blkio: deletion of a cgroup was causes oops
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
35
36 /*
37  * offset from end of service tree
38  */
39 #define CFQ_IDLE_DELAY          (HZ / 5)
40
41 /*
42  * below this threshold, we consider thinktime immediate
43  */
44 #define CFQ_MIN_TT              (2)
45
46 #define CFQ_SLICE_SCALE         (5)
47 #define CFQ_HW_QUEUE_MIN        (5)
48 #define CFQ_SERVICE_SHIFT       12
49
50 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
54
55 #define RQ_CIC(rq)              \
56         ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
59
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
62
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
66
67 static DEFINE_SPINLOCK(cic_index_lock);
68 static DEFINE_IDA(cic_index_ida);
69
70 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
71 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
72 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73
74 #define sample_valid(samples)   ((samples) > 80)
75 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
76
77 /*
78  * Most of our rbtree usage is for sorting with min extraction, so
79  * if we cache the leftmost node we don't have to walk down the tree
80  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81  * move this into the elevator for the rq sorting as well.
82  */
83 struct cfq_rb_root {
84         struct rb_root rb;
85         struct rb_node *left;
86         unsigned count;
87         unsigned total_weight;
88         u64 min_vdisktime;
89         struct rb_node *active;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         atomic_t ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         struct cfq_group *orig_cfqg;
150 };
151
152 /*
153  * First index in the service_trees.
154  * IDLE is handled separately, so it has negative index
155  */
156 enum wl_prio_t {
157         BE_WORKLOAD = 0,
158         RT_WORKLOAD = 1,
159         IDLE_WORKLOAD = 2,
160 };
161
162 /*
163  * Second index in the service_trees.
164  */
165 enum wl_type_t {
166         ASYNC_WORKLOAD = 0,
167         SYNC_NOIDLE_WORKLOAD = 1,
168         SYNC_WORKLOAD = 2
169 };
170
171 /* This is per cgroup per device grouping structure */
172 struct cfq_group {
173         /* group service_tree member */
174         struct rb_node rb_node;
175
176         /* group service_tree key */
177         u64 vdisktime;
178         unsigned int weight;
179         bool on_st;
180
181         /* number of cfqq currently on this group */
182         int nr_cfqq;
183
184         /* Per group busy queus average. Useful for workload slice calc. */
185         unsigned int busy_queues_avg[2];
186         /*
187          * rr lists of queues with requests, onle rr for each priority class.
188          * Counts are embedded in the cfq_rb_root
189          */
190         struct cfq_rb_root service_trees[2][3];
191         struct cfq_rb_root service_tree_idle;
192
193         unsigned long saved_workload_slice;
194         enum wl_type_t saved_workload;
195         enum wl_prio_t saved_serving_prio;
196         struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198         struct hlist_node cfqd_node;
199         atomic_t ref;
200 #endif
201 };
202
203 /*
204  * Per block device queue structure
205  */
206 struct cfq_data {
207         struct request_queue *queue;
208         /* Root service tree for cfq_groups */
209         struct cfq_rb_root grp_service_tree;
210         struct cfq_group root_group;
211
212         /*
213          * The priority currently being served
214          */
215         enum wl_prio_t serving_prio;
216         enum wl_type_t serving_type;
217         unsigned long workload_expires;
218         struct cfq_group *serving_group;
219
220         /*
221          * Each priority tree is sorted by next_request position.  These
222          * trees are used when determining if two or more queues are
223          * interleaving requests (see cfq_close_cooperator).
224          */
225         struct rb_root prio_trees[CFQ_PRIO_LISTS];
226
227         unsigned int busy_queues;
228
229         int rq_in_driver;
230         int rq_in_flight[2];
231
232         /*
233          * queue-depth detection
234          */
235         int rq_queued;
236         int hw_tag;
237         /*
238          * hw_tag can be
239          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
240          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
241          *  0 => no NCQ
242          */
243         int hw_tag_est_depth;
244         unsigned int hw_tag_samples;
245
246         /*
247          * idle window management
248          */
249         struct timer_list idle_slice_timer;
250         struct work_struct unplug_work;
251
252         struct cfq_queue *active_queue;
253         struct cfq_io_context *active_cic;
254
255         /*
256          * async queue for each priority case
257          */
258         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
259         struct cfq_queue *async_idle_cfqq;
260
261         sector_t last_position;
262
263         /*
264          * tunables, see top of file
265          */
266         unsigned int cfq_quantum;
267         unsigned int cfq_fifo_expire[2];
268         unsigned int cfq_back_penalty;
269         unsigned int cfq_back_max;
270         unsigned int cfq_slice[2];
271         unsigned int cfq_slice_async_rq;
272         unsigned int cfq_slice_idle;
273         unsigned int cfq_latency;
274         unsigned int cfq_group_isolation;
275
276         unsigned int cic_index;
277         struct list_head cic_list;
278
279         /*
280          * Fallback dummy cfqq for extreme OOM conditions
281          */
282         struct cfq_queue oom_cfqq;
283
284         unsigned long last_delayed_sync;
285
286         /* List of cfq groups being managed on this device*/
287         struct hlist_head cfqg_list;
288         struct rcu_head rcu;
289 };
290
291 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
292
293 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
294                                             enum wl_prio_t prio,
295                                             enum wl_type_t type)
296 {
297         if (!cfqg)
298                 return NULL;
299
300         if (prio == IDLE_WORKLOAD)
301                 return &cfqg->service_tree_idle;
302
303         return &cfqg->service_trees[prio][type];
304 }
305
306 enum cfqq_state_flags {
307         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
308         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
309         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
310         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
311         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
312         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
313         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
314         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
315         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
316         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
317         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
318         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
319         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
320 };
321
322 #define CFQ_CFQQ_FNS(name)                                              \
323 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
324 {                                                                       \
325         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
326 }                                                                       \
327 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
328 {                                                                       \
329         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
330 }                                                                       \
331 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
332 {                                                                       \
333         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
334 }
335
336 CFQ_CFQQ_FNS(on_rr);
337 CFQ_CFQQ_FNS(wait_request);
338 CFQ_CFQQ_FNS(must_dispatch);
339 CFQ_CFQQ_FNS(must_alloc_slice);
340 CFQ_CFQQ_FNS(fifo_expire);
341 CFQ_CFQQ_FNS(idle_window);
342 CFQ_CFQQ_FNS(prio_changed);
343 CFQ_CFQQ_FNS(slice_new);
344 CFQ_CFQQ_FNS(sync);
345 CFQ_CFQQ_FNS(coop);
346 CFQ_CFQQ_FNS(split_coop);
347 CFQ_CFQQ_FNS(deep);
348 CFQ_CFQQ_FNS(wait_busy);
349 #undef CFQ_CFQQ_FNS
350
351 #ifdef CONFIG_CFQ_GROUP_IOSCHED
352 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
353         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
354                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
355                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
356
357 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
358         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
359                                 blkg_path(&(cfqg)->blkg), ##args);      \
360
361 #else
362 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
363         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
364 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
365 #endif
366 #define cfq_log(cfqd, fmt, args...)     \
367         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
368
369 /* Traverses through cfq group service trees */
370 #define for_each_cfqg_st(cfqg, i, j, st) \
371         for (i = 0; i <= IDLE_WORKLOAD; i++) \
372                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
373                         : &cfqg->service_tree_idle; \
374                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
375                         (i == IDLE_WORKLOAD && j == 0); \
376                         j++, st = i < IDLE_WORKLOAD ? \
377                         &cfqg->service_trees[i][j]: NULL) \
378
379
380 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
381 {
382         if (cfq_class_idle(cfqq))
383                 return IDLE_WORKLOAD;
384         if (cfq_class_rt(cfqq))
385                 return RT_WORKLOAD;
386         return BE_WORKLOAD;
387 }
388
389
390 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
391 {
392         if (!cfq_cfqq_sync(cfqq))
393                 return ASYNC_WORKLOAD;
394         if (!cfq_cfqq_idle_window(cfqq))
395                 return SYNC_NOIDLE_WORKLOAD;
396         return SYNC_WORKLOAD;
397 }
398
399 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
400                                         struct cfq_data *cfqd,
401                                         struct cfq_group *cfqg)
402 {
403         if (wl == IDLE_WORKLOAD)
404                 return cfqg->service_tree_idle.count;
405
406         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
407                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
408                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
409 }
410
411 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
412                                         struct cfq_group *cfqg)
413 {
414         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
415                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
416 }
417
418 static void cfq_dispatch_insert(struct request_queue *, struct request *);
419 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
420                                        struct io_context *, gfp_t);
421 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
422                                                 struct io_context *);
423
424 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
425                                             bool is_sync)
426 {
427         return cic->cfqq[is_sync];
428 }
429
430 static inline void cic_set_cfqq(struct cfq_io_context *cic,
431                                 struct cfq_queue *cfqq, bool is_sync)
432 {
433         cic->cfqq[is_sync] = cfqq;
434 }
435
436 #define CIC_DEAD_KEY    1ul
437 #define CIC_DEAD_INDEX_SHIFT    1
438
439 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
440 {
441         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
442 }
443
444 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
445 {
446         struct cfq_data *cfqd = cic->key;
447
448         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
449                 return NULL;
450
451         return cfqd;
452 }
453
454 /*
455  * We regard a request as SYNC, if it's either a read or has the SYNC bit
456  * set (in which case it could also be direct WRITE).
457  */
458 static inline bool cfq_bio_sync(struct bio *bio)
459 {
460         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
461 }
462
463 /*
464  * scheduler run of queue, if there are requests pending and no one in the
465  * driver that will restart queueing
466  */
467 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
468 {
469         if (cfqd->busy_queues) {
470                 cfq_log(cfqd, "schedule dispatch");
471                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
472         }
473 }
474
475 static int cfq_queue_empty(struct request_queue *q)
476 {
477         struct cfq_data *cfqd = q->elevator->elevator_data;
478
479         return !cfqd->rq_queued;
480 }
481
482 /*
483  * Scale schedule slice based on io priority. Use the sync time slice only
484  * if a queue is marked sync and has sync io queued. A sync queue with async
485  * io only, should not get full sync slice length.
486  */
487 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
488                                  unsigned short prio)
489 {
490         const int base_slice = cfqd->cfq_slice[sync];
491
492         WARN_ON(prio >= IOPRIO_BE_NR);
493
494         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
495 }
496
497 static inline int
498 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
499 {
500         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
501 }
502
503 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
504 {
505         u64 d = delta << CFQ_SERVICE_SHIFT;
506
507         d = d * BLKIO_WEIGHT_DEFAULT;
508         do_div(d, cfqg->weight);
509         return d;
510 }
511
512 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
513 {
514         s64 delta = (s64)(vdisktime - min_vdisktime);
515         if (delta > 0)
516                 min_vdisktime = vdisktime;
517
518         return min_vdisktime;
519 }
520
521 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
522 {
523         s64 delta = (s64)(vdisktime - min_vdisktime);
524         if (delta < 0)
525                 min_vdisktime = vdisktime;
526
527         return min_vdisktime;
528 }
529
530 static void update_min_vdisktime(struct cfq_rb_root *st)
531 {
532         u64 vdisktime = st->min_vdisktime;
533         struct cfq_group *cfqg;
534
535         if (st->active) {
536                 cfqg = rb_entry_cfqg(st->active);
537                 vdisktime = cfqg->vdisktime;
538         }
539
540         if (st->left) {
541                 cfqg = rb_entry_cfqg(st->left);
542                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
543         }
544
545         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
546 }
547
548 /*
549  * get averaged number of queues of RT/BE priority.
550  * average is updated, with a formula that gives more weight to higher numbers,
551  * to quickly follows sudden increases and decrease slowly
552  */
553
554 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
555                                         struct cfq_group *cfqg, bool rt)
556 {
557         unsigned min_q, max_q;
558         unsigned mult  = cfq_hist_divisor - 1;
559         unsigned round = cfq_hist_divisor / 2;
560         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
561
562         min_q = min(cfqg->busy_queues_avg[rt], busy);
563         max_q = max(cfqg->busy_queues_avg[rt], busy);
564         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
565                 cfq_hist_divisor;
566         return cfqg->busy_queues_avg[rt];
567 }
568
569 static inline unsigned
570 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
571 {
572         struct cfq_rb_root *st = &cfqd->grp_service_tree;
573
574         return cfq_target_latency * cfqg->weight / st->total_weight;
575 }
576
577 static inline void
578 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
579 {
580         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
581         if (cfqd->cfq_latency) {
582                 /*
583                  * interested queues (we consider only the ones with the same
584                  * priority class in the cfq group)
585                  */
586                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
587                                                 cfq_class_rt(cfqq));
588                 unsigned sync_slice = cfqd->cfq_slice[1];
589                 unsigned expect_latency = sync_slice * iq;
590                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
591
592                 if (expect_latency > group_slice) {
593                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
594                         /* scale low_slice according to IO priority
595                          * and sync vs async */
596                         unsigned low_slice =
597                                 min(slice, base_low_slice * slice / sync_slice);
598                         /* the adapted slice value is scaled to fit all iqs
599                          * into the target latency */
600                         slice = max(slice * group_slice / expect_latency,
601                                     low_slice);
602                 }
603         }
604         cfqq->slice_start = jiffies;
605         cfqq->slice_end = jiffies + slice;
606         cfqq->allocated_slice = slice;
607         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
608 }
609
610 /*
611  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
612  * isn't valid until the first request from the dispatch is activated
613  * and the slice time set.
614  */
615 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
616 {
617         if (cfq_cfqq_slice_new(cfqq))
618                 return 0;
619         if (time_before(jiffies, cfqq->slice_end))
620                 return 0;
621
622         return 1;
623 }
624
625 /*
626  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
627  * We choose the request that is closest to the head right now. Distance
628  * behind the head is penalized and only allowed to a certain extent.
