cfq-iosched: Always provide group isolation.
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181
182         /* number of cfqq currently on this group */
183         int nr_cfqq;
184
185         /*
186          * Per group busy queus average. Useful for workload slice calc. We
187          * create the array for each prio class but at run time it is used
188          * only for RT and BE class and slot for IDLE class remains unused.
189          * This is primarily done to avoid confusion and a gcc warning.
190          */
191         unsigned int busy_queues_avg[CFQ_PRIO_NR];
192         /*
193          * rr lists of queues with requests. We maintain service trees for
194          * RT and BE classes. These trees are subdivided in subclasses
195          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196          * class there is no subclassification and all the cfq queues go on
197          * a single tree service_tree_idle.
198          * Counts are embedded in the cfq_rb_root
199          */
200         struct cfq_rb_root service_trees[2][3];
201         struct cfq_rb_root service_tree_idle;
202
203         unsigned long saved_workload_slice;
204         enum wl_type_t saved_workload;
205         enum wl_prio_t saved_serving_prio;
206         struct blkio_group blkg;
207 #ifdef CONFIG_CFQ_GROUP_IOSCHED
208         struct hlist_node cfqd_node;
209         int ref;
210 #endif
211         /* number of requests that are on the dispatch list or inside driver */
212         int dispatched;
213 };
214
215 /*
216  * Per block device queue structure
217  */
218 struct cfq_data {
219         struct request_queue *queue;
220         /* Root service tree for cfq_groups */
221         struct cfq_rb_root grp_service_tree;
222         struct cfq_group root_group;
223
224         /*
225          * The priority currently being served
226          */
227         enum wl_prio_t serving_prio;
228         enum wl_type_t serving_type;
229         unsigned long workload_expires;
230         struct cfq_group *serving_group;
231
232         /*
233          * Each priority tree is sorted by next_request position.  These
234          * trees are used when determining if two or more queues are
235          * interleaving requests (see cfq_close_cooperator).
236          */
237         struct rb_root prio_trees[CFQ_PRIO_LISTS];
238
239         unsigned int busy_queues;
240
241         int rq_in_driver;
242         int rq_in_flight[2];
243
244         /*
245          * queue-depth detection
246          */
247         int rq_queued;
248         int hw_tag;
249         /*
250          * hw_tag can be
251          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
252          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
253          *  0 => no NCQ
254          */
255         int hw_tag_est_depth;
256         unsigned int hw_tag_samples;
257
258         /*
259          * idle window management
260          */
261         struct timer_list idle_slice_timer;
262         struct work_struct unplug_work;
263
264         struct cfq_queue *active_queue;
265         struct cfq_io_context *active_cic;
266
267         /*
268          * async queue for each priority case
269          */
270         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
271         struct cfq_queue *async_idle_cfqq;
272
273         sector_t last_position;
274
275         /*
276          * tunables, see top of file
277          */
278         unsigned int cfq_quantum;
279         unsigned int cfq_fifo_expire[2];
280         unsigned int cfq_back_penalty;
281         unsigned int cfq_back_max;
282         unsigned int cfq_slice[2];
283         unsigned int cfq_slice_async_rq;
284         unsigned int cfq_slice_idle;
285         unsigned int cfq_group_idle;
286         unsigned int cfq_latency;
287
288         unsigned int cic_index;
289         struct list_head cic_list;
290
291         /*
292          * Fallback dummy cfqq for extreme OOM conditions
293          */
294         struct cfq_queue oom_cfqq;
295
296         unsigned long last_delayed_sync;
297
298         /* List of cfq groups being managed on this device*/
299         struct hlist_head cfqg_list;
300         struct rcu_head rcu;
301 };
302
303 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
304
305 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
306                                             enum wl_prio_t prio,
307                                             enum wl_type_t type)
308 {
309         if (!cfqg)
310                 return NULL;
311
312         if (prio == IDLE_WORKLOAD)
313                 return &cfqg->service_tree_idle;
314
315         return &cfqg->service_trees[prio][type];
316 }
317
318 enum cfqq_state_flags {
319         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
320         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
321         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
322         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
323         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
324         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
325         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
326         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
327         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
328         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
329         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
330         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
331         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
332 };
333
334 #define CFQ_CFQQ_FNS(name)                                              \
335 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
336 {                                                                       \
337         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
338 }                                                                       \
339 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
340 {                                                                       \
341         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
342 }                                                                       \
343 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
344 {                                                                       \
345         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
346 }
347
348 CFQ_CFQQ_FNS(on_rr);
349 CFQ_CFQQ_FNS(wait_request);
350 CFQ_CFQQ_FNS(must_dispatch);
351 CFQ_CFQQ_FNS(must_alloc_slice);
352 CFQ_CFQQ_FNS(fifo_expire);
353 CFQ_CFQQ_FNS(idle_window);
354 CFQ_CFQQ_FNS(prio_changed);
355 CFQ_CFQQ_FNS(slice_new);
356 CFQ_CFQQ_FNS(sync);
357 CFQ_CFQQ_FNS(coop);
358 CFQ_CFQQ_FNS(split_coop);
359 CFQ_CFQQ_FNS(deep);
360 CFQ_CFQQ_FNS(wait_busy);
361 #undef CFQ_CFQQ_FNS
362
363 #ifdef CONFIG_CFQ_GROUP_IOSCHED
364 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
365         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
366                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
367                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
368
369 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
370         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
371                                 blkg_path(&(cfqg)->blkg), ##args);      \
372
373 #else
374 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
375         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
376 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
377 #endif
378 #define cfq_log(cfqd, fmt, args...)     \
379         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
380
381 /* Traverses through cfq group service trees */
382 #define for_each_cfqg_st(cfqg, i, j, st) \
383         for (i = 0; i <= IDLE_WORKLOAD; i++) \
384                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
385                         : &cfqg->service_tree_idle; \
386                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
387                         (i == IDLE_WORKLOAD && j == 0); \
388                         j++, st = i < IDLE_WORKLOAD ? \
389                         &cfqg->service_trees[i][j]: NULL) \
390
391
392 static inline bool iops_mode(struct cfq_data *cfqd)
393 {
394         /*
395          * If we are not idling on queues and it is a NCQ drive, parallel
396          * execution of requests is on and measuring time is not possible
397          * in most of the cases until and unless we drive shallower queue
398          * depths and that becomes a performance bottleneck. In such cases
399          * switch to start providing fairness in terms of number of IOs.
400          */
401         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
402                 return true;
403         else
404                 return false;
405 }
406
407 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
408 {
409         if (cfq_class_idle(cfqq))
410                 return IDLE_WORKLOAD;
411         if (cfq_class_rt(cfqq))
412                 return RT_WORKLOAD;
413         return BE_WORKLOAD;
414 }
415
416
417 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
418 {
419         if (!cfq_cfqq_sync(cfqq))
420                 return ASYNC_WORKLOAD;
421         if (!cfq_cfqq_idle_window(cfqq))
422                 return SYNC_NOIDLE_WORKLOAD;
423         return SYNC_WORKLOAD;
424 }
425
426 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
427                                         struct cfq_data *cfqd,
428                                         struct cfq_group *cfqg)
429 {
430         if (wl == IDLE_WORKLOAD)
431                 return cfqg->service_tree_idle.count;
432
433         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
434                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
435                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
436 }
437
438 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
439                                         struct cfq_group *cfqg)
440 {
441         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
442                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
443 }
444
445 static void cfq_dispatch_insert(struct request_queue *, struct request *);
446 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
447                                        struct io_context *, gfp_t);
448 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
449                                                 struct io_context *);
450
451 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
452                                             bool is_sync)
453 {
454         return cic->cfqq[is_sync];
455 }
456
457 static inline void cic_set_cfqq(struct cfq_io_context *cic,
458                                 struct cfq_queue *cfqq, bool is_sync)
459 {
460         cic->cfqq[is_sync] = cfqq;
461 }
462
463 #define CIC_DEAD_KEY    1ul
464 #define CIC_DEAD_INDEX_SHIFT    1
465
466 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
467 {
468         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
469 }
470
471 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
472 {
473         struct cfq_data *cfqd = cic->key;
474
475         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
476                 return NULL;
477
478         return cfqd;
479 }
480
481 /*
482  * We regard a request as SYNC, if it's either a read or has the SYNC bit
483  * set (in which case it could also be direct WRITE).
484  */
485 static inline bool cfq_bio_sync(struct bio *bio)
486 {
487         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
488 }
489
490 /*
491  * scheduler run of queue, if there are requests pending and no one in the
492  * driver that will restart queueing
493  */
494 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
495 {
496         if (cfqd->busy_queues) {
497                 cfq_log(cfqd, "schedule dispatch");
498                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
499         }
500 }
501
502 static int cfq_queue_empty(struct request_queue *q)
503 {
504         struct cfq_data *cfqd = q->elevator->elevator_data;
505
506         return !cfqd->rq_queued;
507 }
508
509 /*
510  * Scale schedule slice based on io priority. Use the sync time slice only
511  * if a queue is marked sync and has sync io queued. A sync queue with async
512  * io only, should not get full sync slice length.
513  */
514 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
515                                  unsigned short prio)
516 {
517         const int base_slice = cfqd->cfq_slice[sync];
518
519         WARN_ON(prio >= IOPRIO_BE_NR);
520
521         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
522 }
523
524 static inline int
525 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
526 {
527         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
528 }
529
530 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
531 {
532         u64 d = delta << CFQ_SERVICE_SHIFT;
533
534         d = d * BLKIO_WEIGHT_DEFAULT;
535         do_div(d, cfqg->weight);
536         return d;
537 }
538
539 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
540 {
541         s64 delta = (s64)(vdisktime - min_vdisktime);
542         if (delta > 0)
543                 min_vdisktime = vdisktime;
544
545         return min_vdisktime;
546 }
547
548 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
549 {
550         s64 delta = (s64)(vdisktime - min_vdisktime);
551         if (delta < 0)
552                 min_vdisktime = vdisktime;
553
554         return min_vdisktime;
555 }
556
557 static void update_min_vdisktime(struct cfq_rb_root *st)
558 {
559         u64 vdisktime = st->min_vdisktime;
560         struct cfq_group *cfqg;
561
562         if (st->left) {
563                 cfqg = rb_entry_cfqg(st->left);
564                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
565         }
566
567         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
568 }
569
570 /*
571  * get averaged number of queues of RT/BE priority.
572  * average is updated, with a formula that gives more weight to higher numbers,
573  * to quickly follows sudden increases and decrease slowly
574  */
575
576 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
577                                         struct cfq_group *cfqg, bool rt)
578 {
579         unsigned min_q, max_q;
580         unsigned mult  = cfq_hist_divisor - 1;
581         unsigned round = cfq_hist_divisor / 2;
582         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
583
584         min_q = min(cfqg->busy_queues_avg[rt], busy);
585         max_q = max(cfqg->busy_queues_avg[rt], busy);
586         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
587                 cfq_hist_divisor;
588         return cfqg->busy_queues_avg[rt];
589 }
590
591 static inline unsigned
592 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
593 {
594         struct cfq_rb_root *st = &cfqd->grp_service_tree;
595
596         return cfq_target_latency * cfqg->weight / st->total_weight;
597 }
598
599 static inline unsigned
600 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
601 {
602         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
603         if (cfqd->cfq_latency) {
604                 /*
605                  * interested queues (we consider only the ones with the same
606                  * priority class in the cfq group)
607                  */
608                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
609                                                 cfq_class_rt(cfqq));
610                 unsigned sync_slice = cfqd->cfq_slice[1];
611                 unsigned expect_latency = sync_slice * iq;
612                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
613
614                 if (expect_latency > group_slice) {
615                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
616                         /* scale low_slice according to IO priority
617                          * and sync vs async */
618                         unsigned low_slice =
619                                 min(slice, base_low_slice * slice / sync_slice);
620                         /* the adapted slice value is scaled to fit all iqs
621                          * into the target latency */
622                         slice = max(slice * group_slice / expect_latency,
623                                     low_slice);
624                 }
625         }
626         return slice;
627 }
628
629 static inline void
630 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
631 {
632         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
633
634         cfqq->slice_start = jiffies;
635         cfqq->slice_end = jiffies + slice;
636         cfqq->allocated_slice = slice;
637         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
638 }
639
640 /*
641  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
642  * isn't valid until the first request from the dispatch is activated
643  * and the slice time set.
644  */
645 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
646 {
647         if (cfq_cfqq_slice_new(cfqq))
648                 return false;
649         if (time_before(jiffies, cfqq->slice_end))
650                 return false;
651
652         return true;
653 }
654
655 /*
656  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
657  * We choose the request that is closest to the head right now. Distance
658  * behind the head is penalized and only allowed to a certain extent.
659  */
660 static struct request *
661 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
662 {
663         sector_t s1, s2, d1 = 0, d2 = 0;
664         unsigned long back_max;
665 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
666 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
667         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
668
669         if (rq1 == NULL || rq1 == rq2)
670                 return rq2;
671         if (rq2 == NULL)
672                 return rq1;
673
674         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
675                 return rq1;
676         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
677                 return rq2;
678         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
679                 return rq1;
680         else if ((rq2->cmd_flags & REQ_META) &&
681                  !(rq1->cmd_flags & REQ_META))
682                 return rq2;
683
684         s1 = blk_rq_pos(rq1);
685         s2 = blk_rq_pos(rq2);
686
687         /*
688          * by definition, 1KiB is 2 sectors
689          */
690         back_max = cfqd->cfq_back_max * 2;
691
692         /*
693          * Strict one way elevator _except_ in the case where we allow
694          * short backward seeks which are biased as twice the cost of a
695          * similar forward seek.
696          */
697         if (s1 >= last)
698                 d1 = s1 - last;
699         else if (s1 + back_max >= last)
700                 d1 = (last - s1) * cfqd->cfq_back_penalty;
701         else
702                 wrap |= CFQ_RQ1_WRAP;
703
704         if (s2 >= last)
705                 d2 = s2 - last;
706         else if (s2 + back_max >= last)
707                 d2 = (last - s2) * cfqd->cfq_back_penalty;
708         else
709                 wrap |= CFQ_RQ2_WRAP;
710
711         /* Found required data */
712
713         /*
714          * By doing switch() on the bit mask "wrap" we avoid having to
715          * check two variables for all permutations: --> faster!
