ea83a4f0c27dfda658ee41d87979a2d57c58450d
[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)
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
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         struct cfq_group *orig_cfqg;
150         /* Number of sectors dispatched from queue in single dispatch round */
151         unsigned long nr_sectors;
152 };
153
154 /*
155  * First index in the service_trees.
156  * IDLE is handled separately, so it has negative index
157  */
158 enum wl_prio_t {
159         BE_WORKLOAD = 0,
160         RT_WORKLOAD = 1,
161         IDLE_WORKLOAD = 2,
162         CFQ_PRIO_NR,
163 };
164
165 /*
166  * Second index in the service_trees.
167  */
168 enum wl_type_t {
169         ASYNC_WORKLOAD = 0,
170         SYNC_NOIDLE_WORKLOAD = 1,
171         SYNC_WORKLOAD = 2
172 };
173
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176         /* group service_tree member */
177         struct rb_node rb_node;
178
179         /* group service_tree key */
180         u64 vdisktime;
181         unsigned int weight;
182
183         /* number of cfqq currently on this group */
184         int nr_cfqq;
185
186         /*
187          * Per group busy queus average. Useful for workload slice calc. We
188          * create the array for each prio class but at run time it is used
189          * only for RT and BE class and slot for IDLE class remains unused.
190          * This is primarily done to avoid confusion and a gcc warning.
191          */
192         unsigned int busy_queues_avg[CFQ_PRIO_NR];
193         /*
194          * rr lists of queues with requests. We maintain service trees for
195          * RT and BE classes. These trees are subdivided in subclasses
196          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197          * class there is no subclassification and all the cfq queues go on
198          * a single tree service_tree_idle.
199          * Counts are embedded in the cfq_rb_root
200          */
201         struct cfq_rb_root service_trees[2][3];
202         struct cfq_rb_root service_tree_idle;
203
204         unsigned long saved_workload_slice;
205         enum wl_type_t saved_workload;
206         enum wl_prio_t saved_serving_prio;
207         struct blkio_group blkg;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209         struct hlist_node cfqd_node;
210         int ref;
211 #endif
212         /* number of requests that are on the dispatch list or inside driver */
213         int dispatched;
214 };
215
216 /*
217  * Per block device queue structure
218  */
219 struct cfq_data {
220         struct request_queue *queue;
221         /* Root service tree for cfq_groups */
222         struct cfq_rb_root grp_service_tree;
223         struct cfq_group root_group;
224
225         /*
226          * The priority currently being served
227          */
228         enum wl_prio_t serving_prio;
229         enum wl_type_t serving_type;
230         unsigned long workload_expires;
231         struct cfq_group *serving_group;
232
233         /*
234          * Each priority tree is sorted by next_request position.  These
235          * trees are used when determining if two or more queues are
236          * interleaving requests (see cfq_close_cooperator).
237          */
238         struct rb_root prio_trees[CFQ_PRIO_LISTS];
239
240         unsigned int busy_queues;
241
242         int rq_in_driver;
243         int rq_in_flight[2];
244
245         /*
246          * queue-depth detection
247          */
248         int rq_queued;
249         int hw_tag;
250         /*
251          * hw_tag can be
252          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
253          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
254          *  0 => no NCQ
255          */
256         int hw_tag_est_depth;
257         unsigned int hw_tag_samples;
258
259         /*
260          * idle window management
261          */
262         struct timer_list idle_slice_timer;
263         struct work_struct unplug_work;
264
265         struct cfq_queue *active_queue;
266         struct cfq_io_context *active_cic;
267
268         /*
269          * async queue for each priority case
270          */
271         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
272         struct cfq_queue *async_idle_cfqq;
273
274         sector_t last_position;
275
276         /*
277          * tunables, see top of file
278          */
279         unsigned int cfq_quantum;
280         unsigned int cfq_fifo_expire[2];
281         unsigned int cfq_back_penalty;
282         unsigned int cfq_back_max;
283         unsigned int cfq_slice[2];
284         unsigned int cfq_slice_async_rq;
285         unsigned int cfq_slice_idle;
286         unsigned int cfq_group_idle;
287         unsigned int cfq_latency;
288         unsigned int cfq_group_isolation;
289
290         unsigned int cic_index;
291         struct list_head cic_list;
292
293         /*
294          * Fallback dummy cfqq for extreme OOM conditions
295          */
296         struct cfq_queue oom_cfqq;
297
298         unsigned long last_delayed_sync;
299
300         /* List of cfq groups being managed on this device*/
301         struct hlist_head cfqg_list;
302         struct rcu_head rcu;
303 };
304
305 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
306
307 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
308                                             enum wl_prio_t prio,
309                                             enum wl_type_t type)
310 {
311         if (!cfqg)
312                 return NULL;
313
314         if (prio == IDLE_WORKLOAD)
315                 return &cfqg->service_tree_idle;
316
317         return &cfqg->service_trees[prio][type];
318 }
319
320 enum cfqq_state_flags {
321         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
322         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
323         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
324         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
325         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
326         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
327         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
328         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
329         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
330         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
331         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
332         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
333         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
334 };
335
336 #define CFQ_CFQQ_FNS(name)                                              \
337 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
338 {                                                                       \
339         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
340 }                                                                       \
341 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
342 {                                                                       \
343         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
344 }                                                                       \
345 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
346 {                                                                       \
347         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
348 }
349
350 CFQ_CFQQ_FNS(on_rr);
351 CFQ_CFQQ_FNS(wait_request);
352 CFQ_CFQQ_FNS(must_dispatch);
353 CFQ_CFQQ_FNS(must_alloc_slice);
354 CFQ_CFQQ_FNS(fifo_expire);
355 CFQ_CFQQ_FNS(idle_window);
356 CFQ_CFQQ_FNS(prio_changed);
357 CFQ_CFQQ_FNS(slice_new);
358 CFQ_CFQQ_FNS(sync);
359 CFQ_CFQQ_FNS(coop);
360 CFQ_CFQQ_FNS(split_coop);
361 CFQ_CFQQ_FNS(deep);
362 CFQ_CFQQ_FNS(wait_busy);
363 #undef CFQ_CFQQ_FNS
364
365 #ifdef CONFIG_CFQ_GROUP_IOSCHED
366 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
367         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
368                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
369                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
370
371 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
372         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
373                                 blkg_path(&(cfqg)->blkg), ##args);      \
374
375 #else
376 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
377         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
378 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
379 #endif
380 #define cfq_log(cfqd, fmt, args...)     \
381         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
382
383 /* Traverses through cfq group service trees */
384 #define for_each_cfqg_st(cfqg, i, j, st) \
385         for (i = 0; i <= IDLE_WORKLOAD; i++) \
386                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
387                         : &cfqg->service_tree_idle; \
388                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
389                         (i == IDLE_WORKLOAD && j == 0); \
390                         j++, st = i < IDLE_WORKLOAD ? \
391                         &cfqg->service_trees[i][j]: NULL) \
392
393
394 static inline bool iops_mode(struct cfq_data *cfqd)
395 {
396         /*
397          * If we are not idling on queues and it is a NCQ drive, parallel
398          * execution of requests is on and measuring time is not possible
399          * in most of the cases until and unless we drive shallower queue
400          * depths and that becomes a performance bottleneck. In such cases
401          * switch to start providing fairness in terms of number of IOs.
402          */
403         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
404                 return true;
405         else
406                 return false;
407 }
408
409 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
410 {
411         if (cfq_class_idle(cfqq))
412                 return IDLE_WORKLOAD;
413         if (cfq_class_rt(cfqq))
414                 return RT_WORKLOAD;
415         return BE_WORKLOAD;
416 }
417
418
419 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
420 {
421         if (!cfq_cfqq_sync(cfqq))
422                 return ASYNC_WORKLOAD;
423         if (!cfq_cfqq_idle_window(cfqq))
424                 return SYNC_NOIDLE_WORKLOAD;
425         return SYNC_WORKLOAD;
426 }
427
428 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
429                                         struct cfq_data *cfqd,
430                                         struct cfq_group *cfqg)
431 {
432         if (wl == IDLE_WORKLOAD)
433                 return cfqg->service_tree_idle.count;
434
435         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
436                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
437                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
438 }
439
440 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
441                                         struct cfq_group *cfqg)
442 {
443         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
444                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
445 }
446
447 static void cfq_dispatch_insert(struct request_queue *, struct request *);
448 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
449                                        struct io_context *, gfp_t);
450 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
451                                                 struct io_context *);
452
453 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
454                                             bool is_sync)
455 {
456         return cic->cfqq[is_sync];
457 }
458
459 static inline void cic_set_cfqq(struct cfq_io_context *cic,
460                                 struct cfq_queue *cfqq, bool is_sync)
461 {
462         cic->cfqq[is_sync] = cfqq;
463 }
464
465 #define CIC_DEAD_KEY    1ul
466 #define CIC_DEAD_INDEX_SHIFT    1
467
468 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
469 {
470         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
471 }
472
473 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
474 {
475         struct cfq_data *cfqd = cic->key;
476
477         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
478                 return NULL;
479
480         return cfqd;
481 }
482
483 /*
484  * We regard a request as SYNC, if it's either a read or has the SYNC bit
485  * set (in which case it could also be direct WRITE).
486  */
487 static inline bool cfq_bio_sync(struct bio *bio)
488 {
489         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
490 }
491
492 /*
493  * scheduler run of queue, if there are requests pending and no one in the
494  * driver that will restart queueing
495  */
496 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
497 {
498         if (cfqd->busy_queues) {
499                 cfq_log(cfqd, "schedule dispatch");
500                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
501         }
502 }
503
504 static int cfq_queue_empty(struct request_queue *q)
505 {
506         struct cfq_data *cfqd = q->elevator->elevator_data;
507
508         return !cfqd->rq_queued;
509 }
510
511 /*
512  * Scale schedule slice based on io priority. Use the sync time slice only
513  * if a queue is marked sync and has sync io queued. A sync queue with async
514  * io only, should not get full sync slice length.
515  */
516 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
517                                  unsigned short prio)
518 {
519         const int base_slice = cfqd->cfq_slice[sync];
520
521         WARN_ON(prio >= IOPRIO_BE_NR);
522
523         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
524 }
525
526 static inline int
527 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
528 {
529         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
530 }
531
532 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
533 {
534         u64 d = delta << CFQ_SERVICE_SHIFT;
535
536         d = d * BLKIO_WEIGHT_DEFAULT;
537         do_div(d, cfqg->weight);
538         return d;
539 }
540
541 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
542 {
543         s64 delta = (s64)(vdisktime - min_vdisktime);
544         if (delta > 0)
545                 min_vdisktime = vdisktime;
546
547         return min_vdisktime;
548 }
549
550 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
551 {
552         s64 delta = (s64)(vdisktime - min_vdisktime);
553         if (delta < 0)
554                 min_vdisktime = vdisktime;
555
556         return min_vdisktime;
557 }
558
559 static void update_min_vdisktime(struct cfq_rb_root *st)
560 {
561         u64 vdisktime = st->min_vdisktime;
562         struct cfq_group *cfqg;
563
564         if (st->left) {
565                 cfqg = rb_entry_cfqg(st->left);
566                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
567         }
568
569         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
570 }
571
572 /*
573  * get averaged number of queues of RT/BE priority.
574  * average is updated, with a formula that gives more weight to higher numbers,
575  * to quickly follows sudden increases and decrease slowly
576  */
577
578 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
579                                         struct cfq_group *cfqg, bool rt)
580 {
581         unsigned min_q, max_q;
582         unsigned mult  = cfq_hist_divisor - 1;
583         unsigned round = cfq_hist_divisor / 2;
584         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
585
586         min_q = min(cfqg->busy_queues_avg[rt], busy);
587         max_q = max(cfqg->busy_queues_avg[rt], busy);
588         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
589                 cfq_hist_divisor;
590         return cfqg->busy_queues_avg[rt];
591 }
592
593 static inline unsigned
594 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
595 {
596         struct cfq_rb_root *st = &cfqd->grp_service_tree;
597
598         return cfq_target_latency * cfqg->weight / st->total_weight;
599 }
600
601 static inline unsigned
602 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
603 {
604         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
605         if (cfqd->cfq_latency) {
606                 /*
607                  * interested queues (we consider only the ones with the same
608                  * priority class in the cfq group)
609                  */
610                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
611                                                 cfq_class_rt(cfqq));
612                 unsigned sync_slice = cfqd->cfq_slice[1];
613                 unsigned expect_latency = sync_slice * iq;
614                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
615
616                 if (expect_latency > group_slice) {
617                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
618                         /* scale low_slice according to IO priority
619                          * and sync vs async */
620                         unsigned low_slice =
621                                 min(slice, base_low_slice * slice / sync_slice);
622                         /* the adapted slice value is scaled to fit all iqs
623                          * into the target latency */
624                         slice = max(slice * group_slice / expect_latency,
625                                     low_slice);
626                 }
627         }
628         return slice;
629 }
630
631 static inline void
632 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
633 {
634         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
635
636         cfqq->slice_start = jiffies;
637         cfqq->slice_end = jiffies + slice;
638         cfqq->allocated_slice = slice;
639         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
640 }
641
642 /*
643  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
644  * isn't valid until the first request from the dispatch is activated
645  * and the slice time set.
646  */
647 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
648 {
649         if (cfq_cfqq_slice_new(cfqq))
650                 return false;
651         if (time_before(jiffies, cfqq->slice_end))
652                 return false;
653
654         return true;
655 }
656
657 /*
658  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
659  * We choose the request that is closest to the head right now. Distance
660  * behind the head is penalized and only allowed to a certain extent.
661  */
662 static struct request *
663 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
664 {
665         sector_t s1, s2, d1 = 0, d2 = 0;
666         unsigned long back_max;
667 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
668 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
669         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
670
671         if (rq1 == NULL || rq1 == rq2)
672                 return rq2;
673         if (rq2 == NULL)
674                 return rq1;
675
676         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
677                 return rq1;
678         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
679                 return rq2;
680         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
681                 return rq1;
682         else if ((rq2->cmd_flags & REQ_META) &&
683                  !(rq1->cmd_flags & REQ_META))
684                 return rq2;
685
686         s1 = blk_rq_pos(rq1);
687         s2 = blk_rq_pos(rq2);
688
689         /*
690          * by definition, 1KiB is 2 sectors
691          */
692         back_max = cfqd->cfq_back_max * 2;
693
694         /*
695          * Strict one way elevator _except_ in the case where we allow
696          * short backward seeks which are biased as twice the cost of a
697          * similar forward seek.