629  */
630 static struct request *
631 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
632 {
633         sector_t s1, s2, d1 = 0, d2 = 0;
634         unsigned long back_max;
635 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
636 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
637         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
638
639         if (rq1 == NULL || rq1 == rq2)
640                 return rq2;
641         if (rq2 == NULL)
642                 return rq1;
643
644         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
645                 return rq1;
646         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
647                 return rq2;
648         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
649                 return rq1;
650         else if ((rq2->cmd_flags & REQ_META) &&
651                  !(rq1->cmd_flags & REQ_META))
652                 return rq2;
653
654         s1 = blk_rq_pos(rq1);
655         s2 = blk_rq_pos(rq2);
656
657         /*
658          * by definition, 1KiB is 2 sectors
659          */
660         back_max = cfqd->cfq_back_max * 2;
661
662         /*
663          * Strict one way elevator _except_ in the case where we allow
664          * short backward seeks which are biased as twice the cost of a
665          * similar forward seek.
666          */
667         if (s1 >= last)
668                 d1 = s1 - last;
669         else if (s1 + back_max >= last)
670                 d1 = (last - s1) * cfqd->cfq_back_penalty;
671         else
672                 wrap |= CFQ_RQ1_WRAP;
673
674         if (s2 >= last)
675                 d2 = s2 - last;
676         else if (s2 + back_max >= last)
677                 d2 = (last - s2) * cfqd->cfq_back_penalty;
678         else
679                 wrap |= CFQ_RQ2_WRAP;
680
681         /* Found required data */
682
683         /*
684          * By doing switch() on the bit mask "wrap" we avoid having to
685          * check two variables for all permutations: --> faster!
686          */
687         switch (wrap) {
688         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
689                 if (d1 < d2)
690                         return rq1;
691                 else if (d2 < d1)
692                         return rq2;
693                 else {
694                         if (s1 >= s2)
695                                 return rq1;
696                         else
697                                 return rq2;
698                 }
699
700         case CFQ_RQ2_WRAP:
701                 return rq1;
702         case CFQ_RQ1_WRAP:
703                 return rq2;
704         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
705         default:
706                 /*
707                  * Since both rqs are wrapped,
708                  * start with the one that's further behind head
709                  * (--> only *one* back seek required),
710                  * since back seek takes more time than forward.
711                  */
712                 if (s1 <= s2)
713                         return rq1;
714                 else
715                         return rq2;
716         }
717 }
718
719 /*
720  * The below is leftmost cache rbtree addon
721  */
722 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
723 {
724         /* Service tree is empty */
725         if (!root->count)
726                 return NULL;
727
728         if (!root->left)
729                 root->left = rb_first(&root->rb);
730
731         if (root->left)
732                 return rb_entry(root->left, struct cfq_queue, rb_node);
733
734         return NULL;
735 }
736
737 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
738 {
739         if (!root->left)
740                 root->left = rb_first(&root->rb);
741
742         if (root->left)
743                 return rb_entry_cfqg(root->left);
744
745         return NULL;
746 }
747
748 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
749 {
750         rb_erase(n, root);
751         RB_CLEAR_NODE(n);
752 }
753
754 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
755 {
756         if (root->left == n)
757                 root->left = NULL;
758         rb_erase_init(n, &root->rb);
759         --root->count;
760 }
761
762 /*
763  * would be nice to take fifo expire time into account as well
764  */
765 static struct request *
766 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
767                   struct request *last)
768 {
769         struct rb_node *rbnext = rb_next(&last->rb_node);
770         struct rb_node *rbprev = rb_prev(&last->rb_node);
771         struct request *next = NULL, *prev = NULL;
772
773         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
774
775         if (rbprev)
776                 prev = rb_entry_rq(rbprev);
777
778         if (rbnext)
779                 next = rb_entry_rq(rbnext);
780         else {
781                 rbnext = rb_first(&cfqq->sort_list);
782                 if (rbnext && rbnext != &last->rb_node)
783                         next = rb_entry_rq(rbnext);
784         }
785
786         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
787 }
788
789 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
790                                       struct cfq_queue *cfqq)
791 {
792         /*
793          * just an approximation, should be ok.
794          */
795         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
796                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
797 }
798
799 static inline s64
800 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
801 {
802         return cfqg->vdisktime - st->min_vdisktime;
803 }
804
805 static void
806 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
807 {
808         struct rb_node **node = &st->rb.rb_node;
809         struct rb_node *parent = NULL;
810         struct cfq_group *__cfqg;
811         s64 key = cfqg_key(st, cfqg);
812         int left = 1;
813
814         while (*node != NULL) {
815                 parent = *node;
816                 __cfqg = rb_entry_cfqg(parent);
817
818                 if (key < cfqg_key(st, __cfqg))
819                         node = &parent->rb_left;
820                 else {
821                         node = &parent->rb_right;
822                         left = 0;
823                 }
824         }
825
826         if (left)
827                 st->left = &cfqg->rb_node;
828
829         rb_link_node(&cfqg->rb_node, parent, node);
830         rb_insert_color(&cfqg->rb_node, &st->rb);
831 }
832
833 static void
834 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
835 {
836         struct cfq_rb_root *st = &cfqd->grp_service_tree;
837         struct cfq_group *__cfqg;
838         struct rb_node *n;
839
840         cfqg->nr_cfqq++;
841         if (cfqg->on_st)
842                 return;
843
844         /*
845          * Currently put the group at the end. Later implement something
846          * so that groups get lesser vtime based on their weights, so that
847          * if group does not loose all if it was not continously backlogged.
848          */
849         n = rb_last(&st->rb);
850         if (n) {
851                 __cfqg = rb_entry_cfqg(n);
852                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
853         } else
854                 cfqg->vdisktime = st->min_vdisktime;
855
856         __cfq_group_service_tree_add(st, cfqg);
857         cfqg->on_st = true;
858         st->total_weight += cfqg->weight;
859 }
860
861 static void
862 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
863 {
864         struct cfq_rb_root *st = &cfqd->grp_service_tree;
865
866         if (st->active == &cfqg->rb_node)
867                 st->active = NULL;
868
869         BUG_ON(cfqg->nr_cfqq < 1);
870         cfqg->nr_cfqq--;
871
872         /* If there are other cfq queues under this group, don't delete it */
873         if (cfqg->nr_cfqq)
874                 return;
875
876         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
877         cfqg->on_st = false;
878         st->total_weight -= cfqg->weight;
879         if (!RB_EMPTY_NODE(&cfqg->rb_node))
880                 cfq_rb_erase(&cfqg->rb_node, st);
881         cfqg->saved_workload_slice = 0;
882         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
883 }
884
885 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
886 {
887         unsigned int slice_used;
888
889         /*
890          * Queue got expired before even a single request completed or
891          * got expired immediately after first request completion.
892          */
893         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
894                 /*
895                  * Also charge the seek time incurred to the group, otherwise
896                  * if there are mutiple queues in the group, each can dispatch
897                  * a single request on seeky media and cause lots of seek time
898                  * and group will never know it.
899                  */
900                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
901                                         1);
902         } else {
903                 slice_used = jiffies - cfqq->slice_start;
904                 if (slice_used > cfqq->allocated_slice)
905                         slice_used = cfqq->allocated_slice;
906         }
907
908         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
909         return slice_used;
910 }
911
912 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
913                                 struct cfq_queue *cfqq)
914 {
915         struct cfq_rb_root *st = &cfqd->grp_service_tree;
916         unsigned int used_sl, charge_sl;
917         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
918                         - cfqg->service_tree_idle.count;
919
920         BUG_ON(nr_sync < 0);
921         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
922
923         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
924                 charge_sl = cfqq->allocated_slice;
925
926         /* Can't update vdisktime while group is on service tree */
927         cfq_rb_erase(&cfqg->rb_node, st);
928         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
929         __cfq_group_service_tree_add(st, cfqg);
930
931         /* This group is being expired. Save the context */
932         if (time_after(cfqd->workload_expires, jiffies)) {
933                 cfqg->saved_workload_slice = cfqd->workload_expires
934                                                 - jiffies;
935                 cfqg->saved_workload = cfqd->serving_type;
936                 cfqg->saved_serving_prio = cfqd->serving_prio;
937         } else
938                 cfqg->saved_workload_slice = 0;
939
940         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
941                                         st->min_vdisktime);
942         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
943         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
944 }
945
946 #ifdef CONFIG_CFQ_GROUP_IOSCHED
947 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
948 {
949         if (blkg)
950                 return container_of(blkg, struct cfq_group, blkg);
951         return NULL;
952 }
953
954 void
955 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
956 {
957         cfqg_of_blkg(blkg)->weight = weight;
958 }
959
960 static struct cfq_group *
961 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
962 {
963         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
964         struct cfq_group *cfqg = NULL;
965         void *key = cfqd;
966         int i, j;
967         struct cfq_rb_root *st;
968         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
969         unsigned int major, minor;
970
971         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
972         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
973                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
974                 cfqg->blkg.dev = MKDEV(major, minor);
975                 goto done;
976         }
977         if (cfqg || !create)
978                 goto done;
979
980         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
981         if (!cfqg)
982                 goto done;
983
984         for_each_cfqg_st(cfqg, i, j, st)
985                 *st = CFQ_RB_ROOT;
986         RB_CLEAR_NODE(&cfqg->rb_node);
987
988         /*
989          * Take the initial reference that will be released on destroy
990          * This can be thought of a joint reference by cgroup and
991          * elevator which will be dropped by either elevator exit
992          * or cgroup deletion path depending on who is exiting first.
993          */
994         atomic_set(&cfqg->ref, 1);
995
996         /* Add group onto cgroup list */
997         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
998         cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
999                                         MKDEV(major, minor));
1000         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1001
1002         /* Add group on cfqd list */
1003         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1004
1005 done:
1006         return cfqg;
1007 }
1008
1009 /*
1010  * Search for the cfq group current task belongs to. If create = 1, then also
1011  * create the cfq group if it does not exist. request_queue lock must be held.
1012  */
1013 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1014 {
1015         struct cgroup *cgroup;
1016         struct cfq_group *cfqg = NULL;
1017
1018         rcu_read_lock();
1019         cgroup = task_cgroup(current, blkio_subsys_id);
1020         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1021         if (!cfqg && create)
1022                 cfqg = &cfqd->root_group;
1023         rcu_read_unlock();
1024         return cfqg;
1025 }
1026
1027 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1028 {
1029         atomic_inc(&cfqg->ref);
1030         return cfqg;
1031 }
1032
1033 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1034 {
1035         /* Currently, all async queues are mapped to root group */
1036         if (!cfq_cfqq_sync(cfqq))
1037                 cfqg = &cfqq->cfqd->root_group;
1038
1039         cfqq->cfqg = cfqg;
1040         /* cfqq reference on cfqg */
1041         atomic_inc(&cfqq->cfqg->ref);
1042 }
1043
1044 static void cfq_put_cfqg(struct cfq_group *cfqg)
1045 {
1046         struct cfq_rb_root *st;
1047         int i, j;
1048
1049         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1050         if (!atomic_dec_and_test(&cfqg->ref))
1051                 return;
1052         for_each_cfqg_st(cfqg, i, j, st)
1053                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1054         kfree(cfqg);
1055 }
1056
1057 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1058 {
1059         /* Something wrong if we are trying to remove same group twice */
1060         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1061
1062         hlist_del_init(&cfqg->cfqd_node);
1063
1064         /*
1065          * Put the reference taken at the time of creation so that when all
1066          * queues are gone, group can be destroyed.
1067          */
1068         cfq_put_cfqg(cfqg);
1069 }
1070
1071 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1072 {
1073         struct hlist_node *pos, *n;
1074         struct cfq_group *cfqg;
1075
1076         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1077                 /*
1078                  * If cgroup removal path got to blk_group first and removed
1079                  * it from cgroup list, then it will take care of destroying
1080                  * cfqg also.
1081                  */
1082                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1083                         cfq_destroy_cfqg(cfqd, cfqg);
1084         }
1085 }
1086
1087 /*
1088  * Blk cgroup controller notification saying that blkio_group object is being
1089  * delinked as associated cgroup object is going away. That also means that
1090  * no new IO will come in this group. So get rid of this group as soon as
1091  * any pending IO in the group is finished.
1092  *
1093  * This function is called under rcu_read_lock(). key is the rcu protected
1094  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1095  * read lock.
1096  *
1097  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1098  * it should not be NULL as even if elevator was exiting, cgroup deltion
1099  * path got to it first.
1100  */
1101 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1102 {
1103         unsigned long  flags;
1104         struct cfq_data *cfqd = key;
1105
1106         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1107         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1108         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1109 }
1110
1111 #else /* GROUP_IOSCHED */
1112 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1113 {
1114         return &cfqd->root_group;
1115 }
1116
1117 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1118 {
1119         return cfqg;
1120 }
1121
1122 static inline void
1123 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1124         cfqq->cfqg = cfqg;
1125 }
1126
1127 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1128 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1129
1130 #endif /* GROUP_IOSCHED */
1131
1132 /*
1133  * The cfqd->service_trees holds all pending cfq_queue's that have
1134  * requests waiting to be processed. It is sorted in the order that
1135  * we will service the queues.
1136  */
1137 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1138                                  bool add_front)
1139 {
1140         struct rb_node **p, *parent;
1141         struct cfq_queue *__cfqq;
1142         unsigned long rb_key;
1143         struct cfq_rb_root *service_tree;
1144         int left;
1145         int new_cfqq = 1;
1146         int group_changed = 0;
1147
1148 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1149         if (!cfqd->cfq_group_isolation
1150             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1151             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1152                 /* Move this cfq to root group */
1153                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1154                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1155                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1156                 cfqq->orig_cfqg = cfqq->cfqg;
1157                 cfqq->cfqg = &cfqd->root_group;
1158                 atomic_inc(&cfqd->root_group.ref);
1159                 group_changed = 1;
1160         } else if (!cfqd->cfq_group_isolation
1161                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1162                 /* cfqq is sequential now needs to go to its original group */
1163                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1164                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1165                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1166                 cfq_put_cfqg(cfqq->cfqg);
1167                 cfqq->cfqg = cfqq->orig_cfqg;
1168                 cfqq->orig_cfqg = NULL;
1169                 group_changed = 1;
1170                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1171         }
1172 #endif
1173
1174         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1175                                                 cfqq_type(cfqq));
1176         if (cfq_class_idle(cfqq)) {
1177                 rb_key = CFQ_IDLE_DELAY;
1178                 parent = rb_last(&service_tree->rb);
1179                 if (parent && parent != &cfqq->rb_node) {
1180                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181                         rb_key += __cfqq->rb_key;
1182                 } else
1183                         rb_key += jiffies;
1184         } else if (!add_front) {
1185                 /*
1186                  * Get our rb key offset. Subtract any residual slice
1187                  * value carried from last service. A negative resid
1188                  * count indicates slice overrun, and this should position
1189                  * the next service time further away in the tree.