716          */
717         switch (wrap) {
718         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
719                 if (d1 < d2)
720                         return rq1;
721                 else if (d2 < d1)
722                         return rq2;
723                 else {
724                         if (s1 >= s2)
725                                 return rq1;
726                         else
727                                 return rq2;
728                 }
729
730         case CFQ_RQ2_WRAP:
731                 return rq1;
732         case CFQ_RQ1_WRAP:
733                 return rq2;
734         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
735         default:
736                 /*
737                  * Since both rqs are wrapped,
738                  * start with the one that's further behind head
739                  * (--> only *one* back seek required),
740                  * since back seek takes more time than forward.
741                  */
742                 if (s1 <= s2)
743                         return rq1;
744                 else
745                         return rq2;
746         }
747 }
748
749 /*
750  * The below is leftmost cache rbtree addon
751  */
752 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
753 {
754         /* Service tree is empty */
755         if (!root->count)
756                 return NULL;
757
758         if (!root->left)
759                 root->left = rb_first(&root->rb);
760
761         if (root->left)
762                 return rb_entry(root->left, struct cfq_queue, rb_node);
763
764         return NULL;
765 }
766
767 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
768 {
769         if (!root->left)
770                 root->left = rb_first(&root->rb);
771
772         if (root->left)
773                 return rb_entry_cfqg(root->left);
774
775         return NULL;
776 }
777
778 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
779 {
780         rb_erase(n, root);
781         RB_CLEAR_NODE(n);
782 }
783
784 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
785 {
786         if (root->left == n)
787                 root->left = NULL;
788         rb_erase_init(n, &root->rb);
789         --root->count;
790 }
791
792 /*
793  * would be nice to take fifo expire time into account as well
794  */
795 static struct request *
796 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
797                   struct request *last)
798 {
799         struct rb_node *rbnext = rb_next(&last->rb_node);
800         struct rb_node *rbprev = rb_prev(&last->rb_node);
801         struct request *next = NULL, *prev = NULL;
802
803         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
804
805         if (rbprev)
806                 prev = rb_entry_rq(rbprev);
807
808         if (rbnext)
809                 next = rb_entry_rq(rbnext);
810         else {
811                 rbnext = rb_first(&cfqq->sort_list);
812                 if (rbnext && rbnext != &last->rb_node)
813                         next = rb_entry_rq(rbnext);
814         }
815
816         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
817 }
818
819 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
820                                       struct cfq_queue *cfqq)
821 {
822         /*
823          * just an approximation, should be ok.
824          */
825         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
826                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
827 }
828
829 static inline s64
830 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
831 {
832         return cfqg->vdisktime - st->min_vdisktime;
833 }
834
835 static void
836 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
837 {
838         struct rb_node **node = &st->rb.rb_node;
839         struct rb_node *parent = NULL;
840         struct cfq_group *__cfqg;
841         s64 key = cfqg_key(st, cfqg);
842         int left = 1;
843
844         while (*node != NULL) {
845                 parent = *node;
846                 __cfqg = rb_entry_cfqg(parent);
847
848                 if (key < cfqg_key(st, __cfqg))
849                         node = &parent->rb_left;
850                 else {
851                         node = &parent->rb_right;
852                         left = 0;
853                 }
854         }
855
856         if (left)
857                 st->left = &cfqg->rb_node;
858
859         rb_link_node(&cfqg->rb_node, parent, node);
860         rb_insert_color(&cfqg->rb_node, &st->rb);
861 }
862
863 static void
864 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
865 {
866         struct cfq_rb_root *st = &cfqd->grp_service_tree;
867         struct cfq_group *__cfqg;
868         struct rb_node *n;
869
870         cfqg->nr_cfqq++;
871         if (!RB_EMPTY_NODE(&cfqg->rb_node))
872                 return;
873
874         /*
875          * Currently put the group at the end. Later implement something
876          * so that groups get lesser vtime based on their weights, so that
877          * if group does not loose all if it was not continously backlogged.
878          */
879         n = rb_last(&st->rb);
880         if (n) {
881                 __cfqg = rb_entry_cfqg(n);
882                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
883         } else
884                 cfqg->vdisktime = st->min_vdisktime;
885
886         __cfq_group_service_tree_add(st, cfqg);
887         st->total_weight += cfqg->weight;
888 }
889
890 static void
891 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
892 {
893         struct cfq_rb_root *st = &cfqd->grp_service_tree;
894
895         BUG_ON(cfqg->nr_cfqq < 1);
896         cfqg->nr_cfqq--;
897
898         /* If there are other cfq queues under this group, don't delete it */
899         if (cfqg->nr_cfqq)
900                 return;
901
902         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
903         st->total_weight -= cfqg->weight;
904         if (!RB_EMPTY_NODE(&cfqg->rb_node))
905                 cfq_rb_erase(&cfqg->rb_node, st);
906         cfqg->saved_workload_slice = 0;
907         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
908 }
909
910 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
911 {
912         unsigned int slice_used;
913
914         /*
915          * Queue got expired before even a single request completed or
916          * got expired immediately after first request completion.
917          */
918         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
919                 /*
920                  * Also charge the seek time incurred to the group, otherwise
921                  * if there are mutiple queues in the group, each can dispatch
922                  * a single request on seeky media and cause lots of seek time
923                  * and group will never know it.
924                  */
925                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
926                                         1);
927         } else {
928                 slice_used = jiffies - cfqq->slice_start;
929                 if (slice_used > cfqq->allocated_slice)
930                         slice_used = cfqq->allocated_slice;
931         }
932
933         return slice_used;
934 }
935
936 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
937                                 struct cfq_queue *cfqq)
938 {
939         struct cfq_rb_root *st = &cfqd->grp_service_tree;
940         unsigned int used_sl, charge;
941         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
942                         - cfqg->service_tree_idle.count;
943
944         BUG_ON(nr_sync < 0);
945         used_sl = charge = cfq_cfqq_slice_usage(cfqq);
946
947         if (iops_mode(cfqd))
948                 charge = cfqq->slice_dispatch;
949         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
950                 charge = cfqq->allocated_slice;
951
952         /* Can't update vdisktime while group is on service tree */
953         cfq_rb_erase(&cfqg->rb_node, st);
954         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
955         __cfq_group_service_tree_add(st, cfqg);
956
957         /* This group is being expired. Save the context */
958         if (time_after(cfqd->workload_expires, jiffies)) {
959                 cfqg->saved_workload_slice = cfqd->workload_expires
960                                                 - jiffies;
961                 cfqg->saved_workload = cfqd->serving_type;
962                 cfqg->saved_serving_prio = cfqd->serving_prio;
963         } else
964                 cfqg->saved_workload_slice = 0;
965
966         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
967                                         st->min_vdisktime);
968         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
969                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
970                         iops_mode(cfqd), cfqq->nr_sectors);
971         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
972         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
973 }
974
975 #ifdef CONFIG_CFQ_GROUP_IOSCHED
976 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
977 {
978         if (blkg)
979                 return container_of(blkg, struct cfq_group, blkg);
980         return NULL;
981 }
982
983 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
984                                         unsigned int weight)
985 {
986         cfqg_of_blkg(blkg)->weight = weight;
987 }
988
989 static struct cfq_group *
990 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
991 {
992         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
993         struct cfq_group *cfqg = NULL;
994         void *key = cfqd;
995         int i, j;
996         struct cfq_rb_root *st;
997         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
998         unsigned int major, minor;
999
1000         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1001         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1002                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1003                 cfqg->blkg.dev = MKDEV(major, minor);
1004                 goto done;
1005         }
1006         if (cfqg || !create)
1007                 goto done;
1008
1009         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1010         if (!cfqg)
1011                 goto done;
1012
1013         for_each_cfqg_st(cfqg, i, j, st)
1014                 *st = CFQ_RB_ROOT;
1015         RB_CLEAR_NODE(&cfqg->rb_node);
1016
1017         /*
1018          * Take the initial reference that will be released on destroy
1019          * This can be thought of a joint reference by cgroup and
1020          * elevator which will be dropped by either elevator exit
1021          * or cgroup deletion path depending on who is exiting first.
1022          */
1023         cfqg->ref = 1;
1024
1025         /*
1026          * Add group onto cgroup list. It might happen that bdi->dev is
1027          * not initialized yet. Initialize this new group without major
1028          * and minor info and this info will be filled in once a new thread
1029          * comes for IO. See code above.
1030          */
1031         if (bdi->dev) {
1032                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1033                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1034                                         MKDEV(major, minor));
1035         } else
1036                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1037                                         0);
1038
1039         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1040
1041         /* Add group on cfqd list */
1042         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1043
1044 done:
1045         return cfqg;
1046 }
1047
1048 /*
1049  * Search for the cfq group current task belongs to. If create = 1, then also
1050  * create the cfq group if it does not exist. request_queue lock must be held.
1051  */
1052 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1053 {
1054         struct cgroup *cgroup;
1055         struct cfq_group *cfqg = NULL;
1056
1057         rcu_read_lock();
1058         cgroup = task_cgroup(current, blkio_subsys_id);
1059         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1060         if (!cfqg && create)
1061                 cfqg = &cfqd->root_group;
1062         rcu_read_unlock();
1063         return cfqg;
1064 }
1065
1066 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1067 {
1068         cfqg->ref++;
1069         return cfqg;
1070 }
1071
1072 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1073 {
1074         /* Currently, all async queues are mapped to root group */
1075         if (!cfq_cfqq_sync(cfqq))
1076                 cfqg = &cfqq->cfqd->root_group;
1077
1078         cfqq->cfqg = cfqg;
1079         /* cfqq reference on cfqg */
1080         cfqq->cfqg->ref++;
1081 }
1082
1083 static void cfq_put_cfqg(struct cfq_group *cfqg)
1084 {
1085         struct cfq_rb_root *st;
1086         int i, j;
1087
1088         BUG_ON(cfqg->ref <= 0);
1089         cfqg->ref--;
1090         if (cfqg->ref)
1091                 return;
1092         for_each_cfqg_st(cfqg, i, j, st)
1093                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1094         kfree(cfqg);
1095 }
1096
1097 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1098 {
1099         /* Something wrong if we are trying to remove same group twice */
1100         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1101
1102         hlist_del_init(&cfqg->cfqd_node);
1103
1104         /*
1105          * Put the reference taken at the time of creation so that when all
1106          * queues are gone, group can be destroyed.
1107          */
1108         cfq_put_cfqg(cfqg);
1109 }
1110
1111 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1112 {
1113         struct hlist_node *pos, *n;
1114         struct cfq_group *cfqg;
1115
1116         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1117                 /*
1118                  * If cgroup removal path got to blk_group first and removed
1119                  * it from cgroup list, then it will take care of destroying
1120                  * cfqg also.
1121                  */
1122                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1123                         cfq_destroy_cfqg(cfqd, cfqg);
1124         }
1125 }
1126
1127 /*
1128  * Blk cgroup controller notification saying that blkio_group object is being
1129  * delinked as associated cgroup object is going away. That also means that
1130  * no new IO will come in this group. So get rid of this group as soon as
1131  * any pending IO in the group is finished.
1132  *
1133  * This function is called under rcu_read_lock(). key is the rcu protected
1134  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1135  * read lock.
1136  *
1137  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1138  * it should not be NULL as even if elevator was exiting, cgroup deltion
1139  * path got to it first.
1140  */
1141 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1142 {
1143         unsigned long  flags;
1144         struct cfq_data *cfqd = key;
1145
1146         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1147         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1148         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1149 }
1150
1151 #else /* GROUP_IOSCHED */
1152 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1153 {
1154         return &cfqd->root_group;
1155 }
1156
1157 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1158 {
1159         return cfqg;
1160 }
1161
1162 static inline void
1163 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1164         cfqq->cfqg = cfqg;
1165 }
1166
1167 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1168 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1169
1170 #endif /* GROUP_IOSCHED */
1171
1172 /*
1173  * The cfqd->service_trees holds all pending cfq_queue's that have
1174  * requests waiting to be processed. It is sorted in the order that
1175  * we will service the queues.
1176  */
1177 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1178                                  bool add_front)
1179 {
1180         struct rb_node **p, *parent;
1181         struct cfq_queue *__cfqq;
1182         unsigned long rb_key;
1183         struct cfq_rb_root *service_tree;
1184         int left;
1185         int new_cfqq = 1;
1186         int group_changed = 0;
1187
1188         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1189                                                 cfqq_type(cfqq));
1190         if (cfq_class_idle(cfqq)) {
1191                 rb_key = CFQ_IDLE_DELAY;
1192                 parent = rb_last(&service_tree->rb);
1193                 if (parent && parent != &cfqq->rb_node) {
1194                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1195                         rb_key += __cfqq->rb_key;
1196                 } else
1197                         rb_key += jiffies;
1198         } else if (!add_front) {
1199                 /*
1200                  * Get our rb key offset. Subtract any residual slice
1201                  * value carried from last service. A negative resid
1202                  * count indicates slice overrun, and this should position
1203                  * the next service time further away in the tree.
1204                  */
1205                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1206                 rb_key -= cfqq->slice_resid;
1207                 cfqq->slice_resid = 0;
1208         } else {
1209                 rb_key = -HZ;
1210                 __cfqq = cfq_rb_first(service_tree);
1211                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1212         }
1213
1214         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1215                 new_cfqq = 0;
1216                 /*
1217                  * same position, nothing more to do
1218                  */
1219                 if (rb_key == cfqq->rb_key &&
1220                     cfqq->service_tree == service_tree)
1221                         return;
1222
1223                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1224                 cfqq->service_tree = NULL;
1225         }
1226
1227         left = 1;
1228         parent = NULL;
1229         cfqq->service_tree = service_tree;
1230         p = &service_tree->rb.rb_node;
1231         while (*p) {
1232                 struct rb_node **n;
1233
1234                 parent = *p;
1235                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1236
1237                 /*
1238                  * sort by key, that represents service time.
