 static void
3175 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3176                        struct cfq_io_context *cic)
3177 {
3178         int old_idle, enable_idle;
3179
3180         /*
3181          * Don't idle for async or idle io prio class
3182          */
3183         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3184                 return;
3185
3186         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3187
3188         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3189                 cfq_mark_cfqq_deep(cfqq);
3190
3191         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3192             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3193                 enable_idle = 0;
3194         else if (sample_valid(cic->ttime_samples)) {
3195                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3196                         enable_idle = 0;
3197                 else
3198                         enable_idle = 1;
3199         }
3200
3201         if (old_idle != enable_idle) {
3202                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3203                 if (enable_idle)
3204                         cfq_mark_cfqq_idle_window(cfqq);
3205                 else
3206                         cfq_clear_cfqq_idle_window(cfqq);
3207         }
3208 }
3209
3210 /*
3211  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3212  * no or if we aren't sure, a 1 will cause a preempt.
3213  */
3214 static bool
3215 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3216                    struct request *rq)
3217 {
3218         struct cfq_queue *cfqq;
3219
3220         cfqq = cfqd->active_queue;
3221         if (!cfqq)
3222                 return false;
3223
3224         if (cfq_class_idle(new_cfqq))
3225                 return false;
3226
3227         if (cfq_class_idle(cfqq))
3228                 return true;
3229
3230         /*
3231          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3232          */
3233         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3234                 return false;
3235
3236         /*
3237          * if the new request is sync, but the currently running queue is
3238          * not, let the sync request have priority.
3239          */
3240         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3241                 return true;
3242
3243         if (new_cfqq->cfqg != cfqq->cfqg)
3244                 return false;
3245
3246         if (cfq_slice_used(cfqq))
3247                 return true;
3248
3249         /* Allow preemption only if we are idling on sync-noidle tree */
3250         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3251             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3252             new_cfqq->service_tree->count == 2 &&
3253             RB_EMPTY_ROOT(&cfqq->sort_list))
3254                 return true;
3255
3256         /*
3257          * So both queues are sync. Let the new request get disk time if
3258          * it's a metadata request and the current queue is doing regular IO.
3259          */
3260         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3261                 return true;
3262
3263         /*
3264          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3265          */
3266         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3267                 return true;
3268
3269         /* An idle queue should not be idle now for some reason */
3270         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3271                 return true;
3272
3273         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3274                 return false;
3275
3276         /*
3277          * if this request is as-good as one we would expect from the
3278          * current cfqq, let it preempt
3279          */
3280         if (cfq_rq_close(cfqd, cfqq, rq))
3281                 return true;
3282
3283         return false;
3284 }
3285
3286 /*
3287  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3288  * let it have half of its nominal slice.
3289  */
3290 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3291 {
3292         cfq_log_cfqq(cfqd, cfqq, "preempt");
3293         cfq_slice_expired(cfqd, 1);
3294
3295         /*
3296          * Put the new queue at the front of the of the current list,
3297          * so we know that it will be selected next.
3298          */
3299         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3300
3301         cfq_service_tree_add(cfqd, cfqq, 1);
3302
3303         cfqq->slice_end = 0;
3304         cfq_mark_cfqq_slice_new(cfqq);
3305 }
3306
3307 /*
3308  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3309  * something we should do about it
3310  */
3311 static void
3312 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3313                 struct request *rq)
3314 {
3315         struct cfq_io_context *cic = RQ_CIC(rq);
3316
3317         cfqd->rq_queued++;
3318         if (rq->cmd_flags & REQ_META)
3319                 cfqq->meta_pending++;
3320
3321         cfq_update_io_thinktime(cfqd, cic);
3322         cfq_update_io_seektime(cfqd, cfqq, rq);
3323         cfq_update_idle_window(cfqd, cfqq, cic);
3324
3325         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3326
3327         if (cfqq == cfqd->active_queue) {
3328                 /*
3329                  * Remember that we saw a request from this process, but
3330                  * don't start queuing just yet. Otherwise we risk seeing lots
3331                  * of tiny requests, because we disrupt the normal plugging
3332                  * and merging. If the request is already larger than a single
3333                  * page, let it rip immediately. For that case we assume that
3334                  * merging is already done. Ditto for a busy system that
3335                  * has other work pending, don't risk delaying until the
3336                  * idle timer unplug to continue working.