1190                  */
1191                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1192                 rb_key -= cfqq->slice_resid;
1193                 cfqq->slice_resid = 0;
1194         } else {
1195                 rb_key = -HZ;
1196                 __cfqq = cfq_rb_first(service_tree);
1197                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1198         }
1199
1200         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1201                 new_cfqq = 0;
1202                 /*
1203                  * same position, nothing more to do
1204                  */
1205                 if (rb_key == cfqq->rb_key &&
1206                     cfqq->service_tree == service_tree)
1207                         return;
1208
1209                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1210                 cfqq->service_tree = NULL;
1211         }
1212
1213         left = 1;
1214         parent = NULL;
1215         cfqq->service_tree = service_tree;
1216         p = &service_tree->rb.rb_node;
1217         while (*p) {
1218                 struct rb_node **n;
1219
1220                 parent = *p;
1221                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1222
1223                 /*
1224                  * sort by key, that represents service time.
1225                  */
1226                 if (time_before(rb_key, __cfqq->rb_key))
1227                         n = &(*p)->rb_left;
1228                 else {
1229                         n = &(*p)->rb_right;
1230                         left = 0;
1231                 }
1232
1233                 p = n;
1234         }
1235
1236         if (left)
1237                 service_tree->left = &cfqq->rb_node;
1238
1239         cfqq->rb_key = rb_key;
1240         rb_link_node(&cfqq->rb_node, parent, p);
1241         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1242         service_tree->count++;
1243         if ((add_front || !new_cfqq) && !group_changed)
1244                 return;
1245         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1246 }
1247
1248 static struct cfq_queue *
1249 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1250                      sector_t sector, struct rb_node **ret_parent,
1251                      struct rb_node ***rb_link)
1252 {
1253         struct rb_node **p, *parent;
1254         struct cfq_queue *cfqq = NULL;
1255
1256         parent = NULL;
1257         p = &root->rb_node;
1258         while (*p) {
1259                 struct rb_node **n;
1260
1261                 parent = *p;
1262                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1263
1264                 /*
1265                  * Sort strictly based on sector.  Smallest to the left,
1266                  * largest to the right.
1267                  */
1268                 if (sector > blk_rq_pos(cfqq->next_rq))
1269                         n = &(*p)->rb_right;
1270                 else if (sector < blk_rq_pos(cfqq->next_rq))
1271                         n = &(*p)->rb_left;
1272                 else
1273                         break;
1274                 p = n;
1275                 cfqq = NULL;
1276         }
1277
1278         *ret_parent = parent;
1279         if (rb_link)
1280                 *rb_link = p;
1281         return cfqq;
1282 }
1283
1284 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1285 {
1286         struct rb_node **p, *parent;
1287         struct cfq_queue *__cfqq;
1288
1289         if (cfqq->p_root) {
1290                 rb_erase(&cfqq->p_node, cfqq->p_root);
1291                 cfqq->p_root = NULL;
1292         }
1293
1294         if (cfq_class_idle(cfqq))
1295                 return;
1296         if (!cfqq->next_rq)
1297                 return;
1298
1299         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1300         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1301                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1302         if (!__cfqq) {
1303                 rb_link_node(&cfqq->p_node, parent, p);
1304                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1305         } else
1306                 cfqq->p_root = NULL;
1307 }
1308
1309 /*
1310  * Update cfqq's position in the service tree.
1311  */
1312 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1313 {
1314         /*
1315          * Resorting requires the cfqq to be on the RR list already.
1316          */
1317         if (cfq_cfqq_on_rr(cfqq)) {
1318                 cfq_service_tree_add(cfqd, cfqq, 0);
1319                 cfq_prio_tree_add(cfqd, cfqq);
1320         }
1321 }
1322
1323 /*
1324  * add to busy list of queues for service, trying to be fair in ordering
1325  * the pending list according to last request service
1326  */
1327 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328 {
1329         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1330         BUG_ON(cfq_cfqq_on_rr(cfqq));
1331         cfq_mark_cfqq_on_rr(cfqq);
1332         cfqd->busy_queues++;
1333
1334         cfq_resort_rr_list(cfqd, cfqq);
1335 }
1336
1337 /*
1338  * Called when the cfqq no longer has requests pending, remove it from
1339  * the service tree.
1340  */
1341 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1342 {
1343         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1344         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1345         cfq_clear_cfqq_on_rr(cfqq);
1346
1347         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1348                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1349                 cfqq->service_tree = NULL;
1350         }
1351         if (cfqq->p_root) {
1352                 rb_erase(&cfqq->p_node, cfqq->p_root);
1353                 cfqq->p_root = NULL;
1354         }
1355
1356         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1357         BUG_ON(!cfqd->busy_queues);
1358         cfqd->busy_queues--;
1359 }
1360
1361 /*
1362  * rb tree support functions
1363  */
1364 static void cfq_del_rq_rb(struct request *rq)
1365 {
1366         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1367         const int sync = rq_is_sync(rq);
1368
1369         BUG_ON(!cfqq->queued[sync]);
1370         cfqq->queued[sync]--;
1371
1372         elv_rb_del(&cfqq->sort_list, rq);
1373
1374         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1375                 /*
1376                  * Queue will be deleted from service tree when we actually
1377                  * expire it later. Right now just remove it from prio tree
1378                  * as it is empty.
1379                  */
1380                 if (cfqq->p_root) {
1381                         rb_erase(&cfqq->p_node, cfqq->p_root);
1382                         cfqq->p_root = NULL;
1383                 }
1384         }
1385 }
1386
1387 static void cfq_add_rq_rb(struct request *rq)
1388 {
1389         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1390         struct cfq_data *cfqd = cfqq->cfqd;
1391         struct request *__alias, *prev;
1392
1393         cfqq->queued[rq_is_sync(rq)]++;
1394
1395         /*
1396          * looks a little odd, but the first insert might return an alias.
1397          * if that happens, put the alias on the dispatch list
1398          */
1399         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1400                 cfq_dispatch_insert(cfqd->queue, __alias);
1401
1402         if (!cfq_cfqq_on_rr(cfqq))
1403                 cfq_add_cfqq_rr(cfqd, cfqq);
1404
1405         /*
1406          * check if this request is a better next-serve candidate
1407          */
1408         prev = cfqq->next_rq;
1409         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1410
1411         /*
1412          * adjust priority tree position, if ->next_rq changes
1413          */
1414         if (prev != cfqq->next_rq)
1415                 cfq_prio_tree_add(cfqd, cfqq);
1416
1417         BUG_ON(!cfqq->next_rq);
1418 }
1419
1420 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1421 {
1422         elv_rb_del(&cfqq->sort_list, rq);
1423         cfqq->queued[rq_is_sync(rq)]--;
1424         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1425                                         rq_data_dir(rq), rq_is_sync(rq));
1426         cfq_add_rq_rb(rq);
1427         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1428                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1429                         rq_is_sync(rq));
1430 }
1431
1432 static struct request *
1433 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1434 {
1435         struct task_struct *tsk = current;
1436         struct cfq_io_context *cic;
1437         struct cfq_queue *cfqq;
1438
1439         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1440         if (!cic)
1441                 return NULL;
1442
1443         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1444         if (cfqq) {
1445                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1446
1447                 return elv_rb_find(&cfqq->sort_list, sector);
1448         }
1449
1450         return NULL;
1451 }
1452
1453 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1454 {
1455         struct cfq_data *cfqd = q->elevator->elevator_data;
1456
1457         cfqd->rq_in_driver++;
1458         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1459                                                 cfqd->rq_in_driver);
1460
1461         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1462 }
1463
1464 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1465 {
1466         struct cfq_data *cfqd = q->elevator->elevator_data;
1467
1468         WARN_ON(!cfqd->rq_in_driver);
1469         cfqd->rq_in_driver--;
1470         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1471                                                 cfqd->rq_in_driver);
1472 }
1473
1474 static void cfq_remove_request(struct request *rq)
1475 {
1476         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1477
1478         if (cfqq->next_rq == rq)
1479                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1480
1481         list_del_init(&rq->queuelist);
1482         cfq_del_rq_rb(rq);
1483
1484         cfqq->cfqd->rq_queued--;
1485         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1486                                         rq_data_dir(rq), rq_is_sync(rq));
1487         if (rq->cmd_flags & REQ_META) {
1488                 WARN_ON(!cfqq->meta_pending);
1489                 cfqq->meta_pending--;
1490         }
1491 }
1492
1493 static int cfq_merge(struct request_queue *q, struct request **req,
1494                      struct bio *bio)
1495 {
1496         struct cfq_data *cfqd = q->elevator->elevator_data;
1497         struct request *__rq;
1498
1499         __rq = cfq_find_rq_fmerge(cfqd, bio);
1500         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1501                 *req = __rq;
1502                 return ELEVATOR_FRONT_MERGE;
1503         }
1504
1505         return ELEVATOR_NO_MERGE;
1506 }
1507
1508 static void cfq_merged_request(struct request_queue *q, struct request *req,
1509                                int type)
1510 {
1511         if (type == ELEVATOR_FRONT_MERGE) {
1512                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1513
1514                 cfq_reposition_rq_rb(cfqq, req);
1515         }
1516 }
1517
1518 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1519                                 struct bio *bio)
1520 {
1521         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1522                                         bio_data_dir(bio), cfq_bio_sync(bio));
1523 }
1524
1525 static void
1526 cfq_merged_requests(struct request_queue *q, struct request *rq,
1527                     struct request *next)
1528 {
1529         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1530         /*
1531          * reposition in fifo if next is older than rq
1532          */
1533         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1534             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1535                 list_move(&rq->queuelist, &next->queuelist);
1536                 rq_set_fifo_time(rq, rq_fifo_time(next));
1537         }
1538
1539         if (cfqq->next_rq == next)
1540                 cfqq->next_rq = rq;
1541         cfq_remove_request(next);
1542         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1543                                         rq_data_dir(next), rq_is_sync(next));
1544 }
1545
1546 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1547                            struct bio *bio)
1548 {
1549         struct cfq_data *cfqd = q->elevator->elevator_data;
1550         struct cfq_io_context *cic;
1551         struct cfq_queue *cfqq;
1552
1553         /*
1554          * Disallow merge of a sync bio into an async request.
1555          */
1556         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1557                 return false;
1558
1559         /*
1560          * Lookup the cfqq that this bio will be queued with. Allow
1561          * merge only if rq is queued there.
1562          */
1563         cic = cfq_cic_lookup(cfqd, current->io_context);
1564         if (!cic)
1565                 return false;
1566
1567         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1568         return cfqq == RQ_CFQQ(rq);
1569 }
1570
1571 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1572 {
1573         del_timer(&cfqd->idle_slice_timer);
1574         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1575 }
1576
1577 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1578                                    struct cfq_queue *cfqq)
1579 {
1580         if (cfqq) {
1581                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1582                                 cfqd->serving_prio, cfqd->serving_type);
1583                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1584                 cfqq->slice_start = 0;
1585                 cfqq->dispatch_start = jiffies;
1586                 cfqq->allocated_slice = 0;
1587                 cfqq->slice_end = 0;
1588                 cfqq->slice_dispatch = 0;
1589
1590                 cfq_clear_cfqq_wait_request(cfqq);
1591                 cfq_clear_cfqq_must_dispatch(cfqq);
1592                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1593                 cfq_clear_cfqq_fifo_expire(cfqq);
1594                 cfq_mark_cfqq_slice_new(cfqq);
1595
1596                 cfq_del_timer(cfqd, cfqq);
1597         }
1598
1599         cfqd->active_queue = cfqq;
1600 }
1601
1602 /*
1603  * current cfqq expired its slice (or was too idle), select new one
1604  */
1605 static void
1606 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1607                     bool timed_out)
1608 {
1609         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1610
1611         if (cfq_cfqq_wait_request(cfqq))
1612                 cfq_del_timer(cfqd, cfqq);
1613
1614         cfq_clear_cfqq_wait_request(cfqq);
1615         cfq_clear_cfqq_wait_busy(cfqq);
1616
1617         /*
1618          * If this cfqq is shared between multiple processes, check to
1619          * make sure that those processes are still issuing I/Os within
1620          * the mean seek distance.  If not, it may be time to break the
1621          * queues apart again.
1622          */
1623         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1624                 cfq_mark_cfqq_split_coop(cfqq);
1625
1626         /*
1627          * store what was left of this slice, if the queue idled/timed out
1628          */
1629         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1630                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1631                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1632         }
1633
1634         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1635
1636         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1637                 cfq_del_cfqq_rr(cfqd, cfqq);
1638
1639         cfq_resort_rr_list(cfqd, cfqq);
1640
1641         if (cfqq == cfqd->active_queue)
1642                 cfqd->active_queue = NULL;
1643
1644         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1645                 cfqd->grp_service_tree.active = NULL;
1646
1647         if (cfqd->active_cic) {
1648                 put_io_context(cfqd->active_cic->ioc);
1649                 cfqd->active_cic = NULL;
1650         }
1651 }
1652
1653 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1654 {
1655         struct cfq_queue *cfqq = cfqd->active_queue;
1656
1657         if (cfqq)
1658                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1659 }
1660
1661 /*
1662  * Get next queue for service. Unless we have a queue preemption,
1663  * we'll simply select the first cfqq in the service tree.
