block cfq: select new workload if priority changed
[linux-2.6-block.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         atomic_t ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         struct cfq_group *orig_cfqg;
150         /* 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         atomic_t 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 void
602 cfq_set_prio_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         cfqq->slice_start = jiffies;
629         cfqq->slice_end = jiffies + slice;
630         cfqq->allocated_slice = slice;
631         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
632 }
633
634 /*
635  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
636  * isn't valid until the first request from the dispatch is activated
637  * and the slice time set.
638  */
639 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
640 {
641         if (cfq_cfqq_slice_new(cfqq))
642                 return false;
643         if (time_before(jiffies, cfqq->slice_end))
644                 return false;
645
646         return true;
647 }
648
649 /*
650  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
651  * We choose the request that is closest to the head right now. Distance
652  * behind the head is penalized and only allowed to a certain extent.
653  */
654 static struct request *
655 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
656 {
657         sector_t s1, s2, d1 = 0, d2 = 0;
658         unsigned long back_max;
659 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
660 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
661         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
662
663         if (rq1 == NULL || rq1 == rq2)
664                 return rq2;
665         if (rq2 == NULL)
666                 return rq1;
667
668         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
669                 return rq1;
670         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
671                 return rq2;
672         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
673                 return rq1;
674         else if ((rq2->cmd_flags & REQ_META) &&
675                  !(rq1->cmd_flags & REQ_META))
676                 return rq2;
677
678         s1 = blk_rq_pos(rq1);
679         s2 = blk_rq_pos(rq2);
680
681         /*
682          * by definition, 1KiB is 2 sectors
683          */
684         back_max = cfqd->cfq_back_max * 2;
685
686         /*
687          * Strict one way elevator _except_ in the case where we allow
688          * short backward seeks which are biased as twice the cost of a
689          * similar forward seek.
690          */
691         if (s1 >= last)
692                 d1 = s1 - last;
693         else if (s1 + back_max >= last)
694                 d1 = (last - s1) * cfqd->cfq_back_penalty;
695         else
696                 wrap |= CFQ_RQ1_WRAP;
697
698         if (s2 >= last)
699                 d2 = s2 - last;
700         else if (s2 + back_max >= last)
701                 d2 = (last - s2) * cfqd->cfq_back_penalty;
702         else
703                 wrap |= CFQ_RQ2_WRAP;
704
705         /* Found required data */
706
707         /*
708          * By doing switch() on the bit mask "wrap" we avoid having to
709          * check two variables for all permutations: --> faster!
710          */
711         switch (wrap) {
712         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
713                 if (d1 < d2)
714                         return rq1;
715                 else if (d2 < d1)
716                         return rq2;
717                 else {
718                         if (s1 >= s2)
719                                 return rq1;
720                         else
721                                 return rq2;
722                 }
723
724         case CFQ_RQ2_WRAP:
725                 return rq1;
726         case CFQ_RQ1_WRAP:
727                 return rq2;
728         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
729         default:
730                 /*
731                  * Since both rqs are wrapped,
732                  * start with the one that's further behind head
733                  * (--> only *one* back seek required),
734                  * since back seek takes more time than forward.
735                  */
736                 if (s1 <= s2)
737                         return rq1;
738                 else
739                         return rq2;
740         }
741 }
742
743 /*
744  * The below is leftmost cache rbtree addon
745  */
746 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
747 {
748         /* Service tree is empty */
749         if (!root->count)
750                 return NULL;
751
752         if (!root->left)
753                 root->left = rb_first(&root->rb);
754
755         if (root->left)
756                 return rb_entry(root->left, struct cfq_queue, rb_node);
757
758         return NULL;
759 }
760
761 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
762 {
763         if (!root->left)
764                 root->left = rb_first(&root->rb);
765
766         if (root->left)
767                 return rb_entry_cfqg(root->left);
768
769         return NULL;
770 }
771
772 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
773 {
774         rb_erase(n, root);
775         RB_CLEAR_NODE(n);
776 }
777
778 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
779 {
780         if (root->left == n)
781                 root->left = NULL;
782         rb_erase_init(n, &root->rb);
783         --root->count;
784 }
785
786 /*
787  * would be nice to take fifo expire time into account as well
788  */
789 static struct request *
790 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
791                   struct request *last)
792 {
793         struct rb_node *rbnext = rb_next(&last->rb_node);
794         struct rb_node *rbprev = rb_prev(&last->rb_node);
795         struct request *next = NULL, *prev = NULL;
796
797         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
798
799         if (rbprev)
800                 prev = rb_entry_rq(rbprev);
801
802         if (rbnext)
803                 next = rb_entry_rq(rbnext);
804         else {
805                 rbnext = rb_first(&cfqq->sort_list);
806                 if (rbnext && rbnext != &last->rb_node)
807                         next = rb_entry_rq(rbnext);
808         }
809
810         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
811 }
812
813 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
814                                       struct cfq_queue *cfqq)
815 {
816         /*
817          * just an approximation, should be ok.
818          */
819         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
820                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
821 }
822
823 static inline s64
824 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
825 {
826         return cfqg->vdisktime - st->min_vdisktime;
827 }
828
829 static void
830 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
831 {
832         struct rb_node **node = &st->rb.rb_node;
833         struct rb_node *parent = NULL;
834         struct cfq_group *__cfqg;
835         s64 key = cfqg_key(st, cfqg);
836         int left = 1;
837
838         while (*node != NULL) {
839                 parent = *node;
840                 __cfqg = rb_entry_cfqg(parent);
841
842                 if (key < cfqg_key(st, __cfqg))
843                         node = &parent->rb_left;
844                 else {
845                         node = &parent->rb_right;
846                         left = 0;
847                 }
848         }
849
850         if (left)
851                 st->left = &cfqg->rb_node;
852
853         rb_link_node(&cfqg->rb_node, parent, node);
854         rb_insert_color(&cfqg->rb_node, &st->rb);
855 }
856
857 static void
858 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
859 {
860         struct cfq_rb_root *st = &cfqd->grp_service_tree;
861         struct cfq_group *__cfqg;
862         struct rb_node *n;
863
864         cfqg->nr_cfqq++;
865         if (!RB_EMPTY_NODE(&cfqg->rb_node))
866                 return;
867
868         /*
869          * Currently put the group at the end. Later implement something
870          * so that groups get lesser vtime based on their weights, so that
871          * if group does not loose all if it was not continously backlogged.
872          */
873         n = rb_last(&st->rb);
874         if (n) {
875                 __cfqg = rb_entry_cfqg(n);
876                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
877         } else
878                 cfqg->vdisktime = st->min_vdisktime;
879
880         __cfq_group_service_tree_add(st, cfqg);
881         st->total_weight += cfqg->weight;
882 }
883
884 static void
885 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
886 {
887         struct cfq_rb_root *st = &cfqd->grp_service_tree;
888
889         BUG_ON(cfqg->nr_cfqq < 1);
890         cfqg->nr_cfqq--;
891
892         /* If there are other cfq queues under this group, don't delete it */
893         if (cfqg->nr_cfqq)
894                 return;
895
896         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
897         st->total_weight -= cfqg->weight;
898         if (!RB_EMPTY_NODE(&cfqg->rb_node))
899                 cfq_rb_erase(&cfqg->rb_node, st);
900         cfqg->saved_workload_slice = 0;
901         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
902 }
903
904 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
905 {
906         unsigned int slice_used;
907
908         /*
909          * Queue got expired before even a single request completed or
910          * got expired immediately after first request completion.