3337                  */
3338                 if (cfq_cfqq_wait_request(cfqq)) {
3339                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3340                             cfqd->busy_queues > 1) {
3341                                 cfq_del_timer(cfqd, cfqq);
3342                                 cfq_clear_cfqq_wait_request(cfqq);
3343                                 __blk_run_queue(cfqd->queue);
3344                         } else {
3345                                 cfq_blkiocg_update_idle_time_stats(
3346                                                 &cfqq->cfqg->blkg);
3347                                 cfq_mark_cfqq_must_dispatch(cfqq);
3348                         }
3349                 }
3350         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3351                 /*
3352                  * not the active queue - expire current slice if it is
3353                  * idle and has expired it's mean thinktime or this new queue
3354                  * has some old slice time left and is of higher priority or
3355                  * this new queue is RT and the current one is BE
3356                  */
3357                 cfq_preempt_queue(cfqd, cfqq);
3358                 __blk_run_queue(cfqd->queue);
3359         }
3360 }
3361
3362 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3363 {
3364         struct cfq_data *cfqd = q->elevator->elevator_data;
3365         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3366
3367         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3368         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3369
3370         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3371         list_add_tail(&rq->queuelist, &cfqq->fifo);
3372         cfq_add_rq_rb(rq);
3373         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3374                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3375                         rq_is_sync(rq));
3376         cfq_rq_enqueued(cfqd, cfqq, rq);
3377 }
3378
3379 /*
3380  * Update hw_tag based on peak queue depth over 50 samples under
3381  * sufficient load.
3382  */
3383 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3384 {
3385         struct cfq_queue *cfqq = cfqd->active_queue;
3386
3387         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3388                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3389
3390         if (cfqd->hw_tag == 1)
3391                 return;
3392
3393         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3394             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3395                 return;
3396
3397         /*
3398          * If active queue hasn't enough requests and can idle, cfq might not
3399          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3400          * case
3401          */
3402         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3403             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3404             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3405                 return;
3406
3407         if (cfqd->hw_tag_samples++ < 50)
3408                 return;
3409
3410         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3411                 cfqd->hw_tag = 1;
3412         else
3413                 cfqd->hw_tag = 0;
3414 }
3415
3416 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3417 {
3418         struct cfq_io_context *cic = cfqd->active_cic;
3419
3420         /* If there are other queues in the group, don't wait */
3421         if (cfqq->cfqg->nr_cfqq > 1)
3422                 return false;
3423
3424         if (cfq_slice_used(cfqq))
3425                 return true;
3426
3427         /* if slice left is less than think time, wait busy */
3428         if (cic && sample_valid(cic->ttime_samples)
3429             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3430                 return true;
3431
3432         /*
3433          * If think times is less than a jiffy than ttime_mean=0 and above
3434          * will not be true. It might happen that slice has not expired yet
3435          * but will expire soon (4-5 ns) during select_queue(). To cover the
3436          * case where think time is less than a jiffy, mark the queue wait
3437          * busy if only 1 jiffy is left in the slice.