1664  */
1665 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1666 {
1667         struct cfq_rb_root *service_tree =
1668                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1669                                         cfqd->serving_type);
1670
1671         if (!cfqd->rq_queued)
1672                 return NULL;
1673
1674         /* There is nothing to dispatch */
1675         if (!service_tree)
1676                 return NULL;
1677         if (RB_EMPTY_ROOT(&service_tree->rb))
1678                 return NULL;
1679         return cfq_rb_first(service_tree);
1680 }
1681
1682 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1683 {
1684         struct cfq_group *cfqg;
1685         struct cfq_queue *cfqq;
1686         int i, j;
1687         struct cfq_rb_root *st;
1688
1689         if (!cfqd->rq_queued)
1690                 return NULL;
1691
1692         cfqg = cfq_get_next_cfqg(cfqd);
1693         if (!cfqg)
1694                 return NULL;
1695
1696         for_each_cfqg_st(cfqg, i, j, st)
1697                 if ((cfqq = cfq_rb_first(st)) != NULL)
1698                         return cfqq;
1699         return NULL;
1700 }
1701
1702 /*
1703  * Get and set a new active queue for service.
1704  */
1705 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1706                                               struct cfq_queue *cfqq)
1707 {
1708         if (!cfqq)
1709                 cfqq = cfq_get_next_queue(cfqd);
1710
1711         __cfq_set_active_queue(cfqd, cfqq);
1712         return cfqq;
1713 }
1714
1715 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1716                                           struct request *rq)
1717 {
1718         if (blk_rq_pos(rq) >= cfqd->last_position)
1719                 return blk_rq_pos(rq) - cfqd->last_position;
1720         else
1721                 return cfqd->last_position - blk_rq_pos(rq);
1722 }
1723
1724 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1725                                struct request *rq)
1726 {
1727         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1728 }
1729
1730 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1731                                     struct cfq_queue *cur_cfqq)
1732 {
1733         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1734         struct rb_node *parent, *node;
1735         struct cfq_queue *__cfqq;
1736         sector_t sector = cfqd->last_position;
1737
1738         if (RB_EMPTY_ROOT(root))
1739                 return NULL;
1740
1741         /*
1742          * First, if we find a request starting at the end of the last
1743          * request, choose it.
1744          */
1745         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1746         if (__cfqq)
1747                 return __cfqq;
1748
1749         /*
1750          * If the exact sector wasn't found, the parent of the NULL leaf
1751          * will contain the closest sector.
1752          */
1753         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1754         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1755                 return __cfqq;
1756
1757         if (blk_rq_pos(__cfqq->next_rq) < sector)
1758                 node = rb_next(&__cfqq->p_node);
1759         else
1760                 node = rb_prev(&__cfqq->p_node);
1761         if (!node)
1762                 return NULL;
1763
1764         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1765         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1766                 return __cfqq;
1767
1768         return NULL;
1769 }
1770
1771 /*
1772  * cfqd - obvious
1773  * cur_cfqq - passed in so that we don't decide that the current queue is
1774  *            closely cooperating with itself.
1775  *
1776  * So, basically we're assuming that that cur_cfqq has dispatched at least
1777  * one request, and that cfqd->last_position reflects a position on the disk
1778  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1779  * assumption.
1780  */
1781 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1782                                               struct cfq_queue *cur_cfqq)
1783 {
1784         struct cfq_queue *cfqq;
1785
1786         if (cfq_class_idle(cur_cfqq))
1787                 return NULL;
1788         if (!cfq_cfqq_sync(cur_cfqq))
1789                 return NULL;
1790         if (CFQQ_SEEKY(cur_cfqq))
1791                 return NULL;
1792
1793         /*
1794          * Don't search priority tree if it's the only queue in the group.
1795          */
1796         if (cur_cfqq->cfqg->nr_cfqq == 1)
1797                 return NULL;
1798
1799         /*
1800          * We should notice if some of the queues are cooperating, eg
1801          * working closely on the same area of the disk. In that case,
1802          * we can group them together and don't waste time idling.
1803          */
1804         cfqq = cfqq_close(cfqd, cur_cfqq);
1805         if (!cfqq)
1806                 return NULL;
1807
1808         /* If new queue belongs to different cfq_group, don't choose it */
1809         if (cur_cfqq->cfqg != cfqq->cfqg)
1810                 return NULL;
1811
1812         /*
1813          * It only makes sense to merge sync queues.
1814          */
1815         if (!cfq_cfqq_sync(cfqq))
1816                 return NULL;
1817         if (CFQQ_SEEKY(cfqq))
1818                 return NULL;
1819
1820         /*
1821          * Do not merge queues of different priority classes
1822          */
1823         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1824                 return NULL;
1825
1826         return cfqq;
1827 }
1828
1829 /*
1830  * Determine whether we should enforce idle window for this queue.
1831  */
1832
1833 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1834 {
1835         enum wl_prio_t prio = cfqq_prio(cfqq);
1836         struct cfq_rb_root *service_tree = cfqq->service_tree;
1837
1838         BUG_ON(!service_tree);
1839         BUG_ON(!service_tree->count);
1840
1841         /* We never do for idle class queues. */
1842         if (prio == IDLE_WORKLOAD)
1843                 return false;
1844
1845         /* We do for queues that were marked with idle window flag. */
1846         if (cfq_cfqq_idle_window(cfqq) &&
1847            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1848                 return true;
1849
1850         /*
1851          * Otherwise, we do only if they are the last ones
1852          * in their service tree.
1853          */
1854         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1855                 return 1;
1856         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1857                         service_tree->count);
1858         return 0;
1859 }
1860
1861 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1862 {
1863         struct cfq_queue *cfqq = cfqd->active_queue;
1864         struct cfq_io_context *cic;
1865         unsigned long sl;
1866
1867         /*
1868          * SSD device without seek penalty, disable idling. But only do so
1869          * for devices that support queuing, otherwise we still have a problem
1870          * with sync vs async workloads.
1871          */
1872         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1873                 return;
1874
1875         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1876         WARN_ON(cfq_cfqq_slice_new(cfqq));
1877
1878         /*
1879          * idle is disabled, either manually or by past process history
1880          */
1881         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1882                 return;
1883
1884         /*
1885          * still active requests from this queue, don't idle
1886          */
1887         if (cfqq->dispatched)
1888                 return;
1889
1890         /*
1891          * task has exited, don't wait
1892          */
1893         cic = cfqd->active_cic;
1894         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1895                 return;
1896
1897         /*
1898          * If our average think time is larger than the remaining time
1899          * slice, then don't idle. This avoids overrunning the allotted
1900          * time slice.
1901          */
1902         if (sample_valid(cic->ttime_samples) &&
1903             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1904                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1905                                 cic->ttime_mean);
1906                 return;
1907         }
1908
1909         cfq_mark_cfqq_wait_request(cfqq);
1910
1911         sl = cfqd->cfq_slice_idle;
1912
1913         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1914         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1915         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1916 }
1917
1918 /*
1919  * Move request from internal lists to the request queue dispatch list.
1920  */
1921 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1922 {
1923         struct cfq_data *cfqd = q->elevator->elevator_data;
1924         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1925
1926         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1927
1928         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1929         cfq_remove_request(rq);
1930         cfqq->dispatched++;
1931         elv_dispatch_sort(q, rq);
1932
1933         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1934         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1935                                         rq_data_dir(rq), rq_is_sync(rq));
1936 }
1937
1938 /*
1939  * return expired entry, or NULL to just start from scratch in rbtree
1940  */
1941 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1942 {
1943         struct request *rq = NULL;
1944
1945         if (cfq_cfqq_fifo_expire(cfqq))
1946                 return NULL;
1947
1948         cfq_mark_cfqq_fifo_expire(cfqq);
1949
1950         if (list_empty(&cfqq->fifo))
1951                 return NULL;
1952
1953         rq = rq_entry_fifo(cfqq->fifo.next);
1954         if (time_before(jiffies, rq_fifo_time(rq)))
1955                 rq = NULL;
1956
1957         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1958         return rq;
1959 }
1960
1961 static inline int
1962 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1963 {
1964         const int base_rq = cfqd->cfq_slice_async_rq;
1965
1966         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1967
1968         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1969 }
1970
1971 /*
1972  * Must be called with the queue_lock held.
1973  */
1974 static int cfqq_process_refs(struct cfq_queue *cfqq)
1975 {
1976         int process_refs, io_refs;
1977
1978         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1979         process_refs = atomic_read(&cfqq->ref) - io_refs;
1980         BUG_ON(process_refs < 0);
1981         return process_refs;
1982 }
1983
1984 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1985 {
1986         int process_refs, new_process_refs;
1987         struct cfq_queue *__cfqq;
1988
1989         /*
1990          * If there are no process references on the new_cfqq, then it is
1991          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1992          * chain may have dropped their last reference (not just their
1993          * last process reference).
1994          */
1995         if (!cfqq_process_refs(new_cfqq))
1996                 return;
1997
1998         /* Avoid a circular list and skip interim queue merges */
1999         while ((__cfqq = new_cfqq->new_cfqq)) {
2000                 if (__cfqq == cfqq)
2001                         return;
2002                 new_cfqq = __cfqq;
2003         }
2004
2005         process_refs = cfqq_process_refs(cfqq);
2006         new_process_refs = cfqq_process_refs(new_cfqq);
2007         /*
2008          * If the process for the cfqq has gone away, there is no
2009          * sense in merging the queues.
2010          */
2011         if (process_refs == 0 || new_process_refs == 0)
2012                 return;
2013
2014         /*
2015          * Merge in the direction of the lesser amount of work.
2016          */
2017         if (new_process_refs >= process_refs) {
2018                 cfqq->new_cfqq = new_cfqq;
2019                 atomic_add(process_refs, &new_cfqq->ref);
2020         } else {
2021                 new_cfqq->new_cfqq = cfqq;
2022                 atomic_add(new_process_refs, &cfqq->ref);
2023         }
2024 }
2025
2026 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2027                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2028 {
2029         struct cfq_queue *queue;
2030         int i;
2031         bool key_valid = false;
2032         unsigned long lowest_key = 0;
2033         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2034
2035         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2036                 /* select the one with lowest rb_key */
2037                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2038                 if (queue &&
2039                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2040                         lowest_key = queue->rb_key;
2041                         cur_best = i;
2042                         key_valid = true;
2043                 }
2044         }
2045
2046         return cur_best;
2047 }
2048
2049 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2050 {
2051         unsigned slice;
2052         unsigned count;
2053         struct cfq_rb_root *st;
2054         unsigned group_slice;
2055
2056         if (!cfqg) {
2057                 cfqd->serving_prio = IDLE_WORKLOAD;
2058                 cfqd->workload_expires = jiffies + 1;
2059                 return;
2060         }
2061
2062         /* Choose next priority. RT > BE > IDLE */
2063         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2064                 cfqd->serving_prio = RT_WORKLOAD;
2065         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2066                 cfqd->serving_prio = BE_WORKLOAD;
2067         else {
2068                 cfqd->serving_prio = IDLE_WORKLOAD;
2069                 cfqd->workload_expires = jiffies + 1;
2070                 return;
2071         }
2072
2073         /*
2074          * For RT and BE, we have to choose also the type
2075          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2076          * expiration time
2077          */
2078         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2079         count = st->count;
2080
2081         /*
2082          * check workload expiration, and that we still have other queues ready
2083          */
2084         if (count && !time_after(jiffies, cfqd->workload_expires))
2085                 return;
2086
2087         /* otherwise select new workload type */
2088         cfqd->serving_type =
2089                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2090         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2091         count = st->count;
2092
2093         /*
2094          * the workload slice is computed as a fraction of target latency
2095          * proportional to the number of queues in that workload, over
2096          * all the queues in the same priority class
2097          */
2098         group_slice = cfq_group_slice(cfqd, cfqg);
2099
2100         slice = group_slice * count /
2101                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2102                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2103
2104         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2105                 unsigned int tmp;
2106
2107                 /*
2108                  * Async queues are currently system wide. Just taking
2109                  * proportion of queues with-in same group will lead to higher
2110                  * async ratio system wide as generally root group is going
2111                  * to have higher weight. A more accurate thing would be to
2112                  * calculate system wide asnc/sync ratio.
2113                  */
2114                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2115                 tmp = tmp/cfqd->busy_queues;
2116                 slice = min_t(unsigned, slice, tmp);
2117
2118                 /* async workload slice is scaled down according to
2119                  * the sync/async slice ratio. */
2120                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2121         } else
2122                 /* sync workload slice is at least 2 * cfq_slice_idle */
2123                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2124
2125         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2126         cfq_log(cfqd, "workload slice:%d", slice);
2127         cfqd->workload_expires = jiffies + slice;
2128 }
2129
2130 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2131 {
2132         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2133         struct cfq_group *cfqg;
2134
2135         if (RB_EMPTY_ROOT(&st->rb))
2136                 return NULL;
2137         cfqg = cfq_rb_first_group(st);
2138         st->active = &cfqg->rb_node;
2139         update_min_vdisktime(st);
2140         return cfqg;
2141 }
2142
2143 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2144 {
2145         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2146
2147         cfqd->serving_group = cfqg;
2148
2149         /* Restore the workload type data */
2150         if (cfqg->saved_workload_slice) {
2151                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2152                 cfqd->serving_type = cfqg->saved_workload;
2153                 cfqd->serving_prio = cfqg->saved_serving_prio;
2154         } else
2155                 cfqd->workload_expires = jiffies - 1;
2156
2157         choose_service_tree(cfqd, cfqg);
2158 }
2159
2160 /*
2161  * Select a queue for service. If we have a current active queue,
2162  * check whether to continue servicing it, or retrieve and set a new one.