1239                  */
1240                 if (time_before(rb_key, __cfqq->rb_key))
1241                         n = &(*p)->rb_left;
1242                 else {
1243                         n = &(*p)->rb_right;
1244                         left = 0;
1245                 }
1246
1247                 p = n;
1248         }
1249
1250         if (left)
1251                 service_tree->left = &cfqq->rb_node;
1252
1253         cfqq->rb_key = rb_key;
1254         rb_link_node(&cfqq->rb_node, parent, p);
1255         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1256         service_tree->count++;
1257         if ((add_front || !new_cfqq) && !group_changed)
1258                 return;
1259         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1260 }
1261
1262 static struct cfq_queue *
1263 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1264                      sector_t sector, struct rb_node **ret_parent,
1265                      struct rb_node ***rb_link)
1266 {
1267         struct rb_node **p, *parent;
1268         struct cfq_queue *cfqq = NULL;
1269
1270         parent = NULL;
1271         p = &root->rb_node;
1272         while (*p) {
1273                 struct rb_node **n;
1274
1275                 parent = *p;
1276                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1277
1278                 /*
1279                  * Sort strictly based on sector.  Smallest to the left,
1280                  * largest to the right.
1281                  */
1282                 if (sector > blk_rq_pos(cfqq->next_rq))
1283                         n = &(*p)->rb_right;
1284                 else if (sector < blk_rq_pos(cfqq->next_rq))
1285                         n = &(*p)->rb_left;
1286                 else
1287                         break;
1288                 p = n;
1289                 cfqq = NULL;
1290         }
1291
1292         *ret_parent = parent;
1293         if (rb_link)
1294                 *rb_link = p;
1295         return cfqq;
1296 }
1297
1298 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1299 {
1300         struct rb_node **p, *parent;
1301         struct cfq_queue *__cfqq;
1302
1303         if (cfqq->p_root) {
1304                 rb_erase(&cfqq->p_node, cfqq->p_root);
1305                 cfqq->p_root = NULL;
1306         }
1307
1308         if (cfq_class_idle(cfqq))
1309                 return;
1310         if (!cfqq->next_rq)
1311                 return;
1312
1313         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1314         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1315                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1316         if (!__cfqq) {
1317                 rb_link_node(&cfqq->p_node, parent, p);
1318                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1319         } else
1320                 cfqq->p_root = NULL;
1321 }
1322
1323 /*
1324  * Update cfqq's position in the service tree.
1325  */
1326 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1327 {
1328         /*
1329          * Resorting requires the cfqq to be on the RR list already.
1330          */
1331         if (cfq_cfqq_on_rr(cfqq)) {
1332                 cfq_service_tree_add(cfqd, cfqq, 0);
1333                 cfq_prio_tree_add(cfqd, cfqq);
1334         }
1335 }
1336
1337 /*
1338  * add to busy list of queues for service, trying to be fair in ordering
1339  * the pending list according to last request service
1340  */
1341 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1342 {
1343         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1344         BUG_ON(cfq_cfqq_on_rr(cfqq));
1345         cfq_mark_cfqq_on_rr(cfqq);
1346         cfqd->busy_queues++;
1347
1348         cfq_resort_rr_list(cfqd, cfqq);
1349 }
1350
1351 /*
1352  * Called when the cfqq no longer has requests pending, remove it from
1353  * the service tree.
1354  */
1355 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1356 {
1357         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1358         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1359         cfq_clear_cfqq_on_rr(cfqq);
1360
1361         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1362                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1363                 cfqq->service_tree = NULL;
1364         }
1365         if (cfqq->p_root) {
1366                 rb_erase(&cfqq->p_node, cfqq->p_root);
1367                 cfqq->p_root = NULL;
1368         }
1369
1370         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1371         BUG_ON(!cfqd->busy_queues);
1372         cfqd->busy_queues--;
1373 }
1374
1375 /*
1376  * rb tree support functions
1377  */
1378 static void cfq_del_rq_rb(struct request *rq)
1379 {
1380         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1381         const int sync = rq_is_sync(rq);
1382
1383         BUG_ON(!cfqq->queued[sync]);
1384         cfqq->queued[sync]--;
1385
1386         elv_rb_del(&cfqq->sort_list, rq);
1387
1388         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1389                 /*
1390                  * Queue will be deleted from service tree when we actually
1391                  * expire it later. Right now just remove it from prio tree
1392                  * as it is empty.
1393                  */
1394                 if (cfqq->p_root) {
1395                         rb_erase(&cfqq->p_node, cfqq->p_root);
1396                         cfqq->p_root = NULL;
1397                 }
1398         }
1399 }
1400
1401 static void cfq_add_rq_rb(struct request *rq)
1402 {
1403         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1404         struct cfq_data *cfqd = cfqq->cfqd;
1405         struct request *__alias, *prev;
1406
1407         cfqq->queued[rq_is_sync(rq)]++;
1408
1409         /*
1410          * looks a little odd, but the first insert might return an alias.
1411          * if that happens, put the alias on the dispatch list
1412          */
1413         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1414                 cfq_dispatch_insert(cfqd->queue, __alias);
1415
1416         if (!cfq_cfqq_on_rr(cfqq))
1417                 cfq_add_cfqq_rr(cfqd, cfqq);
1418
1419         /*
1420          * check if this request is a better next-serve candidate
1421          */
1422         prev = cfqq->next_rq;
1423         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1424
1425         /*
1426          * adjust priority tree position, if ->next_rq changes
1427          */
1428         if (prev != cfqq->next_rq)
1429                 cfq_prio_tree_add(cfqd, cfqq);
1430
1431         BUG_ON(!cfqq->next_rq);
1432 }
1433
1434 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1435 {
1436         elv_rb_del(&cfqq->sort_list, rq);
1437         cfqq->queued[rq_is_sync(rq)]--;
1438         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1439                                         rq_data_dir(rq), rq_is_sync(rq));
1440         cfq_add_rq_rb(rq);
1441         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1442                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1443                         rq_is_sync(rq));
1444 }
1445
1446 static struct request *
1447 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1448 {
1449         struct task_struct *tsk = current;
1450         struct cfq_io_context *cic;
1451         struct cfq_queue *cfqq;
1452
1453         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1454         if (!cic)
1455                 return NULL;
1456
1457         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1458         if (cfqq) {
1459                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1460
1461                 return elv_rb_find(&cfqq->sort_list, sector);
1462         }
1463
1464         return NULL;
1465 }
1466
1467 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1468 {
1469         struct cfq_data *cfqd = q->elevator->elevator_data;
1470
1471         cfqd->rq_in_driver++;
1472         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1473                                                 cfqd->rq_in_driver);
1474
1475         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1476 }
1477
1478 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1479 {
1480         struct cfq_data *cfqd = q->elevator->elevator_data;
1481
1482         WARN_ON(!cfqd->rq_in_driver);
1483         cfqd->rq_in_driver--;
1484         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1485                                                 cfqd->rq_in_driver);
1486 }
1487
1488 static void cfq_remove_request(struct request *rq)
1489 {
1490         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1491
1492         if (cfqq->next_rq == rq)
1493                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1494
1495         list_del_init(&rq->queuelist);
1496         cfq_del_rq_rb(rq);
1497
1498         cfqq->cfqd->rq_queued--;
1499         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1500                                         rq_data_dir(rq), rq_is_sync(rq));
1501         if (rq->cmd_flags & REQ_META) {
1502                 WARN_ON(!cfqq->meta_pending);
1503                 cfqq->meta_pending--;
1504         }
1505 }
1506
1507 static int cfq_merge(struct request_queue *q, struct request **req,
1508                      struct bio *bio)
1509 {
1510         struct cfq_data *cfqd = q->elevator->elevator_data;
1511         struct request *__rq;
1512
1513         __rq = cfq_find_rq_fmerge(cfqd, bio);
1514         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1515                 *req = __rq;
1516                 return ELEVATOR_FRONT_MERGE;
1517         }
1518
1519         return ELEVATOR_NO_MERGE;
1520 }
1521
1522 static void cfq_merged_request(struct request_queue *q, struct request *req,
1523                                int type)
1524 {
1525         if (type == ELEVATOR_FRONT_MERGE) {
1526                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1527
1528                 cfq_reposition_rq_rb(cfqq, req);
1529         }
1530 }
1531
1532 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1533                                 struct bio *bio)
1534 {
1535         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1536                                         bio_data_dir(bio), cfq_bio_sync(bio));
1537 }
1538
1539 static void
1540 cfq_merged_requests(struct request_queue *q, struct request *rq,
1541                     struct request *next)
1542 {
1543         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1544         /*
1545          * reposition in fifo if next is older than rq
1546          */
1547         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1548             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1549                 list_move(&rq->queuelist, &next->queuelist);
1550                 rq_set_fifo_time(rq, rq_fifo_time(next));
1551         }
1552
1553         if (cfqq->next_rq == next)
1554                 cfqq->next_rq = rq;
1555         cfq_remove_request(next);
1556         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1557                                         rq_data_dir(next), rq_is_sync(next));
1558 }
1559
1560 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1561                            struct bio *bio)
1562 {
1563         struct cfq_data *cfqd = q->elevator->elevator_data;
1564         struct cfq_io_context *cic;
1565         struct cfq_queue *cfqq;
1566
1567         /*
1568          * Disallow merge of a sync bio into an async request.
1569          */
1570         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1571                 return false;
1572
1573         /*
1574          * Lookup the cfqq that this bio will be queued with. Allow
1575          * merge only if rq is queued there.
1576          */
1577         cic = cfq_cic_lookup(cfqd, current->io_context);
1578         if (!cic)
1579                 return false;
1580
1581         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1582         return cfqq == RQ_CFQQ(rq);
1583 }
1584
1585 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1586 {
1587         del_timer(&cfqd->idle_slice_timer);
1588         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1589 }
1590
1591 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1592                                    struct cfq_queue *cfqq)
1593 {
1594         if (cfqq) {
1595                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1596                                 cfqd->serving_prio, cfqd->serving_type);
1597                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1598                 cfqq->slice_start = 0;
1599                 cfqq->dispatch_start = jiffies;
1600                 cfqq->allocated_slice = 0;
1601                 cfqq->slice_end = 0;
1602                 cfqq->slice_dispatch = 0;
1603                 cfqq->nr_sectors = 0;
1604
1605                 cfq_clear_cfqq_wait_request(cfqq);
1606                 cfq_clear_cfqq_must_dispatch(cfqq);
1607                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1608                 cfq_clear_cfqq_fifo_expire(cfqq);
1609                 cfq_mark_cfqq_slice_new(cfqq);
1610
1611                 cfq_del_timer(cfqd, cfqq);
1612         }
1613
1614         cfqd->active_queue = cfqq;
1615 }
1616
1617 /*
1618  * current cfqq expired its slice (or was too idle), select new one
1619  */
1620 static void
1621 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1622                     bool timed_out)
1623 {
1624         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1625
1626         if (cfq_cfqq_wait_request(cfqq))
1627                 cfq_del_timer(cfqd, cfqq);
1628
1629         cfq_clear_cfqq_wait_request(cfqq);
1630         cfq_clear_cfqq_wait_busy(cfqq);
1631
1632         /*
1633          * If this cfqq is shared between multiple processes, check to
1634          * make sure that those processes are still issuing I/Os within
1635          * the mean seek distance.  If not, it may be time to break the
1636          * queues apart again.
1637          */
1638         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1639                 cfq_mark_cfqq_split_coop(cfqq);
1640
1641         /*
1642          * store what was left of this slice, if the queue idled/timed out
1643          */
1644         if (timed_out) {
1645                 if (cfq_cfqq_slice_new(cfqq))
1646                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1647                 else
1648                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1649                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1650         }
1651
1652         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1653
1654         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1655                 cfq_del_cfqq_rr(cfqd, cfqq);
1656
1657         cfq_resort_rr_list(cfqd, cfqq);
1658
1659         if (cfqq == cfqd->active_queue)
1660                 cfqd->active_queue = NULL;
1661
1662         if (cfqd->active_cic) {
1663                 put_io_context(cfqd->active_cic->ioc);
1664                 cfqd->active_cic = NULL;
1665         }
1666 }
1667
1668 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1669 {
1670         struct cfq_queue *cfqq = cfqd->active_queue;
1671
1672         if (cfqq)
1673                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1674 }
1675
1676 /*
1677  * Get next queue for service. Unless we have a queue preemption,
1678  * we'll simply select the first cfqq in the service tree.
1679  */
1680 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1681 {
1682         struct cfq_rb_root *service_tree =
1683                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1684                                         cfqd->serving_type);
1685
1686         if (!cfqd->rq_queued)
1687                 return NULL;
1688
1689         /* There is nothing to dispatch */
1690         if (!service_tree)
1691                 return NULL;
1692         if (RB_EMPTY_ROOT(&service_tree->rb))
1693                 return NULL;
1694         return cfq_rb_first(service_tree);
1695 }
1696
1697 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1698 {
1699         struct cfq_group *cfqg;
1700         struct cfq_queue *cfqq;
1701         int i, j;
1702         struct cfq_rb_root *st;
1703
1704         if (!cfqd->rq_queued)
1705                 return NULL;
1706
1707         cfqg = cfq_get_next_cfqg(cfqd);
1708         if (!cfqg)
1709                 return NULL;
1710
1711         for_each_cfqg_st(cfqg, i, j, st)
1712                 if ((cfqq = cfq_rb_first(st)) != NULL)
1713                         return cfqq;
1714         return NULL;
1715 }
1716
1717 /*
1718  * Get and set a new active queue for service.