911          */
912         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
913                 /*
914                  * Also charge the seek time incurred to the group, otherwise
915                  * if there are mutiple queues in the group, each can dispatch
916                  * a single request on seeky media and cause lots of seek time
917                  * and group will never know it.
918                  */
919                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
920                                         1);
921         } else {
922                 slice_used = jiffies - cfqq->slice_start;
923                 if (slice_used > cfqq->allocated_slice)
924                         slice_used = cfqq->allocated_slice;
925         }
926
927         return slice_used;
928 }
929
930 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
931                                 struct cfq_queue *cfqq)
932 {
933         struct cfq_rb_root *st = &cfqd->grp_service_tree;
934         unsigned int used_sl, charge;
935         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
936                         - cfqg->service_tree_idle.count;
937
938         BUG_ON(nr_sync < 0);
939         used_sl = charge = cfq_cfqq_slice_usage(cfqq);
940
941         if (iops_mode(cfqd))
942                 charge = cfqq->slice_dispatch;
943         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
944                 charge = cfqq->allocated_slice;
945
946         /* Can't update vdisktime while group is on service tree */
947         cfq_rb_erase(&cfqg->rb_node, st);
948         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
949         __cfq_group_service_tree_add(st, cfqg);
950
951         /* This group is being expired. Save the context */
952         if (time_after(cfqd->workload_expires, jiffies)) {
953                 cfqg->saved_workload_slice = cfqd->workload_expires
954                                                 - jiffies;
955                 cfqg->saved_workload = cfqd->serving_type;
956                 cfqg->saved_serving_prio = cfqd->serving_prio;
957         } else
958                 cfqg->saved_workload_slice = 0;
959
960         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
961                                         st->min_vdisktime);
962         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
963                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
964                         iops_mode(cfqd), cfqq->nr_sectors);
965         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
966         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
967 }
968
969 #ifdef CONFIG_CFQ_GROUP_IOSCHED
970 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
971 {
972         if (blkg)
973                 return container_of(blkg, struct cfq_group, blkg);
974         return NULL;
975 }
976
977 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
978                                         unsigned int weight)
979 {
980         cfqg_of_blkg(blkg)->weight = weight;
981 }
982
983 static struct cfq_group *
984 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
985 {
986         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
987         struct cfq_group *cfqg = NULL;
988         void *key = cfqd;
989         int i, j;
990         struct cfq_rb_root *st;
991         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
992         unsigned int major, minor;
993
994         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
995         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
996                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
997                 cfqg->blkg.dev = MKDEV(major, minor);
998                 goto done;
999         }
1000         if (cfqg || !create)
1001                 goto done;
1002
1003         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1004         if (!cfqg)
1005                 goto done;
1006
1007         for_each_cfqg_st(cfqg, i, j, st)
1008                 *st = CFQ_RB_ROOT;
1009         RB_CLEAR_NODE(&cfqg->rb_node);
1010
1011         /*
1012          * Take the initial reference that will be released on destroy
1013          * This can be thought of a joint reference by cgroup and
1014          * elevator which will be dropped by either elevator exit
1015          * or cgroup deletion path depending on who is exiting first.
1016          */
1017         atomic_set(&cfqg->ref, 1);
1018
1019         /*
1020          * Add group onto cgroup list. It might happen that bdi->dev is
1021          * not initiliazed yet. Initialize this new group without major
1022          * and minor info and this info will be filled in once a new thread
1023          * comes for IO. See code above.
1024          */
1025         if (bdi->dev) {
1026                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1027                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1028                                         MKDEV(major, minor));
1029         } else
1030                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1031                                         0);
1032
1033         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1034
1035         /* Add group on cfqd list */
1036         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1037
1038 done:
1039         return cfqg;
1040 }
1041
1042 /*
1043  * Search for the cfq group current task belongs to. If create = 1, then also
1044  * create the cfq group if it does not exist. request_queue lock must be held.
1045  */
1046 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1047 {
1048         struct cgroup *cgroup;
1049         struct cfq_group *cfqg = NULL;
1050
1051         rcu_read_lock();
1052         cgroup = task_cgroup(current, blkio_subsys_id);
1053         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1054         if (!cfqg && create)
1055                 cfqg = &cfqd->root_group;
1056         rcu_read_unlock();
1057         return cfqg;
1058 }
1059
1060 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1061 {
1062         atomic_inc(&cfqg->ref);
1063         return cfqg;
1064 }
1065
1066 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1067 {
1068         /* Currently, all async queues are mapped to root group */
1069         if (!cfq_cfqq_sync(cfqq))
1070                 cfqg = &cfqq->cfqd->root_group;
1071
1072         cfqq->cfqg = cfqg;
1073         /* cfqq reference on cfqg */
1074         atomic_inc(&cfqq->cfqg->ref);
1075 }
1076
1077 static void cfq_put_cfqg(struct cfq_group *cfqg)
1078 {
1079         struct cfq_rb_root *st;
1080         int i, j;
1081
1082         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1083         if (!atomic_dec_and_test(&cfqg->ref))
1084                 return;
1085         for_each_cfqg_st(cfqg, i, j, st)
1086                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1087         kfree(cfqg);
1088 }
1089
1090 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1091 {
1092         /* Something wrong if we are trying to remove same group twice */
1093         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1094
1095         hlist_del_init(&cfqg->cfqd_node);
1096
1097         /*
1098          * Put the reference taken at the time of creation so that when all
1099          * queues are gone, group can be destroyed.
1100          */
1101         cfq_put_cfqg(cfqg);
1102 }
1103
1104 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1105 {
1106         struct hlist_node *pos, *n;
1107         struct cfq_group *cfqg;
1108
1109         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1110                 /*
1111                  * If cgroup removal path got to blk_group first and removed
1112                  * it from cgroup list, then it will take care of destroying
1113                  * cfqg also.
1114                  */
1115                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1116                         cfq_destroy_cfqg(cfqd, cfqg);
1117         }
1118 }
1119
1120 /*
1121  * Blk cgroup controller notification saying that blkio_group object is being
1122  * delinked as associated cgroup object is going away. That also means that
1123  * no new IO will come in this group. So get rid of this group as soon as
1124  * any pending IO in the group is finished.
1125  *
1126  * This function is called under rcu_read_lock(). key is the rcu protected
1127  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1128  * read lock.
1129  *
1130  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1131  * it should not be NULL as even if elevator was exiting, cgroup deltion
1132  * path got to it first.
1133  */
1134 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1135 {
1136         unsigned long  flags;
1137         struct cfq_data *cfqd = key;
1138
1139         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1140         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1141         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1142 }
1143
1144 #else /* GROUP_IOSCHED */
1145 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1146 {
1147         return &cfqd->root_group;
1148 }
1149
1150 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1151 {
1152         return cfqg;
1153 }
1154
1155 static inline void
1156 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1157         cfqq->cfqg = cfqg;
1158 }
1159
1160 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1161 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1162
1163 #endif /* GROUP_IOSCHED */
1164
1165 /*
1166  * The cfqd->service_trees holds all pending cfq_queue's that have
1167  * requests waiting to be processed. It is sorted in the order that
1168  * we will service the queues.