2163  */
2164 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2165 {
2166         struct cfq_queue *cfqq, *new_cfqq = NULL;
2167
2168         cfqq = cfqd->active_queue;
2169         if (!cfqq)
2170                 goto new_queue;
2171
2172         if (!cfqd->rq_queued)
2173                 return NULL;
2174
2175         /*
2176          * We were waiting for group to get backlogged. Expire the queue
2177          */
2178         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2179                 goto expire;
2180
2181         /*
2182          * The active queue has run out of time, expire it and select new.
2183          */
2184         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2185                 /*
2186                  * If slice had not expired at the completion of last request
2187                  * we might not have turned on wait_busy flag. Don't expire
2188                  * the queue yet. Allow the group to get backlogged.
2189                  *
2190                  * The very fact that we have used the slice, that means we
2191                  * have been idling all along on this queue and it should be
2192                  * ok to wait for this request to complete.
2193                  */
2194                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2195                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2196                         cfqq = NULL;
2197                         goto keep_queue;
2198                 } else
2199                         goto expire;
2200         }
2201
2202         /*
2203          * The active queue has requests and isn't expired, allow it to
2204          * dispatch.
2205          */
2206         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2207                 goto keep_queue;
2208
2209         /*
2210          * If another queue has a request waiting within our mean seek
2211          * distance, let it run.  The expire code will check for close
2212          * cooperators and put the close queue at the front of the service
2213          * tree.  If possible, merge the expiring queue with the new cfqq.
2214          */
2215         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2216         if (new_cfqq) {
2217                 if (!cfqq->new_cfqq)
2218                         cfq_setup_merge(cfqq, new_cfqq);
2219                 goto expire;
2220         }
2221
2222         /*
2223          * No requests pending. If the active queue still has requests in
2224          * flight or is idling for a new request, allow either of these
2225          * conditions to happen (or time out) before selecting a new queue.
2226          */
2227         if (timer_pending(&cfqd->idle_slice_timer) ||
2228             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2229                 cfqq = NULL;
2230                 goto keep_queue;
2231         }
2232
2233 expire:
2234         cfq_slice_expired(cfqd, 0);
2235 new_queue:
2236         /*
2237          * Current queue expired. Check if we have to switch to a new
2238          * service tree
2239          */
2240         if (!new_cfqq)
2241                 cfq_choose_cfqg(cfqd);
2242
2243         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2244 keep_queue:
2245         return cfqq;
2246 }
2247
2248 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2249 {
2250         int dispatched = 0;
2251
2252         while (cfqq->next_rq) {
2253                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2254                 dispatched++;
2255         }
2256
2257         BUG_ON(!list_empty(&cfqq->fifo));
2258
2259         /* By default cfqq is not expired if it is empty. Do it explicitly */
2260         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2261         return dispatched;
2262 }
2263
2264 /*
2265  * Drain our current requests. Used for barriers and when switching
2266  * io schedulers on-the-fly.
2267  */
2268 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2269 {
2270         struct cfq_queue *cfqq;
2271         int dispatched = 0;
2272
2273         /* Expire the timeslice of the current active queue first */
2274         cfq_slice_expired(cfqd, 0);
2275         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2276                 __cfq_set_active_queue(cfqd, cfqq);
2277                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2278         }
2279
2280         BUG_ON(cfqd->busy_queues);
2281
2282         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2283         return dispatched;
2284 }
2285
2286 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2287         struct cfq_queue *cfqq)
2288 {
2289         /* the queue hasn't finished any request, can't estimate */
2290         if (cfq_cfqq_slice_new(cfqq))
2291                 return 1;
2292         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2293                 cfqq->slice_end))
2294                 return 1;
2295
2296         return 0;
2297 }
2298
2299 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2300 {
2301         unsigned int max_dispatch;
2302
2303         /*
2304          * Drain async requests before we start sync IO
2305          */
2306         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2307                 return false;
2308
2309         /*
2310          * If this is an async queue and we have sync IO in flight, let it wait
2311          */
2312         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2313                 return false;
2314
2315         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2316         if (cfq_class_idle(cfqq))
2317                 max_dispatch = 1;
2318
2319         /*
2320          * Does this cfqq already have too much IO in flight?
2321          */
2322         if (cfqq->dispatched >= max_dispatch) {
2323                 /*
2324                  * idle queue must always only have a single IO in flight
2325                  */
2326                 if (cfq_class_idle(cfqq))
2327                         return false;
2328
2329                 /*
2330                  * We have other queues, don't allow more IO from this one
2331                  */
2332                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2333                         return false;
2334
2335                 /*
2336                  * Sole queue user, no limit
2337                  */
2338                 if (cfqd->busy_queues == 1)
2339                         max_dispatch = -1;
2340                 else
2341                         /*
2342                          * Normally we start throttling cfqq when cfq_quantum/2
2343                          * requests have been dispatched. But we can drive
2344                          * deeper queue depths at the beginning of slice
2345                          * subjected to upper limit of cfq_quantum.
2346                          * */
2347                         max_dispatch = cfqd->cfq_quantum;
2348         }
2349
2350         /*
2351          * Async queues must wait a bit before being allowed dispatch.
2352          * We also ramp up the dispatch depth gradually for async IO,
2353          * based on the last sync IO we serviced
2354          */
2355         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2356                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2357                 unsigned int depth;
2358
2359                 depth = last_sync / cfqd->cfq_slice[1];
2360                 if (!depth && !cfqq->dispatched)
2361                         depth = 1;
2362                 if (depth < max_dispatch)
2363                         max_dispatch = depth;
2364         }
2365
2366         /*
2367          * If we're below the current max, allow a dispatch
2368          */
2369         return cfqq->dispatched < max_dispatch;
2370 }
2371
2372 /*
2373  * Dispatch a request from cfqq, moving them to the request queue
2374  * dispatch list.
2375  */
2376 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2377 {
2378         struct request *rq;
2379
2380         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2381
2382         if (!cfq_may_dispatch(cfqd, cfqq))
2383                 return false;
2384
2385         /*
2386          * follow expired path, else get first next available
2387          */
2388         rq = cfq_check_fifo(cfqq);
2389         if (!rq)
2390                 rq = cfqq->next_rq;
2391
2392         /*
2393          * insert request into driver dispatch list
2394          */
2395         cfq_dispatch_insert(cfqd->queue, rq);
2396
2397         if (!cfqd->active_cic) {
2398                 struct cfq_io_context *cic = RQ_CIC(rq);
2399
2400                 atomic_long_inc(&cic->ioc->refcount);
2401                 cfqd->active_cic = cic;
2402         }
2403
2404         return true;
2405 }
2406
2407 /*
2408  * Find the cfqq that we need to service and move a request from that to the
2409  * dispatch list
2410  */
2411 static int cfq_dispatch_requests(struct request_queue *q, int force)
2412 {
2413         struct cfq_data *cfqd = q->elevator->elevator_data;
2414         struct cfq_queue *cfqq;
2415
2416         if (!cfqd->busy_queues)
2417                 return 0;
2418
2419         if (unlikely(force))
2420                 return cfq_forced_dispatch(cfqd);
2421
2422         cfqq = cfq_select_queue(cfqd);
2423         if (!cfqq)
2424                 return 0;
2425
2426         /*
2427          * Dispatch a request from this cfqq, if it is allowed
2428          */
2429         if (!cfq_dispatch_request(cfqd, cfqq))
2430                 return 0;
2431
2432         cfqq->slice_dispatch++;
2433         cfq_clear_cfqq_must_dispatch(cfqq);
2434
2435         /*
2436          * expire an async queue immediately if it has used up its slice. idle
2437          * queue always expire after 1 dispatch round.
2438          */
2439         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2440             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2441             cfq_class_idle(cfqq))) {
2442                 cfqq->slice_end = jiffies + 1;
2443                 cfq_slice_expired(cfqd, 0);
2444         }
2445
2446         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2447         return 1;
2448 }
2449
2450 /*
2451  * task holds one reference to the queue, dropped when task exits. each rq
2452  * in-flight on this queue also holds a reference, dropped when rq is freed.
2453  *
2454  * Each cfq queue took a reference on the parent group. Drop it now.
2455  * queue lock must be held here.
2456  */
2457 static void cfq_put_queue(struct cfq_queue *cfqq)
2458 {
2459         struct cfq_data *cfqd = cfqq->cfqd;
2460         struct cfq_group *cfqg, *orig_cfqg;
2461
2462         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2463
2464         if (!atomic_dec_and_test(&cfqq->ref))
2465                 return;
2466
2467         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2468         BUG_ON(rb_first(&cfqq->sort_list));
2469         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2470         cfqg = cfqq->cfqg;
2471         orig_cfqg = cfqq->orig_cfqg;
2472
2473         if (unlikely(cfqd->active_queue == cfqq)) {
2474                 __cfq_slice_expired(cfqd, cfqq, 0);
2475                 cfq_schedule_dispatch(cfqd);
2476         }
2477
2478         BUG_ON(cfq_cfqq_on_rr(cfqq));
2479         kmem_cache_free(cfq_pool, cfqq);
2480         cfq_put_cfqg(cfqg);
2481         if (orig_cfqg)
2482                 cfq_put_cfqg(orig_cfqg);
2483 }
2484
2485 /*
2486  * Must always be called with the rcu_read_lock() held
2487  */
2488 static void
2489 __call_for_each_cic(struct io_context *ioc,
2490                     void (*func)(struct io_context *, struct cfq_io_context *))
2491 {
2492         struct cfq_io_context *cic;
2493         struct hlist_node *n;
2494
2495         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2496                 func(ioc, cic);
2497 }
2498
2499 /*
2500  * Call func for each cic attached to this ioc.
2501  */
2502 static void
2503 call_for_each_cic(struct io_context *ioc,
2504                   void (*func)(struct io_context *, struct cfq_io_context *))
2505 {
2506         rcu_read_lock();
2507         __call_for_each_cic(ioc, func);
2508         rcu_read_unlock();
2509 }
2510
2511 static void cfq_cic_free_rcu(struct rcu_head *head)
2512 {
2513         struct cfq_io_context *cic;
2514
2515         cic = container_of(head, struct cfq_io_context, rcu_head);
2516
2517         kmem_cache_free(cfq_ioc_pool, cic);
2518         elv_ioc_count_dec(cfq_ioc_count);
2519
2520         if (ioc_gone) {
2521                 /*
2522                  * CFQ scheduler is exiting, grab exit lock and check
2523                  * the pending io context count. If it hits zero,
2524                  * complete ioc_gone and set it back to NULL
2525                  */
2526                 spin_lock(&ioc_gone_lock);
2527                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2528                         complete(ioc_gone);
2529                         ioc_gone = NULL;
2530                 }
2531                 spin_unlock(&ioc_gone_lock);
2532         }
2533 }
2534
2535 static void cfq_cic_free(struct cfq_io_context *cic)
2536 {
2537         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2538 }
2539
2540 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2541 {
2542         unsigned long flags;
2543         unsigned long dead_key = (unsigned long) cic->key;
2544
2545         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2546
2547         spin_lock_irqsave(&ioc->lock, flags);
2548         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2549         hlist_del_rcu(&cic->cic_list);
2550         spin_unlock_irqrestore(&ioc->lock, flags);
2551
2552         cfq_cic_free(cic);
2553 }
2554
2555 /*
2556  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2557  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2558  * and ->trim() which is called with the task lock held
2559  */
2560 static void cfq_free_io_context(struct io_context *ioc)
2561 {
2562         /*
2563          * ioc->refcount is zero here, or we are called from elv_unregister(),
2564          * so no more cic's are allowed to be linked into this ioc.  So it
2565          * should be ok to iterate over the known list, we will see all cic's
2566          * since no new ones are added.
2567          */
2568         __call_for_each_cic(ioc, cic_free_func);
2569 }
2570
2571 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2572 {
2573         struct cfq_queue *__cfqq, *next;
2574
2575         /*
2576          * If this queue was scheduled to merge with another queue, be
2577          * sure to drop the reference taken on that queue (and others in
2578          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2579          */
2580         __cfqq = cfqq->new_cfqq;
2581         while (__cfqq) {
2582                 if (__cfqq == cfqq) {
2583                         WARN(1, "cfqq->new_cfqq loop detected\n");
2584                         break;
2585                 }
2586                 next = __cfqq->new_cfqq;
2587                 cfq_put_queue(__cfqq);
2588                 __cfqq = next;
2589         }
2590 }
2591
2592 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2593 {
2594         if (unlikely(cfqq == cfqd->active_queue)) {
2595                 __cfq_slice_expired(cfqd, cfqq, 0);
2596                 cfq_schedule_dispatch(cfqd);
2597         }
2598
2599         cfq_put_cooperator(cfqq);
2600
2601         cfq_put_queue(cfqq);
2602 }
2603
2604 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2605                                          struct cfq_io_context *cic)
2606 {
2607         struct io_context *ioc = cic->ioc;
2608
2609         list_del_init(&cic->queue_list);
2610
2611         /*
2612          * Make sure dead mark is seen for dead queues
2613          */
2614         smp_wmb();
2615         cic->key = cfqd_dead_key(cfqd);
2616
2617         if (ioc->ioc_data == cic)
2618                 rcu_assign_pointer(ioc->ioc_data, NULL);
2619
2620         if (cic->cfqq[BLK_RW_ASYNC]) {
2621                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2622                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2623         }
2624
2625         if (cic->cfqq[BLK_RW_SYNC]) {
2626                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2627                 cic->cfqq[BLK_RW_SYNC] = NULL;
2628         }
2629 }
2630
2631 static void cfq_exit_single_io_context(struct io_context *ioc,
2632                                        struct cfq_io_context *cic)
2633 {
2634         struct cfq_data *cfqd = cic_to_cfqd(cic);
2635
2636         if (cfqd) {
2637                 struct request_queue *q = cfqd->queue;
2638                 unsigned long flags;
2639
2640                 spin_lock_irqsave(q->queue_lock, flags);
2641
2642                 /*
2643                  * Ensure we get a fresh copy of the ->key to prevent
2644                  * race between exiting task and queue
2645                  */
2646                 smp_read_barrier_depends();
2647                 if (cic->key == cfqd)
2648                         __cfq_exit_single_io_context(cfqd, cic);
2649
2650                 spin_unlock_irqrestore(q->queue_lock, flags);
2651         }
2652 }
2653
2654 /*
2655  * The process that ioc belongs to has exited, we need to clean up
2656  * and put the internal structures we have that belongs to that process.