1719  */
1720 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1721                                               struct cfq_queue *cfqq)
1722 {
1723         if (!cfqq)
1724                 cfqq = cfq_get_next_queue(cfqd);
1725
1726         __cfq_set_active_queue(cfqd, cfqq);
1727         return cfqq;
1728 }
1729
1730 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1731                                           struct request *rq)
1732 {
1733         if (blk_rq_pos(rq) >= cfqd->last_position)
1734                 return blk_rq_pos(rq) - cfqd->last_position;
1735         else
1736                 return cfqd->last_position - blk_rq_pos(rq);
1737 }
1738
1739 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1740                                struct request *rq)
1741 {
1742         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1743 }
1744
1745 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1746                                     struct cfq_queue *cur_cfqq)
1747 {
1748         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1749         struct rb_node *parent, *node;
1750         struct cfq_queue *__cfqq;
1751         sector_t sector = cfqd->last_position;
1752
1753         if (RB_EMPTY_ROOT(root))
1754                 return NULL;
1755
1756         /*
1757          * First, if we find a request starting at the end of the last
1758          * request, choose it.
1759          */
1760         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1761         if (__cfqq)
1762                 return __cfqq;
1763
1764         /*
1765          * If the exact sector wasn't found, the parent of the NULL leaf
1766          * will contain the closest sector.
1767          */
1768         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1769         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1770                 return __cfqq;
1771
1772         if (blk_rq_pos(__cfqq->next_rq) < sector)
1773                 node = rb_next(&__cfqq->p_node);
1774         else
1775                 node = rb_prev(&__cfqq->p_node);
1776         if (!node)
1777                 return NULL;
1778
1779         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1780         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1781                 return __cfqq;
1782
1783         return NULL;
1784 }
1785
1786 /*
1787  * cfqd - obvious
1788  * cur_cfqq - passed in so that we don't decide that the current queue is
1789  *            closely cooperating with itself.
1790  *
1791  * So, basically we're assuming that that cur_cfqq has dispatched at least
1792  * one request, and that cfqd->last_position reflects a position on the disk
1793  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1794  * assumption.
1795  */
1796 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1797                                               struct cfq_queue *cur_cfqq)
1798 {
1799         struct cfq_queue *cfqq;
1800
1801         if (cfq_class_idle(cur_cfqq))
1802                 return NULL;
1803         if (!cfq_cfqq_sync(cur_cfqq))
1804                 return NULL;
1805         if (CFQQ_SEEKY(cur_cfqq))
1806                 return NULL;
1807
1808         /*
1809          * Don't search priority tree if it's the only queue in the group.
1810          */
1811         if (cur_cfqq->cfqg->nr_cfqq == 1)
1812                 return NULL;
1813
1814         /*
1815          * We should notice if some of the queues are cooperating, eg
1816          * working closely on the same area of the disk. In that case,
1817          * we can group them together and don't waste time idling.
1818          */
1819         cfqq = cfqq_close(cfqd, cur_cfqq);
1820         if (!cfqq)
1821                 return NULL;
1822
1823         /* If new queue belongs to different cfq_group, don't choose it */
1824         if (cur_cfqq->cfqg != cfqq->cfqg)
1825                 return NULL;
1826
1827         /*
1828          * It only makes sense to merge sync queues.
1829          */
1830         if (!cfq_cfqq_sync(cfqq))
1831                 return NULL;
1832         if (CFQQ_SEEKY(cfqq))
1833                 return NULL;
1834
1835         /*
1836          * Do not merge queues of different priority classes
1837          */
1838         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1839                 return NULL;
1840
1841         return cfqq;
1842 }
1843
1844 /*
1845  * Determine whether we should enforce idle window for this queue.
1846  */
1847
1848 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1849 {
1850         enum wl_prio_t prio = cfqq_prio(cfqq);
1851         struct cfq_rb_root *service_tree = cfqq->service_tree;
1852
1853         BUG_ON(!service_tree);
1854         BUG_ON(!service_tree->count);
1855
1856         if (!cfqd->cfq_slice_idle)
1857                 return false;
1858
1859         /* We never do for idle class queues. */
1860         if (prio == IDLE_WORKLOAD)
1861                 return false;
1862
1863         /* We do for queues that were marked with idle window flag. */
1864         if (cfq_cfqq_idle_window(cfqq) &&
1865            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1866                 return true;
1867
1868         /*
1869          * Otherwise, we do only if they are the last ones
1870          * in their service tree.
1871          */
1872         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1873                 return true;
1874         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1875                         service_tree->count);
1876         return false;
1877 }
1878
1879 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1880 {
1881         struct cfq_queue *cfqq = cfqd->active_queue;
1882         struct cfq_io_context *cic;
1883         unsigned long sl, group_idle = 0;
1884
1885         /*
1886          * SSD device without seek penalty, disable idling. But only do so
1887          * for devices that support queuing, otherwise we still have a problem
1888          * with sync vs async workloads.
1889          */
1890         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1891                 return;
1892
1893         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1894         WARN_ON(cfq_cfqq_slice_new(cfqq));
1895
1896         /*
1897          * idle is disabled, either manually or by past process history
1898          */
1899         if (!cfq_should_idle(cfqd, cfqq)) {
1900                 /* no queue idling. Check for group idling */
1901                 if (cfqd->cfq_group_idle)
1902                         group_idle = cfqd->cfq_group_idle;
1903                 else
1904                         return;
1905         }
1906
1907         /*
1908          * still active requests from this queue, don't idle
1909          */
1910         if (cfqq->dispatched)
1911                 return;
1912
1913         /*
1914          * task has exited, don't wait
1915          */
1916         cic = cfqd->active_cic;
1917         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1918                 return;
1919
1920         /*
1921          * If our average think time is larger than the remaining time
1922          * slice, then don't idle. This avoids overrunning the allotted
1923          * time slice.
1924          */
1925         if (sample_valid(cic->ttime_samples) &&
1926             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1927                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1928                                 cic->ttime_mean);
1929                 return;
1930         }
1931
1932         /* There are other queues in the group, don't do group idle */
1933         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1934                 return;
1935
1936         cfq_mark_cfqq_wait_request(cfqq);
1937
1938         if (group_idle)
1939                 sl = cfqd->cfq_group_idle;
1940         else
1941                 sl = cfqd->cfq_slice_idle;
1942
1943         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1944         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1945         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1946                         group_idle ? 1 : 0);
1947 }
1948
1949 /*
1950  * Move request from internal lists to the request queue dispatch list.
1951  */
1952 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1953 {
1954         struct cfq_data *cfqd = q->elevator->elevator_data;
1955         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1956
1957         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1958
1959         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1960         cfq_remove_request(rq);
1961         cfqq->dispatched++;
1962         (RQ_CFQG(rq))->dispatched++;
1963         elv_dispatch_sort(q, rq);
1964
1965         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1966         cfqq->nr_sectors += blk_rq_sectors(rq);
1967         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1968                                         rq_data_dir(rq), rq_is_sync(rq));
1969 }
1970
1971 /*
1972  * return expired entry, or NULL to just start from scratch in rbtree
1973  */
1974 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1975 {
1976         struct request *rq = NULL;
1977
1978         if (cfq_cfqq_fifo_expire(cfqq))
1979                 return NULL;
1980
1981         cfq_mark_cfqq_fifo_expire(cfqq);
1982
1983         if (list_empty(&cfqq->fifo))
1984                 return NULL;
1985
1986         rq = rq_entry_fifo(cfqq->fifo.next);
1987         if (time_before(jiffies, rq_fifo_time(rq)))
1988                 rq = NULL;
1989
1990         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1991         return rq;
1992 }
1993
1994 static inline int
1995 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1996 {
1997         const int base_rq = cfqd->cfq_slice_async_rq;
1998
1999         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2000
2001         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2002 }
2003
2004 /*
2005  * Must be called with the queue_lock held.
2006  */
2007 static int cfqq_process_refs(struct cfq_queue *cfqq)
2008 {
2009         int process_refs, io_refs;
2010
2011         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2012         process_refs = cfqq->ref - io_refs;
2013         BUG_ON(process_refs < 0);
2014         return process_refs;
2015 }
2016
2017 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2018 {
2019         int process_refs, new_process_refs;
2020         struct cfq_queue *__cfqq;
2021
2022         /*
2023          * If there are no process references on the new_cfqq, then it is
2024          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2025          * chain may have dropped their last reference (not just their
2026          * last process reference).
2027          */
2028         if (!cfqq_process_refs(new_cfqq))
2029                 return;
2030
2031         /* Avoid a circular list and skip interim queue merges */
2032         while ((__cfqq = new_cfqq->new_cfqq)) {
2033                 if (__cfqq == cfqq)
2034                         return;
2035                 new_cfqq = __cfqq;
2036         }
2037
2038         process_refs = cfqq_process_refs(cfqq);
2039         new_process_refs = cfqq_process_refs(new_cfqq);
2040         /*
2041          * If the process for the cfqq has gone away, there is no
2042          * sense in merging the queues.
2043          */
2044         if (process_refs == 0 || new_process_refs == 0)
2045                 return;
2046
2047         /*
2048          * Merge in the direction of the lesser amount of work.
2049          */
2050         if (new_process_refs >= process_refs) {
2051                 cfqq->new_cfqq = new_cfqq;
2052                 new_cfqq->ref += process_refs;
2053         } else {
2054                 new_cfqq->new_cfqq = cfqq;
2055                 cfqq->ref += new_process_refs;
2056         }
2057 }
2058
2059 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2060                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2061 {
2062         struct cfq_queue *queue;
2063         int i;
2064         bool key_valid = false;
2065         unsigned long lowest_key = 0;
2066         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2067
2068         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2069                 /* select the one with lowest rb_key */
2070                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2071                 if (queue &&
2072                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2073                         lowest_key = queue->rb_key;
2074                         cur_best = i;
2075                         key_valid = true;
2076                 }
2077         }
2078
2079         return cur_best;
2080 }
2081
2082 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2083 {
2084         unsigned slice;
2085         unsigned count;
2086         struct cfq_rb_root *st;
2087         unsigned group_slice;
2088         enum wl_prio_t original_prio = cfqd->serving_prio;
2089
2090         /* Choose next priority. RT > BE > IDLE */
2091         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2092                 cfqd->serving_prio = RT_WORKLOAD;
2093         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2094                 cfqd->serving_prio = BE_WORKLOAD;
2095         else {
2096                 cfqd->serving_prio = IDLE_WORKLOAD;
2097                 cfqd->workload_expires = jiffies + 1;
2098                 return;
2099         }
2100
2101         if (original_prio != cfqd->serving_prio)
2102                 goto new_workload;
2103
2104         /*
2105          * For RT and BE, we have to choose also the type
2106          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2107          * expiration time
2108          */
2109         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2110         count = st->count;
2111
2112         /*
2113          * check workload expiration, and that we still have other queues ready
2114          */
2115         if (count && !time_after(jiffies, cfqd->workload_expires))
2116                 return;
2117
2118 new_workload:
2119         /* otherwise select new workload type */
2120         cfqd->serving_type =
2121                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2122         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2123         count = st->count;
2124
2125         /*
2126          * the workload slice is computed as a fraction of target latency
2127          * proportional to the number of queues in that workload, over
2128          * all the queues in the same priority class
2129          */
2130         group_slice = cfq_group_slice(cfqd, cfqg);
2131
2132         slice = group_slice * count /
2133                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2134                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2135
2136         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2137                 unsigned int tmp;
2138
2139                 /*
2140                  * Async queues are currently system wide. Just taking
2141                  * proportion of queues with-in same group will lead to higher
2142                  * async ratio system wide as generally root group is going
2143                  * to have higher weight. A more accurate thing would be to
2144                  * calculate system wide asnc/sync ratio.
2145                  */
2146                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2147                 tmp = tmp/cfqd->busy_queues;
2148                 slice = min_t(unsigned, slice, tmp);
2149
2150                 /* async workload slice is scaled down according to
2151                  * the sync/async slice ratio. */
2152                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2153         } else
2154                 /* sync workload slice is at least 2 * cfq_slice_idle */
2155                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2156
2157         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2158         cfq_log(cfqd, "workload slice:%d", slice);
2159         cfqd->workload_expires = jiffies + slice;
2160 }
2161
2162 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2163 {
2164         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2165         struct cfq_group *cfqg;
2166
2167         if (RB_EMPTY_ROOT(&st->rb))
2168                 return NULL;
2169         cfqg = cfq_rb_first_group(st);
2170         update_min_vdisktime(st);
2171         return cfqg;
2172 }
2173
2174 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2175 {
2176         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2177
2178         cfqd->serving_group = cfqg;
2179
2180         /* Restore the workload type data */
2181         if (cfqg->saved_workload_slice) {
2182                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2183                 cfqd->serving_type = cfqg->saved_workload;
2184                 cfqd->serving_prio = cfqg->saved_serving_prio;
2185         } else
2186                 cfqd->workload_expires = jiffies - 1;
2187
2188         choose_service_tree(cfqd, cfqg);
2189 }
2190
2191 /*
2192  * Select a queue for service. If we have a current active queue,
2193  * check whether to continue servicing it, or retrieve and set a new one.
2194  */
2195 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2196 {
2197         struct cfq_queue *cfqq, *new_cfqq = NULL;
2198
2199         cfqq = cfqd->active_queue;
2200         if (!cfqq)
2201                 goto new_queue;
2202
2203         if (!cfqd->rq_queued)
2204                 return NULL;
2205
2206         /*
2207          * We were waiting for group to get backlogged. Expire the queue
2208          */
2209         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2210                 goto expire;
2211
2212         /*
2213          * The active queue has run out of time, expire it and select new.
2214          */
2215         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2216                 /*
2217                  * If slice had not expired at the completion of last request
2218                  * we might not have turned on wait_busy flag. Don't expire
2219                  * the queue yet. Allow the group to get backlogged.
2220                  *
2221                  * The very fact that we have used the slice, that means we
2222                  * have been idling all along on this queue and it should be
2223                  * ok to wait for this request to complete.
2224                  */
2225                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2226                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2227                         cfqq = NULL;
2228                         goto keep_queue;
2229                 } else
2230                         goto check_group_idle;
2231         }
2232
2233         /*
2234          * The active queue has requests and isn't expired, allow it to
2235          * dispatch.
2236          */
2237         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2238                 goto keep_queue;
2239
2240         /*
2241          * If another queue has a request waiting within our mean seek
2242          * distance, let it run.  The expire code will check for close
2243          * cooperators and put the close queue at the front of the service
2244          * tree.  If possible, merge the expiring queue with the new cfqq.