698          */
699         if (s1 >= last)
700                 d1 = s1 - last;
701         else if (s1 + back_max >= last)
702                 d1 = (last - s1) * cfqd->cfq_back_penalty;
703         else
704                 wrap |= CFQ_RQ1_WRAP;
705
706         if (s2 >= last)
707                 d2 = s2 - last;
708         else if (s2 + back_max >= last)
709                 d2 = (last - s2) * cfqd->cfq_back_penalty;
710         else
711                 wrap |= CFQ_RQ2_WRAP;
712
713         /* Found required data */
714
715         /*
716          * By doing switch() on the bit mask "wrap" we avoid having to
717          * check two variables for all permutations: --> faster!
718          */
719         switch (wrap) {
720         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
721                 if (d1 < d2)
722                         return rq1;
723                 else if (d2 < d1)
724                         return rq2;
725                 else {
726                         if (s1 >= s2)
727                                 return rq1;
728                         else
729                                 return rq2;
730                 }
731
732         case CFQ_RQ2_WRAP:
733                 return rq1;
734         case CFQ_RQ1_WRAP:
735                 return rq2;
736         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
737         default:
738                 /*
739                  * Since both rqs are wrapped,
740                  * start with the one that's further behind head
741                  * (--> only *one* back seek required),
742                  * since back seek takes more time than forward.
743                  */
744                 if (s1 <= s2)
745                         return rq1;
746                 else
747                         return rq2;
748         }
749 }
750
751 /*
752  * The below is leftmost cache rbtree addon
753  */
754 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
755 {
756         /* Service tree is empty */
757         if (!root->count)
758                 return NULL;
759
760         if (!root->left)
761                 root->left = rb_first(&root->rb);
762
763         if (root->left)
764                 return rb_entry(root->left, struct cfq_queue, rb_node);
765
766         return NULL;
767 }
768
769 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
770 {
771         if (!root->left)
772                 root->left = rb_first(&root->rb);
773
774         if (root->left)
775                 return rb_entry_cfqg(root->left);
776
777         return NULL;
778 }
779
780 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
781 {
782         rb_erase(n, root);
783         RB_CLEAR_NODE(n);
784 }
785
786 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
787 {
788         if (root->left == n)
789                 root->left = NULL;
790         rb_erase_init(n, &root->rb);
791         --root->count;
792 }
793
794 /*
795  * would be nice to take fifo expire time into account as well
796  */
797 static struct request *
798 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
799                   struct request *last)
800 {
801         struct rb_node *rbnext = rb_next(&last->rb_node);
802         struct rb_node *rbprev = rb_prev(&last->rb_node);
803         struct request *next = NULL, *prev = NULL;
804
805         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
806
807         if (rbprev)
808                 prev = rb_entry_rq(rbprev);
809
810         if (rbnext)
811                 next = rb_entry_rq(rbnext);
812         else {
813                 rbnext = rb_first(&cfqq->sort_list);
814                 if (rbnext && rbnext != &last->rb_node)
815                         next = rb_entry_rq(rbnext);
816         }
817
818         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
819 }
820
821 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
822                                       struct cfq_queue *cfqq)
823 {
824         /*
825          * just an approximation, should be ok.
826          */
827         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
828                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
829 }
830
831 static inline s64
832 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
833 {
834         return cfqg->vdisktime - st->min_vdisktime;
835 }
836
837 static void
838 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
839 {
840         struct rb_node **node = &st->rb.rb_node;
841         struct rb_node *parent = NULL;
842         struct cfq_group *__cfqg;
843         s64 key = cfqg_key(st, cfqg);
844         int left = 1;
845
846         while (*node != NULL) {
847                 parent = *node;
848                 __cfqg = rb_entry_cfqg(parent);
849
850                 if (key < cfqg_key(st, __cfqg))
851                         node = &parent->rb_left;
852                 else {
853                         node = &parent->rb_right;
854                         left = 0;
855                 }
856         }
857
858         if (left)
859                 st->left = &cfqg->rb_node;
860
861         rb_link_node(&cfqg->rb_node, parent, node);
862         rb_insert_color(&cfqg->rb_node, &st->rb);
863 }
864
865 static void
866 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
867 {
868         struct cfq_rb_root *st = &cfqd->grp_service_tree;
869         struct cfq_group *__cfqg;
870         struct rb_node *n;
871
872         cfqg->nr_cfqq++;
873         if (!RB_EMPTY_NODE(&cfqg->rb_node))
874                 return;
875
876         /*
877          * Currently put the group at the end. Later implement something
878          * so that groups get lesser vtime based on their weights, so that
879          * if group does not loose all if it was not continously backlogged.
880          */
881         n = rb_last(&st->rb);
882         if (n) {
883                 __cfqg = rb_entry_cfqg(n);
884                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
885         } else
886                 cfqg->vdisktime = st->min_vdisktime;
887
888         __cfq_group_service_tree_add(st, cfqg);
889         st->total_weight += cfqg->weight;
890 }
891
892 static void
893 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
894 {
895         struct cfq_rb_root *st = &cfqd->grp_service_tree;
896
897         BUG_ON(cfqg->nr_cfqq < 1);
898         cfqg->nr_cfqq--;
899
900         /* If there are other cfq queues under this group, don't delete it */
901         if (cfqg->nr_cfqq)
902                 return;
903
904         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
905         st->total_weight -= cfqg->weight;
906         if (!RB_EMPTY_NODE(&cfqg->rb_node))
907                 cfq_rb_erase(&cfqg->rb_node, st);
908         cfqg->saved_workload_slice = 0;
909         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
910 }
911
912 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
913 {
914         unsigned int slice_used;
915
916         /*
917          * Queue got expired before even a single request completed or
918          * got expired immediately after first request completion.
919          */
920         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
921                 /*
922                  * Also charge the seek time incurred to the group, otherwise
923                  * if there are mutiple queues in the group, each can dispatch
924                  * a single request on seeky media and cause lots of seek time
925                  * and group will never know it.
926                  */
927                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
928                                         1);
929         } else {
930                 slice_used = jiffies - cfqq->slice_start;
931                 if (slice_used > cfqq->allocated_slice)
932                         slice_used = cfqq->allocated_slice;
933         }
934
935         return slice_used;
936 }
937
938 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
939                                 struct cfq_queue *cfqq)
940 {
941         struct cfq_rb_root *st = &cfqd->grp_service_tree;
942         unsigned int used_sl, charge;
943         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
944                         - cfqg->service_tree_idle.count;
945
946         BUG_ON(nr_sync < 0);
947         used_sl = charge = cfq_cfqq_slice_usage(cfqq);
948
949         if (iops_mode(cfqd))
950                 charge = cfqq->slice_dispatch;
951         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
952                 charge = cfqq->allocated_slice;
953
954         /* Can't update vdisktime while group is on service tree */
955         cfq_rb_erase(&cfqg->rb_node, st);
956         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
957         __cfq_group_service_tree_add(st, cfqg);
958
959         /* This group is being expired. Save the context */
960         if (time_after(cfqd->workload_expires, jiffies)) {
961                 cfqg->saved_workload_slice = cfqd->workload_expires
962                                                 - jiffies;
963                 cfqg->saved_workload = cfqd->serving_type;
964                 cfqg->saved_serving_prio = cfqd->serving_prio;
965         } else
966                 cfqg->saved_workload_slice = 0;
967
968         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
969                                         st->min_vdisktime);
970         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
971                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
972                         iops_mode(cfqd), cfqq->nr_sectors);
973         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
974         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
975 }
976
977 #ifdef CONFIG_CFQ_GROUP_IOSCHED
978 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
979 {
980         if (blkg)
981                 return container_of(blkg, struct cfq_group, blkg);
982         return NULL;
983 }
984
985 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
986                                         unsigned int weight)
987 {
988         cfqg_of_blkg(blkg)->weight = weight;
989 }
990
991 static struct cfq_group *
992 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
993 {
994         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
995         struct cfq_group *cfqg = NULL;
996         void *key = cfqd;
997         int i, j;
998         struct cfq_rb_root *st;
999         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1000         unsigned int major, minor;
1001
1002         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1003         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1004                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1005                 cfqg->blkg.dev = MKDEV(major, minor);
1006                 goto done;
1007         }
1008         if (cfqg || !create)
1009                 goto done;
1010
1011         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1012         if (!cfqg)
1013                 goto done;
1014
1015         for_each_cfqg_st(cfqg, i, j, st)
1016                 *st = CFQ_RB_ROOT;
1017         RB_CLEAR_NODE(&cfqg->rb_node);
1018
1019         /*
1020          * Take the initial reference that will be released on destroy
1021          * This can be thought of a joint reference by cgroup and
1022          * elevator which will be dropped by either elevator exit
1023          * or cgroup deletion path depending on who is exiting first.
1024          */
1025         cfqg->ref = 1;
1026
1027         /*
1028          * Add group onto cgroup list. It might happen that bdi->dev is
1029          * not initialized yet. Initialize this new group without major
1030          * and minor info and this info will be filled in once a new thread
1031          * comes for IO. See code above.
1032          */
1033         if (bdi->dev) {
1034                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1035                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1036                                         MKDEV(major, minor));
1037         } else
1038                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1039                                         0);
1040
1041         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1042
1043         /* Add group on cfqd list */
1044         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1045
1046 done:
1047         return cfqg;
1048 }
1049
1050 /*
1051  * Search for the cfq group current task belongs to. If create = 1, then also
1052  * create the cfq group if it does not exist. request_queue lock must be held.
1053  */
1054 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1055 {
1056         struct cgroup *cgroup;
1057         struct cfq_group *cfqg = NULL;
1058
1059         rcu_read_lock();
1060         cgroup = task_cgroup(current, blkio_subsys_id);
1061         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1062         if (!cfqg && create)
1063                 cfqg = &cfqd->root_group;
1064         rcu_read_unlock();
1065         return cfqg;
1066 }
1067
1068 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1069 {
1070         cfqg->ref++;
1071         return cfqg;
1072 }
1073
1074 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1075 {
1076         /* Currently, all async queues are mapped to root group */
1077         if (!cfq_cfqq_sync(cfqq))
1078                 cfqg = &cfqq->cfqd->root_group;
1079
1080         cfqq->cfqg = cfqg;
1081         /* cfqq reference on cfqg */
1082         cfqq->cfqg->ref++;
1083 }
1084
1085 static void cfq_put_cfqg(struct cfq_group *cfqg)
1086 {
1087         struct cfq_rb_root *st;
1088         int i, j;
1089
1090         BUG_ON(cfqg->ref <= 0);
1091         cfqg->ref--;
1092         if (cfqg->ref)
1093                 return;
1094         for_each_cfqg_st(cfqg, i, j, st)
1095                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1096         kfree(cfqg);
1097 }
1098
1099 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1100 {
1101         /* Something wrong if we are trying to remove same group twice */
1102         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1103
1104         hlist_del_init(&cfqg->cfqd_node);
1105
1106         /*
1107          * Put the reference taken at the time of creation so that when all
1108          * queues are gone, group can be destroyed.
1109          */
1110         cfq_put_cfqg(cfqg);
1111 }
1112
1113 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1114 {
1115         struct hlist_node *pos, *n;
1116         struct cfq_group *cfqg;
1117
1118         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1119                 /*
1120                  * If cgroup removal path got to blk_group first and removed
1121                  * it from cgroup list, then it will take care of destroying
1122                  * cfqg also.
1123                  */
1124                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1125                         cfq_destroy_cfqg(cfqd, cfqg);
1126         }
1127 }
1128
1129 /*
1130  * Blk cgroup controller notification saying that blkio_group object is being
1131  * delinked as associated cgroup object is going away. That also means that
1132  * no new IO will come in this group. So get rid of this group as soon as
1133  * any pending IO in the group is finished.
1134  *
1135  * This function is called under rcu_read_lock(). key is the rcu protected
1136  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1137  * read lock.
1138  *
1139  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1140  * it should not be NULL as even if elevator was exiting, cgroup deltion
1141  * path got to it first.
1142  */
1143 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1144 {
1145         unsigned long  flags;
1146         struct cfq_data *cfqd = key;
1147
1148         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1149         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1150         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1151 }
1152
1153 #else /* GROUP_IOSCHED */
1154 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1155 {
1156         return &cfqd->root_group;
1157 }
1158
1159 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1160 {
1161         return cfqg;
1162 }
1163
1164 static inline void
1165 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1166         cfqq->cfqg = cfqg;
1167 }
1168
1169 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1170 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1171
1172 #endif /* GROUP_IOSCHED */
1173
1174 /*
1175  * The cfqd->service_trees holds all pending cfq_queue's that have
1176  * requests waiting to be processed. It is sorted in the order that
1177  * we will service the queues.
1178  */
1179 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1180                                  bool add_front)
1181 {
1182         struct rb_node **p, *parent;
1183         struct cfq_queue *__cfqq;
1184         unsigned long rb_key;
1185         struct cfq_rb_root *service_tree;
1186         int left;
1187         int new_cfqq = 1;
1188         int group_changed = 0;
1189
1190 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1191         if (!cfqd->cfq_group_isolation
1192             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1193             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1194                 /* Move this cfq to root group */
1195                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1196                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1197                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1198                 cfqq->orig_cfqg = cfqq->cfqg;
1199                 cfqq->cfqg = &cfqd->root_group;
1200                 cfqd->root_group.ref++;
1201                 group_changed = 1;
1202         } else if (!cfqd->cfq_group_isolation
1203                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1204                 /* cfqq is sequential now needs to go to its original group */
1205                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1206                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1207                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1208                 cfq_put_cfqg(cfqq->cfqg);
1209                 cfqq->cfqg = cfqq->orig_cfqg;
1210                 cfqq->orig_cfqg = NULL;
1211                 group_changed = 1;
1212                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1213         }
1214 #endif
1215
1216         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1217                                                 cfqq_type(cfqq));
1218         if (cfq_class_idle(cfqq)) {
1219                 rb_key = CFQ_IDLE_DELAY;
1220                 parent = rb_last(&service_tree->rb);
1221                 if (parent && parent != &cfqq->rb_node) {
1222                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1223                         rb_key += __cfqq->rb_key;
1224                 } else
1225                         rb_key += jiffies;
1226         } else if (!add_front) {
1227                 /*
1228                  * Get our rb key offset. Subtract any residual slice
1229                  * value carried from last service. A negative resid
1230                  * count indicates slice overrun, and this should position
1231                  * the next service time further away in the tree.