1169  */
1170 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1171                                  bool add_front)
1172 {
1173         struct rb_node **p, *parent;
1174         struct cfq_queue *__cfqq;
1175         unsigned long rb_key;
1176         struct cfq_rb_root *service_tree;
1177         int left;
1178         int new_cfqq = 1;
1179         int group_changed = 0;
1180
1181 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1182         if (!cfqd->cfq_group_isolation
1183             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1184             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1185                 /* Move this cfq to root group */
1186                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1187                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1188                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1189                 cfqq->orig_cfqg = cfqq->cfqg;
1190                 cfqq->cfqg = &cfqd->root_group;
1191                 atomic_inc(&cfqd->root_group.ref);
1192                 group_changed = 1;
1193         } else if (!cfqd->cfq_group_isolation
1194                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1195                 /* cfqq is sequential now needs to go to its original group */
1196                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1197                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1198                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1199                 cfq_put_cfqg(cfqq->cfqg);
1200                 cfqq->cfqg = cfqq->orig_cfqg;
1201                 cfqq->orig_cfqg = NULL;
1202                 group_changed = 1;
1203                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1204         }
1205 #endif
1206
1207         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1208                                                 cfqq_type(cfqq));
1209         if (cfq_class_idle(cfqq)) {
1210                 rb_key = CFQ_IDLE_DELAY;
1211                 parent = rb_last(&service_tree->rb);
1212                 if (parent && parent != &cfqq->rb_node) {
1213                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1214                         rb_key += __cfqq->rb_key;
1215                 } else
1216                         rb_key += jiffies;
1217         } else if (!add_front) {
1218                 /*
1219                  * Get our rb key offset. Subtract any residual slice
1220                  * value carried from last service. A negative resid
1221                  * count indicates slice overrun, and this should position
1222                  * the next service time further away in the tree.
1223                  */
1224                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1225                 rb_key -= cfqq->slice_resid;
1226                 cfqq->slice_resid = 0;
1227         } else {
1228                 rb_key = -HZ;
1229                 __cfqq = cfq_rb_first(service_tree);
1230                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1231         }
1232
1233         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1234                 new_cfqq = 0;
1235                 /*
1236                  * same position, nothing more to do
1237                  */
1238                 if (rb_key == cfqq->rb_key &&
1239                     cfqq->service_tree == service_tree)
1240                         return;
1241
1242                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1243                 cfqq->service_tree = NULL;
1244         }
1245
1246         left = 1;
1247         parent = NULL;
1248         cfqq->service_tree = service_tree;
1249         p = &service_tree->rb.rb_node;
1250         while (*p) {
1251                 struct rb_node **n;
1252
1253                 parent = *p;
1254                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1255
1256                 /*
1257                  * sort by key, that represents service time.
1258                  */
1259                 if (time_before(rb_key, __cfqq->rb_key))
1260                         n = &(*p)->rb_left;
1261                 else {
1262                         n = &(*p)->rb_right;
1263                         left = 0;
1264                 }
1265
1266                 p = n;
1267         }
1268
1269         if (left)
1270                 service_tree->left = &cfqq->rb_node;
1271
1272         cfqq->rb_key = rb_key;
1273         rb_link_node(&cfqq->rb_node, parent, p);
1274         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1275         service_tree->count++;
1276         if ((add_front || !new_cfqq) && !group_changed)
1277                 return;
1278         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1279 }
1280
1281 static struct cfq_queue *
1282 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1283                      sector_t sector, struct rb_node **ret_parent,
1284                      struct rb_node ***rb_link)
1285 {
1286         struct rb_node **p, *parent;
1287         struct cfq_queue *cfqq = NULL;
1288
1289         parent = NULL;
1290         p = &root->rb_node;
1291         while (*p) {
1292                 struct rb_node **n;
1293
1294                 parent = *p;
1295                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1296
1297                 /*
1298                  * Sort strictly based on sector.  Smallest to the left,
1299                  * largest to the right.
1300                  */
1301                 if (sector > blk_rq_pos(cfqq->next_rq))
1302                         n = &(*p)->rb_right;
1303                 else if (sector < blk_rq_pos(cfqq->next_rq))
1304                         n = &(*p)->rb_left;
1305                 else
1306                         break;
1307                 p = n;
1308                 cfqq = NULL;
1309         }
1310
1311         *ret_parent = parent;
1312         if (rb_link)
1313                 *rb_link = p;
1314         return cfqq;
1315 }
1316
1317 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1318 {
1319         struct rb_node **p, *parent;
1320         struct cfq_queue *__cfqq;
1321
1322         if (cfqq->p_root) {
1323                 rb_erase(&cfqq->p_node, cfqq->p_root);
1324                 cfqq->p_root = NULL;
1325         }
1326
1327         if (cfq_class_idle(cfqq))
1328                 return;
1329         if (!cfqq->next_rq)
1330                 return;
1331
1332         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1333         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1334                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1335         if (!__cfqq) {
1336                 rb_link_node(&cfqq->p_node, parent, p);
1337                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1338         } else
1339                 cfqq->p_root = NULL;
1340 }
1341
1342 /*
1343  * Update cfqq's position in the service tree.
1344  */
1345 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1346 {
1347         /*
1348          * Resorting requires the cfqq to be on the RR list already.
1349          */
1350         if (cfq_cfqq_on_rr(cfqq)) {
1351                 cfq_service_tree_add(cfqd, cfqq, 0);
1352                 cfq_prio_tree_add(cfqd, cfqq);
1353         }
1354 }
1355
1356 /*
1357  * add to busy list of queues for service, trying to be fair in ordering
1358  * the pending list according to last request service
1359  */
1360 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1361 {
1362         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1363         BUG_ON(cfq_cfqq_on_rr(cfqq));
1364         cfq_mark_cfqq_on_rr(cfqq);
1365         cfqd->busy_queues++;
1366
1367         cfq_resort_rr_list(cfqd, cfqq);
1368 }
1369
1370 /*
1371  * Called when the cfqq no longer has requests pending, remove it from
1372  * the service tree.
1373  */
1374 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1375 {
1376         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1377         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1378         cfq_clear_cfqq_on_rr(cfqq);
1379
1380         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1381                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1382                 cfqq->service_tree = NULL;
1383         }
1384         if (cfqq->p_root) {
1385                 rb_erase(&cfqq->p_node, cfqq->p_root);
1386                 cfqq->p_root = NULL;
1387         }
1388
1389         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1390         BUG_ON(!cfqd->busy_queues);
1391         cfqd->busy_queues--;
1392 }
1393
1394 /*
1395  * rb tree support functions
1396  */
1397 static void cfq_del_rq_rb(struct request *rq)
1398 {
1399         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1400         const int sync = rq_is_sync(rq);
1401
1402         BUG_ON(!cfqq->queued[sync]);
1403         cfqq->queued[sync]--;
1404
1405         elv_rb_del(&cfqq->sort_list, rq);
1406
1407         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1408                 /*
1409                  * Queue will be deleted from service tree when we actually
1410                  * expire it later. Right now just remove it from prio tree
1411                  * as it is empty.
1412                  */
1413                 if (cfqq->p_root) {
1414                         rb_erase(&cfqq->p_node, cfqq->p_root);
1415                         cfqq->p_root = NULL;
1416                 }
1417         }
1418 }
1419
1420 static void cfq_add_rq_rb(struct request *rq)
1421 {
1422         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1423         struct cfq_data *cfqd = cfqq->cfqd;
1424         struct request *__alias, *prev;
1425
1426         cfqq->queued[rq_is_sync(rq)]++;
1427
1428         /*
1429          * looks a little odd, but the first insert might return an alias.