2245          */
2246         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2247         if (new_cfqq) {
2248                 if (!cfqq->new_cfqq)
2249                         cfq_setup_merge(cfqq, new_cfqq);
2250                 goto expire;
2251         }
2252
2253         /*
2254          * No requests pending. If the active queue still has requests in
2255          * flight or is idling for a new request, allow either of these
2256          * conditions to happen (or time out) before selecting a new queue.
2257          */
2258         if (timer_pending(&cfqd->idle_slice_timer)) {
2259                 cfqq = NULL;
2260                 goto keep_queue;
2261         }
2262
2263         /*
2264          * This is a deep seek queue, but the device is much faster than
2265          * the queue can deliver, don't idle
2266          **/
2267         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2268             (cfq_cfqq_slice_new(cfqq) ||
2269             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2270                 cfq_clear_cfqq_deep(cfqq);
2271                 cfq_clear_cfqq_idle_window(cfqq);
2272         }
2273
2274         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2275                 cfqq = NULL;
2276                 goto keep_queue;
2277         }
2278
2279         /*
2280          * If group idle is enabled and there are requests dispatched from
2281          * this group, wait for requests to complete.
2282          */
2283 check_group_idle:
2284         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2285             && cfqq->cfqg->dispatched) {
2286                 cfqq = NULL;
2287                 goto keep_queue;
2288         }
2289
2290 expire:
2291         cfq_slice_expired(cfqd, 0);
2292 new_queue:
2293         /*
2294          * Current queue expired. Check if we have to switch to a new
2295          * service tree
2296          */
2297         if (!new_cfqq)
2298                 cfq_choose_cfqg(cfqd);
2299
2300         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2301 keep_queue:
2302         return cfqq;
2303 }
2304
2305 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2306 {
2307         int dispatched = 0;
2308
2309         while (cfqq->next_rq) {
2310                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2311                 dispatched++;
2312         }
2313
2314         BUG_ON(!list_empty(&cfqq->fifo));
2315
2316         /* By default cfqq is not expired if it is empty. Do it explicitly */
2317         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2318         return dispatched;
2319 }
2320
2321 /*
2322  * Drain our current requests. Used for barriers and when switching
2323  * io schedulers on-the-fly.
2324  */
2325 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2326 {
2327         struct cfq_queue *cfqq;
2328         int dispatched = 0;
2329
2330         /* Expire the timeslice of the current active queue first */
2331         cfq_slice_expired(cfqd, 0);
2332         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2333                 __cfq_set_active_queue(cfqd, cfqq);
2334                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2335         }
2336
2337         BUG_ON(cfqd->busy_queues);
2338
2339         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2340         return dispatched;
2341 }
2342
2343 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2344         struct cfq_queue *cfqq)
2345 {
2346         /* the queue hasn't finished any request, can't estimate */
2347         if (cfq_cfqq_slice_new(cfqq))
2348                 return true;
2349         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2350                 cfqq->slice_end))
2351                 return true;
2352
2353         return false;
2354 }
2355
2356 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2357 {
2358         unsigned int max_dispatch;
2359
2360         /*
2361          * Drain async requests before we start sync IO
2362          */
2363         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2364                 return false;
2365
2366         /*
2367          * If this is an async queue and we have sync IO in flight, let it wait
2368          */
2369         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2370                 return false;
2371
2372         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2373         if (cfq_class_idle(cfqq))
2374                 max_dispatch = 1;
2375
2376         /*
2377          * Does this cfqq already have too much IO in flight?
2378          */
2379         if (cfqq->dispatched >= max_dispatch) {
2380                 /*
2381                  * idle queue must always only have a single IO in flight
2382                  */
2383                 if (cfq_class_idle(cfqq))
2384                         return false;
2385
2386                 /*
2387                  * We have other queues, don't allow more IO from this one
2388                  */
2389                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2390                         return false;
2391
2392                 /*
2393                  * Sole queue user, no limit
2394                  */
2395                 if (cfqd->busy_queues == 1)
2396                         max_dispatch = -1;
2397                 else
2398                         /*
2399                          * Normally we start throttling cfqq when cfq_quantum/2
2400                          * requests have been dispatched. But we can drive
2401                          * deeper queue depths at the beginning of slice
2402                          * subjected to upper limit of cfq_quantum.
2403                          * */
2404                         max_dispatch = cfqd->cfq_quantum;
2405         }
2406
2407         /*
2408          * Async queues must wait a bit before being allowed dispatch.
2409          * We also ramp up the dispatch depth gradually for async IO,
2410          * based on the last sync IO we serviced
2411          */
2412         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2413                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2414                 unsigned int depth;
2415
2416                 depth = last_sync / cfqd->cfq_slice[1];
2417                 if (!depth && !cfqq->dispatched)
2418                         depth = 1;
2419                 if (depth < max_dispatch)
2420                         max_dispatch = depth;
2421         }
2422
2423         /*
2424          * If we're below the current max, allow a dispatch
2425          */
2426         return cfqq->dispatched < max_dispatch;
2427 }
2428
2429 /*
2430  * Dispatch a request from cfqq, moving them to the request queue
2431  * dispatch list.
2432  */
2433 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2434 {
2435         struct request *rq;
2436
2437         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2438
2439         if (!cfq_may_dispatch(cfqd, cfqq))
2440                 return false;
2441
2442         /*
2443          * follow expired path, else get first next available
2444          */
2445         rq = cfq_check_fifo(cfqq);
2446         if (!rq)
2447                 rq = cfqq->next_rq;
2448
2449         /*
2450          * insert request into driver dispatch list
2451          */
2452         cfq_dispatch_insert(cfqd->queue, rq);
2453
2454         if (!cfqd->active_cic) {
2455                 struct cfq_io_context *cic = RQ_CIC(rq);
2456
2457                 atomic_long_inc(&cic->ioc->refcount);
2458                 cfqd->active_cic = cic;
2459         }
2460
2461         return true;
2462 }
2463
2464 /*
2465  * Find the cfqq that we need to service and move a request from that to the
2466  * dispatch list
2467  */
2468 static int cfq_dispatch_requests(struct request_queue *q, int force)
2469 {
2470         struct cfq_data *cfqd = q->elevator->elevator_data;
2471         struct cfq_queue *cfqq;
2472
2473         if (!cfqd->busy_queues)
2474                 return 0;
2475
2476         if (unlikely(force))
2477                 return cfq_forced_dispatch(cfqd);
2478
2479         cfqq = cfq_select_queue(cfqd);
2480         if (!cfqq)
2481                 return 0;
2482
2483         /*
2484          * Dispatch a request from this cfqq, if it is allowed
2485          */
2486         if (!cfq_dispatch_request(cfqd, cfqq))
2487                 return 0;
2488
2489         cfqq->slice_dispatch++;
2490         cfq_clear_cfqq_must_dispatch(cfqq);
2491
2492         /*
2493          * expire an async queue immediately if it has used up its slice. idle
2494          * queue always expire after 1 dispatch round.
2495          */
2496         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2497             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2498             cfq_class_idle(cfqq))) {
2499                 cfqq->slice_end = jiffies + 1;
2500                 cfq_slice_expired(cfqd, 0);
2501         }
2502
2503         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2504         return 1;
2505 }
2506
2507 /*
2508  * task holds one reference to the queue, dropped when task exits. each rq
2509  * in-flight on this queue also holds a reference, dropped when rq is freed.
2510  *
2511  * Each cfq queue took a reference on the parent group. Drop it now.
2512  * queue lock must be held here.
2513  */
2514 static void cfq_put_queue(struct cfq_queue *cfqq)
2515 {
2516         struct cfq_data *cfqd = cfqq->cfqd;
2517         struct cfq_group *cfqg;
2518
2519         BUG_ON(cfqq->ref <= 0);
2520
2521         cfqq->ref--;
2522         if (cfqq->ref)
2523                 return;
2524
2525         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2526         BUG_ON(rb_first(&cfqq->sort_list));
2527         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2528         cfqg = cfqq->cfqg;
2529
2530         if (unlikely(cfqd->active_queue == cfqq)) {
2531                 __cfq_slice_expired(cfqd, cfqq, 0);
2532                 cfq_schedule_dispatch(cfqd);
2533         }
2534
2535         BUG_ON(cfq_cfqq_on_rr(cfqq));
2536         kmem_cache_free(cfq_pool, cfqq);
2537         cfq_put_cfqg(cfqg);
2538 }
2539
2540 /*
2541  * Must always be called with the rcu_read_lock() held
2542  */
2543 static void
2544 __call_for_each_cic(struct io_context *ioc,
2545                     void (*func)(struct io_context *, struct cfq_io_context *))
2546 {
2547         struct cfq_io_context *cic;
2548         struct hlist_node *n;
2549
2550         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2551                 func(ioc, cic);
2552 }
2553
2554 /*
2555  * Call func for each cic attached to this ioc.
2556  */
2557 static void
2558 call_for_each_cic(struct io_context *ioc,
2559                   void (*func)(struct io_context *, struct cfq_io_context *))
2560 {
2561         rcu_read_lock();
2562         __call_for_each_cic(ioc, func);
2563         rcu_read_unlock();
2564 }
2565
2566 static void cfq_cic_free_rcu(struct rcu_head *head)
2567 {
2568         struct cfq_io_context *cic;
2569
2570         cic = container_of(head, struct cfq_io_context, rcu_head);
2571
2572         kmem_cache_free(cfq_ioc_pool, cic);
2573         elv_ioc_count_dec(cfq_ioc_count);
2574
2575         if (ioc_gone) {
2576                 /*
2577                  * CFQ scheduler is exiting, grab exit lock and check
2578                  * the pending io context count. If it hits zero,
2579                  * complete ioc_gone and set it back to NULL
2580                  */
2581                 spin_lock(&ioc_gone_lock);
2582                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2583                         complete(ioc_gone);
2584                         ioc_gone = NULL;
2585                 }
2586                 spin_unlock(&ioc_gone_lock);
2587         }
2588 }
2589
2590 static void cfq_cic_free(struct cfq_io_context *cic)
2591 {
2592         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2593 }
2594
2595 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2596 {
2597         unsigned long flags;
2598         unsigned long dead_key = (unsigned long) cic->key;
2599
2600         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2601
2602         spin_lock_irqsave(&ioc->lock, flags);
2603         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2604         hlist_del_rcu(&cic->cic_list);
2605         spin_unlock_irqrestore(&ioc->lock, flags);
2606
2607         cfq_cic_free(cic);
2608 }
2609
2610 /*
2611  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2612  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2613  * and ->trim() which is called with the task lock held
2614  */
2615 static void cfq_free_io_context(struct io_context *ioc)
2616 {
2617         /*
2618          * ioc->refcount is zero here, or we are called from elv_unregister(),
2619          * so no more cic's are allowed to be linked into this ioc.  So it
2620          * should be ok to iterate over the known list, we will see all cic's
2621          * since no new ones are added.
2622          */
2623         __call_for_each_cic(ioc, cic_free_func);
2624 }
2625
2626 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2627 {
2628         struct cfq_queue *__cfqq, *next;
2629
2630         /*
2631          * If this queue was scheduled to merge with another queue, be
2632          * sure to drop the reference taken on that queue (and others in
2633          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2634          */
2635         __cfqq = cfqq->new_cfqq;
2636         while (__cfqq) {
2637                 if (__cfqq == cfqq) {
2638                         WARN(1, "cfqq->new_cfqq loop detected\n");
2639                         break;
2640                 }
2641                 next = __cfqq->new_cfqq;
2642                 cfq_put_queue(__cfqq);
2643                 __cfqq = next;
2644         }
2645 }
2646
2647 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2648 {
2649         if (unlikely(cfqq == cfqd->active_queue)) {
2650                 __cfq_slice_expired(cfqd, cfqq, 0);
2651                 cfq_schedule_dispatch(cfqd);
2652         }
2653
2654         cfq_put_cooperator(cfqq);
2655
2656         cfq_put_queue(cfqq);
2657 }
2658
2659 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2660                                          struct cfq_io_context *cic)
2661 {
2662         struct io_context *ioc = cic->ioc;
2663
2664         list_del_init(&cic->queue_list);
2665
2666         /*
2667          * Make sure dead mark is seen for dead queues
2668          */
2669         smp_wmb();
2670         cic->key = cfqd_dead_key(cfqd);
2671
2672         if (ioc->ioc_data == cic)
2673                 rcu_assign_pointer(ioc->ioc_data, NULL);
2674
2675         if (cic->cfqq[BLK_RW_ASYNC]) {
2676                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2677                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2678         }
2679
2680         if (cic->cfqq[BLK_RW_SYNC]) {
2681                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2682                 cic->cfqq[BLK_RW_SYNC] = NULL;
2683         }
2684 }
2685
2686 static void cfq_exit_single_io_context(struct io_context *ioc,
2687                                        struct cfq_io_context *cic)
2688 {
2689         struct cfq_data *cfqd = cic_to_cfqd(cic);
2690
2691         if (cfqd) {
2692                 struct request_queue *q = cfqd->queue;
2693                 unsigned long flags;
2694
2695                 spin_lock_irqsave(q->queue_lock, flags);
2696
2697                 /*
2698                  * Ensure we get a fresh copy of the ->key to prevent
2699                  * race between exiting task and queue
2700                  */
2701                 smp_read_barrier_depends();
2702                 if (cic->key == cfqd)
2703                         __cfq_exit_single_io_context(cfqd, cic);
2704
2705                 spin_unlock_irqrestore(q->queue_lock, flags);
2706         }
2707 }
2708
2709 /*
2710  * The process that ioc belongs to has exited, we need to clean up
2711  * and put the internal structures we have that belongs to that process.