1232                  */
1233                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1234                 rb_key -= cfqq->slice_resid;
1235                 cfqq->slice_resid = 0;
1236         } else {
1237                 rb_key = -HZ;
1238                 __cfqq = cfq_rb_first(service_tree);
1239                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1240         }
1241
1242         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1243                 new_cfqq = 0;
1244                 /*
1245                  * same position, nothing more to do
1246                  */
1247                 if (rb_key == cfqq->rb_key &&
1248                     cfqq->service_tree == service_tree)
1249                         return;
1250
1251                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1252                 cfqq->service_tree = NULL;
1253         }
1254
1255         left = 1;
1256         parent = NULL;
1257         cfqq->service_tree = service_tree;
1258         p = &service_tree->rb.rb_node;
1259         while (*p) {
1260                 struct rb_node **n;
1261
1262                 parent = *p;
1263                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1264
1265                 /*
1266                  * sort by key, that represents service time.
1267                  */
1268                 if (time_before(rb_key, __cfqq->rb_key))
1269                         n = &(*p)->rb_left;
1270                 else {
1271                         n = &(*p)->rb_right;
1272                         left = 0;
1273                 }
1274
1275                 p = n;
1276         }
1277
1278         if (left)
1279                 service_tree->left = &cfqq->rb_node;
1280
1281         cfqq->rb_key = rb_key;
1282         rb_link_node(&cfqq->rb_node, parent, p);
1283         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1284         service_tree->count++;
1285         if ((add_front || !new_cfqq) && !group_changed)
1286                 return;
1287         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1288 }
1289
1290 static struct cfq_queue *
1291 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1292                      sector_t sector, struct rb_node **ret_parent,
1293                      struct rb_node ***rb_link)
1294 {
1295         struct rb_node **p, *parent;
1296         struct cfq_queue *cfqq = NULL;
1297
1298         parent = NULL;
1299         p = &root->rb_node;
1300         while (*p) {
1301                 struct rb_node **n;
1302
1303                 parent = *p;
1304                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1305
1306                 /*
1307                  * Sort strictly based on sector.  Smallest to the left,
1308                  * largest to the right.
1309                  */
1310                 if (sector > blk_rq_pos(cfqq->next_rq))
1311                         n = &(*p)->rb_right;
1312                 else if (sector < blk_rq_pos(cfqq->next_rq))
1313                         n = &(*p)->rb_left;
1314                 else
1315                         break;
1316                 p = n;
1317                 cfqq = NULL;
1318         }
1319
1320         *ret_parent = parent;
1321         if (rb_link)
1322                 *rb_link = p;
1323         return cfqq;
1324 }
1325
1326 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1327 {
1328         struct rb_node **p, *parent;
1329         struct cfq_queue *__cfqq;
1330
1331         if (cfqq->p_root) {
1332                 rb_erase(&cfqq->p_node, cfqq->p_root);
1333                 cfqq->p_root = NULL;
1334         }
1335
1336         if (cfq_class_idle(cfqq))
1337                 return;
1338         if (!cfqq->next_rq)
1339                 return;
1340
1341         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1342         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1343                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1344         if (!__cfqq) {
1345                 rb_link_node(&cfqq->p_node, parent, p);
1346                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1347         } else
1348                 cfqq->p_root = NULL;
1349 }
1350
1351 /*
1352  * Update cfqq's position in the service tree.
1353  */
1354 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1355 {
1356         /*
1357          * Resorting requires the cfqq to be on the RR list already.
1358          */
1359         if (cfq_cfqq_on_rr(cfqq)) {
1360                 cfq_service_tree_add(cfqd, cfqq, 0);
1361                 cfq_prio_tree_add(cfqd, cfqq);
1362         }
1363 }
1364
1365 /*
1366  * add to busy list of queues for service, trying to be fair in ordering
1367  * the pending list according to last request service
1368  */
1369 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1370 {
1371         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1372         BUG_ON(cfq_cfqq_on_rr(cfqq));
1373         cfq_mark_cfqq_on_rr(cfqq);
1374         cfqd->busy_queues++;
1375
1376         cfq_resort_rr_list(cfqd, cfqq);
1377 }
1378
1379 /*
1380  * Called when the cfqq no longer has requests pending, remove it from
1381  * the service tree.
1382  */
1383 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1384 {
1385         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1386         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1387         cfq_clear_cfqq_on_rr(cfqq);
1388
1389         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1390                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1391                 cfqq->service_tree = NULL;
1392         }
1393         if (cfqq->p_root) {
1394                 rb_erase(&cfqq->p_node, cfqq->p_root);
1395                 cfqq->p_root = NULL;
1396         }
1397
1398         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1399         BUG_ON(!cfqd->busy_queues);
1400         cfqd->busy_queues--;
1401 }
1402
1403 /*
1404  * rb tree support functions
1405  */
1406 static void cfq_del_rq_rb(struct request *rq)
1407 {
1408         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1409         const int sync = rq_is_sync(rq);
1410
1411         BUG_ON(!cfqq->queued[sync]);
1412         cfqq->queued[sync]--;
1413
1414         elv_rb_del(&cfqq->sort_list, rq);
1415
1416         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1417                 /*
1418                  * Queue will be deleted from service tree when we actually
1419                  * expire it later. Right now just remove it from prio tree
1420                  * as it is empty.
1421                  */
1422                 if (cfqq->p_root) {
1423                         rb_erase(&cfqq->p_node, cfqq->p_root);
1424                         cfqq->p_root = NULL;
1425                 }
1426         }
1427 }
1428
1429 static void cfq_add_rq_rb(struct request *rq)
1430 {
1431         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1432         struct cfq_data *cfqd = cfqq->cfqd;
1433         struct request *__alias, *prev;
1434
1435         cfqq->queued[rq_is_sync(rq)]++;
1436
1437         /*
1438          * looks a little odd, but the first insert might return an alias.
1439          * if that happens, put the alias on the dispatch list
1440          */
1441         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1442                 cfq_dispatch_insert(cfqd->queue, __alias);
1443
1444         if (!cfq_cfqq_on_rr(cfqq))
1445                 cfq_add_cfqq_rr(cfqd, cfqq);
1446
1447         /*
1448          * check if this request is a better next-serve candidate
1449          */
1450         prev = cfqq->next_rq;
1451         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1452
1453         /*
1454          * adjust priority tree position, if ->next_rq changes
1455          */
1456         if (prev != cfqq->next_rq)
1457                 cfq_prio_tree_add(cfqd, cfqq);
1458
1459         BUG_ON(!cfqq->next_rq);
1460 }
1461
1462 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1463 {
1464         elv_rb_del(&cfqq->sort_list, rq);
1465         cfqq->queued[rq_is_sync(rq)]--;
1466         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1467                                         rq_data_dir(rq), rq_is_sync(rq));
1468         cfq_add_rq_rb(rq);
1469         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1470                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1471                         rq_is_sync(rq));
1472 }
1473
1474 static struct request *
1475 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1476 {
1477         struct task_struct *tsk = current;
1478         struct cfq_io_context *cic;
1479         struct cfq_queue *cfqq;
1480
1481         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1482         if (!cic)
1483                 return NULL;
1484
1485         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1486         if (cfqq) {
1487                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1488
1489                 return elv_rb_find(&cfqq->sort_list, sector);
1490         }
1491
1492         return NULL;
1493 }
1494
1495 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1496 {
1497         struct cfq_data *cfqd = q->elevator->elevator_data;
1498
1499         cfqd->rq_in_driver++;
1500         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1501                                                 cfqd->rq_in_driver);
1502
1503         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1504 }
1505
1506 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1507 {
1508         struct cfq_data *cfqd = q->elevator->elevator_data;
1509
1510         WARN_ON(!cfqd->rq_in_driver);
1511         cfqd->rq_in_driver--;
1512         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1513                                                 cfqd->rq_in_driver);
1514 }
1515
1516 static void cfq_remove_request(struct request *rq)
1517 {
1518         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1519
1520         if (cfqq->next_rq == rq)
1521                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1522
1523         list_del_init(&rq->queuelist);
1524         cfq_del_rq_rb(rq);
1525
1526         cfqq->cfqd->rq_queued--;
1527         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1528                                         rq_data_dir(rq), rq_is_sync(rq));
1529         if (rq->cmd_flags & REQ_META) {
1530                 WARN_ON(!cfqq->meta_pending);
1531                 cfqq->meta_pending--;
1532         }
1533 }
1534
1535 static int cfq_merge(struct request_queue *q, struct request **req,
1536                      struct bio *bio)
1537 {
1538         struct cfq_data *cfqd = q->elevator->elevator_data;
1539         struct request *__rq;
1540
1541         __rq = cfq_find_rq_fmerge(cfqd, bio);
1542         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1543                 *req = __rq;
1544                 return ELEVATOR_FRONT_MERGE;
1545         }
1546
1547         return ELEVATOR_NO_MERGE;
1548 }
1549
1550 static void cfq_merged_request(struct request_queue *q, struct request *req,
1551                                int type)
1552 {
1553         if (type == ELEVATOR_FRONT_MERGE) {
1554                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1555
1556                 cfq_reposition_rq_rb(cfqq, req);
1557         }
1558 }
1559
1560 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1561                                 struct bio *bio)
1562 {
1563         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1564                                         bio_data_dir(bio), cfq_bio_sync(bio));
1565 }
1566
1567 static void
1568 cfq_merged_requests(struct request_queue *q, struct request *rq,
1569                     struct request *next)
1570 {
1571         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1572         /*
1573          * reposition in fifo if next is older than rq
1574          */
1575         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1576             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1577                 list_move(&rq->queuelist, &next->queuelist);
1578                 rq_set_fifo_time(rq, rq_fifo_time(next));
1579         }
1580
1581         if (cfqq->next_rq == next)
1582                 cfqq->next_rq = rq;
1583         cfq_remove_request(next);
1584         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1585                                         rq_data_dir(next), rq_is_sync(next));
1586 }
1587
1588 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1589                            struct bio *bio)
1590 {
1591         struct cfq_data *cfqd = q->elevator->elevator_data;
1592         struct cfq_io_context *cic;
1593         struct cfq_queue *cfqq;
1594
1595         /*
1596          * Disallow merge of a sync bio into an async request.
1597          */
1598         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1599                 return false;
1600
1601         /*
1602          * Lookup the cfqq that this bio will be queued with. Allow
1603          * merge only if rq is queued there.
1604          */
1605         cic = cfq_cic_lookup(cfqd, current->io_context);
1606         if (!cic)
1607                 return false;
1608
1609         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1610         return cfqq == RQ_CFQQ(rq);
1611 }
1612
1613 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1614 {
1615         del_timer(&cfqd->idle_slice_timer);
1616         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1617 }
1618
1619 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1620                                    struct cfq_queue *cfqq)
1621 {
1622         if (cfqq) {
1623                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1624                                 cfqd->serving_prio, cfqd->serving_type);
1625                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1626                 cfqq->slice_start = 0;
1627                 cfqq->dispatch_start = jiffies;
1628                 cfqq->allocated_slice = 0;
1629                 cfqq->slice_end = 0;
1630                 cfqq->slice_dispatch = 0;
1631                 cfqq->nr_sectors = 0;
1632
1633                 cfq_clear_cfqq_wait_request(cfqq);
1634                 cfq_clear_cfqq_must_dispatch(cfqq);
1635                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1636                 cfq_clear_cfqq_fifo_expire(cfqq);
1637                 cfq_mark_cfqq_slice_new(cfqq);
1638
1639                 cfq_del_timer(cfqd, cfqq);
1640         }
1641
1642         cfqd->active_queue = cfqq;
1643 }
1644
1645 /*
1646  * current cfqq expired its slice (or was too idle), select new one
1647  */
1648 static void
1649 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1650                     bool timed_out)
1651 {
1652         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1653
1654         if (cfq_cfqq_wait_request(cfqq))
1655                 cfq_del_timer(cfqd, cfqq);
1656
1657         cfq_clear_cfqq_wait_request(cfqq);
1658         cfq_clear_cfqq_wait_busy(cfqq);
1659
1660         /*
1661          * If this cfqq is shared between multiple processes, check to
1662          * make sure that those processes are still issuing I/Os within
1663          * the mean seek distance.  If not, it may be time to break the
1664          * queues apart again.
1665          */
1666         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1667                 cfq_mark_cfqq_split_coop(cfqq);
1668
1669         /*
1670          * store what was left of this slice, if the queue idled/timed out
1671          */
1672         if (timed_out) {
1673                 if (cfq_cfqq_slice_new(cfqq))
1674                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1675                 else
1676                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1677                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1678         }
1679
1680         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1681
1682         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1683                 cfq_del_cfqq_rr(cfqd, cfqq);
1684
1685         cfq_resort_rr_list(cfqd, cfqq);
1686
1687         if (cfqq == cfqd->active_queue)
1688                 cfqd->active_queue = NULL;
1689
1690         if (cfqd->active_cic) {
1691                 put_io_context(cfqd->active_cic->ioc);
1692                 cfqd->active_cic = NULL;
1693         }
1694 }
1695
1696 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1697 {
1698         struct cfq_queue *cfqq = cfqd->active_queue;
1699
1700         if (cfqq)
1701                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1702 }
1703
1704 /*
1705  * Get next queue for service. Unless we have a queue preemption,
1706  * we'll simply select the first cfqq in the service tree.
1707  */
1708 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1709 {
1710         struct cfq_rb_root *service_tree =
1711                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1712                                         cfqd->serving_type);
1713
1714         if (!cfqd->rq_queued)
1715                 return NULL;
1716
1717         /* There is nothing to dispatch */
1718         if (!service_tree)
1719                 return NULL;
1720         if (RB_EMPTY_ROOT(&service_tree->rb))
1721                 return NULL;
1722         return cfq_rb_first(service_tree);
1723 }
1724
1725 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1726 {
1727         struct cfq_group *cfqg;
1728         struct cfq_queue *cfqq;
1729         int i, j;
1730         struct cfq_rb_root *st;
1731
1732         if (!cfqd->rq_queued)
1733                 return NULL;
1734
1735         cfqg = cfq_get_next_cfqg(cfqd);
1736         if (!cfqg)
1737                 return NULL;
1738
1739         for_each_cfqg_st(cfqg, i, j, st)
1740                 if ((cfqq = cfq_rb_first(st)) != NULL)
1741                         return cfqq;
1742         return NULL;
1743 }
1744
1745 /*
1746  * Get and set a new active queue for service.