3438          */
3439         if (cfqq->slice_end - jiffies == 1)
3440                 return true;
3441
3442         return false;
3443 }
3444
3445 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3446 {
3447         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3448         struct cfq_data *cfqd = cfqq->cfqd;
3449         const int sync = rq_is_sync(rq);
3450         unsigned long now;
3451
3452         now = jiffies;
3453         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3454                      !!(rq->cmd_flags & REQ_NOIDLE));
3455
3456         cfq_update_hw_tag(cfqd);
3457
3458         WARN_ON(!cfqd->rq_in_driver);
3459         WARN_ON(!cfqq->dispatched);
3460         cfqd->rq_in_driver--;
3461         cfqq->dispatched--;
3462         (RQ_CFQG(rq))->dispatched--;
3463         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3464                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3465                         rq_data_dir(rq), rq_is_sync(rq));
3466
3467         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3468
3469         if (sync) {
3470                 RQ_CIC(rq)->last_end_request = now;
3471                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3472                         cfqd->last_delayed_sync = now;
3473         }
3474
3475         /*
3476          * If this is the active queue, check if it needs to be expired,
3477          * or if we want to idle in case it has no pending requests.
3478          */
3479         if (cfqd->active_queue == cfqq) {
3480                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3481
3482                 if (cfq_cfqq_slice_new(cfqq)) {
3483                         cfq_set_prio_slice(cfqd, cfqq);
3484                         cfq_clear_cfqq_slice_new(cfqq);
3485                 }
3486
3487                 /*
3488                  * Should we wait for next request to come in before we expire
3489                  * the queue.
3490                  */
3491                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3492                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3493                         if (!cfqd->cfq_slice_idle)
3494                                 extend_sl = cfqd->cfq_group_idle;
3495                         cfqq->slice_end = jiffies + extend_sl;
3496                         cfq_mark_cfqq_wait_busy(cfqq);
3497                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3498                 }
3499
3500                 /*
3501                  * Idling is not enabled on:
3502                  * - expired queues
3503                  * - idle-priority queues
3504                  * - async queues
3505                  * - queues with still some requests queued
3506                  * - when there is a close cooperator
3507                  */
3508                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3509                         cfq_slice_expired(cfqd, 1);
3510                 else if (sync && cfqq_empty &&
3511                          !cfq_close_cooperator(cfqd, cfqq)) {
3512                         cfqd->noidle_tree_requires_idle |=
3513                                 !(rq->cmd_flags & REQ_NOIDLE);
3514                         /*
3515                          * Idling is enabled for SYNC_WORKLOAD.
3516                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3517                          * only if we processed at least one !REQ_NOIDLE request
3518                          */
3519                         if (cfqd->serving_type == SYNC_WORKLOAD
3520                             || cfqd->noidle_tree_requires_idle
3521                             || cfqq->cfqg->nr_cfqq == 1)
3522                                 cfq_arm_slice_timer(cfqd);
3523                 }
3524         }
3525
3526         if (!cfqd->rq_in_driver) {
3527                 cfq_schedule_dispatch(cfqd);
3528                 return;
3529         }
3530         /*
3531          * A queue is idle at cfq_dispatch_requests(), but it gets noidle
3532          * later. We schedule a dispatch if the queue has no requests,
3533          * otherwise the disk is actually in idle till all requests
3534          * are finished even cfq_arm_slice_timer doesn't make the queue idle
3535          * */
3536         cfqq = cfqd->active_queue;
3537         if (!cfqq)
3538                 return;
3539
3540         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq) &&
3541             (!cfqd->cfq_group_idle || cfqq->cfqg->nr_cfqq > 1)) {
3542                 cfq_del_timer(cfqd, cfqq);
3543                 cfq_schedule_dispatch(cfqd);
3544         }
3545 }
3546
3547 /*
3548  * we temporarily boost lower priority queues if they are holding fs exclusive
3549  * resources. they are boosted to normal prio (CLASS_BE/4)
3550  */
3551 static void cfq_prio_boost(struct cfq_queue *cfqq)
3552 {
3553         if (has_fs_excl()) {
3554                 /*
3555                  * boost idle prio on transactions that would lock out other
3556                  * users of the filesystem
3557                  */
3558                 if (cfq_class_idle(cfqq))
3559                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3560                 if (cfqq->ioprio > IOPRIO_NORM)
3561                         cfqq->ioprio = IOPRIO_NORM;
3562         } else {
3563                 /*
3564                  * unboost the queue (if needed)
3565                  */
3566                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3567                 cfqq->ioprio = cfqq->org_ioprio;
3568         }
3569 }
3570
3571 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3572 {
3573         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3574                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3575                 return ELV_MQUEUE_MUST;
3576         }
3577
3578         return ELV_MQUEUE_MAY;
3579 }
3580
3581 static int cfq_may_queue(struct request_queue *q, int rw)
3582 {
3583         struct cfq_data *cfqd = q->elevator->elevator_data;
3584         struct task_struct *tsk = current;
3585         struct cfq_io_context *cic;
3586         struct cfq_queue *cfqq;
3587
3588         /*
3589          * don't force setup of a queue from here, as a call to may_queue
3590          * does not necessarily imply that a request actually will be queued.