1430          * if that happens, put the alias on the dispatch list
1431          */
1432         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1433                 cfq_dispatch_insert(cfqd->queue, __alias);
1434
1435         if (!cfq_cfqq_on_rr(cfqq))
1436                 cfq_add_cfqq_rr(cfqd, cfqq);
1437
1438         /*
1439          * check if this request is a better next-serve candidate
1440          */
1441         prev = cfqq->next_rq;
1442         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1443
1444         /*
1445          * adjust priority tree position, if ->next_rq changes
1446          */
1447         if (prev != cfqq->next_rq)
1448                 cfq_prio_tree_add(cfqd, cfqq);
1449
1450         BUG_ON(!cfqq->next_rq);
1451 }
1452
1453 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1454 {
1455         elv_rb_del(&cfqq->sort_list, rq);
1456         cfqq->queued[rq_is_sync(rq)]--;
1457         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1458                                         rq_data_dir(rq), rq_is_sync(rq));
1459         cfq_add_rq_rb(rq);
1460         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1461                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1462                         rq_is_sync(rq));
1463 }
1464
1465 static struct request *
1466 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1467 {
1468         struct task_struct *tsk = current;
1469         struct cfq_io_context *cic;
1470         struct cfq_queue *cfqq;
1471
1472         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1473         if (!cic)
1474                 return NULL;
1475
1476         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1477         if (cfqq) {
1478                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1479
1480                 return elv_rb_find(&cfqq->sort_list, sector);
1481         }
1482
1483         return NULL;
1484 }
1485
1486 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1487 {
1488         struct cfq_data *cfqd = q->elevator->elevator_data;
1489
1490         cfqd->rq_in_driver++;
1491         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1492                                                 cfqd->rq_in_driver);
1493
1494         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1495 }
1496
1497 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1498 {
1499         struct cfq_data *cfqd = q->elevator->elevator_data;
1500
1501         WARN_ON(!cfqd->rq_in_driver);
1502         cfqd->rq_in_driver--;
1503         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1504                                                 cfqd->rq_in_driver);
1505 }
1506
1507 static void cfq_remove_request(struct request *rq)
1508 {
1509         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1510
1511         if (cfqq->next_rq == rq)
1512                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1513
1514         list_del_init(&rq->queuelist);
1515         cfq_del_rq_rb(rq);
1516
1517         cfqq->cfqd->rq_queued--;
1518         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1519                                         rq_data_dir(rq), rq_is_sync(rq));
1520         if (rq->cmd_flags & REQ_META) {
1521                 WARN_ON(!cfqq->meta_pending);
1522                 cfqq->meta_pending--;
1523         }
1524 }
1525
1526 static int cfq_merge(struct request_queue *q, struct request **req,
1527                      struct bio *bio)
1528 {
1529         struct cfq_data *cfqd = q->elevator->elevator_data;
1530         struct request *__rq;
1531
1532         __rq = cfq_find_rq_fmerge(cfqd, bio);
1533         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1534                 *req = __rq;
1535                 return ELEVATOR_FRONT_MERGE;
1536         }
1537
1538         return ELEVATOR_NO_MERGE;
1539 }
1540
1541 static void cfq_merged_request(struct request_queue *q, struct request *req,
1542                                int type)
1543 {
1544         if (type == ELEVATOR_FRONT_MERGE) {
1545                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1546
1547                 cfq_reposition_rq_rb(cfqq, req);
1548         }
1549 }
1550
1551 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1552                                 struct bio *bio)
1553 {
1554         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1555                                         bio_data_dir(bio), cfq_bio_sync(bio));
1556 }
1557
1558 static void
1559 cfq_merged_requests(struct request_queue *q, struct request *rq,
1560                     struct request *next)
1561 {
1562         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1563         /*
1564          * reposition in fifo if next is older than rq
1565          */
1566         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1567             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1568                 list_move(&rq->queuelist, &next->queuelist);
1569                 rq_set_fifo_time(rq, rq_fifo_time(next));
1570         }
1571
1572         if (cfqq->next_rq == next)
1573                 cfqq->next_rq = rq;
1574         cfq_remove_request(next);
1575         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1576                                         rq_data_dir(next), rq_is_sync(next));
1577 }
1578
1579 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1580                            struct bio *bio)
1581 {
1582         struct cfq_data *cfqd = q->elevator->elevator_data;
1583         struct cfq_io_context *cic;
1584         struct cfq_queue *cfqq;
1585
1586         /*
1587          * Disallow merge of a sync bio into an async request.
1588          */
1589         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1590                 return false;
1591
1592         /*
1593          * Lookup the cfqq that this bio will be queued with. Allow
1594          * merge only if rq is queued there.
1595          */
1596         cic = cfq_cic_lookup(cfqd, current->io_context);
1597         if (!cic)
1598                 return false;
1599
1600         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1601         return cfqq == RQ_CFQQ(rq);
1602 }
1603
1604 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1605 {
1606         del_timer(&cfqd->idle_slice_timer);
1607         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1608 }
1609
1610 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1611                                    struct cfq_queue *cfqq)
1612 {
1613         if (cfqq) {
1614                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1615                                 cfqd->serving_prio, cfqd->serving_type);
1616                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1617                 cfqq->slice_start = 0;
1618                 cfqq->dispatch_start = jiffies;
1619                 cfqq->allocated_slice = 0;
1620                 cfqq->slice_end = 0;
1621                 cfqq->slice_dispatch = 0;
1622                 cfqq->nr_sectors = 0;
1623
1624                 cfq_clear_cfqq_wait_request(cfqq);
1625                 cfq_clear_cfqq_must_dispatch(cfqq);
1626                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1627                 cfq_clear_cfqq_fifo_expire(cfqq);
1628                 cfq_mark_cfqq_slice_new(cfqq);
1629
1630                 cfq_del_timer(cfqd, cfqq);
1631         }
1632
1633         cfqd->active_queue = cfqq;
1634 }
1635
1636 /*
1637  * current cfqq expired its slice (or was too idle), select new one
1638  */
1639 static void
1640 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1641                     bool timed_out)
1642 {
1643         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1644
1645         if (cfq_cfqq_wait_request(cfqq))
1646                 cfq_del_timer(cfqd, cfqq);
1647
1648         cfq_clear_cfqq_wait_request(cfqq);
1649         cfq_clear_cfqq_wait_busy(cfqq);
1650
1651         /*
1652          * If this cfqq is shared between multiple processes, check to
1653          * make sure that those processes are still issuing I/Os within
1654          * the mean seek distance.  If not, it may be time to break the
1655          * queues apart again.
1656          */
1657         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1658                 cfq_mark_cfqq_split_coop(cfqq);
1659
1660         /*
1661          * store what was left of this slice, if the queue idled/timed out
1662          */
1663         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1664                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1665                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1666         }
1667
1668         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1669
1670         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1671                 cfq_del_cfqq_rr(cfqd, cfqq);
1672
1673         cfq_resort_rr_list(cfqd, cfqq);
1674
1675         if (cfqq == cfqd->active_queue)
1676                 cfqd->active_queue = NULL;
1677
1678         if (cfqd->active_cic) {
1679                 put_io_context(cfqd->active_cic->ioc);
1680                 cfqd->active_cic = NULL;
1681         }
1682 }
1683
1684 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1685 {
1686         struct cfq_queue *cfqq = cfqd->active_queue;
1687
1688         if (cfqq)
1689                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1690 }
1691
1692 /*
1693  * Get next queue for service. Unless we have a queue preemption,
1694  * we'll simply select the first cfqq in the service tree.