1747  */
1748 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1749                                               struct cfq_queue *cfqq)
1750 {
1751         if (!cfqq)
1752                 cfqq = cfq_get_next_queue(cfqd);
1753
1754         __cfq_set_active_queue(cfqd, cfqq);
1755         return cfqq;
1756 }
1757
1758 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1759                                           struct request *rq)
1760 {
1761         if (blk_rq_pos(rq) >= cfqd->last_position)
1762                 return blk_rq_pos(rq) - cfqd->last_position;
1763         else
1764                 return cfqd->last_position - blk_rq_pos(rq);
1765 }
1766
1767 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1768                                struct request *rq)
1769 {
1770         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1771 }
1772
1773 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1774                                     struct cfq_queue *cur_cfqq)
1775 {
1776         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1777         struct rb_node *parent, *node;
1778         struct cfq_queue *__cfqq;
1779         sector_t sector = cfqd->last_position;
1780
1781         if (RB_EMPTY_ROOT(root))
1782                 return NULL;
1783
1784         /*
1785          * First, if we find a request starting at the end of the last
1786          * request, choose it.
1787          */
1788         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1789         if (__cfqq)
1790                 return __cfqq;
1791
1792         /*
1793          * If the exact sector wasn't found, the parent of the NULL leaf
1794          * will contain the closest sector.
1795          */
1796         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1797         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1798                 return __cfqq;
1799
1800         if (blk_rq_pos(__cfqq->next_rq) < sector)
1801                 node = rb_next(&__cfqq->p_node);
1802         else
1803                 node = rb_prev(&__cfqq->p_node);
1804         if (!node)
1805                 return NULL;
1806
1807         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1808         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1809                 return __cfqq;
1810
1811         return NULL;
1812 }
1813
1814 /*
1815  * cfqd - obvious
1816  * cur_cfqq - passed in so that we don't decide that the current queue is
1817  *            closely cooperating with itself.
1818  *
1819  * So, basically we're assuming that that cur_cfqq has dispatched at least
1820  * one request, and that cfqd->last_position reflects a position on the disk
1821  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1822  * assumption.
1823  */
1824 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1825                                               struct cfq_queue *cur_cfqq)
1826 {
1827         struct cfq_queue *cfqq;
1828
1829         if (cfq_class_idle(cur_cfqq))
1830                 return NULL;
1831         if (!cfq_cfqq_sync(cur_cfqq))
1832                 return NULL;
1833         if (CFQQ_SEEKY(cur_cfqq))
1834                 return NULL;
1835
1836         /*
1837          * Don't search priority tree if it's the only queue in the group.
1838          */
1839         if (cur_cfqq->cfqg->nr_cfqq == 1)
1840                 return NULL;
1841
1842         /*
1843          * We should notice if some of the queues are cooperating, eg
1844          * working closely on the same area of the disk. In that case,
1845          * we can group them together and don't waste time idling.
1846          */
1847         cfqq = cfqq_close(cfqd, cur_cfqq);
1848         if (!cfqq)
1849                 return NULL;
1850
1851         /* If new queue belongs to different cfq_group, don't choose it */
1852         if (cur_cfqq->cfqg != cfqq->cfqg)
1853                 return NULL;
1854
1855         /*
1856          * It only makes sense to merge sync queues.
1857          */
1858         if (!cfq_cfqq_sync(cfqq))
1859                 return NULL;
1860         if (CFQQ_SEEKY(cfqq))
1861                 return NULL;
1862
1863         /*
1864          * Do not merge queues of different priority classes
1865          */
1866         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1867                 return NULL;
1868
1869         return cfqq;
1870 }
1871
1872 /*
1873  * Determine whether we should enforce idle window for this queue.
1874  */
1875
1876 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1877 {
1878         enum wl_prio_t prio = cfqq_prio(cfqq);
1879         struct cfq_rb_root *service_tree = cfqq->service_tree;
1880
1881         BUG_ON(!service_tree);
1882         BUG_ON(!service_tree->count);
1883
1884         if (!cfqd->cfq_slice_idle)
1885                 return false;
1886
1887         /* We never do for idle class queues. */
1888         if (prio == IDLE_WORKLOAD)
1889                 return false;
1890
1891         /* We do for queues that were marked with idle window flag. */
1892         if (cfq_cfqq_idle_window(cfqq) &&
1893            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1894                 return true;
1895
1896         /*
1897          * Otherwise, we do only if they are the last ones
1898          * in their service tree.
1899          */
1900         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1901                 return true;
1902         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1903                         service_tree->count);
1904         return false;
1905 }
1906
1907 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1908 {
1909         struct cfq_queue *cfqq = cfqd->active_queue;
1910         struct cfq_io_context *cic;
1911         unsigned long sl, group_idle = 0;
1912
1913         /*
1914          * SSD device without seek penalty, disable idling. But only do so
1915          * for devices that support queuing, otherwise we still have a problem
1916          * with sync vs async workloads.
1917          */
1918         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1919                 return;
1920
1921         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1922         WARN_ON(cfq_cfqq_slice_new(cfqq));
1923
1924         /*
1925          * idle is disabled, either manually or by past process history
1926          */
1927         if (!cfq_should_idle(cfqd, cfqq)) {
1928                 /* no queue idling. Check for group idling */
1929                 if (cfqd->cfq_group_idle)
1930                         group_idle = cfqd->cfq_group_idle;
1931                 else
1932                         return;
1933         }
1934
1935         /*
1936          * still active requests from this queue, don't idle
1937          */
1938         if (cfqq->dispatched)
1939                 return;
1940
1941         /*
1942          * task has exited, don't wait
1943          */
1944         cic = cfqd->active_cic;
1945         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1946                 return;
1947
1948         /*
1949          * If our average think time is larger than the remaining time
1950          * slice, then don't idle. This avoids overrunning the allotted
1951          * time slice.
1952          */
1953         if (sample_valid(cic->ttime_samples) &&
1954             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1955                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1956                                 cic->ttime_mean);
1957                 return;
1958         }
1959
1960         /* There are other queues in the group, don't do group idle */
1961         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1962                 return;
1963
1964         cfq_mark_cfqq_wait_request(cfqq);
1965
1966         if (group_idle)
1967                 sl = cfqd->cfq_group_idle;
1968         else
1969                 sl = cfqd->cfq_slice_idle;
1970
1971         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1972         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1973         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1974                         group_idle ? 1 : 0);
1975 }
1976
1977 /*
1978  * Move request from internal lists to the request queue dispatch list.
1979  */
1980 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1981 {
1982         struct cfq_data *cfqd = q->elevator->elevator_data;
1983         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1984
1985         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1986
1987         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1988         cfq_remove_request(rq);
1989         cfqq->dispatched++;
1990         (RQ_CFQG(rq))->dispatched++;
1991         elv_dispatch_sort(q, rq);
1992
1993         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1994         cfqq->nr_sectors += blk_rq_sectors(rq);
1995         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1996                                         rq_data_dir(rq), rq_is_sync(rq));
1997 }
1998
1999 /*
2000  * return expired entry, or NULL to just start from scratch in rbtree
2001  */
2002 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2003 {
2004         struct request *rq = NULL;
2005
2006         if (cfq_cfqq_fifo_expire(cfqq))
2007                 return NULL;
2008
2009         cfq_mark_cfqq_fifo_expire(cfqq);
2010
2011         if (list_empty(&cfqq->fifo))
2012                 return NULL;
2013
2014         rq = rq_entry_fifo(cfqq->fifo.next);
2015         if (time_before(jiffies, rq_fifo_time(rq)))
2016                 rq = NULL;
2017
2018         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2019         return rq;
2020 }
2021
2022 static inline int
2023 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2024 {
2025         const int base_rq = cfqd->cfq_slice_async_rq;
2026
2027         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2028
2029         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2030 }
2031
2032 /*
2033  * Must be called with the queue_lock held.
2034  */
2035 static int cfqq_process_refs(struct cfq_queue *cfqq)
2036 {
2037         int process_refs, io_refs;
2038
2039         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2040         process_refs = cfqq->ref - io_refs;
2041         BUG_ON(process_refs < 0);
2042         return process_refs;
2043 }
2044
2045 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2046 {
2047         int process_refs, new_process_refs;
2048         struct cfq_queue *__cfqq;
2049
2050         /*
2051          * If there are no process references on the new_cfqq, then it is
2052          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2053          * chain may have dropped their last reference (not just their
2054          * last process reference).
2055          */
2056         if (!cfqq_process_refs(new_cfqq))
2057                 return;
2058
2059         /* Avoid a circular list and skip interim queue merges */
2060         while ((__cfqq = new_cfqq->new_cfqq)) {
2061                 if (__cfqq == cfqq)
2062                         return;
2063                 new_cfqq = __cfqq;
2064         }
2065
2066         process_refs = cfqq_process_refs(cfqq);
2067         new_process_refs = cfqq_process_refs(new_cfqq);
2068         /*
2069          * If the process for the cfqq has gone away, there is no
2070          * sense in merging the queues.
2071          */
2072         if (process_refs == 0 || new_process_refs == 0)
2073                 return;
2074
2075         /*
2076          * Merge in the direction of the lesser amount of work.
2077          */
2078         if (new_process_refs >= process_refs) {
2079                 cfqq->new_cfqq = new_cfqq;
2080                 new_cfqq->ref += process_refs;
2081         } else {
2082                 new_cfqq->new_cfqq = cfqq;
2083                 cfqq->ref += new_process_refs;
2084         }
2085 }
2086
2087 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2088                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2089 {
2090         struct cfq_queue *queue;
2091         int i;
2092         bool key_valid = false;
2093         unsigned long lowest_key = 0;
2094         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2095
2096         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2097                 /* select the one with lowest rb_key */
2098                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2099                 if (queue &&
2100                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2101                         lowest_key = queue->rb_key;
2102                         cur_best = i;
2103                         key_valid = true;
2104                 }
2105         }
2106
2107         return cur_best;
2108 }
2109
2110 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2111 {
2112         unsigned slice;
2113         unsigned count;
2114         struct cfq_rb_root *st;
2115         unsigned group_slice;
2116         enum wl_prio_t original_prio = cfqd->serving_prio;
2117
2118         /* Choose next priority. RT > BE > IDLE */
2119         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2120                 cfqd->serving_prio = RT_WORKLOAD;
2121         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2122                 cfqd->serving_prio = BE_WORKLOAD;
2123         else {
2124                 cfqd->serving_prio = IDLE_WORKLOAD;
2125                 cfqd->workload_expires = jiffies + 1;
2126                 return;
2127         }
2128
2129         if (original_prio != cfqd->serving_prio)
2130                 goto new_workload;
2131
2132         /*
2133          * For RT and BE, we have to choose also the type
2134          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2135          * expiration time
2136          */
2137         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2138         count = st->count;
2139
2140         /*
2141          * check workload expiration, and that we still have other queues ready
2142          */
2143         if (count && !time_after(jiffies, cfqd->workload_expires))
2144                 return;
2145
2146 new_workload:
2147         /* otherwise select new workload type */
2148         cfqd->serving_type =
2149                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2150         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2151         count = st->count;
2152
2153         /*
2154          * the workload slice is computed as a fraction of target latency
2155          * proportional to the number of queues in that workload, over
2156          * all the queues in the same priority class
2157          */
2158         group_slice = cfq_group_slice(cfqd, cfqg);
2159
2160         slice = group_slice * count /
2161                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2162                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2163
2164         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2165                 unsigned int tmp;
2166
2167                 /*
2168                  * Async queues are currently system wide. Just taking
2169                  * proportion of queues with-in same group will lead to higher
2170                  * async ratio system wide as generally root group is going
2171                  * to have higher weight. A more accurate thing would be to
2172                  * calculate system wide asnc/sync ratio.
2173                  */
2174                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2175                 tmp = tmp/cfqd->busy_queues;
2176                 slice = min_t(unsigned, slice, tmp);
2177
2178                 /* async workload slice is scaled down according to
2179                  * the sync/async slice ratio. */
2180                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2181         } else
2182                 /* sync workload slice is at least 2 * cfq_slice_idle */
2183                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2184
2185         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2186         cfq_log(cfqd, "workload slice:%d", slice);
2187         cfqd->workload_expires = jiffies + slice;
2188 }
2189
2190 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2191 {
2192         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2193         struct cfq_group *cfqg;
2194
2195         if (RB_EMPTY_ROOT(&st->rb))
2196                 return NULL;
2197         cfqg = cfq_rb_first_group(st);
2198         update_min_vdisktime(st);
2199         return cfqg;
2200 }
2201
2202 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2203 {
2204         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2205
2206         cfqd->serving_group = cfqg;
2207
2208         /* Restore the workload type data */
2209         if (cfqg->saved_workload_slice) {
2210                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2211                 cfqd->serving_type = cfqg->saved_workload;
2212                 cfqd->serving_prio = cfqg->saved_serving_prio;
2213         } else
2214                 cfqd->workload_expires = jiffies - 1;
2215
2216         choose_service_tree(cfqd, cfqg);
2217 }
2218
2219 /*
2220  * Select a queue for service. If we have a current active queue,
2221  * check whether to continue servicing it, or retrieve and set a new one.
2222  */
2223 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2224 {
2225         struct cfq_queue *cfqq, *new_cfqq = NULL;
2226
2227         cfqq = cfqd->active_queue;
2228         if (!cfqq)
2229                 goto new_queue;
2230
2231         if (!cfqd->rq_queued)
2232                 return NULL;
2233
2234         /*
2235          * We were waiting for group to get backlogged. Expire the queue
2236          */
2237         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2238                 goto expire;
2239
2240         /*
2241          * The active queue has run out of time, expire it and select new.