3591          * so just lookup a possibly existing queue, or return 'may queue'
3592          * if that fails
3593          */
3594         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3595         if (!cic)
3596                 return ELV_MQUEUE_MAY;
3597
3598         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3599         if (cfqq) {
3600                 cfq_init_prio_data(cfqq, cic->ioc);
3601                 cfq_prio_boost(cfqq);
3602
3603                 return __cfq_may_queue(cfqq);
3604         }
3605
3606         return ELV_MQUEUE_MAY;
3607 }
3608
3609 /*
3610  * queue lock held here
3611  */
3612 static void cfq_put_request(struct request *rq)
3613 {
3614         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3615
3616         if (cfqq) {
3617                 const int rw = rq_data_dir(rq);
3618
3619                 BUG_ON(!cfqq->allocated[rw]);
3620                 cfqq->allocated[rw]--;
3621
3622                 put_io_context(RQ_CIC(rq)->ioc);
3623
3624                 rq->elevator_private = NULL;
3625                 rq->elevator_private2 = NULL;
3626
3627                 /* Put down rq reference on cfqg */
3628                 cfq_put_cfqg(RQ_CFQG(rq));
3629                 rq->elevator_private3 = NULL;
3630
3631                 cfq_put_queue(cfqq);
3632         }
3633 }
3634
3635 static struct cfq_queue *
3636 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3637                 struct cfq_queue *cfqq)
3638 {
3639         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3640         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3641         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3642         cfq_put_queue(cfqq);
3643         return cic_to_cfqq(cic, 1);
3644 }
3645
3646 /*
3647  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3648  * was the last process referring to said cfqq.
3649  */
3650 static struct cfq_queue *
3651 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3652 {
3653         if (cfqq_process_refs(cfqq) == 1) {
3654                 cfqq->pid = current->pid;
3655                 cfq_clear_cfqq_coop(cfqq);
3656                 cfq_clear_cfqq_split_coop(cfqq);
3657                 return cfqq;
3658         }
3659
3660         cic_set_cfqq(cic, NULL, 1);
3661
3662         cfq_put_cooperator(cfqq);
3663
3664         cfq_put_queue(cfqq);
3665         return NULL;
3666 }
3667 /*
3668  * Allocate cfq data structures associated with this request.
3669  */
3670 static int
3671 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3672 {
3673         struct cfq_data *cfqd = q->elevator->elevator_data;
3674         struct cfq_io_context *cic;
3675         const int rw = rq_data_dir(rq);
3676         const bool is_sync = rq_is_sync(rq);
3677         struct cfq_queue *cfqq;
3678         unsigned long flags;
3679
3680         might_sleep_if(gfp_mask & __GFP_WAIT);
3681
3682         cic = cfq_get_io_context(cfqd, gfp_mask);
3683
3684         spin_lock_irqsave(q->queue_lock, flags);
3685
3686         if (!cic)
3687                 goto queue_fail;
3688
3689 new_queue:
3690         cfqq = cic_to_cfqq(cic, is_sync);
3691         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3692                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3693                 cic_set_cfqq(cic, cfqq, is_sync);
3694         } else {
3695                 /*
3696                  * If the queue was seeky for too long, break it apart.