1695  */
1696 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1697 {
1698         struct cfq_rb_root *service_tree =
1699                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1700                                         cfqd->serving_type);
1701
1702         if (!cfqd->rq_queued)
1703                 return NULL;
1704
1705         /* There is nothing to dispatch */
1706         if (!service_tree)
1707                 return NULL;
1708         if (RB_EMPTY_ROOT(&service_tree->rb))
1709                 return NULL;
1710         return cfq_rb_first(service_tree);
1711 }
1712
1713 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1714 {
1715         struct cfq_group *cfqg;
1716         struct cfq_queue *cfqq;
1717         int i, j;
1718         struct cfq_rb_root *st;
1719
1720         if (!cfqd->rq_queued)
1721                 return NULL;
1722
1723         cfqg = cfq_get_next_cfqg(cfqd);
1724         if (!cfqg)
1725                 return NULL;
1726
1727         for_each_cfqg_st(cfqg, i, j, st)
1728                 if ((cfqq = cfq_rb_first(st)) != NULL)
1729                         return cfqq;
1730         return NULL;
1731 }
1732
1733 /*
1734  * Get and set a new active queue for service.
1735  */
1736 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1737                                               struct cfq_queue *cfqq)
1738 {
1739         if (!cfqq)
1740                 cfqq = cfq_get_next_queue(cfqd);
1741
1742         __cfq_set_active_queue(cfqd, cfqq);
1743         return cfqq;
1744 }
1745
1746 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1747                                           struct request *rq)
1748 {
1749         if (blk_rq_pos(rq) >= cfqd->last_position)
1750                 return blk_rq_pos(rq) - cfqd->last_position;
1751         else
1752                 return cfqd->last_position - blk_rq_pos(rq);
1753 }
1754
1755 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1756                                struct request *rq)
1757 {
1758         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1759 }
1760
1761 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1762                                     struct cfq_queue *cur_cfqq)
1763 {
1764         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1765         struct rb_node *parent, *node;
1766         struct cfq_queue *__cfqq;
1767         sector_t sector = cfqd->last_position;
1768
1769         if (RB_EMPTY_ROOT(root))
1770                 return NULL;
1771
1772         /*
1773          * First, if we find a request starting at the end of the last
1774          * request, choose it.
1775          */
1776         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1777         if (__cfqq)
1778                 return __cfqq;
1779
1780         /*
1781          * If the exact sector wasn't found, the parent of the NULL leaf
1782          * will contain the closest sector.
1783          */
1784         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1785         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1786                 return __cfqq;
1787
1788         if (blk_rq_pos(__cfqq->next_rq) < sector)
1789                 node = rb_next(&__cfqq->p_node);
1790         else
1791                 node = rb_prev(&__cfqq->p_node);
1792         if (!node)
1793                 return NULL;
1794
1795         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1796         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1797                 return __cfqq;
1798
1799         return NULL;
1800 }
1801
1802 /*
1803  * cfqd - obvious
1804  * cur_cfqq - passed in so that we don't decide that the current queue is
1805  *            closely cooperating with itself.
1806  *
1807  * So, basically we're assuming that that cur_cfqq has dispatched at least
1808  * one request, and that cfqd->last_position reflects a position on the disk
1809  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1810  * assumption.
1811  */
1812 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1813                                               struct cfq_queue *cur_cfqq)
1814 {
1815         struct cfq_queue *cfqq;
1816
1817         if (cfq_class_idle(cur_cfqq))
1818                 return NULL;
1819         if (!cfq_cfqq_sync(cur_cfqq))
1820                 return NULL;
1821         if (CFQQ_SEEKY(cur_cfqq))
1822                 return NULL;
1823
1824         /*
1825          * Don't search priority tree if it's the only queue in the group.
1826          */
1827         if (cur_cfqq->cfqg->nr_cfqq == 1)
1828                 return NULL;
1829
1830         /*
1831          * We should notice if some of the queues are cooperating, eg
1832          * working closely on the same area of the disk. In that case,
1833          * we can group them together and don't waste time idling.
1834          */
1835         cfqq = cfqq_close(cfqd, cur_cfqq);
1836         if (!cfqq)
1837                 return NULL;
1838
1839         /* If new queue belongs to different cfq_group, don't choose it */
1840         if (cur_cfqq->cfqg != cfqq->cfqg)
1841                 return NULL;
1842
1843         /*
1844          * It only makes sense to merge sync queues.
1845          */
1846         if (!cfq_cfqq_sync(cfqq))
1847                 return NULL;
1848         if (CFQQ_SEEKY(cfqq))
1849                 return NULL;
1850
1851         /*
1852          * Do not merge queues of different priority classes
1853          */
1854         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1855                 return NULL;
1856
1857         return cfqq;
1858 }
1859
1860 /*
1861  * Determine whether we should enforce idle window for this queue.
1862  */
1863
1864 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1865 {
1866         enum wl_prio_t prio = cfqq_prio(cfqq);
1867         struct cfq_rb_root *service_tree = cfqq->service_tree;
1868
1869         BUG_ON(!service_tree);
1870         BUG_ON(!service_tree->count);
1871
1872         if (!cfqd->cfq_slice_idle)
1873                 return false;
1874
1875         /* We never do for idle class queues. */
1876         if (prio == IDLE_WORKLOAD)
1877                 return false;
1878
1879         /* We do for queues that were marked with idle window flag. */
1880         if (cfq_cfqq_idle_window(cfqq) &&
1881            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1882                 return true;
1883
1884         /*
1885          * Otherwise, we do only if they are the last ones
1886          * in their service tree.
1887          */
1888         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1889                 return true;
1890         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1891                         service_tree->count);
1892         return false;
1893 }
1894
1895 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1896 {
1897         struct cfq_queue *cfqq = cfqd->active_queue;
1898         struct cfq_io_context *cic;
1899         unsigned long sl, group_idle = 0;
1900
1901         /*
1902          * SSD device without seek penalty, disable idling. But only do so
1903          * for devices that support queuing, otherwise we still have a problem
1904          * with sync vs async workloads.
1905          */
1906         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1907                 return;
1908
1909         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1910         WARN_ON(cfq_cfqq_slice_new(cfqq));
1911
1912         /*
1913          * idle is disabled, either manually or by past process history
1914          */
1915         if (!cfq_should_idle(cfqd, cfqq)) {
1916                 /* no queue idling. Check for group idling */
1917                 if (cfqd->cfq_group_idle)
1918                         group_idle = cfqd->cfq_group_idle;
1919                 else
1920                         return;
1921         }
1922
1923         /*
1924          * still active requests from this queue, don't idle
1925          */
1926         if (cfqq->dispatched)
1927                 return;
1928
1929         /*
1930          * task has exited, don't wait
1931          */
1932         cic = cfqd->active_cic;
1933         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1934                 return;
1935
1936         /*
1937          * If our average think time is larger than the remaining time
1938          * slice, then don't idle. This avoids overrunning the allotted
1939          * time slice.