2712  */
2713 static void cfq_exit_io_context(struct io_context *ioc)
2714 {
2715         call_for_each_cic(ioc, cfq_exit_single_io_context);
2716 }
2717
2718 static struct cfq_io_context *
2719 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2720 {
2721         struct cfq_io_context *cic;
2722
2723         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2724                                                         cfqd->queue->node);
2725         if (cic) {
2726                 cic->last_end_request = jiffies;
2727                 INIT_LIST_HEAD(&cic->queue_list);
2728                 INIT_HLIST_NODE(&cic->cic_list);
2729                 cic->dtor = cfq_free_io_context;
2730                 cic->exit = cfq_exit_io_context;
2731                 elv_ioc_count_inc(cfq_ioc_count);
2732         }
2733
2734         return cic;
2735 }
2736
2737 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2738 {
2739         struct task_struct *tsk = current;
2740         int ioprio_class;
2741
2742         if (!cfq_cfqq_prio_changed(cfqq))
2743                 return;
2744
2745         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2746         switch (ioprio_class) {
2747         default:
2748                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2749         case IOPRIO_CLASS_NONE:
2750                 /*
2751                  * no prio set, inherit CPU scheduling settings
2752                  */
2753                 cfqq->ioprio = task_nice_ioprio(tsk);
2754                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2755                 break;
2756         case IOPRIO_CLASS_RT:
2757                 cfqq->ioprio = task_ioprio(ioc);
2758                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2759                 break;
2760         case IOPRIO_CLASS_BE:
2761                 cfqq->ioprio = task_ioprio(ioc);
2762                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2763                 break;
2764         case IOPRIO_CLASS_IDLE:
2765                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2766                 cfqq->ioprio = 7;
2767                 cfq_clear_cfqq_idle_window(cfqq);
2768                 break;
2769         }
2770
2771         /*
2772          * keep track of original prio settings in case we have to temporarily
2773          * elevate the priority of this queue
2774          */
2775         cfqq->org_ioprio = cfqq->ioprio;
2776         cfqq->org_ioprio_class = cfqq->ioprio_class;
2777         cfq_clear_cfqq_prio_changed(cfqq);
2778 }
2779
2780 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2781 {
2782         struct cfq_data *cfqd = cic_to_cfqd(cic);
2783         struct cfq_queue *cfqq;
2784         unsigned long flags;
2785
2786         if (unlikely(!cfqd))
2787                 return;
2788
2789         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2790
2791         cfqq = cic->cfqq[BLK_RW_ASYNC];
2792         if (cfqq) {
2793                 struct cfq_queue *new_cfqq;
2794                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2795                                                 GFP_ATOMIC);
2796                 if (new_cfqq) {
2797                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2798                         cfq_put_queue(cfqq);
2799                 }
2800         }
2801
2802         cfqq = cic->cfqq[BLK_RW_SYNC];
2803         if (cfqq)
2804                 cfq_mark_cfqq_prio_changed(cfqq);
2805
2806         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2807 }
2808
2809 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2810 {
2811         call_for_each_cic(ioc, changed_ioprio);
2812         ioc->ioprio_changed = 0;
2813 }
2814
2815 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2816                           pid_t pid, bool is_sync)
2817 {
2818         RB_CLEAR_NODE(&cfqq->rb_node);
2819         RB_CLEAR_NODE(&cfqq->p_node);
2820         INIT_LIST_HEAD(&cfqq->fifo);
2821
2822         cfqq->ref = 0;
2823         cfqq->cfqd = cfqd;
2824
2825         cfq_mark_cfqq_prio_changed(cfqq);
2826
2827         if (is_sync) {
2828                 if (!cfq_class_idle(cfqq))
2829                         cfq_mark_cfqq_idle_window(cfqq);
2830                 cfq_mark_cfqq_sync(cfqq);
2831         }
2832         cfqq->pid = pid;
2833 }
2834
2835 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2836 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2837 {
2838         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2839         struct cfq_data *cfqd = cic_to_cfqd(cic);
2840         unsigned long flags;
2841         struct request_queue *q;
2842
2843         if (unlikely(!cfqd))
2844                 return;
2845
2846         q = cfqd->queue;
2847
2848         spin_lock_irqsave(q->queue_lock, flags);
2849
2850         if (sync_cfqq) {
2851                 /*
2852                  * Drop reference to sync queue. A new sync queue will be
2853                  * assigned in new group upon arrival of a fresh request.
2854                  */
2855                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2856                 cic_set_cfqq(cic, NULL, 1);
2857                 cfq_put_queue(sync_cfqq);
2858         }
2859
2860         spin_unlock_irqrestore(q->queue_lock, flags);
2861 }
2862
2863 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2864 {
2865         call_for_each_cic(ioc, changed_cgroup);
2866         ioc->cgroup_changed = 0;
2867 }
2868 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2869
2870 static struct cfq_queue *
2871 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2872                      struct io_context *ioc, gfp_t gfp_mask)
2873 {
2874         struct cfq_queue *cfqq, *new_cfqq = NULL;
2875         struct cfq_io_context *cic;
2876         struct cfq_group *cfqg;
2877
2878 retry:
2879         cfqg = cfq_get_cfqg(cfqd, 1);
2880         cic = cfq_cic_lookup(cfqd, ioc);
2881         /* cic always exists here */
2882         cfqq = cic_to_cfqq(cic, is_sync);
2883
2884         /*
2885          * Always try a new alloc if we fell back to the OOM cfqq
2886          * originally, since it should just be a temporary situation.
2887          */
2888         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2889                 cfqq = NULL;
2890                 if (new_cfqq) {
2891                         cfqq = new_cfqq;
2892                         new_cfqq = NULL;
2893                 } else if (gfp_mask & __GFP_WAIT) {
2894                         spin_unlock_irq(cfqd->queue->queue_lock);
2895                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2896                                         gfp_mask | __GFP_ZERO,
2897                                         cfqd->queue->node);
2898                         spin_lock_irq(cfqd->queue->queue_lock);
2899                         if (new_cfqq)
2900                                 goto retry;
2901                 } else {
2902                         cfqq = kmem_cache_alloc_node(cfq_pool,
2903                                         gfp_mask | __GFP_ZERO,
2904                                         cfqd->queue->node);
2905                 }
2906
2907                 if (cfqq) {
2908                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2909                         cfq_init_prio_data(cfqq, ioc);
2910                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2911                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2912                 } else
2913                         cfqq = &cfqd->oom_cfqq;
2914         }
2915
2916         if (new_cfqq)
2917                 kmem_cache_free(cfq_pool, new_cfqq);
2918
2919         return cfqq;
2920 }
2921
2922 static struct cfq_queue **
2923 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2924 {
2925         switch (ioprio_class) {
2926         case IOPRIO_CLASS_RT:
2927                 return &cfqd->async_cfqq[0][ioprio];
2928         case IOPRIO_CLASS_BE:
2929                 return &cfqd->async_cfqq[1][ioprio];
2930         case IOPRIO_CLASS_IDLE:
2931                 return &cfqd->async_idle_cfqq;
2932         default:
2933                 BUG();
2934         }
2935 }
2936
2937 static struct cfq_queue *
2938 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2939               gfp_t gfp_mask)
2940 {
2941         const int ioprio = task_ioprio(ioc);
2942         const int ioprio_class = task_ioprio_class(ioc);
2943         struct cfq_queue **async_cfqq = NULL;
2944         struct cfq_queue *cfqq = NULL;
2945
2946         if (!is_sync) {
2947                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2948                 cfqq = *async_cfqq;
2949         }
2950
2951         if (!cfqq)
2952                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2953
2954         /*
2955          * pin the queue now that it's allocated, scheduler exit will prune it
2956          */
2957         if (!is_sync && !(*async_cfqq)) {
2958                 cfqq->ref++;
2959                 *async_cfqq = cfqq;
2960         }
2961
2962         cfqq->ref++;
2963         return cfqq;
2964 }
2965
2966 /*
2967  * We drop cfq io contexts lazily, so we may find a dead one.
2968  */
2969 static void
2970 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2971                   struct cfq_io_context *cic)
2972 {
2973         unsigned long flags;
2974
2975         WARN_ON(!list_empty(&cic->queue_list));
2976         BUG_ON(cic->key != cfqd_dead_key(cfqd));
2977
2978         spin_lock_irqsave(&ioc->lock, flags);
2979
2980         BUG_ON(ioc->ioc_data == cic);
2981
2982         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2983         hlist_del_rcu(&cic->cic_list);
2984         spin_unlock_irqrestore(&ioc->lock, flags);
2985
2986         cfq_cic_free(cic);
2987 }
2988
2989 static struct cfq_io_context *
2990 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2991 {
2992         struct cfq_io_context *cic;
2993         unsigned long flags;
2994
2995         if (unlikely(!ioc))
2996                 return NULL;
2997
2998         rcu_read_lock();
2999
3000         /*
3001          * we maintain a last-hit cache, to avoid browsing over the tree
3002          */
3003         cic = rcu_dereference(ioc->ioc_data);
3004         if (cic && cic->key == cfqd) {
3005                 rcu_read_unlock();
3006                 return cic;
3007         }
3008
3009         do {
3010                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3011                 rcu_read_unlock();
3012                 if (!cic)
3013                         break;
3014                 if (unlikely(cic->key != cfqd)) {
3015                         cfq_drop_dead_cic(cfqd, ioc, cic);
3016                         rcu_read_lock();
3017                         continue;
3018                 }
3019
3020                 spin_lock_irqsave(&ioc->lock, flags);
3021                 rcu_assign_pointer(ioc->ioc_data, cic);
3022                 spin_unlock_irqrestore(&ioc->lock, flags);
3023                 break;
3024         } while (1);
3025
3026         return cic;
3027 }
3028
3029 /*
3030  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3031  * the process specific cfq io context when entered from the block layer.
3032  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3033  */
3034 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3035                         struct cfq_io_context *cic, gfp_t gfp_mask)
3036 {
3037         unsigned long flags;
3038         int ret;
3039
3040         ret = radix_tree_preload(gfp_mask);
3041         if (!ret) {
3042                 cic->ioc = ioc;
3043                 cic->key = cfqd;
3044
3045                 spin_lock_irqsave(&ioc->lock, flags);
3046                 ret = radix_tree_insert(&ioc->radix_root,
3047                                                 cfqd->cic_index, cic);
3048                 if (!ret)
3049                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3050                 spin_unlock_irqrestore(&ioc->lock, flags);
3051
3052                 radix_tree_preload_end();
3053
3054                 if (!ret) {
3055                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3056                         list_add(&cic->queue_list, &cfqd->cic_list);
3057                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3058                 }
3059         }
3060
3061         if (ret)
3062                 printk(KERN_ERR "cfq: cic link failed!\n");
3063
3064         return ret;
3065 }
3066
3067 /*
3068  * Setup general io context and cfq io context. There can be several cfq
3069  * io contexts per general io context, if this process is doing io to more
3070  * than one device managed by cfq.
3071  */
3072 static struct cfq_io_context *
3073 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3074 {
3075         struct io_context *ioc = NULL;
3076         struct cfq_io_context *cic;
3077
3078         might_sleep_if(gfp_mask & __GFP_WAIT);
3079
3080         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3081         if (!ioc)
3082                 return NULL;
3083
3084         cic = cfq_cic_lookup(cfqd, ioc);
3085         if (cic)
3086                 goto out;
3087
3088         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3089         if (cic == NULL)
3090                 goto err;
3091
3092         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3093                 goto err_free;
3094
3095 out:
3096         smp_read_barrier_depends();
3097         if (unlikely(ioc->ioprio_changed))
3098                 cfq_ioc_set_ioprio(ioc);
3099
3100 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3101         if (unlikely(ioc->cgroup_changed))
3102                 cfq_ioc_set_cgroup(ioc);
3103 #endif
3104         return cic;
3105 err_free:
3106         cfq_cic_free(cic);
3107 err:
3108         put_io_context(ioc);
3109         return NULL;
3110 }
3111
3112 static void
3113 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3114 {
3115         unsigned long elapsed = jiffies - cic->last_end_request;
3116         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3117
3118         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3119         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3120         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3121 }
3122
3123 static void
3124 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3125                        struct request *rq)
3126 {
3127         sector_t sdist = 0;
3128         sector_t n_sec = blk_rq_sectors(rq);
3129         if (cfqq->last_request_pos) {
3130                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3131                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3132                 else
3133                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3134         }
3135
3136         cfqq->seek_history <<= 1;
3137         if (blk_queue_nonrot(cfqd->queue))
3138                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3139         else
3140                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3141 }
3142
3143 /*
3144  * Disable idle window if the process thinks too long or seeks so much that
3145  * it doesn't matter
3146  */
3147 static void
3148 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3149                        struct cfq_io_context *cic)
3150 {
3151         int old_idle, enable_idle;
3152
3153         /*
3154          * Don't idle for async or idle io prio class
3155          */
3156         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3157                 return;
3158
3159         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3160
3161         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3162                 cfq_mark_cfqq_deep(cfqq);
3163
3164         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3165                 enable_idle = 0;
3166         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3167             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3168                 enable_idle = 0;
3169         else if (sample_valid(cic->ttime_samples)) {
3170                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3171                         enable_idle = 0;
3172                 else
3173                         enable_idle = 1;
3174         }
3175
3176         if (old_idle != enable_idle) {
3177                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3178                 if (enable_idle)
3179                         cfq_mark_cfqq_idle_window(cfqq);
3180                 else
3181                         cfq_clear_cfqq_idle_window(cfqq);
3182         }
3183 }
3184
3185 /*
3186  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3187  * no or if we aren't sure, a 1 will cause a preempt.
3188  */
3189 static bool
3190 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3191                    struct request *rq)
3192 {
3193         struct cfq_queue *cfqq;
3194
3195         cfqq = cfqd->active_queue;
3196         if (!cfqq)
3197                 return false;
3198
3199         if (cfq_class_idle(new_cfqq))
3200                 return false;
3201
3202         if (cfq_class_idle(cfqq))
3203                 return true;
3204
3205         /*
3206          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3207          */
3208         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3209                 return false;
3210
3211         /*
3212          * if the new request is sync, but the currently running queue is
3213          * not, let the sync request have priority.