3697                  */
3698                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3699                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3700                         cfqq = split_cfqq(cic, cfqq);
3701                         if (!cfqq)
3702                                 goto new_queue;
3703                 }
3704
3705                 /*
3706                  * Check to see if this queue is scheduled to merge with
3707                  * another, closely cooperating queue.  The merging of
3708                  * queues happens here as it must be done in process context.
3709                  * The reference on new_cfqq was taken in merge_cfqqs.
3710                  */
3711                 if (cfqq->new_cfqq)
3712                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3713         }
3714
3715         cfqq->allocated[rw]++;
3716         atomic_inc(&cfqq->ref);
3717
3718         spin_unlock_irqrestore(q->queue_lock, flags);
3719
3720         rq->elevator_private = cic;
3721         rq->elevator_private2 = cfqq;
3722         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3723         return 0;
3724
3725 queue_fail:
3726         if (cic)
3727                 put_io_context(cic->ioc);
3728
3729         cfq_schedule_dispatch(cfqd);
3730         spin_unlock_irqrestore(q->queue_lock, flags);
3731         cfq_log(cfqd, "set_request fail");
3732         return 1;
3733 }
3734
3735 static void cfq_kick_queue(struct work_struct *work)
3736 {
3737         struct cfq_data *cfqd =
3738                 container_of(work, struct cfq_data, unplug_work);
3739         struct request_queue *q = cfqd->queue;
3740
3741         spin_lock_irq(q->queue_lock);
3742         __blk_run_queue(cfqd->queue);
3743         spin_unlock_irq(q->queue_lock);
3744 }
3745
3746 /*
3747  * Timer running if the active_queue is currently idling inside its time slice
3748  */
3749 static void cfq_idle_slice_timer(unsigned long data)
3750 {
3751         struct cfq_data *cfqd = (struct cfq_data *) data;
3752         struct cfq_queue *cfqq;
3753         unsigned long flags;
3754         int timed_out = 1;
3755
3756         cfq_log(cfqd, "idle timer fired");
3757
3758         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3759
3760         cfqq = cfqd->active_queue;
3761         if (cfqq) {
3762                 timed_out = 0;
3763
3764                 /*
3765                  * We saw a request before the queue expired, let it through
3766                  */
3767                 if (cfq_cfqq_must_dispatch(cfqq))
3768                         goto out_kick;
3769
3770                 /*
3771                  * expired
3772                  */
3773                 if (cfq_slice_used(cfqq))
3774                         goto expire;
3775
3776                 /*
3777                  * only expire and reinvoke request handler, if there are
3778                  * other queues with pending requests
3779                  */
3780                 if (!cfqd->busy_queues)
3781                         goto out_cont;
3782
3783                 /*
3784                  * not expired and it has a request pending, let it dispatch
3785                  */
3786                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3787                         goto out_kick;
3788
3789                 /*
3790                  * Queue depth flag is reset only when the idle didn't succeed
3791                  */
3792                 cfq_clear_cfqq_deep(cfqq);
3793         }
3794 expire:
3795         cfq_slice_expired(cfqd, timed_out);
3796 out_kick:
3797         cfq_schedule_dispatch(cfqd);
3798 out_cont:
3799         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3800 }
3801
3802 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3803 {
3804         del_timer_sync(&cfqd->idle_slice_timer);
3805         cancel_work_sync(&cfqd->unplug_work);
3806 }
3807
3808 static void cfq_put_async_queues(struct cfq_data *cfqd)
3809 {
3810         int i;
3811
3812         for (i = 0; i < IOPRIO_BE_NR; i++) {
3813                 if (cfqd->async_cfqq[0][i])
3814                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3815                 if (cfqd->async_cfqq[1][i])