2657  */
2658 static void cfq_exit_io_context(struct io_context *ioc)
2659 {
2660         call_for_each_cic(ioc, cfq_exit_single_io_context);
2661 }
2662
2663 static struct cfq_io_context *
2664 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2665 {
2666         struct cfq_io_context *cic;
2667
2668         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2669                                                         cfqd->queue->node);
2670         if (cic) {
2671                 cic->last_end_request = jiffies;
2672                 INIT_LIST_HEAD(&cic->queue_list);
2673                 INIT_HLIST_NODE(&cic->cic_list);
2674                 cic->dtor = cfq_free_io_context;
2675                 cic->exit = cfq_exit_io_context;
2676                 elv_ioc_count_inc(cfq_ioc_count);
2677         }
2678
2679         return cic;
2680 }
2681
2682 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2683 {
2684         struct task_struct *tsk = current;
2685         int ioprio_class;
2686
2687         if (!cfq_cfqq_prio_changed(cfqq))
2688                 return;
2689
2690         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2691         switch (ioprio_class) {
2692         default:
2693                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2694         case IOPRIO_CLASS_NONE:
2695                 /*
2696                  * no prio set, inherit CPU scheduling settings
2697                  */
2698                 cfqq->ioprio = task_nice_ioprio(tsk);
2699                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2700                 break;
2701         case IOPRIO_CLASS_RT:
2702                 cfqq->ioprio = task_ioprio(ioc);
2703                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2704                 break;
2705         case IOPRIO_CLASS_BE:
2706                 cfqq->ioprio = task_ioprio(ioc);
2707                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2708                 break;
2709         case IOPRIO_CLASS_IDLE:
2710                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2711                 cfqq->ioprio = 7;
2712                 cfq_clear_cfqq_idle_window(cfqq);
2713                 break;
2714         }
2715
2716         /*
2717          * keep track of original prio settings in case we have to temporarily
2718          * elevate the priority of this queue
2719          */
2720         cfqq->org_ioprio = cfqq->ioprio;
2721         cfqq->org_ioprio_class = cfqq->ioprio_class;
2722         cfq_clear_cfqq_prio_changed(cfqq);
2723 }
2724
2725 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2726 {
2727         struct cfq_data *cfqd = cic_to_cfqd(cic);
2728         struct cfq_queue *cfqq;
2729         unsigned long flags;
2730
2731         if (unlikely(!cfqd))
2732                 return;
2733
2734         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2735
2736         cfqq = cic->cfqq[BLK_RW_ASYNC];
2737         if (cfqq) {
2738                 struct cfq_queue *new_cfqq;
2739                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2740                                                 GFP_ATOMIC);
2741                 if (new_cfqq) {
2742                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2743                         cfq_put_queue(cfqq);
2744                 }
2745         }
2746
2747         cfqq = cic->cfqq[BLK_RW_SYNC];
2748         if (cfqq)
2749                 cfq_mark_cfqq_prio_changed(cfqq);
2750
2751         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2752 }
2753
2754 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2755 {
2756         call_for_each_cic(ioc, changed_ioprio);
2757         ioc->ioprio_changed = 0;
2758 }
2759
2760 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2761                           pid_t pid, bool is_sync)
2762 {
2763         RB_CLEAR_NODE(&cfqq->rb_node);
2764         RB_CLEAR_NODE(&cfqq->p_node);
2765         INIT_LIST_HEAD(&cfqq->fifo);
2766
2767         atomic_set(&cfqq->ref, 0);
2768         cfqq->cfqd = cfqd;
2769
2770         cfq_mark_cfqq_prio_changed(cfqq);
2771
2772         if (is_sync) {
2773                 if (!cfq_class_idle(cfqq))
2774                         cfq_mark_cfqq_idle_window(cfqq);
2775                 cfq_mark_cfqq_sync(cfqq);
2776         }
2777         cfqq->pid = pid;
2778 }
2779
2780 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2781 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2782 {
2783         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2784         struct cfq_data *cfqd = cic_to_cfqd(cic);
2785         unsigned long flags;
2786         struct request_queue *q;
2787
2788         if (unlikely(!cfqd))
2789                 return;
2790
2791         q = cfqd->queue;
2792
2793         spin_lock_irqsave(q->queue_lock, flags);
2794
2795         if (sync_cfqq) {
2796                 /*
2797                  * Drop reference to sync queue. A new sync queue will be
2798                  * assigned in new group upon arrival of a fresh request.
2799                  */
2800                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2801                 cic_set_cfqq(cic, NULL, 1);
2802                 cfq_put_queue(sync_cfqq);
2803         }
2804
2805         spin_unlock_irqrestore(q->queue_lock, flags);
2806 }
2807
2808 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2809 {
2810         call_for_each_cic(ioc, changed_cgroup);
2811         ioc->cgroup_changed = 0;
2812 }
2813 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2814
2815 static struct cfq_queue *
2816 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2817                      struct io_context *ioc, gfp_t gfp_mask)
2818 {
2819         struct cfq_queue *cfqq, *new_cfqq = NULL;
2820         struct cfq_io_context *cic;
2821         struct cfq_group *cfqg;
2822
2823 retry:
2824         cfqg = cfq_get_cfqg(cfqd, 1);
2825         cic = cfq_cic_lookup(cfqd, ioc);
2826         /* cic always exists here */
2827         cfqq = cic_to_cfqq(cic, is_sync);
2828
2829         /*
2830          * Always try a new alloc if we fell back to the OOM cfqq
2831          * originally, since it should just be a temporary situation.
2832          */
2833         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2834                 cfqq = NULL;
2835                 if (new_cfqq) {
2836                         cfqq = new_cfqq;
2837                         new_cfqq = NULL;
2838                 } else if (gfp_mask & __GFP_WAIT) {
2839                         spin_unlock_irq(cfqd->queue->queue_lock);
2840                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2841                                         gfp_mask | __GFP_ZERO,
2842                                         cfqd->queue->node);
2843                         spin_lock_irq(cfqd->queue->queue_lock);
2844                         if (new_cfqq)
2845                                 goto retry;
2846                 } else {
2847                         cfqq = kmem_cache_alloc_node(cfq_pool,
2848                                         gfp_mask | __GFP_ZERO,
2849                                         cfqd->queue->node);
2850                 }
2851
2852                 if (cfqq) {
2853                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2854                         cfq_init_prio_data(cfqq, ioc);
2855                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2856                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2857                 } else
2858                         cfqq = &cfqd->oom_cfqq;
2859         }
2860
2861         if (new_cfqq)
2862                 kmem_cache_free(cfq_pool, new_cfqq);
2863
2864         return cfqq;
2865 }
2866
2867 static struct cfq_queue **
2868 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2869 {
2870         switch (ioprio_class) {
2871         case IOPRIO_CLASS_RT:
2872                 return &cfqd->async_cfqq[0][ioprio];
2873         case IOPRIO_CLASS_BE:
2874                 return &cfqd->async_cfqq[1][ioprio];
2875         case IOPRIO_CLASS_IDLE:
2876                 return &cfqd->async_idle_cfqq;
2877         default:
2878                 BUG();
2879         }
2880 }
2881
2882 static struct cfq_queue *
2883 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2884               gfp_t gfp_mask)
2885 {
2886         const int ioprio = task_ioprio(ioc);
2887         const int ioprio_class = task_ioprio_class(ioc);
2888         struct cfq_queue **async_cfqq = NULL;
2889         struct cfq_queue *cfqq = NULL;
2890
2891         if (!is_sync) {
2892                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2893                 cfqq = *async_cfqq;
2894         }
2895
2896         if (!cfqq)
2897                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2898
2899         /*
2900          * pin the queue now that it's allocated, scheduler exit will prune it
2901          */
2902         if (!is_sync && !(*async_cfqq)) {
2903                 atomic_inc(&cfqq->ref);
2904                 *async_cfqq = cfqq;
2905         }
2906
2907         atomic_inc(&cfqq->ref);
2908         return cfqq;
2909 }
2910
2911 /*
2912  * We drop cfq io contexts lazily, so we may find a dead one.
2913  */
2914 static void
2915 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2916                   struct cfq_io_context *cic)
2917 {
2918         unsigned long flags;
2919
2920         WARN_ON(!list_empty(&cic->queue_list));
2921         BUG_ON(cic->key != cfqd_dead_key(cfqd));
2922
2923         spin_lock_irqsave(&ioc->lock, flags);
2924
2925         BUG_ON(ioc->ioc_data == cic);
2926
2927         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2928         hlist_del_rcu(&cic->cic_list);
2929         spin_unlock_irqrestore(&ioc->lock, flags);
2930
2931         cfq_cic_free(cic);
2932 }
2933
2934 static struct cfq_io_context *
2935 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2936 {
2937         struct cfq_io_context *cic;
2938         unsigned long flags;
2939
2940         if (unlikely(!ioc))
2941                 return NULL;
2942
2943         rcu_read_lock();
2944
2945         /*
2946          * we maintain a last-hit cache, to avoid browsing over the tree
2947          */
2948         cic = rcu_dereference(ioc->ioc_data);
2949         if (cic && cic->key == cfqd) {
2950                 rcu_read_unlock();
2951                 return cic;
2952         }
2953
2954         do {
2955                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
2956                 rcu_read_unlock();
2957                 if (!cic)
2958                         break;
2959                 if (unlikely(cic->key != cfqd)) {
2960                         cfq_drop_dead_cic(cfqd, ioc, cic);
2961                         rcu_read_lock();
2962                         continue;
2963                 }
2964
2965                 spin_lock_irqsave(&ioc->lock, flags);
2966                 rcu_assign_pointer(ioc->ioc_data, cic);
2967                 spin_unlock_irqrestore(&ioc->lock, flags);
2968                 break;
2969         } while (1);
2970
2971         return cic;
2972 }
2973
2974 /*
2975  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2976  * the process specific cfq io context when entered from the block layer.
2977  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2978  */
2979 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2980                         struct cfq_io_context *cic, gfp_t gfp_mask)
2981 {
2982         unsigned long flags;
2983         int ret;
2984
2985         ret = radix_tree_preload(gfp_mask);
2986         if (!ret) {
2987                 cic->ioc = ioc;
2988                 cic->key = cfqd;
2989
2990                 spin_lock_irqsave(&ioc->lock, flags);
2991                 ret = radix_tree_insert(&ioc->radix_root,
2992                                                 cfqd->cic_index, cic);
2993                 if (!ret)
2994                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2995                 spin_unlock_irqrestore(&ioc->lock, flags);
2996
2997                 radix_tree_preload_end();
2998
2999                 if (!ret) {
3000                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3001                         list_add(&cic->queue_list, &cfqd->cic_list);
3002                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3003                 }
3004         }
3005
3006         if (ret)
3007                 printk(KERN_ERR "cfq: cic link failed!\n");
3008
3009         return ret;
3010 }
3011
3012 /*
3013  * Setup general io context and cfq io context. There can be several cfq
3014  * io contexts per general io context, if this process is doing io to more
3015  * than one device managed by cfq.
3016  */
3017 static struct cfq_io_context *
3018 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3019 {
3020         struct io_context *ioc = NULL;
3021         struct cfq_io_context *cic;
3022
3023         might_sleep_if(gfp_mask & __GFP_WAIT);
3024
3025         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3026         if (!ioc)
3027                 return NULL;
3028
3029         cic = cfq_cic_lookup(cfqd, ioc);
3030         if (cic)
3031                 goto out;
3032
3033         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3034         if (cic == NULL)
3035                 goto err;
3036
3037         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3038                 goto err_free;
3039
3040 out:
3041         smp_read_barrier_depends();
3042         if (unlikely(ioc->ioprio_changed))
3043                 cfq_ioc_set_ioprio(ioc);
3044
3045 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3046         if (unlikely(ioc->cgroup_changed))
3047                 cfq_ioc_set_cgroup(ioc);
3048 #endif
3049         return cic;
3050 err_free:
3051         cfq_cic_free(cic);
3052 err:
3053         put_io_context(ioc);
3054         return NULL;
3055 }
3056
3057 static void
3058 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3059 {
3060         unsigned long elapsed = jiffies - cic->last_end_request;
3061         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3062
3063         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3064         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3065         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3066 }
3067
3068 static void
3069 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3070                        struct request *rq)
3071 {
3072         sector_t sdist = 0;
3073         sector_t n_sec = blk_rq_sectors(rq);
3074         if (cfqq->last_request_pos) {
3075                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3076                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3077                 else
3078                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3079         }
3080
3081         cfqq->seek_history <<= 1;
3082         if (blk_queue_nonrot(cfqd->queue))
3083                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3084         else
3085                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3086 }
3087
3088 /*
3089  * Disable idle window if the process thinks too long or seeks so much that
3090  * it doesn't matter
3091  */
3092 static void
3093 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3094                        struct cfq_io_context *cic)
3095 {
3096         int old_idle, enable_idle;
3097
3098         /*
3099          * Don't idle for async or idle io prio class
3100          */
3101         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3102                 return;
3103
3104         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3105
3106         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3107                 cfq_mark_cfqq_deep(cfqq);
3108
3109         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3110                 enable_idle = 0;
3111         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3112             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3113                 enable_idle = 0;
3114         else if (sample_valid(cic->ttime_samples)) {
3115                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3116                         enable_idle = 0;
3117                 else
3118                         enable_idle = 1;
3119         }
3120
3121         if (old_idle != enable_idle) {
3122                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3123                 if (enable_idle)
3124                         cfq_mark_cfqq_idle_window(cfqq);
3125                 else
3126                         cfq_clear_cfqq_idle_window(cfqq);
3127         }
3128 }
3129
3130 /*
3131  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3132  * no or if we aren't sure, a 1 will cause a preempt.