1940          */
1941         if (sample_valid(cic->ttime_samples) &&
1942             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1943                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1944                                 cic->ttime_mean);
1945                 return;
1946         }
1947
1948         /* There are other queues in the group, don't do group idle */
1949         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1950                 return;
1951
1952         cfq_mark_cfqq_wait_request(cfqq);
1953
1954         if (group_idle)
1955                 sl = cfqd->cfq_group_idle;
1956         else
1957                 sl = cfqd->cfq_slice_idle;
1958
1959         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1960         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1961         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1962                         group_idle ? 1 : 0);
1963 }
1964
1965 /*
1966  * Move request from internal lists to the request queue dispatch list.
1967  */
1968 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1969 {
1970         struct cfq_data *cfqd = q->elevator->elevator_data;
1971         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1972
1973         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1974
1975         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1976         cfq_remove_request(rq);
1977         cfqq->dispatched++;
1978         (RQ_CFQG(rq))->dispatched++;
1979         elv_dispatch_sort(q, rq);
1980
1981         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1982         cfqq->nr_sectors += blk_rq_sectors(rq);
1983         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1984                                         rq_data_dir(rq), rq_is_sync(rq));
1985 }
1986
1987 /*
1988  * return expired entry, or NULL to just start from scratch in rbtree
1989  */
1990 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1991 {
1992         struct request *rq = NULL;
1993
1994         if (cfq_cfqq_fifo_expire(cfqq))
1995                 return NULL;
1996
1997         cfq_mark_cfqq_fifo_expire(cfqq);
1998
1999         if (list_empty(&cfqq->fifo))
2000                 return NULL;
2001
2002         rq = rq_entry_fifo(cfqq->fifo.next);
2003         if (time_before(jiffies, rq_fifo_time(rq)))
2004                 rq = NULL;
2005
2006         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2007         return rq;
2008 }
2009
2010 static inline int
2011 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2012 {
2013         const int base_rq = cfqd->cfq_slice_async_rq;
2014
2015         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2016
2017         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2018 }
2019
2020 /*
2021  * Must be called with the queue_lock held.
2022  */
2023 static int cfqq_process_refs(struct cfq_queue *cfqq)
2024 {
2025         int process_refs, io_refs;
2026
2027         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2028         process_refs = atomic_read(&cfqq->ref) - io_refs;
2029         BUG_ON(process_refs < 0);
2030         return process_refs;
2031 }
2032
2033 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2034 {
2035         int process_refs, new_process_refs;
2036         struct cfq_queue *__cfqq;
2037
2038         /*
2039          * If there are no process references on the new_cfqq, then it is
2040          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2041          * chain may have dropped their last reference (not just their
2042          * last process reference).
2043          */
2044         if (!cfqq_process_refs(new_cfqq))
2045                 return;
2046
2047         /* Avoid a circular list and skip interim queue merges */
2048         while ((__cfqq = new_cfqq->new_cfqq)) {
2049                 if (__cfqq == cfqq)
2050                         return;
2051                 new_cfqq = __cfqq;
2052         }
2053
2054         process_refs = cfqq_process_refs(cfqq);
2055         new_process_refs = cfqq_process_refs(new_cfqq);
2056         /*
2057          * If the process for the cfqq has gone away, there is no
2058          * sense in merging the queues.
2059          */
2060         if (process_refs == 0 || new_process_refs == 0)
2061                 return;
2062
2063         /*
2064          * Merge in the direction of the lesser amount of work.
2065          */
2066         if (new_process_refs >= process_refs) {
2067                 cfqq->new_cfqq = new_cfqq;
2068                 atomic_add(process_refs, &new_cfqq->ref);
2069         } else {
2070                 new_cfqq->new_cfqq = cfqq;
2071                 atomic_add(new_process_refs, &cfqq->ref);
2072         }
2073 }
2074
2075 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2076                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2077 {
2078         struct cfq_queue *queue;
2079         int i;
2080         bool key_valid = false;
2081         unsigned long lowest_key = 0;
2082         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2083
2084         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2085                 /* select the one with lowest rb_key */
2086                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2087                 if (queue &&
2088                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2089                         lowest_key = queue->rb_key;
2090                         cur_best = i;
2091                         key_valid = true;
2092                 }
2093         }
2094
2095         return cur_best;
2096 }
2097
2098 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2099 {
2100         unsigned slice;
2101         unsigned count;
2102         struct cfq_rb_root *st;
2103         unsigned group_slice;
2104         enum wl_prio_t original_prio = cfqd->serving_prio;
2105
2106         if (!cfqg) {
2107                 cfqd->serving_prio = IDLE_WORKLOAD;
2108                 cfqd->workload_expires = jiffies + 1;
2109                 return;
2110         }
2111
2112         /* Choose next priority. RT > BE > IDLE */
2113         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2114                 cfqd->serving_prio = RT_WORKLOAD;
2115         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2116                 cfqd->serving_prio = BE_WORKLOAD;
2117         else {
2118                 cfqd->serving_prio = IDLE_WORKLOAD;
2119                 cfqd->workload_expires = jiffies + 1;
2120                 return;
2121         }
2122
2123         if (original_prio != cfqd->serving_prio)
2124                 goto new_workload;
2125
2126         /*
2127          * For RT and BE, we have to choose also the type
2128          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2129          * expiration time
2130          */
2131         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2132         count = st->count;
2133
2134         /*
2135          * check workload expiration, and that we still have other queues ready
2136          */
2137         if (count && !time_after(jiffies, cfqd->workload_expires))
2138                 return;
2139
2140 new_workload:
2141         /* otherwise select new workload type */
2142         cfqd->serving_type =
2143                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2144         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2145         count = st->count;
2146
2147         /*
2148          * the workload slice is computed as a fraction of target latency
2149          * proportional to the number of queues in that workload, over
2150          * all the queues in the same priority class
2151          */
2152         group_slice = cfq_group_slice(cfqd, cfqg);
2153
2154         slice = group_slice * count /
2155                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2156                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2157
2158         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2159                 unsigned int tmp;
2160
2161                 /*
2162                  * Async queues are currently system wide. Just taking
2163                  * proportion of queues with-in same group will lead to higher
2164                  * async ratio system wide as generally root group is going
2165                  * to have higher weight. A more accurate thing would be to
2166                  * calculate system wide asnc/sync ratio.
2167                  */
2168                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2169                 tmp = tmp/cfqd->busy_queues;
2170                 slice = min_t(unsigned, slice, tmp);
2171
2172                 /* async workload slice is scaled down according to
2173                  * the sync/async slice ratio. */
2174                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2175         } else
2176                 /* sync workload slice is at least 2 * cfq_slice_idle */
2177                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2178
2179         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2180         cfq_log(cfqd, "workload slice:%d", slice);
2181         cfqd->workload_expires = jiffies + slice;
2182 }
2183
2184 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2185 {
2186         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2187         struct cfq_group *cfqg;
2188
2189         if (RB_EMPTY_ROOT(&st->rb))
2190                 return NULL;
2191         cfqg = cfq_rb_first_group(st);
2192         update_min_vdisktime(st);
2193         return cfqg;
2194 }
2195
2196 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2197 {
2198         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2199
2200         cfqd->serving_group = cfqg;
2201
2202         /* Restore the workload type data */
2203         if (cfqg->saved_workload_slice) {
2204                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2205                 cfqd->serving_type = cfqg->saved_workload;
2206                 cfqd->serving_prio = cfqg->saved_serving_prio;
2207         } else
2208                 cfqd->workload_expires = jiffies - 1;
2209
2210         choose_service_tree(cfqd, cfqg);
2211 }
2212
2213 /*
2214  * Select a queue for service. If we have a current active queue,
2215  * check whether to continue servicing it, or retrieve and set a new one.