3816                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3817         }
3818
3819         if (cfqd->async_idle_cfqq)
3820                 cfq_put_queue(cfqd->async_idle_cfqq);
3821 }
3822
3823 static void cfq_cfqd_free(struct rcu_head *head)
3824 {
3825         kfree(container_of(head, struct cfq_data, rcu));
3826 }
3827
3828 static void cfq_exit_queue(struct elevator_queue *e)
3829 {
3830         struct cfq_data *cfqd = e->elevator_data;
3831         struct request_queue *q = cfqd->queue;
3832
3833         cfq_shutdown_timer_wq(cfqd);
3834
3835         spin_lock_irq(q->queue_lock);
3836
3837         if (cfqd->active_queue)
3838                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3839
3840         while (!list_empty(&cfqd->cic_list)) {
3841                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3842                                                         struct cfq_io_context,
3843                                                         queue_list);
3844
3845                 __cfq_exit_single_io_context(cfqd, cic);
3846         }
3847
3848         cfq_put_async_queues(cfqd);
3849         cfq_release_cfq_groups(cfqd);
3850         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3851
3852         spin_unlock_irq(q->queue_lock);
3853
3854         cfq_shutdown_timer_wq(cfqd);
3855
3856         spin_lock(&cic_index_lock);
3857         ida_remove(&cic_index_ida, cfqd->cic_index);
3858         spin_unlock(&cic_index_lock);
3859
3860         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3861         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3862 }
3863
3864 static int cfq_alloc_cic_index(void)
3865 {
3866         int index, error;
3867
3868         do {
3869                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3870                         return -ENOMEM;
3871
3872                 spin_lock(&cic_index_lock);
3873                 error = ida_get_new(&cic_index_ida, &index);
3874                 spin_unlock(&cic_index_lock);
3875                 if (error && error != -EAGAIN)
3876                         return error;
3877         } while (error);
3878
3879         return index;
3880 }
3881
3882 static void *cfq_init_queue(struct request_queue *q)
3883 {
3884         struct cfq_data *cfqd;
3885         int i, j;
3886         struct cfq_group *cfqg;
3887         struct cfq_rb_root *st;
3888
3889         i = cfq_alloc_cic_index();
3890         if (i < 0)
3891                 return NULL;
3892
3893         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3894         if (!cfqd)
3895                 return NULL;
3896
3897         cfqd->cic_index = i;
3898
3899         /* Init root service tree */
3900         cfqd->grp_service_tree = CFQ_RB_ROOT;
3901
3902         /* Init root group */
3903         cfqg = &cfqd->root_group;
3904         for_each_cfqg_st(cfqg, i, j, st)
3905                 *st = CFQ_RB_ROOT;
3906         RB_CLEAR_NODE(&cfqg->rb_node);
3907
3908         /* Give preference to root group over other groups */
3909         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3910
3911 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3912         /*
3913          * Take a reference to root group which we never drop. This is just
3914          * to make sure that cfq_put_cfqg() does not try to kfree root group
3915          */
3916         atomic_set(&cfqg->ref, 1);
3917         rcu_read_lock();
3918         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3919                                         (void *)cfqd, 0);
3920         rcu_read_unlock();
3921 #endif
3922         /*
3923          * Not strictly needed (since RB_ROOT just clears the node and we
3924          * zeroed cfqd on alloc), but better be safe in case someone decides
3925          * to add magic to the rb code
3926          */
3927         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3928                 cfqd->prio_trees[i] = RB_ROOT;
3929
3930         /*
3931          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3932          * Grab a permanent reference to it, so that the normal code flow
3933          * will not attempt to free it.