3133  */
3134 static bool
3135 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3136                    struct request *rq)
3137 {
3138         struct cfq_queue *cfqq;
3139
3140         cfqq = cfqd->active_queue;
3141         if (!cfqq)
3142                 return false;
3143
3144         if (cfq_class_idle(new_cfqq))
3145                 return false;
3146
3147         if (cfq_class_idle(cfqq))
3148                 return true;
3149
3150         /*
3151          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3152          */
3153         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3154                 return false;
3155
3156         /*
3157          * if the new request is sync, but the currently running queue is
3158          * not, let the sync request have priority.
3159          */
3160         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3161                 return true;
3162
3163         if (new_cfqq->cfqg != cfqq->cfqg)
3164                 return false;
3165
3166         if (cfq_slice_used(cfqq))
3167                 return true;
3168
3169         /* Allow preemption only if we are idling on sync-noidle tree */
3170         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3171             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3172             new_cfqq->service_tree->count == 2 &&
3173             RB_EMPTY_ROOT(&cfqq->sort_list))
3174                 return true;
3175
3176         /*
3177          * So both queues are sync. Let the new request get disk time if
3178          * it's a metadata request and the current queue is doing regular IO.
3179          */
3180         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3181                 return true;
3182
3183         /*
3184          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3185          */
3186         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3187                 return true;
3188
3189         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3190                 return false;
3191
3192         /*
3193          * if this request is as-good as one we would expect from the
3194          * current cfqq, let it preempt
3195          */
3196         if (cfq_rq_close(cfqd, cfqq, rq))
3197                 return true;
3198
3199         return false;
3200 }
3201
3202 /*
3203  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3204  * let it have half of its nominal slice.
3205  */
3206 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3207 {
3208         cfq_log_cfqq(cfqd, cfqq, "preempt");
3209         cfq_slice_expired(cfqd, 1);
3210
3211         /*
3212          * Put the new queue at the front of the of the current list,
3213          * so we know that it will be selected next.
3214          */
3215         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3216
3217         cfq_service_tree_add(cfqd, cfqq, 1);
3218
3219         cfqq->slice_end = 0;
3220         cfq_mark_cfqq_slice_new(cfqq);
3221 }
3222
3223 /*
3224  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3225  * something we should do about it
3226  */
3227 static void
3228 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3229                 struct request *rq)
3230 {
3231         struct cfq_io_context *cic = RQ_CIC(rq);
3232
3233         cfqd->rq_queued++;
3234         if (rq->cmd_flags & REQ_META)
3235                 cfqq->meta_pending++;
3236
3237         cfq_update_io_thinktime(cfqd, cic);
3238         cfq_update_io_seektime(cfqd, cfqq, rq);
3239         cfq_update_idle_window(cfqd, cfqq, cic);
3240
3241         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3242
3243         if (cfqq == cfqd->active_queue) {
3244                 /*
3245                  * Remember that we saw a request from this process, but
3246                  * don't start queuing just yet. Otherwise we risk seeing lots
3247                  * of tiny requests, because we disrupt the normal plugging
3248                  * and merging. If the request is already larger than a single
3249                  * page, let it rip immediately. For that case we assume that
3250                  * merging is already done. Ditto for a busy system that
3251                  * has other work pending, don't risk delaying until the
3252                  * idle timer unplug to continue working.
3253                  */
3254                 if (cfq_cfqq_wait_request(cfqq)) {
3255                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3256                             cfqd->busy_queues > 1) {
3257                                 cfq_del_timer(cfqd, cfqq);
3258                                 cfq_clear_cfqq_wait_request(cfqq);
3259                                 __blk_run_queue(cfqd->queue);
3260                         } else {
3261                                 cfq_blkiocg_update_idle_time_stats(
3262                                                 &cfqq->cfqg->blkg);
3263                                 cfq_mark_cfqq_must_dispatch(cfqq);
3264                         }
3265                 }
3266         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3267                 /*
3268                  * not the active queue - expire current slice if it is
3269                  * idle and has expired it's mean thinktime or this new queue
3270                  * has some old slice time left and is of higher priority or
3271                  * this new queue is RT and the current one is BE
3272                  */
3273                 cfq_preempt_queue(cfqd, cfqq);
3274                 __blk_run_queue(cfqd->queue);
3275         }
3276 }
3277
3278 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3279 {
3280         struct cfq_data *cfqd = q->elevator->elevator_data;
3281         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3282
3283         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3284         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3285
3286         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3287         list_add_tail(&rq->queuelist, &cfqq->fifo);
3288         cfq_add_rq_rb(rq);
3289         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3290                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3291                         rq_is_sync(rq));
3292         cfq_rq_enqueued(cfqd, cfqq, rq);
3293 }
3294
3295 /*
3296  * Update hw_tag based on peak queue depth over 50 samples under
3297  * sufficient load.
3298  */
3299 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3300 {
3301         struct cfq_queue *cfqq = cfqd->active_queue;
3302
3303         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3304                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3305
3306         if (cfqd->hw_tag == 1)
3307                 return;
3308
3309         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3310             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3311                 return;
3312
3313         /*
3314          * If active queue hasn't enough requests and can idle, cfq might not
3315          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3316          * case
3317          */
3318         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3319             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3320             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3321                 return;
3322
3323         if (cfqd->hw_tag_samples++ < 50)
3324                 return;
3325
3326         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3327                 cfqd->hw_tag = 1;
3328         else
3329                 cfqd->hw_tag = 0;
3330 }
3331
3332 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3333 {
3334         struct cfq_io_context *cic = cfqd->active_cic;
3335
3336         /* If there are other queues in the group, don't wait */
3337         if (cfqq->cfqg->nr_cfqq > 1)
3338                 return false;
3339
3340         if (cfq_slice_used(cfqq))
3341                 return true;
3342
3343         /* if slice left is less than think time, wait busy */
3344         if (cic && sample_valid(cic->ttime_samples)
3345             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3346                 return true;
3347
3348         /*
3349          * If think times is less than a jiffy than ttime_mean=0 and above
3350          * will not be true. It might happen that slice has not expired yet
3351          * but will expire soon (4-5 ns) during select_queue(). To cover the
3352          * case where think time is less than a jiffy, mark the queue wait
3353          * busy if only 1 jiffy is left in the slice.
3354          */
3355         if (cfqq->slice_end - jiffies == 1)
3356                 return true;
3357
3358         return false;
3359 }
3360
3361 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3362 {
3363         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3364         struct cfq_data *cfqd = cfqq->cfqd;
3365         const int sync = rq_is_sync(rq);
3366         unsigned long now;
3367
3368         now = jiffies;
3369         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3370                      !!(rq->cmd_flags & REQ_NOIDLE));
3371
3372         cfq_update_hw_tag(cfqd);
3373
3374         WARN_ON(!cfqd->rq_in_driver);
3375         WARN_ON(!cfqq->dispatched);
3376         cfqd->rq_in_driver--;
3377         cfqq->dispatched--;
3378         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3379                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3380                         rq_data_dir(rq), rq_is_sync(rq));
3381
3382         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3383
3384         if (sync) {
3385                 RQ_CIC(rq)->last_end_request = now;
3386                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3387                         cfqd->last_delayed_sync = now;
3388         }
3389
3390         /*
3391          * If this is the active queue, check if it needs to be expired,
3392          * or if we want to idle in case it has no pending requests.
3393          */
3394         if (cfqd->active_queue == cfqq) {
3395                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3396
3397                 if (cfq_cfqq_slice_new(cfqq)) {
3398                         cfq_set_prio_slice(cfqd, cfqq);
3399                         cfq_clear_cfqq_slice_new(cfqq);
3400                 }
3401
3402                 /*
3403                  * Should we wait for next request to come in before we expire
3404                  * the queue.
3405                  */
3406                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3407                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3408                         cfq_mark_cfqq_wait_busy(cfqq);
3409                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3410                 }
3411
3412                 /*
3413                  * Idling is not enabled on:
3414                  * - expired queues
3415                  * - idle-priority queues
3416                  * - async queues
3417                  * - queues with still some requests queued
3418                  * - when there is a close cooperator
3419                  */
3420                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3421                         cfq_slice_expired(cfqd, 1);
3422                 else if (sync && cfqq_empty &&
3423                          !cfq_close_cooperator(cfqd, cfqq)) {
3424                         cfq_arm_slice_timer(cfqd);
3425                 }
3426         }
3427
3428         if (!cfqd->rq_in_driver)
3429                 cfq_schedule_dispatch(cfqd);
3430 }
3431
3432 /*
3433  * we temporarily boost lower priority queues if they are holding fs exclusive
3434  * resources. they are boosted to normal prio (CLASS_BE/4)
3435  */
3436 static void cfq_prio_boost(struct cfq_queue *cfqq)
3437 {
3438         if (has_fs_excl()) {
3439                 /*
3440                  * boost idle prio on transactions that would lock out other
3441                  * users of the filesystem
3442                  */
3443                 if (cfq_class_idle(cfqq))
3444                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3445                 if (cfqq->ioprio > IOPRIO_NORM)
3446                         cfqq->ioprio = IOPRIO_NORM;
3447         } else {
3448                 /*
3449                  * unboost the queue (if needed)
3450                  */
3451                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3452                 cfqq->ioprio = cfqq->org_ioprio;
3453         }
3454 }
3455
3456 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3457 {
3458         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3459                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3460                 return ELV_MQUEUE_MUST;
3461         }
3462
3463         return ELV_MQUEUE_MAY;
3464 }
3465
3466 static int cfq_may_queue(struct request_queue *q, int rw)
3467 {
3468         struct cfq_data *cfqd = q->elevator->elevator_data;
3469         struct task_struct *tsk = current;
3470         struct cfq_io_context *cic;
3471         struct cfq_queue *cfqq;
3472
3473         /*
3474          * don't force setup of a queue from here, as a call to may_queue
3475          * does not necessarily imply that a request actually will be queued.
3476          * so just lookup a possibly existing queue, or return 'may queue'
3477          * if that fails
3478          */
3479         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3480         if (!cic)
3481                 return ELV_MQUEUE_MAY;
3482
3483         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3484         if (cfqq) {
3485                 cfq_init_prio_data(cfqq, cic->ioc);
3486                 cfq_prio_boost(cfqq);
3487
3488                 return __cfq_may_queue(cfqq);
3489         }
3490
3491         return ELV_MQUEUE_MAY;
3492 }
3493
3494 /*
3495  * queue lock held here
3496  */
3497 static void cfq_put_request(struct request *rq)
3498 {
3499         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3500
3501         if (cfqq) {
3502                 const int rw = rq_data_dir(rq);
3503
3504                 BUG_ON(!cfqq->allocated[rw]);
3505                 cfqq->allocated[rw]--;
3506
3507                 put_io_context(RQ_CIC(rq)->ioc);
3508
3509                 rq->elevator_private = NULL;
3510                 rq->elevator_private2 = NULL;
3511
3512                 /* Put down rq reference on cfqg */
3513                 cfq_put_cfqg(RQ_CFQG(rq));
3514                 rq->elevator_private3 = NULL;
3515
3516                 cfq_put_queue(cfqq);
3517         }
3518 }
3519
3520 static struct cfq_queue *
3521 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3522                 struct cfq_queue *cfqq)
3523 {
3524         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3525         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3526         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3527         cfq_put_queue(cfqq);
3528         return cic_to_cfqq(cic, 1);
3529 }
3530
3531 /*
3532  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3533  * was the last process referring to said cfqq.
3534  */
3535 static struct cfq_queue *
3536 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3537 {
3538         if (cfqq_process_refs(cfqq) == 1) {
3539                 cfqq->pid = current->pid;
3540                 cfq_clear_cfqq_coop(cfqq);
3541                 cfq_clear_cfqq_split_coop(cfqq);
3542                 return cfqq;
3543         }
3544
3545         cic_set_cfqq(cic, NULL, 1);
3546
3547         cfq_put_cooperator(cfqq);
3548
3549         cfq_put_queue(cfqq);
3550         return NULL;
3551 }
3552 /*
3553  * Allocate cfq data structures associated with this request.
3554  */
3555 static int
3556 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3557 {
3558         struct cfq_data *cfqd = q->elevator->elevator_data;
3559         struct cfq_io_context *cic;
3560         const int rw = rq_data_dir(rq);
3561         const bool is_sync = rq_is_sync(rq);
3562         struct cfq_queue *cfqq;
3563         unsigned long flags;
3564
3565         might_sleep_if(gfp_mask & __GFP_WAIT);
3566
3567         cic = cfq_get_io_context(cfqd, gfp_mask);
3568
3569         spin_lock_irqsave(q->queue_lock, flags);
3570
3571         if (!cic)
3572                 goto queue_fail;
3573
3574 new_queue:
3575         cfqq = cic_to_cfqq(cic, is_sync);
3576         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3577                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3578                 cic_set_cfqq(cic, cfqq, is_sync);
3579         } else {
3580                 /*
3581                  * If the queue was seeky for too long, break it apart.
3582                  */
3583                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3584                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3585                         cfqq = split_cfqq(cic, cfqq);
3586                         if (!cfqq)
3587                                 goto new_queue;
3588                 }
3589
3590                 /*
3591                  * Check to see if this queue is scheduled to merge with
3592                  * another, closely cooperating queue.  The merging of
3593                  * queues happens here as it must be done in process context.
3594                  * The reference on new_cfqq was taken in merge_cfqqs.