2216  */
2217 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2218 {
2219         struct cfq_queue *cfqq, *new_cfqq = NULL;
2220
2221         cfqq = cfqd->active_queue;
2222         if (!cfqq)
2223                 goto new_queue;
2224
2225         if (!cfqd->rq_queued)
2226                 return NULL;
2227
2228         /*
2229          * We were waiting for group to get backlogged. Expire the queue
2230          */
2231         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2232                 goto expire;
2233
2234         /*
2235          * The active queue has run out of time, expire it and select new.
2236          */
2237         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2238                 /*
2239                  * If slice had not expired at the completion of last request
2240                  * we might not have turned on wait_busy flag. Don't expire
2241                  * the queue yet. Allow the group to get backlogged.
2242                  *
2243                  * The very fact that we have used the slice, that means we
2244                  * have been idling all along on this queue and it should be
2245                  * ok to wait for this request to complete.
2246                  */
2247                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2248                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2249                         cfqq = NULL;
2250                         goto keep_queue;
2251                 } else
2252                         goto check_group_idle;
2253         }
2254
2255         /*
2256          * The active queue has requests and isn't expired, allow it to
2257          * dispatch.
2258          */
2259         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2260                 goto keep_queue;
2261
2262         /*
2263          * If another queue has a request waiting within our mean seek
2264          * distance, let it run.  The expire code will check for close
2265          * cooperators and put the close queue at the front of the service
2266          * tree.  If possible, merge the expiring queue with the new cfqq.
2267          */
2268         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2269         if (new_cfqq) {
2270                 if (!cfqq->new_cfqq)
2271                         cfq_setup_merge(cfqq, new_cfqq);
2272                 goto expire;
2273         }
2274
2275         /*
2276          * No requests pending. If the active queue still has requests in
2277          * flight or is idling for a new request, allow either of these
2278          * conditions to happen (or time out) before selecting a new queue.
2279          */
2280         if (timer_pending(&cfqd->idle_slice_timer)) {
2281                 cfqq = NULL;
2282                 goto keep_queue;
2283         }
2284
2285         /*
2286          * This is a deep seek queue, but the device is much faster than
2287          * the queue can deliver, don't idle
2288          **/
2289         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2290             (cfq_cfqq_slice_new(cfqq) ||
2291             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2292                 cfq_clear_cfqq_deep(cfqq);
2293                 cfq_clear_cfqq_idle_window(cfqq);
2294         }
2295
2296         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2297                 cfqq = NULL;
2298                 goto keep_queue;
2299         }
2300
2301         /*
2302          * If group idle is enabled and there are requests dispatched from
2303          * this group, wait for requests to complete.
2304          */
2305 check_group_idle:
2306         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2307             && cfqq->cfqg->dispatched) {
2308                 cfqq = NULL;
2309                 goto keep_queue;
2310         }
2311
2312 expire:
2313         cfq_slice_expired(cfqd, 0);
2314 new_queue:
2315         /*
2316          * Current queue expired. Check if we have to switch to a new
2317          * service tree
2318          */
2319         if (!new_cfqq)
2320                 cfq_choose_cfqg(cfqd);
2321
2322         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2323 keep_queue:
2324         return cfqq;
2325 }
2326
2327 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2328 {
2329         int dispatched = 0;
2330
2331         while (cfqq->next_rq) {
2332                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2333                 dispatched++;
2334         }
2335
2336         BUG_ON(!list_empty(&cfqq->fifo));
2337
2338         /* By default cfqq is not expired if it is empty. Do it explicitly */
2339         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2340         return dispatched;
2341 }
2342
2343 /*
2344  * Drain our current requests. Used for barriers and when switching
2345  * io schedulers on-the-fly.
2346  */
2347 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2348 {
2349         struct cfq_queue *cfqq;
2350         int dispatched = 0;
2351
2352         /* Expire the timeslice of the current active queue first */
2353         cfq_slice_expired(cfqd, 0);
2354         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2355                 __cfq_set_active_queue(cfqd, cfqq);
2356                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2357         }
2358
2359         BUG_ON(cfqd->busy_queues);
2360
2361         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2362         return dispatched;
2363 }
2364
2365 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2366         struct cfq_queue *cfqq)
2367 {
2368         /* the queue hasn't finished any request, can't estimate */
2369         if (cfq_cfqq_slice_new(cfqq))
2370                 return true;
2371         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2372                 cfqq->slice_end))
2373                 return true;
2374
2375         return false;
2376 }
2377
2378 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2379 {
2380         unsigned int max_dispatch;
2381
2382         /*
2383          * Drain async requests before we start sync IO
2384          */
2385         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2386                 return false;
2387
2388         /*
2389          * If this is an async queue and we have sync IO in flight, let it wait
2390          */
2391         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2392                 return false;
2393
2394         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2395         if (cfq_class_idle(cfqq))
2396                 max_dispatch = 1;
2397
2398         /*
2399          * Does this cfqq already have too much IO in flight?
2400          */
2401         if (cfqq->dispatched >= max_dispatch) {
2402                 /*
2403                  * idle queue must always only have a single IO in flight
2404                  */
2405                 if (cfq_class_idle(cfqq))
2406                         return false;
2407
2408                 /*
2409                  * We have other queues, don't allow more IO from this one
2410                  */
2411                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2412                         return false;
2413
2414                 /*
2415                  * Sole queue user, no limit
2416                  */
2417                 if (cfqd->busy_queues == 1)
2418                         max_dispatch = -1;
2419                 else
2420                         /*
2421                          * Normally we start throttling cfqq when cfq_quantum/2
2422                          * requests have been dispatched. But we can drive
2423                          * deeper queue depths at the beginning of slice
2424                          * subjected to upper limit of cfq_quantum.
2425                          * */
2426                         max_dispatch = cfqd->cfq_quantum;
2427         }
2428
2429         /*
2430          * Async queues must wait a bit before being allowed dispatch.
2431          * We also ramp up the dispatch depth gradually for async IO,
2432          * based on the last sync IO we serviced
2433          */
2434         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2435                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2436                 unsigned int depth;
2437
2438                 depth = last_sync / cfqd->cfq_slice[1];
2439                 if (!depth && !cfqq->dispatched)
2440                         depth = 1;
2441                 if (depth < max_dispatch)
2442                         max_dispatch = depth;
2443         }
2444
2445         /*
2446          * If we're below the current max, allow a dispatch
2447          */
2448         return cfqq->dispatched < max_dispatch;
2449 }
2450
2451 /*
2452  * Dispatch a request from cfqq, moving them to the request queue
2453  * dispatch list.