3934          */
3935         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3936         atomic_inc(&cfqd->oom_cfqq.ref);
3937         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3938
3939         INIT_LIST_HEAD(&cfqd->cic_list);
3940
3941         cfqd->queue = q;
3942
3943         init_timer(&cfqd->idle_slice_timer);
3944         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3945         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3946
3947         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3948
3949         cfqd->cfq_quantum = cfq_quantum;
3950         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3951         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3952         cfqd->cfq_back_max = cfq_back_max;
3953         cfqd->cfq_back_penalty = cfq_back_penalty;
3954         cfqd->cfq_slice[0] = cfq_slice_async;
3955         cfqd->cfq_slice[1] = cfq_slice_sync;
3956         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3957         cfqd->cfq_slice_idle = cfq_slice_idle;
3958         cfqd->cfq_group_idle = cfq_group_idle;
3959         cfqd->cfq_latency = 1;
3960         cfqd->cfq_group_isolation = 0;
3961         cfqd->hw_tag = -1;
3962         /*
3963          * we optimistically start assuming sync ops weren't delayed in last
3964          * second, in order to have larger depth for async operations.
3965          */
3966         cfqd->last_delayed_sync = jiffies - HZ;
3967         return cfqd;
3968 }
3969
3970 static void cfq_slab_kill(void)
3971 {
3972         /*
3973          * Caller already ensured that pending RCU callbacks are completed,
3974          * so we should have no busy allocations at this point.
3975          */
3976         if (cfq_pool)
3977                 kmem_cache_destroy(cfq_pool);
3978         if (cfq_ioc_pool)
3979                 kmem_cache_destroy(cfq_ioc_pool);
3980 }
3981
3982 static int __init cfq_slab_setup(void)
3983 {
3984         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3985         if (!cfq_pool)
3986                 goto fail;
3987
3988         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3989         if (!cfq_ioc_pool)
3990                 goto fail;
3991
3992         return 0;
3993 fail:
3994         cfq_slab_kill();
3995         return -ENOMEM;
3996 }
3997
3998 /*
3999  * sysfs parts below -->
4000  */
4001 static ssize_t
4002 cfq_var_show(unsigned int var, char *page)
4003 {
4004         return sprintf(page, "%d\n", var);
4005 }
4006
4007 static ssize_t
4008 cfq_var_store(unsigned int *var, const char *page, size_t count)
4009 {
4010         char *p = (char *) page;
4011
4012         *var = simple_strtoul(p, &p, 10);
4013         return count;
4014 }
4015
4016 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4017 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4018 {                                                                       \
4019         struct cfq_data *cfqd = e->elevator_data;                       \
4020         unsigned int __data = __VAR;                                    \
4021         if (__CONV)                                                     \
4022                 __data = jiffies_to_msecs(__data);                      \
4023         return cfq_var_show(__data, (page));                            \
4024 }
4025 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4026 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4027 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4028 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4029 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4030 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4031 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4032 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4033 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4034 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4035 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4036 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4037 #undef SHOW_FUNCTION
4038
4039 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4040 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4041 {                                                                       \
4042         struct cfq_data *cfqd = e->elevator_data;                       \
4043         unsigned int __data;                                            \
4044         int ret = cfq_var_store(&__data, (page), count);                \
4045         if (__data < (MIN))                                             \
4046                 __data = (MIN);                                         \
4047         else if (__data > (MAX))                                        \
4048                 __data = (MAX);                                         \
4049         if (__CONV)                                                     \
4050                 *(__PTR) = msecs_to_jiffies(__data);                    \
4051         else                                                            \
4052                 *(__PTR) = __data;                                      \