2242          */
2243         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2244                 /*
2245                  * If slice had not expired at the completion of last request
2246                  * we might not have turned on wait_busy flag. Don't expire
2247                  * the queue yet. Allow the group to get backlogged.
2248                  *
2249                  * The very fact that we have used the slice, that means we
2250                  * have been idling all along on this queue and it should be
2251                  * ok to wait for this request to complete.
2252                  */
2253                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2254                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2255                         cfqq = NULL;
2256                         goto keep_queue;
2257                 } else
2258                         goto check_group_idle;
2259         }
2260
2261         /*
2262          * The active queue has requests and isn't expired, allow it to
2263          * dispatch.
2264          */
2265         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2266                 goto keep_queue;
2267
2268         /*
2269          * If another queue has a request waiting within our mean seek
2270          * distance, let it run.  The expire code will check for close
2271          * cooperators and put the close queue at the front of the service
2272          * tree.  If possible, merge the expiring queue with the new cfqq.
2273          */
2274         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2275         if (new_cfqq) {
2276                 if (!cfqq->new_cfqq)
2277                         cfq_setup_merge(cfqq, new_cfqq);
2278                 goto expire;
2279         }
2280
2281         /*
2282          * No requests pending. If the active queue still has requests in
2283          * flight or is idling for a new request, allow either of these
2284          * conditions to happen (or time out) before selecting a new queue.
2285          */
2286         if (timer_pending(&cfqd->idle_slice_timer)) {
2287                 cfqq = NULL;
2288                 goto keep_queue;
2289         }
2290
2291         /*
2292          * This is a deep seek queue, but the device is much faster than
2293          * the queue can deliver, don't idle
2294          **/
2295         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2296             (cfq_cfqq_slice_new(cfqq) ||
2297             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2298                 cfq_clear_cfqq_deep(cfqq);
2299                 cfq_clear_cfqq_idle_window(cfqq);
2300         }
2301
2302         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2303                 cfqq = NULL;
2304                 goto keep_queue;
2305         }
2306
2307         /*
2308          * If group idle is enabled and there are requests dispatched from
2309          * this group, wait for requests to complete.
2310          */
2311 check_group_idle:
2312         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2313             && cfqq->cfqg->dispatched) {
2314                 cfqq = NULL;
2315                 goto keep_queue;
2316         }
2317
2318 expire:
2319         cfq_slice_expired(cfqd, 0);
2320 new_queue:
2321         /*
2322          * Current queue expired. Check if we have to switch to a new
2323          * service tree
2324          */
2325         if (!new_cfqq)
2326                 cfq_choose_cfqg(cfqd);
2327
2328         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2329 keep_queue:
2330         return cfqq;
2331 }
2332
2333 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2334 {
2335         int dispatched = 0;
2336
2337         while (cfqq->next_rq) {
2338                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2339                 dispatched++;
2340         }
2341
2342         BUG_ON(!list_empty(&cfqq->fifo));
2343
2344         /* By default cfqq is not expired if it is empty. Do it explicitly */
2345         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2346         return dispatched;
2347 }
2348
2349 /*
2350  * Drain our current requests. Used for barriers and when switching
2351  * io schedulers on-the-fly.
2352  */
2353 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2354 {
2355         struct cfq_queue *cfqq;
2356         int dispatched = 0;
2357
2358         /* Expire the timeslice of the current active queue first */
2359         cfq_slice_expired(cfqd, 0);
2360         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2361                 __cfq_set_active_queue(cfqd, cfqq);
2362                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2363         }
2364
2365         BUG_ON(cfqd->busy_queues);
2366
2367         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2368         return dispatched;
2369 }
2370
2371 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2372         struct cfq_queue *cfqq)
2373 {
2374         /* the queue hasn't finished any request, can't estimate */
2375         if (cfq_cfqq_slice_new(cfqq))
2376                 return true;
2377         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2378                 cfqq->slice_end))
2379                 return true;
2380
2381         return false;
2382 }
2383
2384 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2385 {
2386         unsigned int max_dispatch;
2387
2388         /*
2389          * Drain async requests before we start sync IO
2390          */
2391         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2392                 return false;
2393
2394         /*
2395          * If this is an async queue and we have sync IO in flight, let it wait
2396          */
2397         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2398                 return false;
2399
2400         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2401         if (cfq_class_idle(cfqq))
2402                 max_dispatch = 1;
2403
2404         /*
2405          * Does this cfqq already have too much IO in flight?
2406          */
2407         if (cfqq->dispatched >= max_dispatch) {
2408                 /*
2409                  * idle queue must always only have a single IO in flight
2410                  */
2411                 if (cfq_class_idle(cfqq))
2412                         return false;
2413
2414                 /*
2415                  * We have other queues, don't allow more IO from this one
2416                  */
2417                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2418                         return false;
2419
2420                 /*
2421                  * Sole queue user, no limit
2422                  */
2423                 if (cfqd->busy_queues == 1)
2424                         max_dispatch = -1;
2425                 else
2426                         /*
2427                          * Normally we start throttling cfqq when cfq_quantum/2
2428                          * requests have been dispatched. But we can drive
2429                          * deeper queue depths at the beginning of slice
2430                          * subjected to upper limit of cfq_quantum.
2431                          * */
2432                         max_dispatch = cfqd->cfq_quantum;
2433         }
2434
2435         /*
2436          * Async queues must wait a bit before being allowed dispatch.
2437          * We also ramp up the dispatch depth gradually for async IO,
2438          * based on the last sync IO we serviced
2439          */
2440         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2441                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2442                 unsigned int depth;
2443
2444                 depth = last_sync / cfqd->cfq_slice[1];
2445                 if (!depth && !cfqq->dispatched)
2446                         depth = 1;
2447                 if (depth < max_dispatch)
2448                         max_dispatch = depth;
2449         }
2450
2451         /*
2452          * If we're below the current max, allow a dispatch
2453          */
2454         return cfqq->dispatched < max_dispatch;
2455 }
2456
2457 /*
2458  * Dispatch a request from cfqq, moving them to the request queue
2459  * dispatch list.
2460  */
2461 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2462 {
2463         struct request *rq;
2464
2465         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2466
2467         if (!cfq_may_dispatch(cfqd, cfqq))
2468                 return false;
2469
2470         /*
2471          * follow expired path, else get first next available
2472          */
2473         rq = cfq_check_fifo(cfqq);
2474         if (!rq)
2475                 rq = cfqq->next_rq;
2476
2477         /*
2478          * insert request into driver dispatch list
2479          */
2480         cfq_dispatch_insert(cfqd->queue, rq);
2481
2482         if (!cfqd->active_cic) {
2483                 struct cfq_io_context *cic = RQ_CIC(rq);
2484
2485                 atomic_long_inc(&cic->ioc->refcount);
2486                 cfqd->active_cic = cic;
2487         }
2488
2489         return true;
2490 }
2491
2492 /*
2493  * Find the cfqq that we need to service and move a request from that to the
2494  * dispatch list
2495  */
2496 static int cfq_dispatch_requests(struct request_queue *q, int force)
2497 {
2498         struct cfq_data *cfqd = q->elevator->elevator_data;
2499         struct cfq_queue *cfqq;
2500
2501         if (!cfqd->busy_queues)
2502                 return 0;
2503
2504         if (unlikely(force))
2505                 return cfq_forced_dispatch(cfqd);
2506
2507         cfqq = cfq_select_queue(cfqd);
2508         if (!cfqq)
2509                 return 0;
2510
2511         /*
2512          * Dispatch a request from this cfqq, if it is allowed
2513          */
2514         if (!cfq_dispatch_request(cfqd, cfqq))
2515                 return 0;
2516
2517         cfqq->slice_dispatch++;
2518         cfq_clear_cfqq_must_dispatch(cfqq);
2519
2520         /*
2521          * expire an async queue immediately if it has used up its slice. idle
2522          * queue always expire after 1 dispatch round.
2523          */
2524         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2525             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2526             cfq_class_idle(cfqq))) {
2527                 cfqq->slice_end = jiffies + 1;
2528                 cfq_slice_expired(cfqd, 0);
2529         }
2530
2531         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2532         return 1;
2533 }
2534
2535 /*
2536  * task holds one reference to the queue, dropped when task exits. each rq
2537  * in-flight on this queue also holds a reference, dropped when rq is freed.
2538  *
2539  * Each cfq queue took a reference on the parent group. Drop it now.
2540  * queue lock must be held here.
2541  */
2542 static void cfq_put_queue(struct cfq_queue *cfqq)
2543 {
2544         struct cfq_data *cfqd = cfqq->cfqd;
2545         struct cfq_group *cfqg, *orig_cfqg;
2546
2547         BUG_ON(cfqq->ref <= 0);
2548
2549         cfqq->ref--;
2550         if (cfqq->ref)
2551                 return;
2552
2553         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2554         BUG_ON(rb_first(&cfqq->sort_list));
2555         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2556         cfqg = cfqq->cfqg;
2557         orig_cfqg = cfqq->orig_cfqg;
2558
2559         if (unlikely(cfqd->active_queue == cfqq)) {
2560                 __cfq_slice_expired(cfqd, cfqq, 0);
2561                 cfq_schedule_dispatch(cfqd);
2562         }
2563
2564         BUG_ON(cfq_cfqq_on_rr(cfqq));
2565         kmem_cache_free(cfq_pool, cfqq);
2566         cfq_put_cfqg(cfqg);
2567         if (orig_cfqg)
2568                 cfq_put_cfqg(orig_cfqg);
2569 }
2570
2571 /*
2572  * Must always be called with the rcu_read_lock() held
2573  */
2574 static void
2575 __call_for_each_cic(struct io_context *ioc,
2576                     void (*func)(struct io_context *, struct cfq_io_context *))
2577 {
2578         struct cfq_io_context *cic;
2579         struct hlist_node *n;
2580
2581         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2582                 func(ioc, cic);
2583 }
2584
2585 /*
2586  * Call func for each cic attached to this ioc.
2587  */
2588 static void
2589 call_for_each_cic(struct io_context *ioc,
2590                   void (*func)(struct io_context *, struct cfq_io_context *))
2591 {
2592         rcu_read_lock();
2593         __call_for_each_cic(ioc, func);
2594         rcu_read_unlock();
2595 }
2596
2597 static void cfq_cic_free_rcu(struct rcu_head *head)
2598 {
2599         struct cfq_io_context *cic;
2600
2601         cic = container_of(head, struct cfq_io_context, rcu_head);
2602
2603         kmem_cache_free(cfq_ioc_pool, cic);
2604         elv_ioc_count_dec(cfq_ioc_count);
2605
2606         if (ioc_gone) {
2607                 /*
2608                  * CFQ scheduler is exiting, grab exit lock and check
2609                  * the pending io context count. If it hits zero,
2610                  * complete ioc_gone and set it back to NULL
2611                  */
2612                 spin_lock(&ioc_gone_lock);
2613                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2614                         complete(ioc_gone);
2615                         ioc_gone = NULL;
2616                 }
2617                 spin_unlock(&ioc_gone_lock);
2618         }
2619 }
2620
2621 static void cfq_cic_free(struct cfq_io_context *cic)
2622 {
2623         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2624 }
2625
2626 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2627 {
2628         unsigned long flags;
2629         unsigned long dead_key = (unsigned long) cic->key;
2630
2631         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2632
2633         spin_lock_irqsave(&ioc->lock, flags);
2634         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2635         hlist_del_rcu(&cic->cic_list);
2636         spin_unlock_irqrestore(&ioc->lock, flags);
2637
2638         cfq_cic_free(cic);
2639 }
2640
2641 /*
2642  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2643  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2644  * and ->trim() which is called with the task lock held
2645  */
2646 static void cfq_free_io_context(struct io_context *ioc)
2647 {
2648         /*
2649          * ioc->refcount is zero here, or we are called from elv_unregister(),
2650          * so no more cic's are allowed to be linked into this ioc.  So it
2651          * should be ok to iterate over the known list, we will see all cic's
2652          * since no new ones are added.
2653          */
2654         __call_for_each_cic(ioc, cic_free_func);
2655 }
2656
2657 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2658 {
2659         struct cfq_queue *__cfqq, *next;
2660
2661         /*
2662          * If this queue was scheduled to merge with another queue, be
2663          * sure to drop the reference taken on that queue (and others in
2664          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2665          */
2666         __cfqq = cfqq->new_cfqq;
2667         while (__cfqq) {
2668                 if (__cfqq == cfqq) {
2669                         WARN(1, "cfqq->new_cfqq loop detected\n");
2670                         break;
2671                 }
2672                 next = __cfqq->new_cfqq;
2673                 cfq_put_queue(__cfqq);
2674                 __cfqq = next;
2675         }
2676 }
2677
2678 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2679 {
2680         if (unlikely(cfqq == cfqd->active_queue)) {
2681                 __cfq_slice_expired(cfqd, cfqq, 0);
2682                 cfq_schedule_dispatch(cfqd);
2683         }
2684
2685         cfq_put_cooperator(cfqq);
2686
2687         cfq_put_queue(cfqq);
2688 }
2689
2690 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2691                                          struct cfq_io_context *cic)
2692 {
2693         struct io_context *ioc = cic->ioc;
2694
2695         list_del_init(&cic->queue_list);
2696
2697         /*
2698          * Make sure dead mark is seen for dead queues
2699          */
2700         smp_wmb();
2701         cic->key = cfqd_dead_key(cfqd);
2702
2703         if (ioc->ioc_data == cic)
2704                 rcu_assign_pointer(ioc->ioc_data, NULL);
2705
2706         if (cic->cfqq[BLK_RW_ASYNC]) {
2707                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2708                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2709         }
2710
2711         if (cic->cfqq[BLK_RW_SYNC]) {
2712                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2713                 cic->cfqq[BLK_RW_SYNC] = NULL;
2714         }
2715 }
2716
2717 static void cfq_exit_single_io_context(struct io_context *ioc,
2718                                        struct cfq_io_context *cic)
2719 {
2720         struct cfq_data *cfqd = cic_to_cfqd(cic);
2721
2722         if (cfqd) {
2723                 struct request_queue *q = cfqd->queue;
2724                 unsigned long flags;
2725
2726                 spin_lock_irqsave(q->queue_lock, flags);
2727
2728                 /*
2729                  * Ensure we get a fresh copy of the ->key to prevent
2730                  * race between exiting task and queue
2731                  */
2732                 smp_read_barrier_depends();
2733                 if (cic->key == cfqd)
2734                         __cfq_exit_single_io_context(cfqd, cic);
2735
2736                 spin_unlock_irqrestore(q->queue_lock, flags);
2737         }
2738 }
2739
2740 /*
2741  * The process that ioc belongs to has exited, we need to clean up
2742  * and put the internal structures we have that belongs to that process.