3214          */
3215         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3216                 return true;
3217
3218         if (new_cfqq->cfqg != cfqq->cfqg)
3219                 return false;
3220
3221         if (cfq_slice_used(cfqq))
3222                 return true;
3223
3224         /* Allow preemption only if we are idling on sync-noidle tree */
3225         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3226             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3227             new_cfqq->service_tree->count == 2 &&
3228             RB_EMPTY_ROOT(&cfqq->sort_list))
3229                 return true;
3230
3231         /*
3232          * So both queues are sync. Let the new request get disk time if
3233          * it's a metadata request and the current queue is doing regular IO.
3234          */
3235         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3236                 return true;
3237
3238         /*
3239          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3240          */
3241         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3242                 return true;
3243
3244         /* An idle queue should not be idle now for some reason */
3245         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3246                 return true;
3247
3248         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3249                 return false;
3250
3251         /*
3252          * if this request is as-good as one we would expect from the
3253          * current cfqq, let it preempt
3254          */
3255         if (cfq_rq_close(cfqd, cfqq, rq))
3256                 return true;
3257
3258         return false;
3259 }
3260
3261 /*
3262  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3263  * let it have half of its nominal slice.
3264  */
3265 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3266 {
3267         struct cfq_queue *old_cfqq = cfqd->active_queue;
3268
3269         cfq_log_cfqq(cfqd, cfqq, "preempt");
3270         cfq_slice_expired(cfqd, 1);
3271
3272         /*
3273          * workload type is changed, don't save slice, otherwise preempt
3274          * doesn't happen
3275          */
3276         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3277                 cfqq->cfqg->saved_workload_slice = 0;
3278
3279         /*
3280          * Put the new queue at the front of the of the current list,
3281          * so we know that it will be selected next.
3282          */
3283         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3284
3285         cfq_service_tree_add(cfqd, cfqq, 1);
3286
3287         cfqq->slice_end = 0;
3288         cfq_mark_cfqq_slice_new(cfqq);
3289 }
3290
3291 /*
3292  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3293  * something we should do about it
3294  */
3295 static void
3296 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3297                 struct request *rq)
3298 {
3299         struct cfq_io_context *cic = RQ_CIC(rq);
3300
3301         cfqd->rq_queued++;
3302         if (rq->cmd_flags & REQ_META)
3303                 cfqq->meta_pending++;
3304
3305         cfq_update_io_thinktime(cfqd, cic);
3306         cfq_update_io_seektime(cfqd, cfqq, rq);
3307         cfq_update_idle_window(cfqd, cfqq, cic);
3308
3309         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3310
3311         if (cfqq == cfqd->active_queue) {
3312                 /*
3313                  * Remember that we saw a request from this process, but
3314                  * don't start queuing just yet. Otherwise we risk seeing lots
3315                  * of tiny requests, because we disrupt the normal plugging
3316                  * and merging. If the request is already larger than a single
3317                  * page, let it rip immediately. For that case we assume that
3318                  * merging is already done. Ditto for a busy system that
3319                  * has other work pending, don't risk delaying until the
3320                  * idle timer unplug to continue working.
3321                  */
3322                 if (cfq_cfqq_wait_request(cfqq)) {
3323                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3324                             cfqd->busy_queues > 1) {
3325                                 cfq_del_timer(cfqd, cfqq);
3326                                 cfq_clear_cfqq_wait_request(cfqq);
3327                                 __blk_run_queue(cfqd->queue);
3328                         } else {
3329                                 cfq_blkiocg_update_idle_time_stats(
3330                                                 &cfqq->cfqg->blkg);
3331                                 cfq_mark_cfqq_must_dispatch(cfqq);
3332                         }
3333                 }
3334         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3335                 /*
3336                  * not the active queue - expire current slice if it is
3337                  * idle and has expired it's mean thinktime or this new queue
3338                  * has some old slice time left and is of higher priority or
3339                  * this new queue is RT and the current one is BE
3340                  */
3341                 cfq_preempt_queue(cfqd, cfqq);
3342                 __blk_run_queue(cfqd->queue);
3343         }
3344 }
3345
3346 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3347 {
3348         struct cfq_data *cfqd = q->elevator->elevator_data;
3349         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3350
3351         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3352         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3353
3354         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3355         list_add_tail(&rq->queuelist, &cfqq->fifo);
3356         cfq_add_rq_rb(rq);
3357         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3358                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3359                         rq_is_sync(rq));
3360         cfq_rq_enqueued(cfqd, cfqq, rq);
3361 }
3362
3363 /*
3364  * Update hw_tag based on peak queue depth over 50 samples under
3365  * sufficient load.
3366  */
3367 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3368 {
3369         struct cfq_queue *cfqq = cfqd->active_queue;
3370
3371         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3372                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3373
3374         if (cfqd->hw_tag == 1)
3375                 return;
3376
3377         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3378             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3379                 return;
3380
3381         /*
3382          * If active queue hasn't enough requests and can idle, cfq might not
3383          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3384          * case
3385          */
3386         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3387             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3388             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3389                 return;
3390
3391         if (cfqd->hw_tag_samples++ < 50)
3392                 return;
3393
3394         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3395                 cfqd->hw_tag = 1;
3396         else
3397                 cfqd->hw_tag = 0;
3398 }
3399
3400 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3401 {
3402         struct cfq_io_context *cic = cfqd->active_cic;
3403
3404         /* If the queue already has requests, don't wait */
3405         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3406                 return false;
3407
3408         /* If there are other queues in the group, don't wait */
3409         if (cfqq->cfqg->nr_cfqq > 1)
3410                 return false;
3411
3412         if (cfq_slice_used(cfqq))
3413                 return true;
3414
3415         /* if slice left is less than think time, wait busy */
3416         if (cic && sample_valid(cic->ttime_samples)
3417             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3418                 return true;
3419
3420         /*
3421          * If think times is less than a jiffy than ttime_mean=0 and above
3422          * will not be true. It might happen that slice has not expired yet
3423          * but will expire soon (4-5 ns) during select_queue(). To cover the
3424          * case where think time is less than a jiffy, mark the queue wait
3425          * busy if only 1 jiffy is left in the slice.
3426          */
3427         if (cfqq->slice_end - jiffies == 1)
3428                 return true;
3429
3430         return false;
3431 }
3432
3433 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3434 {
3435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3436         struct cfq_data *cfqd = cfqq->cfqd;
3437         const int sync = rq_is_sync(rq);
3438         unsigned long now;
3439
3440         now = jiffies;
3441         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3442                      !!(rq->cmd_flags & REQ_NOIDLE));
3443
3444         cfq_update_hw_tag(cfqd);
3445
3446         WARN_ON(!cfqd->rq_in_driver);
3447         WARN_ON(!cfqq->dispatched);
3448         cfqd->rq_in_driver--;
3449         cfqq->dispatched--;
3450         (RQ_CFQG(rq))->dispatched--;
3451         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3452                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3453                         rq_data_dir(rq), rq_is_sync(rq));
3454
3455         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3456
3457         if (sync) {
3458                 RQ_CIC(rq)->last_end_request = now;
3459                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3460                         cfqd->last_delayed_sync = now;
3461         }
3462
3463         /*
3464          * If this is the active queue, check if it needs to be expired,
3465          * or if we want to idle in case it has no pending requests.
3466          */
3467         if (cfqd->active_queue == cfqq) {
3468                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3469
3470                 if (cfq_cfqq_slice_new(cfqq)) {
3471                         cfq_set_prio_slice(cfqd, cfqq);
3472                         cfq_clear_cfqq_slice_new(cfqq);
3473                 }
3474
3475                 /*
3476                  * Should we wait for next request to come in before we expire
3477                  * the queue.
3478                  */
3479                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3480                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3481                         if (!cfqd->cfq_slice_idle)
3482                                 extend_sl = cfqd->cfq_group_idle;
3483                         cfqq->slice_end = jiffies + extend_sl;
3484                         cfq_mark_cfqq_wait_busy(cfqq);
3485                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3486                 }
3487
3488                 /*
3489                  * Idling is not enabled on:
3490                  * - expired queues
3491                  * - idle-priority queues
3492                  * - async queues
3493                  * - queues with still some requests queued
3494                  * - when there is a close cooperator
3495                  */
3496                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3497                         cfq_slice_expired(cfqd, 1);
3498                 else if (sync && cfqq_empty &&
3499                          !cfq_close_cooperator(cfqd, cfqq)) {
3500                         cfq_arm_slice_timer(cfqd);
3501                 }
3502         }
3503
3504         if (!cfqd->rq_in_driver)
3505                 cfq_schedule_dispatch(cfqd);
3506 }
3507
3508 /*
3509  * we temporarily boost lower priority queues if they are holding fs exclusive
3510  * resources. they are boosted to normal prio (CLASS_BE/4)
3511  */
3512 static void cfq_prio_boost(struct cfq_queue *cfqq)
3513 {
3514         if (has_fs_excl()) {
3515                 /*
3516                  * boost idle prio on transactions that would lock out other
3517                  * users of the filesystem
3518                  */
3519                 if (cfq_class_idle(cfqq))
3520                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3521                 if (cfqq->ioprio > IOPRIO_NORM)
3522                         cfqq->ioprio = IOPRIO_NORM;
3523         } else {
3524                 /*
3525                  * unboost the queue (if needed)
3526                  */
3527                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3528                 cfqq->ioprio = cfqq->org_ioprio;
3529         }
3530 }
3531
3532 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3533 {
3534         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3535                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3536                 return ELV_MQUEUE_MUST;
3537         }
3538
3539         return ELV_MQUEUE_MAY;
3540 }
3541
3542 static int cfq_may_queue(struct request_queue *q, int rw)
3543 {
3544         struct cfq_data *cfqd = q->elevator->elevator_data;
3545         struct task_struct *tsk = current;
3546         struct cfq_io_context *cic;
3547         struct cfq_queue *cfqq;
3548
3549         /*
3550          * don't force setup of a queue from here, as a call to may_queue
3551          * does not necessarily imply that a request actually will be queued.
3552          * so just lookup a possibly existing queue, or return 'may queue'
3553          * if that fails
3554          */
3555         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3556         if (!cic)
3557                 return ELV_MQUEUE_MAY;
3558
3559         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3560         if (cfqq) {
3561                 cfq_init_prio_data(cfqq, cic->ioc);
3562                 cfq_prio_boost(cfqq);
3563
3564                 return __cfq_may_queue(cfqq);
3565         }
3566
3567         return ELV_MQUEUE_MAY;
3568 }
3569
3570 /*
3571  * queue lock held here
3572  */
3573 static void cfq_put_request(struct request *rq)
3574 {
3575         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3576
3577         if (cfqq) {
3578                 const int rw = rq_data_dir(rq);
3579
3580                 BUG_ON(!cfqq->allocated[rw]);
3581                 cfqq->allocated[rw]--;
3582
3583                 put_io_context(RQ_CIC(rq)->ioc);
3584
3585                 rq->elevator_private[0] = NULL;
3586                 rq->elevator_private[1] = NULL;
3587
3588                 /* Put down rq reference on cfqg */
3589                 cfq_put_cfqg(RQ_CFQG(rq));
3590                 rq->elevator_private[2] = NULL;
3591
3592                 cfq_put_queue(cfqq);
3593         }
3594 }
3595
3596 static struct cfq_queue *
3597 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3598                 struct cfq_queue *cfqq)
3599 {
3600         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3601         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3602         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3603         cfq_put_queue(cfqq);
3604         return cic_to_cfqq(cic, 1);
3605 }
3606
3607 /*
3608  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3609  * was the last process referring to said cfqq.
3610  */
3611 static struct cfq_queue *
3612 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3613 {
3614         if (cfqq_process_refs(cfqq) == 1) {
3615                 cfqq->pid = current->pid;
3616                 cfq_clear_cfqq_coop(cfqq);
3617                 cfq_clear_cfqq_split_coop(cfqq);
3618                 return cfqq;
3619         }
3620
3621         cic_set_cfqq(cic, NULL, 1);
3622
3623         cfq_put_cooperator(cfqq);
3624
3625         cfq_put_queue(cfqq);
3626         return NULL;
3627 }
3628 /*
3629  * Allocate cfq data structures associated with this request.
3630  */
3631 static int
3632 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3633 {
3634         struct cfq_data *cfqd = q->elevator->elevator_data;
3635         struct cfq_io_context *cic;
3636         const int rw = rq_data_dir(rq);
3637         const bool is_sync = rq_is_sync(rq);
3638         struct cfq_queue *cfqq;
3639         unsigned long flags;
3640
3641         might_sleep_if(gfp_mask & __GFP_WAIT);
3642
3643         cic = cfq_get_io_context(cfqd, gfp_mask);
3644
3645         spin_lock_irqsave(q->queue_lock, flags);
3646
3647         if (!cic)
3648                 goto queue_fail;
3649
3650 new_queue:
3651         cfqq = cic_to_cfqq(cic, is_sync);
3652         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3653                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3654                 cic_set_cfqq(cic, cfqq, is_sync);
3655         } else {
3656                 /*
3657                  * If the queue was seeky for too long, break it apart.
3658                  */
3659                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3660                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3661                         cfqq = split_cfqq(cic, cfqq);
3662                         if (!cfqq)
3663                                 goto new_queue;
3664                 }
3665
3666                 /*
3667                  * Check to see if this queue is scheduled to merge with
3668                  * another, closely cooperating queue.  The merging of
3669                  * queues happens here as it must be done in process context.
3670                  * The reference on new_cfqq was taken in merge_cfqqs.