4053         return ret;                                                     \
4054 }
4055 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4056 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4057                 UINT_MAX, 1);
4058 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4059                 UINT_MAX, 1);
4060 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4061 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4062                 UINT_MAX, 0);
4063 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4064 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4065 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4066 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4067 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4068                 UINT_MAX, 0);
4069 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4070 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4071 #undef STORE_FUNCTION
4072
4073 #define CFQ_ATTR(name) \
4074         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4075
4076 static struct elv_fs_entry cfq_attrs[] = {
4077         CFQ_ATTR(quantum),
4078         CFQ_ATTR(fifo_expire_sync),
4079         CFQ_ATTR(fifo_expire_async),
4080         CFQ_ATTR(back_seek_max),
4081         CFQ_ATTR(back_seek_penalty),
4082         CFQ_ATTR(slice_sync),
4083         CFQ_ATTR(slice_async),
4084         CFQ_ATTR(slice_async_rq),
4085         CFQ_ATTR(slice_idle),
4086         CFQ_ATTR(group_idle),
4087         CFQ_ATTR(low_latency),
4088         CFQ_ATTR(group_isolation),
4089         __ATTR_NULL
4090 };
4091
4092 static struct elevator_type iosched_cfq = {
4093         .ops = {
4094                 .elevator_merge_fn =            cfq_merge,
4095                 .elevator_merged_fn =           cfq_merged_request,
4096                 .elevator_merge_req_fn =        cfq_merged_requests,
4097                 .elevator_allow_merge_fn =      cfq_allow_merge,
4098                 .elevator_bio_merged_fn =       cfq_bio_merged,
4099                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4100                 .elevator_add_req_fn =          cfq_insert_request,
4101                 .elevator_activate_req_fn =     cfq_activate_request,
4102                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4103                 .elevator_queue_empty_fn =      cfq_queue_empty,
4104                 .elevator_completed_req_fn =    cfq_completed_request,
4105                 .elevator_former_req_fn =       elv_rb_former_request,
4106                 .elevator_latter_req_fn =       elv_rb_latter_request,
4107                 .elevator_set_req_fn =          cfq_set_request,
4108                 .elevator_put_req_fn =          cfq_put_request,
4109                 .elevator_may_queue_fn =        cfq_may_queue,
4110                 .elevator_init_fn =             cfq_init_queue,
4111                 .elevator_exit_fn =             cfq_exit_queue,
4112                 .trim =                         cfq_free_io_context,
4113         },
4114         .elevator_attrs =       cfq_attrs,
4115         .elevator_name =        "cfq",
4116         .elevator_owner =       THIS_MODULE,
4117 };
4118
4119 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4120 static struct blkio_policy_type blkio_policy_cfq = {
4121         .ops = {
4122                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4123                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4124         },
4125 };
4126 #else
4127 static struct blkio_policy_type blkio_policy_cfq;
4128 #endif
4129
4130 static int __init cfq_init(void)
4131 {
4132         /*
4133          * could be 0 on HZ < 1000 setups
4134          */
4135         if (!cfq_slice_async)
4136                 cfq_slice_async = 1;
4137         if (!cfq_slice_idle)
4138                 cfq_slice_idle = 1;
4139
4140 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4141         if (!cfq_group_idle)
4142                 cfq_group_idle = 1;
4143 #else
4144                 cfq_group_idle = 0;
4145 #endif
4146         if (cfq_slab_setup())
4147                 return -ENOMEM;
4148
4149         elv_register(&iosched_cfq);
4150         blkio_policy_register(&blkio_policy_cfq);
4151
4152         return 0;
4153 }
4154
4155 static void __exit cfq_exit(void)
4156 {
4157         DECLARE_COMPLETION_ONSTACK(all_gone);
4158         blkio_policy_unregister(&blkio_policy_cfq);
4159         elv_unregister(&iosched_cfq);
4160         ioc_gone = &all_gone;
4161         /* ioc_gone's update must be visible before reading ioc_count */
4162         smp_wmb();
4163
4164         /*
4165          * this also protects us from entering cfq_slab_kill() with
4166          * pending RCU callbacks
4167          */
4168         if (elv_ioc_count_read(cfq_ioc_count))
4169                 wait_for_completion(&all_gone);
4170         ida_destroy(&cic_index_ida);
4171         cfq_slab_kill();
4172 }
4173
4174 module_init(cfq_init);
4175 module_exit(cfq_exit);
4176
4177 MODULE_AUTHOR("Jens Axboe");
4178 MODULE_LICENSE("GPL");
4179 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");