2743  */
2744 static void cfq_exit_io_context(struct io_context *ioc)
2745 {
2746         call_for_each_cic(ioc, cfq_exit_single_io_context);
2747 }
2748
2749 static struct cfq_io_context *
2750 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2751 {
2752         struct cfq_io_context *cic;
2753
2754         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2755                                                         cfqd->queue->node);
2756         if (cic) {
2757                 cic->last_end_request = jiffies;
2758                 INIT_LIST_HEAD(&cic->queue_list);
2759                 INIT_HLIST_NODE(&cic->cic_list);
2760                 cic->dtor = cfq_free_io_context;
2761                 cic->exit = cfq_exit_io_context;
2762                 elv_ioc_count_inc(cfq_ioc_count);
2763         }
2764
2765         return cic;
2766 }
2767
2768 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2769 {
2770         struct task_struct *tsk = current;
2771         int ioprio_class;
2772
2773         if (!cfq_cfqq_prio_changed(cfqq))
2774                 return;
2775
2776         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2777         switch (ioprio_class) {
2778         default:
2779                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2780         case IOPRIO_CLASS_NONE:
2781                 /*
2782                  * no prio set, inherit CPU scheduling settings
2783                  */
2784                 cfqq->ioprio = task_nice_ioprio(tsk);
2785                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2786                 break;
2787         case IOPRIO_CLASS_RT:
2788                 cfqq->ioprio = task_ioprio(ioc);
2789                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2790                 break;
2791         case IOPRIO_CLASS_BE:
2792                 cfqq->ioprio = task_ioprio(ioc);
2793                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2794                 break;
2795         case IOPRIO_CLASS_IDLE:
2796                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2797                 cfqq->ioprio = 7;
2798                 cfq_clear_cfqq_idle_window(cfqq);
2799                 break;
2800         }
2801
2802         /*
2803          * keep track of original prio settings in case we have to temporarily
2804          * elevate the priority of this queue
2805          */
2806         cfqq->org_ioprio = cfqq->ioprio;
2807         cfqq->org_ioprio_class = cfqq->ioprio_class;
2808         cfq_clear_cfqq_prio_changed(cfqq);
2809 }
2810
2811 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2812 {
2813         struct cfq_data *cfqd = cic_to_cfqd(cic);
2814         struct cfq_queue *cfqq;
2815         unsigned long flags;
2816
2817         if (unlikely(!cfqd))
2818                 return;
2819
2820         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2821
2822         cfqq = cic->cfqq[BLK_RW_ASYNC];
2823         if (cfqq) {
2824                 struct cfq_queue *new_cfqq;
2825                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2826                                                 GFP_ATOMIC);
2827                 if (new_cfqq) {
2828                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2829                         cfq_put_queue(cfqq);
2830                 }
2831         }
2832
2833         cfqq = cic->cfqq[BLK_RW_SYNC];
2834         if (cfqq)
2835                 cfq_mark_cfqq_prio_changed(cfqq);
2836
2837         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2838 }
2839
2840 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2841 {
2842         call_for_each_cic(ioc, changed_ioprio);
2843         ioc->ioprio_changed = 0;
2844 }
2845
2846 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2847                           pid_t pid, bool is_sync)
2848 {
2849         RB_CLEAR_NODE(&cfqq->rb_node);
2850         RB_CLEAR_NODE(&cfqq->p_node);
2851         INIT_LIST_HEAD(&cfqq->fifo);
2852
2853         cfqq->ref = 0;
2854         cfqq->cfqd = cfqd;
2855
2856         cfq_mark_cfqq_prio_changed(cfqq);
2857
2858         if (is_sync) {
2859                 if (!cfq_class_idle(cfqq))
2860                         cfq_mark_cfqq_idle_window(cfqq);
2861                 cfq_mark_cfqq_sync(cfqq);
2862         }
2863         cfqq->pid = pid;
2864 }
2865
2866 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2867 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2868 {
2869         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2870         struct cfq_data *cfqd = cic_to_cfqd(cic);
2871         unsigned long flags;
2872         struct request_queue *q;
2873
2874         if (unlikely(!cfqd))
2875                 return;
2876
2877         q = cfqd->queue;
2878
2879         spin_lock_irqsave(q->queue_lock, flags);
2880
2881         if (sync_cfqq) {
2882                 /*
2883                  * Drop reference to sync queue. A new sync queue will be
2884                  * assigned in new group upon arrival of a fresh request.
2885                  */
2886                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2887                 cic_set_cfqq(cic, NULL, 1);
2888                 cfq_put_queue(sync_cfqq);
2889         }
2890
2891         spin_unlock_irqrestore(q->queue_lock, flags);
2892 }
2893
2894 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2895 {
2896         call_for_each_cic(ioc, changed_cgroup);
2897         ioc->cgroup_changed = 0;
2898 }
2899 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2900
2901 static struct cfq_queue *
2902 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2903                      struct io_context *ioc, gfp_t gfp_mask)
2904 {
2905         struct cfq_queue *cfqq, *new_cfqq = NULL;
2906         struct cfq_io_context *cic;
2907         struct cfq_group *cfqg;
2908
2909 retry:
2910         cfqg = cfq_get_cfqg(cfqd, 1);
2911         cic = cfq_cic_lookup(cfqd, ioc);
2912         /* cic always exists here */
2913         cfqq = cic_to_cfqq(cic, is_sync);
2914
2915         /*
2916          * Always try a new alloc if we fell back to the OOM cfqq
2917          * originally, since it should just be a temporary situation.
2918          */
2919         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2920                 cfqq = NULL;
2921                 if (new_cfqq) {
2922                         cfqq = new_cfqq;
2923                         new_cfqq = NULL;
2924                 } else if (gfp_mask & __GFP_WAIT) {
2925                         spin_unlock_irq(cfqd->queue->queue_lock);
2926                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2927                                         gfp_mask | __GFP_ZERO,
2928                                         cfqd->queue->node);
2929                         spin_lock_irq(cfqd->queue->queue_lock);
2930                         if (new_cfqq)
2931                                 goto retry;
2932                 } else {
2933                         cfqq = kmem_cache_alloc_node(cfq_pool,
2934                                         gfp_mask | __GFP_ZERO,
2935                                         cfqd->queue->node);
2936                 }
2937
2938                 if (cfqq) {
2939                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2940                         cfq_init_prio_data(cfqq, ioc);
2941                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2942                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2943                 } else
2944                         cfqq = &cfqd->oom_cfqq;
2945         }
2946
2947         if (new_cfqq)
2948                 kmem_cache_free(cfq_pool, new_cfqq);
2949
2950         return cfqq;
2951 }
2952
2953 static struct cfq_queue **
2954 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2955 {
2956         switch (ioprio_class) {
2957         case IOPRIO_CLASS_RT:
2958                 return &cfqd->async_cfqq[0][ioprio];
2959         case IOPRIO_CLASS_BE:
2960                 return &cfqd->async_cfqq[1][ioprio];
2961         case IOPRIO_CLASS_IDLE:
2962                 return &cfqd->async_idle_cfqq;
2963         default:
2964                 BUG();
2965         }
2966 }
2967
2968 static struct cfq_queue *
2969 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2970               gfp_t gfp_mask)
2971 {
2972         const int ioprio = task_ioprio(ioc);
2973         const int ioprio_class = task_ioprio_class(ioc);
2974         struct cfq_queue **async_cfqq = NULL;
2975         struct cfq_queue *cfqq = NULL;
2976
2977         if (!is_sync) {
2978                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2979                 cfqq = *async_cfqq;
2980         }
2981
2982         if (!cfqq)
2983                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2984
2985         /*
2986          * pin the queue now that it's allocated, scheduler exit will prune it
2987          */
2988         if (!is_sync && !(*async_cfqq)) {
2989                 cfqq->ref++;
2990                 *async_cfqq = cfqq;
2991         }
2992
2993         cfqq->ref++;
2994         return cfqq;
2995 }
2996
2997 /*
2998  * We drop cfq io contexts lazily, so we may find a dead one.
2999  */
3000 static void
3001 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3002                   struct cfq_io_context *cic)
3003 {
3004         unsigned long flags;
3005
3006         WARN_ON(!list_empty(&cic->queue_list));
3007         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3008
3009         spin_lock_irqsave(&ioc->lock, flags);
3010
3011         BUG_ON(ioc->ioc_data == cic);
3012
3013         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3014         hlist_del_rcu(&cic->cic_list);
3015         spin_unlock_irqrestore(&ioc->lock, flags);
3016
3017         cfq_cic_free(cic);
3018 }
3019
3020 static struct cfq_io_context *
3021 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3022 {
3023         struct cfq_io_context *cic;
3024         unsigned long flags;
3025
3026         if (unlikely(!ioc))
3027                 return NULL;
3028
3029         rcu_read_lock();
3030
3031         /*
3032          * we maintain a last-hit cache, to avoid browsing over the tree
3033          */
3034         cic = rcu_dereference(ioc->ioc_data);
3035         if (cic && cic->key == cfqd) {
3036                 rcu_read_unlock();
3037                 return cic;
3038         }
3039
3040         do {
3041                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3042                 rcu_read_unlock();
3043                 if (!cic)
3044                         break;
3045                 if (unlikely(cic->key != cfqd)) {
3046                         cfq_drop_dead_cic(cfqd, ioc, cic);
3047                         rcu_read_lock();
3048                         continue;
3049                 }
3050
3051                 spin_lock_irqsave(&ioc->lock, flags);
3052                 rcu_assign_pointer(ioc->ioc_data, cic);
3053                 spin_unlock_irqrestore(&ioc->lock, flags);
3054                 break;
3055         } while (1);
3056
3057         return cic;
3058 }
3059
3060 /*
3061  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3062  * the process specific cfq io context when entered from the block layer.
3063  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3064  */
3065 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3066                         struct cfq_io_context *cic, gfp_t gfp_mask)
3067 {
3068         unsigned long flags;
3069         int ret;
3070
3071         ret = radix_tree_preload(gfp_mask);
3072         if (!ret) {
3073                 cic->ioc = ioc;
3074                 cic->key = cfqd;
3075
3076                 spin_lock_irqsave(&ioc->lock, flags);
3077                 ret = radix_tree_insert(&ioc->radix_root,
3078                                                 cfqd->cic_index, cic);
3079                 if (!ret)
3080                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3081                 spin_unlock_irqrestore(&ioc->lock, flags);
3082
3083                 radix_tree_preload_end();
3084
3085                 if (!ret) {
3086                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3087                         list_add(&cic->queue_list, &cfqd->cic_list);
3088                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3089                 }
3090         }
3091
3092         if (ret)
3093                 printk(KERN_ERR "cfq: cic link failed!\n");
3094
3095         return ret;
3096 }
3097
3098 /*
3099  * Setup general io context and cfq io context. There can be several cfq
3100  * io contexts per general io context, if this process is doing io to more
3101  * than one device managed by cfq.
3102  */
3103 static struct cfq_io_context *
3104 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3105 {
3106         struct io_context *ioc = NULL;
3107         struct cfq_io_context *cic;
3108
3109         might_sleep_if(gfp_mask & __GFP_WAIT);
3110
3111         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3112         if (!ioc)
3113                 return NULL;
3114
3115         cic = cfq_cic_lookup(cfqd, ioc);
3116         if (cic)
3117                 goto out;
3118
3119         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3120         if (cic == NULL)
3121                 goto err;
3122
3123         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3124                 goto err_free;
3125
3126 out:
3127         smp_read_barrier_depends();
3128         if (unlikely(ioc->ioprio_changed))
3129                 cfq_ioc_set_ioprio(ioc);
3130
3131 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3132         if (unlikely(ioc->cgroup_changed))
3133                 cfq_ioc_set_cgroup(ioc);
3134 #endif
3135         return cic;
3136 err_free:
3137         cfq_cic_free(cic);
3138 err:
3139         put_io_context(ioc);
3140         return NULL;
3141 }
3142
3143 static void
3144 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3145 {
3146         unsigned long elapsed = jiffies - cic->last_end_request;
3147         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3148
3149         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3150         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3151         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3152 }
3153
3154 static void
3155 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3156                        struct request *rq)
3157 {
3158         sector_t sdist = 0;
3159         sector_t n_sec = blk_rq_sectors(rq);
3160         if (cfqq->last_request_pos) {
3161                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3162                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3163                 else
3164                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3165         }
3166
3167         cfqq->seek_history <<= 1;
3168         if (blk_queue_nonrot(cfqd->queue))
3169                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3170         else
3171                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3172 }
3173
3174 /*
3175  * Disable idle window if the process thinks too long or seeks so much that
3176  * it doesn't matter
3177  */
3178 static void
3179 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3180                        struct cfq_io_context *cic)
3181 {
3182         int old_idle, enable_idle;
3183
3184         /*
3185          * Don't idle for async or idle io prio class
3186          */
3187         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3188                 return;
3189
3190         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3191
3192         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3193                 cfq_mark_cfqq_deep(cfqq);
3194
3195         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3196                 enable_idle = 0;
3197         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3198             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3199                 enable_idle = 0;
3200         else if (sample_valid(cic->ttime_samples)) {
3201                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3202                         enable_idle = 0;
3203                 else
3204                         enable_idle = 1;
3205         }
3206
3207         if (old_idle != enable_idle) {
3208                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3209                 if (enable_idle)
3210                         cfq_mark_cfqq_idle_window(cfqq);
3211                 else
3212                         cfq_clear_cfqq_idle_window(cfqq);
3213         }
3214 }
3215
3216 /*
3217  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3218  * no or if we aren't sure, a 1 will cause a preempt.
3219  */
3220 static bool
3221 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3222                    struct request *rq)
3223 {
3224         struct cfq_queue *cfqq;
3225
3226         cfqq = cfqd->active_queue;
3227         if (!cfqq)
3228                 return false;
3229
3230         if (cfq_class_idle(new_cfqq))
3231                 return false;
3232
3233         if (cfq_class_idle(cfqq))
3234                 return true;
3235
3236         /*
3237          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3238          */
3239         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3240                 return false;
3241
3242         /*
3243          * if the new request is sync, but the currently running queue is
3244          * not, let the sync request have priority.