3595                  */
3596                 if (cfqq->new_cfqq)
3597                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3598         }
3599
3600         cfqq->allocated[rw]++;
3601         atomic_inc(&cfqq->ref);
3602
3603         spin_unlock_irqrestore(q->queue_lock, flags);
3604
3605         rq->elevator_private = cic;
3606         rq->elevator_private2 = cfqq;
3607         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3608         return 0;
3609
3610 queue_fail:
3611         if (cic)
3612                 put_io_context(cic->ioc);
3613
3614         cfq_schedule_dispatch(cfqd);
3615         spin_unlock_irqrestore(q->queue_lock, flags);
3616         cfq_log(cfqd, "set_request fail");
3617         return 1;
3618 }
3619
3620 static void cfq_kick_queue(struct work_struct *work)
3621 {
3622         struct cfq_data *cfqd =
3623                 container_of(work, struct cfq_data, unplug_work);
3624         struct request_queue *q = cfqd->queue;
3625
3626         spin_lock_irq(q->queue_lock);
3627         __blk_run_queue(cfqd->queue);
3628         spin_unlock_irq(q->queue_lock);
3629 }
3630
3631 /*
3632  * Timer running if the active_queue is currently idling inside its time slice
3633  */
3634 static void cfq_idle_slice_timer(unsigned long data)
3635 {
3636         struct cfq_data *cfqd = (struct cfq_data *) data;
3637         struct cfq_queue *cfqq;
3638         unsigned long flags;
3639         int timed_out = 1;
3640
3641         cfq_log(cfqd, "idle timer fired");
3642
3643         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3644
3645         cfqq = cfqd->active_queue;
3646         if (cfqq) {
3647                 timed_out = 0;
3648
3649                 /*
3650                  * We saw a request before the queue expired, let it through
3651                  */
3652                 if (cfq_cfqq_must_dispatch(cfqq))
3653                         goto out_kick;
3654
3655                 /*
3656                  * expired
3657                  */
3658                 if (cfq_slice_used(cfqq))
3659                         goto expire;
3660
3661                 /*
3662                  * only expire and reinvoke request handler, if there are
3663                  * other queues with pending requests
3664                  */
3665                 if (!cfqd->busy_queues)
3666                         goto out_cont;
3667
3668                 /*
3669                  * not expired and it has a request pending, let it dispatch
3670                  */
3671                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3672                         goto out_kick;
3673
3674                 /*
3675                  * Queue depth flag is reset only when the idle didn't succeed
3676                  */
3677                 cfq_clear_cfqq_deep(cfqq);
3678         }
3679 expire:
3680         cfq_slice_expired(cfqd, timed_out);
3681 out_kick:
3682         cfq_schedule_dispatch(cfqd);
3683 out_cont:
3684         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3685 }
3686
3687 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3688 {
3689         del_timer_sync(&cfqd->idle_slice_timer);
3690         cancel_work_sync(&cfqd->unplug_work);
3691 }
3692
3693 static void cfq_put_async_queues(struct cfq_data *cfqd)
3694 {
3695         int i;
3696
3697         for (i = 0; i < IOPRIO_BE_NR; i++) {
3698                 if (cfqd->async_cfqq[0][i])
3699                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3700                 if (cfqd->async_cfqq[1][i])
3701                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3702         }
3703
3704         if (cfqd->async_idle_cfqq)
3705                 cfq_put_queue(cfqd->async_idle_cfqq);
3706 }
3707
3708 static void cfq_cfqd_free(struct rcu_head *head)
3709 {
3710         kfree(container_of(head, struct cfq_data, rcu));
3711 }
3712
3713 static void cfq_exit_queue(struct elevator_queue *e)
3714 {
3715         struct cfq_data *cfqd = e->elevator_data;
3716         struct request_queue *q = cfqd->queue;
3717
3718         cfq_shutdown_timer_wq(cfqd);
3719
3720         spin_lock_irq(q->queue_lock);
3721
3722         if (cfqd->active_queue)
3723                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3724
3725         while (!list_empty(&cfqd->cic_list)) {
3726                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3727                                                         struct cfq_io_context,
3728                                                         queue_list);
3729
3730                 __cfq_exit_single_io_context(cfqd, cic);
3731         }
3732
3733         cfq_put_async_queues(cfqd);
3734         cfq_release_cfq_groups(cfqd);
3735         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3736
3737         spin_unlock_irq(q->queue_lock);
3738
3739         cfq_shutdown_timer_wq(cfqd);
3740
3741         spin_lock(&cic_index_lock);
3742         ida_remove(&cic_index_ida, cfqd->cic_index);
3743         spin_unlock(&cic_index_lock);
3744
3745         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3746         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3747 }
3748
3749 static int cfq_alloc_cic_index(void)
3750 {
3751         int index, error;
3752
3753         do {
3754                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3755                         return -ENOMEM;
3756
3757                 spin_lock(&cic_index_lock);
3758                 error = ida_get_new(&cic_index_ida, &index);
3759                 spin_unlock(&cic_index_lock);
3760                 if (error && error != -EAGAIN)
3761                         return error;
3762         } while (error);
3763
3764         return index;
3765 }
3766
3767 static void *cfq_init_queue(struct request_queue *q)
3768 {
3769         struct cfq_data *cfqd;
3770         int i, j;
3771         struct cfq_group *cfqg;
3772         struct cfq_rb_root *st;
3773
3774         i = cfq_alloc_cic_index();
3775         if (i < 0)
3776                 return NULL;
3777
3778         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3779         if (!cfqd)
3780                 return NULL;
3781
3782         cfqd->cic_index = i;
3783
3784         /* Init root service tree */
3785         cfqd->grp_service_tree = CFQ_RB_ROOT;
3786
3787         /* Init root group */
3788         cfqg = &cfqd->root_group;
3789         for_each_cfqg_st(cfqg, i, j, st)
3790                 *st = CFQ_RB_ROOT;
3791         RB_CLEAR_NODE(&cfqg->rb_node);
3792
3793         /* Give preference to root group over other groups */
3794         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3795
3796 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3797         /*
3798          * Take a reference to root group which we never drop. This is just
3799          * to make sure that cfq_put_cfqg() does not try to kfree root group
3800          */
3801         atomic_set(&cfqg->ref, 1);
3802         rcu_read_lock();
3803         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3804                                         (void *)cfqd, 0);
3805         rcu_read_unlock();
3806 #endif
3807         /*
3808          * Not strictly needed (since RB_ROOT just clears the node and we
3809          * zeroed cfqd on alloc), but better be safe in case someone decides
3810          * to add magic to the rb code
3811          */
3812         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3813                 cfqd->prio_trees[i] = RB_ROOT;
3814
3815         /*
3816          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3817          * Grab a permanent reference to it, so that the normal code flow
3818          * will not attempt to free it.
3819          */
3820         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3821         atomic_inc(&cfqd->oom_cfqq.ref);
3822         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3823
3824         INIT_LIST_HEAD(&cfqd->cic_list);
3825
3826         cfqd->queue = q;
3827
3828         init_timer(&cfqd->idle_slice_timer);
3829         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3830         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3831
3832         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3833
3834         cfqd->cfq_quantum = cfq_quantum;
3835         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3836         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3837         cfqd->cfq_back_max = cfq_back_max;
3838         cfqd->cfq_back_penalty = cfq_back_penalty;
3839         cfqd->cfq_slice[0] = cfq_slice_async;
3840         cfqd->cfq_slice[1] = cfq_slice_sync;
3841         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3842         cfqd->cfq_slice_idle = cfq_slice_idle;
3843         cfqd->cfq_latency = 1;
3844         cfqd->cfq_group_isolation = 0;
3845         cfqd->hw_tag = -1;
3846         /*
3847          * we optimistically start assuming sync ops weren't delayed in last
3848          * second, in order to have larger depth for async operations.
3849          */
3850         cfqd->last_delayed_sync = jiffies - HZ;
3851         return cfqd;
3852 }
3853
3854 static void cfq_slab_kill(void)
3855 {
3856         /*
3857          * Caller already ensured that pending RCU callbacks are completed,
3858          * so we should have no busy allocations at this point.
3859          */
3860         if (cfq_pool)
3861                 kmem_cache_destroy(cfq_pool);
3862         if (cfq_ioc_pool)
3863                 kmem_cache_destroy(cfq_ioc_pool);
3864 }
3865
3866 static int __init cfq_slab_setup(void)
3867 {
3868         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3869         if (!cfq_pool)
3870                 goto fail;
3871
3872         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3873         if (!cfq_ioc_pool)
3874                 goto fail;
3875
3876         return 0;
3877 fail:
3878         cfq_slab_kill();
3879         return -ENOMEM;
3880 }
3881
3882 /*
3883  * sysfs parts below -->
3884  */
3885 static ssize_t
3886 cfq_var_show(unsigned int var, char *page)
3887 {
3888         return sprintf(page, "%d\n", var);
3889 }
3890
3891 static ssize_t
3892 cfq_var_store(unsigned int *var, const char *page, size_t count)
3893 {
3894         char *p = (char *) page;
3895
3896         *var = simple_strtoul(p, &p, 10);
3897         return count;
3898 }
3899
3900 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3901 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3902 {                                                                       \
3903         struct cfq_data *cfqd = e->elevator_data;                       \
3904         unsigned int __data = __VAR;                                    \
3905         if (__CONV)                                                     \
3906                 __data = jiffies_to_msecs(__data);                      \
3907         return cfq_var_show(__data, (page));                            \
3908 }
3909 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3910 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3911 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3912 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3913 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3914 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3915 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3916 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3917 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3918 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3919 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3920 #undef SHOW_FUNCTION
3921
3922 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3923 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3924 {                                                                       \
3925         struct cfq_data *cfqd = e->elevator_data;                       \
3926         unsigned int __data;                                            \
3927         int ret = cfq_var_store(&__data, (page), count);                \
3928         if (__data < (MIN))                                             \
3929                 __data = (MIN);                                         \
3930         else if (__data > (MAX))                                        \
3931                 __data = (MAX);                                         \
3932         if (__CONV)                                                     \
3933                 *(__PTR) = msecs_to_jiffies(__data);                    \
3934         else                                                            \
3935                 *(__PTR) = __data;                                      \
3936         return ret;                                                     \
3937 }
3938 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3939 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3940                 UINT_MAX, 1);
3941 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3942                 UINT_MAX, 1);
3943 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3944 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3945                 UINT_MAX, 0);
3946 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3947 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3948 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3949 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3950                 UINT_MAX, 0);
3951 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3952 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3953 #undef STORE_FUNCTION
3954
3955 #define CFQ_ATTR(name) \
3956         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3957
3958 static struct elv_fs_entry cfq_attrs[] = {
3959         CFQ_ATTR(quantum),
3960         CFQ_ATTR(fifo_expire_sync),
3961         CFQ_ATTR(fifo_expire_async),
3962         CFQ_ATTR(back_seek_max),
3963         CFQ_ATTR(back_seek_penalty),
3964         CFQ_ATTR(slice_sync),
3965         CFQ_ATTR(slice_async),
3966         CFQ_ATTR(slice_async_rq),
3967         CFQ_ATTR(slice_idle),
3968         CFQ_ATTR(low_latency),
3969         CFQ_ATTR(group_isolation),
3970         __ATTR_NULL
3971 };
3972
3973 static struct elevator_type iosched_cfq = {
3974         .ops = {
3975                 .elevator_merge_fn =            cfq_merge,
3976                 .elevator_merged_fn =           cfq_merged_request,
3977                 .elevator_merge_req_fn =        cfq_merged_requests,
3978                 .elevator_allow_merge_fn =      cfq_allow_merge,
3979                 .elevator_bio_merged_fn =       cfq_bio_merged,
3980                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3981                 .elevator_add_req_fn =          cfq_insert_request,
3982                 .elevator_activate_req_fn =     cfq_activate_request,
3983                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3984                 .elevator_queue_empty_fn =      cfq_queue_empty,
3985                 .elevator_completed_req_fn =    cfq_completed_request,
3986                 .elevator_former_req_fn =       elv_rb_former_request,
3987                 .elevator_latter_req_fn =       elv_rb_latter_request,
3988                 .elevator_set_req_fn =          cfq_set_request,
3989                 .elevator_put_req_fn =          cfq_put_request,
3990                 .elevator_may_queue_fn =        cfq_may_queue,
3991                 .elevator_init_fn =             cfq_init_queue,
3992                 .elevator_exit_fn =             cfq_exit_queue,
3993                 .trim =                         cfq_free_io_context,
3994         },
3995         .elevator_attrs =       cfq_attrs,
3996         .elevator_name =        "cfq",
3997         .elevator_owner =       THIS_MODULE,
3998 };
3999
4000 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4001 static struct blkio_policy_type blkio_policy_cfq = {
4002         .ops = {
4003                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4004                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4005         },
4006         .plid = BLKIO_POLICY_PROP,
4007 };
4008 #else
4009 static struct blkio_policy_type blkio_policy_cfq;
4010 #endif
4011
4012 static int __init cfq_init(void)
4013 {
4014         /*
4015          * could be 0 on HZ < 1000 setups
4016          */
4017         if (!cfq_slice_async)
4018                 cfq_slice_async = 1;
4019         if (!cfq_slice_idle)
4020                 cfq_slice_idle = 1;
4021
4022         if (cfq_slab_setup())
4023                 return -ENOMEM;
4024
4025         elv_register(&iosched_cfq);
4026         blkio_policy_register(&blkio_policy_cfq);
4027
4028         return 0;
4029 }
4030
4031 static void __exit cfq_exit(void)
4032 {
4033         DECLARE_COMPLETION_ONSTACK(all_gone);
4034         blkio_policy_unregister(&blkio_policy_cfq);
4035         elv_unregister(&iosched_cfq);
4036         ioc_gone = &all_gone;
4037         /* ioc_gone's update must be visible before reading ioc_count */
4038         smp_wmb();
4039
4040         /*
4041          * this also protects us from entering cfq_slab_kill() with
4042          * pending RCU callbacks
4043          */
4044         if (elv_ioc_count_read(cfq_ioc_count))
4045                 wait_for_completion(&all_gone);
4046         ida_destroy(&cic_index_ida);
4047         cfq_slab_kill();
4048 }
4049
4050 module_init(cfq_init);
4051 module_exit(cfq_exit);
4052
4053 MODULE_AUTHOR("Jens Axboe");
4054 MODULE_LICENSE("GPL");
4055 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");