3671                  */
3672                 if (cfqq->new_cfqq)
3673                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3674         }
3675
3676         cfqq->allocated[rw]++;
3677
3678         spin_unlock_irqrestore(q->queue_lock, flags);
3679
3680         cfqq->ref++;
3681         rq->elevator_private[0] = cic;
3682         rq->elevator_private[1] = cfqq;
3683         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3684         return 0;
3685
3686 queue_fail:
3687         if (cic)
3688                 put_io_context(cic->ioc);
3689
3690         cfq_schedule_dispatch(cfqd);
3691         spin_unlock_irqrestore(q->queue_lock, flags);
3692         cfq_log(cfqd, "set_request fail");
3693         return 1;
3694 }
3695
3696 static void cfq_kick_queue(struct work_struct *work)
3697 {
3698         struct cfq_data *cfqd =
3699                 container_of(work, struct cfq_data, unplug_work);
3700         struct request_queue *q = cfqd->queue;
3701
3702         spin_lock_irq(q->queue_lock);
3703         __blk_run_queue(cfqd->queue);
3704         spin_unlock_irq(q->queue_lock);
3705 }
3706
3707 /*
3708  * Timer running if the active_queue is currently idling inside its time slice
3709  */
3710 static void cfq_idle_slice_timer(unsigned long data)
3711 {
3712         struct cfq_data *cfqd = (struct cfq_data *) data;
3713         struct cfq_queue *cfqq;
3714         unsigned long flags;
3715         int timed_out = 1;
3716
3717         cfq_log(cfqd, "idle timer fired");
3718
3719         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3720
3721         cfqq = cfqd->active_queue;
3722         if (cfqq) {
3723                 timed_out = 0;
3724
3725                 /*
3726                  * We saw a request before the queue expired, let it through
3727                  */
3728                 if (cfq_cfqq_must_dispatch(cfqq))
3729                         goto out_kick;
3730
3731                 /*
3732                  * expired
3733                  */
3734                 if (cfq_slice_used(cfqq))
3735                         goto expire;
3736
3737                 /*
3738                  * only expire and reinvoke request handler, if there are
3739                  * other queues with pending requests
3740                  */
3741                 if (!cfqd->busy_queues)
3742                         goto out_cont;
3743
3744                 /*
3745                  * not expired and it has a request pending, let it dispatch
3746                  */
3747                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3748                         goto out_kick;
3749
3750                 /*
3751                  * Queue depth flag is reset only when the idle didn't succeed
3752                  */
3753                 cfq_clear_cfqq_deep(cfqq);
3754         }
3755 expire:
3756         cfq_slice_expired(cfqd, timed_out);
3757 out_kick:
3758         cfq_schedule_dispatch(cfqd);
3759 out_cont:
3760         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3761 }
3762
3763 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3764 {
3765         del_timer_sync(&cfqd->idle_slice_timer);
3766         cancel_work_sync(&cfqd->unplug_work);
3767 }
3768
3769 static void cfq_put_async_queues(struct cfq_data *cfqd)
3770 {
3771         int i;
3772
3773         for (i = 0; i < IOPRIO_BE_NR; i++) {
3774                 if (cfqd->async_cfqq[0][i])
3775                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3776                 if (cfqd->async_cfqq[1][i])
3777                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3778         }
3779
3780         if (cfqd->async_idle_cfqq)
3781                 cfq_put_queue(cfqd->async_idle_cfqq);
3782 }
3783
3784 static void cfq_cfqd_free(struct rcu_head *head)
3785 {
3786         kfree(container_of(head, struct cfq_data, rcu));
3787 }
3788
3789 static void cfq_exit_queue(struct elevator_queue *e)
3790 {
3791         struct cfq_data *cfqd = e->elevator_data;
3792         struct request_queue *q = cfqd->queue;
3793
3794         cfq_shutdown_timer_wq(cfqd);
3795
3796         spin_lock_irq(q->queue_lock);
3797
3798         if (cfqd->active_queue)
3799                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3800
3801         while (!list_empty(&cfqd->cic_list)) {
3802                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3803                                                         struct cfq_io_context,
3804                                                         queue_list);
3805
3806                 __cfq_exit_single_io_context(cfqd, cic);
3807         }
3808
3809         cfq_put_async_queues(cfqd);
3810         cfq_release_cfq_groups(cfqd);
3811         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3812
3813         spin_unlock_irq(q->queue_lock);
3814
3815         cfq_shutdown_timer_wq(cfqd);
3816
3817         spin_lock(&cic_index_lock);
3818         ida_remove(&cic_index_ida, cfqd->cic_index);
3819         spin_unlock(&cic_index_lock);
3820
3821         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3822         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3823 }
3824
3825 static int cfq_alloc_cic_index(void)
3826 {
3827         int index, error;
3828
3829         do {
3830                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3831                         return -ENOMEM;
3832
3833                 spin_lock(&cic_index_lock);
3834                 error = ida_get_new(&cic_index_ida, &index);
3835                 spin_unlock(&cic_index_lock);
3836                 if (error && error != -EAGAIN)
3837                         return error;
3838         } while (error);
3839
3840         return index;
3841 }
3842
3843 static void *cfq_init_queue(struct request_queue *q)
3844 {
3845         struct cfq_data *cfqd;
3846         int i, j;
3847         struct cfq_group *cfqg;
3848         struct cfq_rb_root *st;
3849
3850         i = cfq_alloc_cic_index();
3851         if (i < 0)
3852                 return NULL;
3853
3854         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3855         if (!cfqd)
3856                 return NULL;
3857
3858         /*
3859          * Don't need take queue_lock in the routine, since we are
3860          * initializing the ioscheduler, and nobody is using cfqd
3861          */
3862         cfqd->cic_index = i;
3863
3864         /* Init root service tree */
3865         cfqd->grp_service_tree = CFQ_RB_ROOT;
3866
3867         /* Init root group */
3868         cfqg = &cfqd->root_group;
3869         for_each_cfqg_st(cfqg, i, j, st)
3870                 *st = CFQ_RB_ROOT;
3871         RB_CLEAR_NODE(&cfqg->rb_node);
3872
3873         /* Give preference to root group over other groups */
3874         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3875
3876 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3877         /*
3878          * Take a reference to root group which we never drop. This is just
3879          * to make sure that cfq_put_cfqg() does not try to kfree root group
3880          */
3881         cfqg->ref = 1;
3882         rcu_read_lock();
3883         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3884                                         (void *)cfqd, 0);
3885         rcu_read_unlock();
3886 #endif
3887         /*
3888          * Not strictly needed (since RB_ROOT just clears the node and we
3889          * zeroed cfqd on alloc), but better be safe in case someone decides
3890          * to add magic to the rb code
3891          */
3892         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3893                 cfqd->prio_trees[i] = RB_ROOT;
3894
3895         /*
3896          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3897          * Grab a permanent reference to it, so that the normal code flow
3898          * will not attempt to free it.
3899          */
3900         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3901         cfqd->oom_cfqq.ref++;
3902         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3903
3904         INIT_LIST_HEAD(&cfqd->cic_list);
3905
3906         cfqd->queue = q;
3907
3908         init_timer(&cfqd->idle_slice_timer);
3909         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3910         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3911
3912         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3913
3914         cfqd->cfq_quantum = cfq_quantum;
3915         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3916         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3917         cfqd->cfq_back_max = cfq_back_max;
3918         cfqd->cfq_back_penalty = cfq_back_penalty;
3919         cfqd->cfq_slice[0] = cfq_slice_async;
3920         cfqd->cfq_slice[1] = cfq_slice_sync;
3921         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3922         cfqd->cfq_slice_idle = cfq_slice_idle;
3923         cfqd->cfq_group_idle = cfq_group_idle;
3924         cfqd->cfq_latency = 1;
3925         cfqd->hw_tag = -1;
3926         /*
3927          * we optimistically start assuming sync ops weren't delayed in last
3928          * second, in order to have larger depth for async operations.
3929          */
3930         cfqd->last_delayed_sync = jiffies - HZ;
3931         return cfqd;
3932 }
3933
3934 static void cfq_slab_kill(void)
3935 {
3936         /*
3937          * Caller already ensured that pending RCU callbacks are completed,
3938          * so we should have no busy allocations at this point.
3939          */
3940         if (cfq_pool)
3941                 kmem_cache_destroy(cfq_pool);
3942         if (cfq_ioc_pool)
3943                 kmem_cache_destroy(cfq_ioc_pool);
3944 }
3945
3946 static int __init cfq_slab_setup(void)
3947 {
3948         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3949         if (!cfq_pool)
3950                 goto fail;
3951
3952         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3953         if (!cfq_ioc_pool)
3954                 goto fail;
3955
3956         return 0;
3957 fail:
3958         cfq_slab_kill();
3959         return -ENOMEM;
3960 }
3961
3962 /*
3963  * sysfs parts below -->
3964  */
3965 static ssize_t
3966 cfq_var_show(unsigned int var, char *page)
3967 {
3968         return sprintf(page, "%d\n", var);
3969 }
3970
3971 static ssize_t
3972 cfq_var_store(unsigned int *var, const char *page, size_t count)
3973 {
3974         char *p = (char *) page;
3975
3976         *var = simple_strtoul(p, &p, 10);
3977         return count;
3978 }
3979
3980 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3981 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3982 {                                                                       \
3983         struct cfq_data *cfqd = e->elevator_data;                       \
3984         unsigned int __data = __VAR;                                    \
3985         if (__CONV)                                                     \
3986                 __data = jiffies_to_msecs(__data);                      \
3987         return cfq_var_show(__data, (page));                            \
3988 }
3989 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3990 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3991 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3992 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3993 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3994 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3995 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3996 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3997 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3998 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3999 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4000 #undef SHOW_FUNCTION
4001
4002 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4003 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4004 {                                                                       \
4005         struct cfq_data *cfqd = e->elevator_data;                       \
4006         unsigned int __data;                                            \
4007         int ret = cfq_var_store(&__data, (page), count);                \
4008         if (__data < (MIN))                                             \
4009                 __data = (MIN);                                         \
4010         else if (__data > (MAX))                                        \
4011                 __data = (MAX);                                         \
4012         if (__CONV)                                                     \
4013                 *(__PTR) = msecs_to_jiffies(__data);                    \
4014         else                                                            \
4015                 *(__PTR) = __data;                                      \
4016         return ret;                                                     \
4017 }
4018 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4019 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4020                 UINT_MAX, 1);
4021 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4022                 UINT_MAX, 1);
4023 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4024 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4025                 UINT_MAX, 0);
4026 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4027 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4028 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4029 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4030 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4031                 UINT_MAX, 0);
4032 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4033 #undef STORE_FUNCTION
4034
4035 #define CFQ_ATTR(name) \
4036         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4037
4038 static struct elv_fs_entry cfq_attrs[] = {
4039         CFQ_ATTR(quantum),
4040         CFQ_ATTR(fifo_expire_sync),
4041         CFQ_ATTR(fifo_expire_async),
4042         CFQ_ATTR(back_seek_max),
4043         CFQ_ATTR(back_seek_penalty),
4044         CFQ_ATTR(slice_sync),
4045         CFQ_ATTR(slice_async),
4046         CFQ_ATTR(slice_async_rq),
4047         CFQ_ATTR(slice_idle),
4048         CFQ_ATTR(group_idle),
4049         CFQ_ATTR(low_latency),
4050         __ATTR_NULL
4051 };
4052
4053 static struct elevator_type iosched_cfq = {
4054         .ops = {
4055                 .elevator_merge_fn =            cfq_merge,
4056                 .elevator_merged_fn =           cfq_merged_request,
4057                 .elevator_merge_req_fn =        cfq_merged_requests,
4058                 .elevator_allow_merge_fn =      cfq_allow_merge,
4059                 .elevator_bio_merged_fn =       cfq_bio_merged,
4060                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4061                 .elevator_add_req_fn =          cfq_insert_request,
4062                 .elevator_activate_req_fn =     cfq_activate_request,
4063                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4064                 .elevator_queue_empty_fn =      cfq_queue_empty,
4065                 .elevator_completed_req_fn =    cfq_completed_request,
4066                 .elevator_former_req_fn =       elv_rb_former_request,
4067                 .elevator_latter_req_fn =       elv_rb_latter_request,
4068                 .elevator_set_req_fn =          cfq_set_request,
4069                 .elevator_put_req_fn =          cfq_put_request,
4070                 .elevator_may_queue_fn =        cfq_may_queue,
4071                 .elevator_init_fn =             cfq_init_queue,
4072                 .elevator_exit_fn =             cfq_exit_queue,
4073                 .trim =                         cfq_free_io_context,
4074         },
4075         .elevator_attrs =       cfq_attrs,
4076         .elevator_name =        "cfq",
4077         .elevator_owner =       THIS_MODULE,
4078 };
4079
4080 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4081 static struct blkio_policy_type blkio_policy_cfq = {
4082         .ops = {
4083                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4084                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4085         },
4086         .plid = BLKIO_POLICY_PROP,
4087 };
4088 #else
4089 static struct blkio_policy_type blkio_policy_cfq;
4090 #endif
4091
4092 static int __init cfq_init(void)
4093 {
4094         /*
4095          * could be 0 on HZ < 1000 setups
4096          */
4097         if (!cfq_slice_async)
4098                 cfq_slice_async = 1;
4099         if (!cfq_slice_idle)
4100                 cfq_slice_idle = 1;
4101
4102 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4103         if (!cfq_group_idle)
4104                 cfq_group_idle = 1;
4105 #else
4106                 cfq_group_idle = 0;
4107 #endif
4108         if (cfq_slab_setup())
4109                 return -ENOMEM;
4110
4111         elv_register(&iosched_cfq);
4112         blkio_policy_register(&blkio_policy_cfq);
4113
4114         return 0;
4115 }
4116
4117 static void __exit cfq_exit(void)
4118 {
4119         DECLARE_COMPLETION_ONSTACK(all_gone);
4120         blkio_policy_unregister(&blkio_policy_cfq);
4121         elv_unregister(&iosched_cfq);
4122         ioc_gone = &all_gone;
4123         /* ioc_gone's update must be visible before reading ioc_count */
4124         smp_wmb();
4125
4126         /*
4127          * this also protects us from entering cfq_slab_kill() with
4128          * pending RCU callbacks
4129          */
4130         if (elv_ioc_count_read(cfq_ioc_count))
4131                 wait_for_completion(&all_gone);
4132         ida_destroy(&cic_index_ida);
4133         cfq_slab_kill();
4134 }
4135
4136 module_init(cfq_init);
4137 module_exit(cfq_exit);
4138
4139 MODULE_AUTHOR("Jens Axboe");
4140 MODULE_LICENSE("GPL");
4141 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");