3245          */
3246         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3247                 return true;
3248
3249         if (new_cfqq->cfqg != cfqq->cfqg)
3250                 return false;
3251
3252         if (cfq_slice_used(cfqq))
3253                 return true;
3254
3255         /* Allow preemption only if we are idling on sync-noidle tree */
3256         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3257             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3258             new_cfqq->service_tree->count == 2 &&
3259             RB_EMPTY_ROOT(&cfqq->sort_list))
3260                 return true;
3261
3262         /*
3263          * So both queues are sync. Let the new request get disk time if
3264          * it's a metadata request and the current queue is doing regular IO.
3265          */
3266         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3267                 return true;
3268
3269         /*
3270          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3271          */
3272         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3273                 return true;
3274
3275         /* An idle queue should not be idle now for some reason */
3276         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3277                 return true;
3278
3279         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3280                 return false;
3281
3282         /*
3283          * if this request is as-good as one we would expect from the
3284          * current cfqq, let it preempt
3285          */
3286         if (cfq_rq_close(cfqd, cfqq, rq))
3287                 return true;
3288
3289         return false;
3290 }
3291
3292 /*
3293  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3294  * let it have half of its nominal slice.
3295  */
3296 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3297 {
3298         struct cfq_queue *old_cfqq = cfqd->active_queue;
3299
3300         cfq_log_cfqq(cfqd, cfqq, "preempt");
3301         cfq_slice_expired(cfqd, 1);
3302
3303         /*
3304          * workload type is changed, don't save slice, otherwise preempt
3305          * doesn't happen
3306          */
3307         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3308                 cfqq->cfqg->saved_workload_slice = 0;
3309
3310         /*
3311          * Put the new queue at the front of the of the current list,
3312          * so we know that it will be selected next.
3313          */
3314         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3315
3316         cfq_service_tree_add(cfqd, cfqq, 1);
3317
3318         cfqq->slice_end = 0;
3319         cfq_mark_cfqq_slice_new(cfqq);
3320 }
3321
3322 /*
3323  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3324  * something we should do about it
3325  */
3326 static void
3327 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3328                 struct request *rq)
3329 {
3330         struct cfq_io_context *cic = RQ_CIC(rq);
3331
3332         cfqd->rq_queued++;
3333         if (rq->cmd_flags & REQ_META)
3334                 cfqq->meta_pending++;
3335
3336         cfq_update_io_thinktime(cfqd, cic);
3337         cfq_update_io_seektime(cfqd, cfqq, rq);
3338         cfq_update_idle_window(cfqd, cfqq, cic);
3339
3340         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3341
3342         if (cfqq == cfqd->active_queue) {
3343                 /*
3344                  * Remember that we saw a request from this process, but
3345                  * don't start queuing just yet. Otherwise we risk seeing lots
3346                  * of tiny requests, because we disrupt the normal plugging
3347                  * and merging. If the request is already larger than a single
3348                  * page, let it rip immediately. For that case we assume that
3349                  * merging is already done. Ditto for a busy system that
3350                  * has other work pending, don't risk delaying until the
3351                  * idle timer unplug to continue working.
3352                  */
3353                 if (cfq_cfqq_wait_request(cfqq)) {
3354                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3355                             cfqd->busy_queues > 1) {
3356                                 cfq_del_timer(cfqd, cfqq);
3357                                 cfq_clear_cfqq_wait_request(cfqq);
3358                                 __blk_run_queue(cfqd->queue, false);
3359                         } else {
3360                                 cfq_blkiocg_update_idle_time_stats(
3361                                                 &cfqq->cfqg->blkg);
3362                                 cfq_mark_cfqq_must_dispatch(cfqq);
3363                         }
3364                 }
3365         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3366                 /*
3367                  * not the active queue - expire current slice if it is
3368                  * idle and has expired it's mean thinktime or this new queue
3369                  * has some old slice time left and is of higher priority or
3370                  * this new queue is RT and the current one is BE
3371                  */
3372                 cfq_preempt_queue(cfqd, cfqq);
3373                 __blk_run_queue(cfqd->queue, false);
3374         }
3375 }
3376
3377 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3378 {
3379         struct cfq_data *cfqd = q->elevator->elevator_data;
3380         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3381
3382         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3383         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3384
3385         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3386         list_add_tail(&rq->queuelist, &cfqq->fifo);
3387         cfq_add_rq_rb(rq);
3388         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3389                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3390                         rq_is_sync(rq));
3391         cfq_rq_enqueued(cfqd, cfqq, rq);
3392 }
3393
3394 /*
3395  * Update hw_tag based on peak queue depth over 50 samples under
3396  * sufficient load.
3397  */
3398 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3399 {
3400         struct cfq_queue *cfqq = cfqd->active_queue;
3401
3402         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3403                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3404
3405         if (cfqd->hw_tag == 1)
3406                 return;
3407
3408         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3409             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3410                 return;
3411
3412         /*
3413          * If active queue hasn't enough requests and can idle, cfq might not
3414          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3415          * case
3416          */
3417         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3418             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3419             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3420                 return;
3421
3422         if (cfqd->hw_tag_samples++ < 50)
3423                 return;
3424
3425         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3426                 cfqd->hw_tag = 1;
3427         else
3428                 cfqd->hw_tag = 0;
3429 }
3430
3431 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3432 {
3433         struct cfq_io_context *cic = cfqd->active_cic;
3434
3435         /* If the queue already has requests, don't wait */
3436         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3437                 return false;
3438
3439         /* If there are other queues in the group, don't wait */
3440         if (cfqq->cfqg->nr_cfqq > 1)
3441                 return false;
3442
3443         if (cfq_slice_used(cfqq))
3444                 return true;
3445
3446         /* if slice left is less than think time, wait busy */
3447         if (cic && sample_valid(cic->ttime_samples)
3448             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3449                 return true;
3450
3451         /*
3452          * If think times is less than a jiffy than ttime_mean=0 and above
3453          * will not be true. It might happen that slice has not expired yet
3454          * but will expire soon (4-5 ns) during select_queue(). To cover the
3455          * case where think time is less than a jiffy, mark the queue wait
3456          * busy if only 1 jiffy is left in the slice.
3457          */
3458         if (cfqq->slice_end - jiffies == 1)
3459                 return true;
3460
3461         return false;
3462 }
3463
3464 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3465 {
3466         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3467         struct cfq_data *cfqd = cfqq->cfqd;
3468         const int sync = rq_is_sync(rq);
3469         unsigned long now;
3470
3471         now = jiffies;
3472         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3473                      !!(rq->cmd_flags & REQ_NOIDLE));
3474
3475         cfq_update_hw_tag(cfqd);
3476
3477         WARN_ON(!cfqd->rq_in_driver);
3478         WARN_ON(!cfqq->dispatched);
3479         cfqd->rq_in_driver--;
3480         cfqq->dispatched--;
3481         (RQ_CFQG(rq))->dispatched--;
3482         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3483                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3484                         rq_data_dir(rq), rq_is_sync(rq));
3485
3486         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3487
3488         if (sync) {
3489                 RQ_CIC(rq)->last_end_request = now;
3490                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3491                         cfqd->last_delayed_sync = now;
3492         }
3493
3494         /*
3495          * If this is the active queue, check if it needs to be expired,
3496          * or if we want to idle in case it has no pending requests.
3497          */
3498         if (cfqd->active_queue == cfqq) {
3499                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3500
3501                 if (cfq_cfqq_slice_new(cfqq)) {
3502                         cfq_set_prio_slice(cfqd, cfqq);
3503                         cfq_clear_cfqq_slice_new(cfqq);
3504                 }
3505
3506                 /*
3507                  * Should we wait for next request to come in before we expire
3508                  * the queue.
3509                  */
3510                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3511                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3512                         if (!cfqd->cfq_slice_idle)
3513                                 extend_sl = cfqd->cfq_group_idle;
3514                         cfqq->slice_end = jiffies + extend_sl;
3515                         cfq_mark_cfqq_wait_busy(cfqq);
3516                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3517                 }
3518
3519                 /*
3520                  * Idling is not enabled on:
3521                  * - expired queues
3522                  * - idle-priority queues
3523                  * - async queues
3524                  * - queues with still some requests queued
3525                  * - when there is a close cooperator
3526                  */
3527                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3528                         cfq_slice_expired(cfqd, 1);
3529                 else if (sync && cfqq_empty &&
3530                          !cfq_close_cooperator(cfqd, cfqq)) {
3531                         cfq_arm_slice_timer(cfqd);
3532                 }
3533         }
3534
3535         if (!cfqd->rq_in_driver)
3536                 cfq_schedule_dispatch(cfqd);
3537 }
3538
3539 /*
3540  * we temporarily boost lower priority queues if they are holding fs exclusive
3541  * resources. they are boosted to normal prio (CLASS_BE/4)
3542  */
3543 static void cfq_prio_boost(struct cfq_queue *cfqq)
3544 {
3545         if (has_fs_excl()) {
3546                 /*
3547                  * boost idle prio on transactions that would lock out other
3548                  * users of the filesystem
3549                  */
3550                 if (cfq_class_idle(cfqq))
3551                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3552                 if (cfqq->ioprio > IOPRIO_NORM)
3553                         cfqq->ioprio = IOPRIO_NORM;
3554         } else {
3555                 /*
3556                  * unboost the queue (if needed)
3557                  */
3558                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3559                 cfqq->ioprio = cfqq->org_ioprio;
3560         }
3561 }
3562
3563 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3564 {
3565         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3566                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3567                 return ELV_MQUEUE_MUST;
3568         }
3569
3570         return ELV_MQUEUE_MAY;
3571 }
3572
3573 static int cfq_may_queue(struct request_queue *q, int rw)
3574 {
3575         struct cfq_data *cfqd = q->elevator->elevator_data;
3576         struct task_struct *tsk = current;
3577         struct cfq_io_context *cic;
3578         struct cfq_queue *cfqq;
3579
3580         /*
3581          * don't force setup of a queue from here, as a call to may_queue
3582          * does not necessarily imply that a request actually will be queued.
3583          * so just lookup a possibly existing queue, or return 'may queue'
3584          * if that fails
3585          */
3586         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3587         if (!cic)
3588                 return ELV_MQUEUE_MAY;
3589
3590         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3591         if (cfqq) {
3592                 cfq_init_prio_data(cfqq, cic->ioc);
3593                 cfq_prio_boost(cfqq);
3594
3595                 return __cfq_may_queue(cfqq);
3596         }
3597
3598         return ELV_MQUEUE_MAY;
3599 }
3600
3601 /*
3602  * queue lock held here
3603  */
3604 static void cfq_put_request(struct request *rq)
3605 {
3606         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3607
3608         if (cfqq) {
3609                 const int rw = rq_data_dir(rq);
3610
3611                 BUG_ON(!cfqq->allocated[rw]);
3612                 cfqq->allocated[rw]--;
3613
3614                 put_io_context(RQ_CIC(rq)->ioc);
3615
3616                 rq->elevator_private = NULL;
3617                 rq->elevator_private2 = NULL;
3618
3619                 /* Put down rq reference on cfqg */
3620                 cfq_put_cfqg(RQ_CFQG(rq));
3621                 rq->elevator_private3 = NULL;
3622
3623                 cfq_put_queue(cfqq);
3624         }
3625 }
3626
3627 static struct cfq_queue *
3628 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3629                 struct cfq_queue *cfqq)
3630 {
3631         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3632         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3633         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3634         cfq_put_queue(cfqq);
3635         return cic_to_cfqq(cic, 1);
3636 }
3637
3638 /*
3639  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3640  * was the last process referring to said cfqq.
3641  */
3642 static struct cfq_queue *
3643 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3644 {
3645         if (cfqq_process_refs(cfqq) == 1) {
3646                 cfqq->pid = current->pid;
3647                 cfq_clear_cfqq_coop(cfqq);
3648                 cfq_clear_cfqq_split_coop(cfqq);
3649                 return cfqq;
3650         }
3651
3652         cic_set_cfqq(cic, NULL, 1);
3653
3654         cfq_put_cooperator(cfqq);
3655
3656         cfq_put_queue(cfqq);
3657         return NULL;
3658 }
3659 /*
3660  * Allocate cfq data structures associated with this request.
3661  */
3662 static int
3663 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3664 {
3665         struct cfq_data *cfqd = q->elevator->elevator_data;
3666         struct cfq_io_context *cic;
3667         const int rw = rq_data_dir(rq);
3668         const bool is_sync = rq_is_sync(rq);
3669         struct cfq_queue *cfqq;
3670         unsigned long flags;
3671
3672         might_sleep_if(gfp_mask & __GFP_WAIT);
3673
3674         cic = cfq_get_io_context(cfqd, gfp_mask);
3675
3676         spin_lock_irqsave(q->queue_lock, flags);
3677
3678         if (!cic)
3679                 goto queue_fail;
3680
3681 new_queue:
3682         cfqq = cic_to_cfqq(cic, is_sync);
3683         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3684                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3685                 cic_set_cfqq(cic, cfqq, is_sync);
3686         } else {
3687                 /*
3688                  * If the queue was seeky for too long, break it apart.
3689                  */
3690                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3691                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3692                         cfqq = split_cfqq(cic, cfqq);
3693                         if (!cfqq)
3694                                 goto new_queue;
3695                 }
3696
3697                 /*
3698                  * Check to see if this queue is scheduled to merge with
3699                  * another, closely cooperating queue.  The merging of
3700                  * queues happens here as it must be done in process context.
3701                  * The reference on new_cfqq was taken in merge_cfqqs.
3702                  */
3703                 if (cfqq->new_cfqq)
3704                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3705         }
3706
3707         cfqq->allocated[rw]++;
3708         cfqq->ref++;
3709         rq->elevator_private = cic;
3710         rq->elevator_private2 = cfqq;
3711         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3712
3713         spin_unlock_irqrestore(q->queue_lock, flags);
3714
3715         return 0;
3716
3717 queue_fail:
3718         if (cic)
3719                 put_io_context(cic->ioc);
3720
3721         cfq_schedule_dispatch(cfqd);
3722         spin_unlock_irqrestore(q->queue_lock, flags);
3723         cfq_log(cfqd, "set_request fail");
3724         return 1;
3725 }
3726
3727 static void cfq_kick_queue(struct work_struct *work)
3728 {
3729         struct cfq_data *cfqd =
3730                 container_of(work, struct cfq_data, unplug_work);
3731         struct request_queue *q = cfqd->queue;
3732
3733         spin_lock_irq(q->queue_lock);
3734         __blk_run_queue(cfqd->queue, false);
3735         spin_unlock_irq(q->queue_lock);
3736 }
3737
3738 /*
3739  * Timer running if the active_queue is currently idling inside its time slice
3740  */
3741 static void cfq_idle_slice_timer(unsigned long data)
3742 {
3743         struct cfq_data *cfqd = (struct cfq_data *) data;
3744         struct cfq_queue *cfqq;
3745         unsigned long flags;
3746         int timed_out = 1;
3747
3748         cfq_log(cfqd, "idle timer fired");
3749
3750         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3751
3752         cfqq = cfqd->active_queue;
3753         if (cfqq) {
3754                 timed_out = 0;
3755
3756                 /*
3757                  * We saw a request before the queue expired, let it through
3758                  */
3759                 if (cfq_cfqq_must_dispatch(cfqq))
3760                         goto out_kick;
3761
3762                 /*
3763                  * expired
3764                  */
3765                 if (cfq_slice_used(cfqq))
3766                         goto expire;
3767
3768                 /*
3769                  * only expire and reinvoke request handler, if there are
3770                  * other queues with pending requests
3771                  */
3772                 if (!cfqd->busy_queues)
3773                         goto out_cont;
3774
3775                 /*
3776                  * not expired and it has a request pending, let it dispatch
3777                  */
3778                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3779                         goto out_kick;
3780
3781                 /*
3782                  * Queue depth flag is reset only when the idle didn't succeed
3783                  */
3784                 cfq_clear_cfqq_deep(cfqq);
3785         }
3786 expire:
3787         cfq_slice_expired(cfqd, timed_out);
3788 out_kick:
3789         cfq_schedule_dispatch(cfqd);
3790 out_cont:
3791         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3792 }
3793
3794 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3795 {
3796         del_timer_sync(&cfqd->idle_slice_timer);
3797         cancel_work_sync(&cfqd->unplug_work);
3798 }
3799
3800 static void cfq_put_async_queues(struct cfq_data *cfqd)
3801 {
3802         int i;
3803
3804         for (i = 0; i < IOPRIO_BE_NR; i++) {
3805                 if (cfqd->async_cfqq[0][i])
3806                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3807                 if (cfqd->async_cfqq[1][i])
3808                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3809         }
3810
3811         if (cfqd->async_idle_cfqq)
3812                 cfq_put_queue(cfqd->async_idle_cfqq);
3813 }
3814
3815 static void cfq_cfqd_free(struct rcu_head *head)
3816 {
3817         kfree(container_of(head, struct cfq_data, rcu));
3818 }
3819
3820 static void cfq_exit_queue(struct elevator_queue *e)
3821 {
3822         struct cfq_data *cfqd = e->elevator_data;
3823         struct request_queue *q = cfqd->queue;
3824
3825         cfq_shutdown_timer_wq(cfqd);
3826
3827         spin_lock_irq(q->queue_lock);
3828
3829         if (cfqd->active_queue)
3830                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3831
3832         while (!list_empty(&cfqd->cic_list)) {
3833                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3834                                                         struct cfq_io_context,
3835                                                         queue_list);
3836
3837                 __cfq_exit_single_io_context(cfqd, cic);
3838         }
3839
3840         cfq_put_async_queues(cfqd);
3841         cfq_release_cfq_groups(cfqd);
3842         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3843
3844         spin_unlock_irq(q->queue_lock);
3845
3846         cfq_shutdown_timer_wq(cfqd);
3847
3848         spin_lock(&cic_index_lock);
3849         ida_remove(&cic_index_ida, cfqd->cic_index);
3850         spin_unlock(&cic_index_lock);
3851
3852         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3853         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3854 }
3855
3856 static int cfq_alloc_cic_index(void)
3857 {
3858         int index, error;
3859
3860         do {
3861                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3862                         return -ENOMEM;
3863
3864                 spin_lock(&cic_index_lock);
3865                 error = ida_get_new(&cic_index_ida, &index);
3866                 spin_unlock(&cic_index_lock);
3867                 if (error && error != -EAGAIN)
3868                         return error;
3869         } while (error);
3870
3871         return index;
3872 }
3873
3874 static void *cfq_init_queue(struct request_queue *q)
3875 {
3876         struct cfq_data *cfqd;
3877         int i, j;
3878         struct cfq_group *cfqg;
3879         struct cfq_rb_root *st;
3880
3881         i = cfq_alloc_cic_index();
3882         if (i < 0)
3883                 return NULL;
3884
3885         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3886         if (!cfqd)
3887                 return NULL;
3888
3889         /*
3890          * Don't need take queue_lock in the routine, since we are
3891          * initializing the ioscheduler, and nobody is using cfqd
3892          */
3893         cfqd->cic_index = i;
3894
3895         /* Init root service tree */
3896         cfqd->grp_service_tree = CFQ_RB_ROOT;
3897
3898         /* Init root group */
3899         cfqg = &cfqd->root_group;
3900         for_each_cfqg_st(cfqg, i, j, st)
3901                 *st = CFQ_RB_ROOT;
3902         RB_CLEAR_NODE(&cfqg->rb_node);
3903
3904         /* Give preference to root group over other groups */
3905         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3906
3907 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3908         /*
3909          * Take a reference to root group which we never drop. This is just
3910          * to make sure that cfq_put_cfqg() does not try to kfree root group
3911          */
3912         cfqg->ref = 1;
3913         rcu_read_lock();
3914         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3915                                         (void *)cfqd, 0);
3916         rcu_read_unlock();
3917 #endif
3918         /*
3919          * Not strictly needed (since RB_ROOT just clears the node and we
3920          * zeroed cfqd on alloc), but better be safe in case someone decides
3921          * to add magic to the rb code
3922          */
3923         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3924                 cfqd->prio_trees[i] = RB_ROOT;
3925
3926         /*
3927          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3928          * Grab a permanent reference to it, so that the normal code flow
3929          * will not attempt to free it.
3930          */
3931         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3932         cfqd->oom_cfqq.ref++;
3933         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3934
3935         INIT_LIST_HEAD(&cfqd->cic_list);
3936
3937         cfqd->queue = q;
3938
3939         init_timer(&cfqd->idle_slice_timer);
3940         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3941         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3942
3943         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3944
3945         cfqd->cfq_quantum = cfq_quantum;
3946         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3947         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3948         cfqd->cfq_back_max = cfq_back_max;
3949         cfqd->cfq_back_penalty = cfq_back_penalty;
3950         cfqd->cfq_slice[0] = cfq_slice_async;
3951         cfqd->cfq_slice[1] = cfq_slice_sync;
3952         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3953         cfqd->cfq_slice_idle = cfq_slice_idle;
3954         cfqd->cfq_group_idle = cfq_group_idle;
3955         cfqd->cfq_latency = 1;
3956         cfqd->cfq_group_isolation = 0;
3957         cfqd->hw_tag = -1;
3958         /*
3959          * we optimistically start assuming sync ops weren't delayed in last
3960          * second, in order to have larger depth for async operations.
3961          */
3962         cfqd->last_delayed_sync = jiffies - HZ;
3963         return cfqd;
3964 }
3965
3966 static void cfq_slab_kill(void)
3967 {
3968         /*
3969          * Caller already ensured that pending RCU callbacks are completed,
3970          * so we should have no busy allocations at this point.
3971          */
3972         if (cfq_pool)
3973                 kmem_cache_destroy(cfq_pool);
3974         if (cfq_ioc_pool)
3975                 kmem_cache_destroy(cfq_ioc_pool);
3976 }
3977
3978 static int __init cfq_slab_setup(void)
3979 {
3980         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3981         if (!cfq_pool)
3982                 goto fail;
3983
3984         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3985         if (!cfq_ioc_pool)
3986                 goto fail;
3987
3988         return 0;
3989 fail:
3990         cfq_slab_kill();
3991         return -ENOMEM;
3992 }
3993
3994 /*
3995  * sysfs parts below -->
3996  */
3997 static ssize_t
3998 cfq_var_show(unsigned int var, char *page)
3999 {
4000         return sprintf(page, "%d\n", var);
4001 }
4002
4003 static ssize_t
4004 cfq_var_store(unsigned int *var, const char *page, size_t count)
4005 {
4006         char *p = (char *) page;
4007
4008         *var = simple_strtoul(p, &p, 10);
4009         return count;
4010 }
4011
4012 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4013 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4014 {                                                                       \
4015         struct cfq_data *cfqd = e->elevator_data;                       \
4016         unsigned int __data = __VAR;                                    \
4017         if (__CONV)                                                     \
4018                 __data = jiffies_to_msecs(__data);                      \
4019         return cfq_var_show(__data, (page));                            \
4020 }
4021 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4022 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4023 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4024 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4025 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4026 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4027 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4028 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4029 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4030 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4031 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4032 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4033 #undef SHOW_FUNCTION
4034
4035 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4036 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4037 {                                                                       \
4038         struct cfq_data *cfqd = e->elevator_data;                       \
4039         unsigned int __data;                                            \
4040         int ret = cfq_var_store(&__data, (page), count);                \
4041         if (__data < (MIN))                                             \
4042                 __data = (MIN);                                         \
4043         else if (__data > (MAX))                                        \
4044                 __data = (MAX);                                         \
4045         if (__CONV)                                                     \
4046                 *(__PTR) = msecs_to_jiffies(__data);                    \
4047         else                                                            \
4048                 *(__PTR) = __data;                                      \
4049         return ret;                                                     \
4050 }
4051 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4052 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4053                 UINT_MAX, 1);
4054 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4055                 UINT_MAX, 1);
4056 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4057 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4058                 UINT_MAX, 0);
4059 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4060 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4061 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4062 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4063 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4064                 UINT_MAX, 0);
4065 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4066 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4067 #undef STORE_FUNCTION
4068
4069 #define CFQ_ATTR(name) \
4070         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4071
4072 static struct elv_fs_entry cfq_attrs[] = {
4073         CFQ_ATTR(quantum),
4074         CFQ_ATTR(fifo_expire_sync),
4075         CFQ_ATTR(fifo_expire_async),
4076         CFQ_ATTR(back_seek_max),
4077         CFQ_ATTR(back_seek_penalty),
4078         CFQ_ATTR(slice_sync),
4079         CFQ_ATTR(slice_async),
4080         CFQ_ATTR(slice_async_rq),
4081         CFQ_ATTR(slice_idle),
4082         CFQ_ATTR(group_idle),
4083         CFQ_ATTR(low_latency),
4084         CFQ_ATTR(group_isolation),
4085         __ATTR_NULL
4086 };
4087
4088 static struct elevator_type iosched_cfq = {
4089         .ops = {
4090                 .elevator_merge_fn =            cfq_merge,
4091                 .elevator_merged_fn =           cfq_merged_request,
4092                 .elevator_merge_req_fn =        cfq_merged_requests,
4093                 .elevator_allow_merge_fn =      cfq_allow_merge,
4094                 .elevator_bio_merged_fn =       cfq_bio_merged,
4095                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4096                 .elevator_add_req_fn =          cfq_insert_request,
4097                 .elevator_activate_req_fn =     cfq_activate_request,
4098                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4099                 .elevator_queue_empty_fn =      cfq_queue_empty,
4100                 .elevator_completed_req_fn =    cfq_completed_request,
4101                 .elevator_former_req_fn =       elv_rb_former_request,
4102                 .elevator_latter_req_fn =       elv_rb_latter_request,
4103                 .elevator_set_req_fn =          cfq_set_request,
4104                 .elevator_put_req_fn =          cfq_put_request,
4105                 .elevator_may_queue_fn =        cfq_may_queue,
4106                 .elevator_init_fn =             cfq_init_queue,
4107                 .elevator_exit_fn =             cfq_exit_queue,
4108                 .trim =                         cfq_free_io_context,
4109         },
4110         .elevator_attrs =       cfq_attrs,
4111         .elevator_name =        "cfq",
4112         .elevator_owner =       THIS_MODULE,
4113 };
4114
4115 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4116 static struct blkio_policy_type blkio_policy_cfq = {
4117         .ops = {
4118                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4119                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4120         },
4121         .plid = BLKIO_POLICY_PROP,
4122 };
4123 #else
4124 static struct blkio_policy_type blkio_policy_cfq;
4125 #endif
4126
4127 static int __init cfq_init(void)
4128 {
4129         /*
4130          * could be 0 on HZ < 1000 setups
4131          */
4132         if (!cfq_slice_async)
4133                 cfq_slice_async = 1;
4134         if (!cfq_slice_idle)
4135                 cfq_slice_idle = 1;
4136
4137 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4138         if (!cfq_group_idle)
4139                 cfq_group_idle = 1;
4140 #else
4141                 cfq_group_idle = 0;
4142 #endif
4143         if (cfq_slab_setup())
4144                 return -ENOMEM;
4145
4146         elv_register(&iosched_cfq);
4147         blkio_policy_register(&blkio_policy_cfq);
4148
4149         return 0;
4150 }
4151
4152 static void __exit cfq_exit(void)
4153 {
4154         DECLARE_COMPLETION_ONSTACK(all_gone);
4155         blkio_policy_unregister(&blkio_policy_cfq);
4156         elv_unregister(&iosched_cfq);
4157         ioc_gone = &all_gone;
4158         /* ioc_gone's update must be visible before reading ioc_count */
4159         smp_wmb();
4160
4161         /*
4162          * this also protects us from entering cfq_slab_kill() with
4163          * pending RCU callbacks
4164          */
4165         if (elv_ioc_count_read(cfq_ioc_count))
4166                 wait_for_completion(&all_gone);
4167         ida_destroy(&cic_index_ida);
4168         cfq_slab_kill();
4169 }
4170
4171 module_init(cfq_init);
4172 module_exit(cfq_exit);
4173
4174 MODULE_AUTHOR("Jens Axboe");
4175 MODULE_LICENSE("GPL");
4176 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");