2454  */
2455 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2456 {
2457         struct request *rq;
2458
2459         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2460
2461         if (!cfq_may_dispatch(cfqd, cfqq))
2462                 return false;
2463
2464         /*
2465          * follow expired path, else get first next available
2466          */
2467         rq = cfq_check_fifo(cfqq);
2468         if (!rq)
2469                 rq = cfqq->next_rq;
2470
2471         /*
2472          * insert request into driver dispatch list
2473          */
2474         cfq_dispatch_insert(cfqd->queue, rq);
2475
2476         if (!cfqd->active_cic) {
2477                 struct cfq_io_context *cic = RQ_CIC(rq);
2478
2479                 atomic_long_inc(&cic->ioc->refcount);
2480                 cfqd->active_cic = cic;
2481         }
2482
2483         return true;
2484 }
2485
2486 /*
2487  * Find the cfqq that we need to service and move a request from that to the
2488  * dispatch list
2489  */
2490 static int cfq_dispatch_requests(struct request_queue *q, int force)
2491 {
2492         struct cfq_data *cfqd = q->elevator->elevator_data;
2493         struct cfq_queue *cfqq;
2494
2495         if (!cfqd->busy_queues)
2496                 return 0;
2497
2498         if (unlikely(force))
2499                 return cfq_forced_dispatch(cfqd);
2500
2501         cfqq = cfq_select_queue(cfqd);
2502         if (!cfqq)
2503                 return 0;
2504
2505         /*
2506          * Dispatch a request from this cfqq, if it is allowed
2507          */
2508         if (!cfq_dispatch_request(cfqd, cfqq))
2509                 return 0;
2510
2511         cfqq->slice_dispatch++;
2512         cfq_clear_cfqq_must_dispatch(cfqq);
2513
2514         /*
2515          * expire an async queue immediately if it has used up its slice. idle
2516          * queue always expire after 1 dispatch round.
2517          */
2518         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2519             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2520             cfq_class_idle(cfqq))) {
2521                 cfqq->slice_end = jiffies + 1;
2522                 cfq_slice_expired(cfqd, 0);
2523         }
2524
2525         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2526         return 1;
2527 }
2528
2529 /*
2530  * task holds one reference to the queue, dropped when task exits. each rq
2531  * in-flight on this queue also holds a reference, dropped when rq is freed.
2532  *
2533  * Each cfq queue took a reference on the parent group. Drop it now.
2534  * queue lock must be held here.
2535  */
2536 static void cfq_put_queue(struct cfq_queue *cfqq)
2537 {
2538         struct cfq_data *cfqd = cfqq->cfqd;
2539         struct cfq_group *cfqg, *orig_cfqg;
2540
2541         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2542
2543         if (!atomic_dec_and_test(&cfqq->ref))
2544                 return;
2545
2546         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2547         BUG_ON(rb_first(&cfqq->sort_list));
2548         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2549         cfqg = cfqq->cfqg;
2550         orig_cfqg = cfqq->orig_cfqg;
2551
2552         if (unlikely(cfqd->active_queue == cfqq)) {
2553                 __cfq_slice_expired(cfqd, cfqq, 0);
2554                 cfq_schedule_dispatch(cfqd);
2555         }
2556
2557         BUG_ON(cfq_cfqq_on_rr(cfqq));
2558         kmem_cache_free(cfq_pool, cfqq);
2559         cfq_put_cfqg(cfqg);
2560         if (orig_cfqg)
2561                 cfq_put_cfqg(orig_cfqg);
2562 }
2563
2564 /*
2565  * Must always be called with the rcu_read_lock() held
2566  */
2567 static void
2568 __call_for_each_cic(struct io_context *ioc,
2569                     void (*func)(struct io_context *, struct cfq_io_context *))
2570 {
2571         struct cfq_io_context *cic;
2572         struct hlist_node *n;
2573
2574         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2575                 func(ioc, cic);
2576 }
2577
2578 /*
2579  * Call func for each cic attached to this ioc.
2580  */
2581 static void
2582 call_for_each_cic(struct io_context *ioc,
2583                   void (*func)(struct io_context *, struct cfq_io_context *))
2584 {
2585         rcu_read_lock();
2586         __call_for_each_cic(ioc, func);
2587         rcu_read_unlock();
2588 }
2589
2590 static void cfq_cic_free_rcu(struct rcu_head *head)
2591 {
2592         struct cfq_io_context *cic;
2593
2594         cic = container_of(head, struct cfq_io_context, rcu_head);
2595
2596         kmem_cache_free(cfq_ioc_pool, cic);
2597         elv_ioc_count_dec(cfq_ioc_count);
2598
2599         if (ioc_gone) {
2600                 /*
2601                  * CFQ scheduler is exiting, grab exit lock and check
2602                  * the pending io context count. If it hits zero,
2603                  * complete ioc_gone and set it back to NULL
2604                  */
2605                 spin_lock(&ioc_gone_lock);
2606                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2607                         complete(ioc_gone);
2608                         ioc_gone = NULL;
2609                 }
2610                 spin_unlock(&ioc_gone_lock);
2611         }
2612 }
2613
2614 static void cfq_cic_free(struct cfq_io_context *cic)
2615 {
2616         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2617 }
2618
2619 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2620 {
2621         unsigned long flags;
2622         unsigned long dead_key = (unsigned long) cic->key;
2623
2624         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2625
2626         spin_lock_irqsave(&ioc->lock, flags);
2627         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2628         hlist_del_rcu(&cic->cic_list);
2629         spin_unlock_irqrestore(&ioc->lock, flags);
2630
2631         cfq_cic_free(cic);
2632 }
2633
2634 /*
2635  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2636  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2637  * and ->trim() which is called with the task lock held
2638  */
2639 static void cfq_free_io_context(struct io_context *ioc)
2640 {
2641         /*
2642          * ioc->refcount is zero here, or we are called from elv_unregister(),
2643          * so no more cic's are allowed to be linked into this ioc.  So it
2644          * should be ok to iterate over the known list, we will see all cic's
2645          * since no new ones are added.
2646          */
2647         __call_for_each_cic(ioc, cic_free_func);
2648 }
2649
2650 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2651 {
2652         struct cfq_queue *__cfqq, *next;
2653
2654         /*
2655          * If this queue was scheduled to merge with another queue, be
2656          * sure to drop the reference taken on that queue (and others in
2657          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2658          */
2659         __cfqq = cfqq->new_cfqq;
2660         while (__cfqq) {
2661                 if (__cfqq == cfqq) {
2662                         WARN(1, "cfqq->new_cfqq loop detected\n");
2663                         break;
2664                 }
2665                 next = __cfqq->new_cfqq;
2666                 cfq_put_queue(__cfqq);
2667                 __cfqq = next;
2668         }
2669 }
2670
2671 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2672 {
2673         if (unlikely(cfqq == cfqd->active_queue)) {
2674                 __cfq_slice_expired(cfqd, cfqq, 0);
2675                 cfq_schedule_dispatch(cfqd);
2676         }
2677
2678         cfq_put_cooperator(cfqq);
2679
2680         cfq_put_queue(cfqq);
2681 }
2682
2683 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2684                                          struct cfq_io_context *cic)
2685 {
2686         struct io_context *ioc = cic->ioc;
2687
2688         list_del_init(&cic->queue_list);
2689
2690         /*
2691          * Make sure dead mark is seen for dead queues
2692          */
2693         smp_wmb();
2694         cic->key = cfqd_dead_key(cfqd);
2695
2696         if (ioc->ioc_data == cic)
2697                 rcu_assign_pointer(ioc->ioc_data, NULL);
2698
2699         if (cic->cfqq[BLK_RW_ASYNC]) {
2700                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2701                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2702         }
2703
2704         if (cic->cfqq[BLK_RW_SYNC]) {
2705                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2706                 cic->cfqq[BLK_RW_SYNC] = NULL;
2707         }
2708 }
2709
2710 static void cfq_exit_single_io_context(struct io_context *ioc,
2711                                        struct cfq_io_context *cic)
2712 {
2713         struct cfq_data *cfqd = cic_to_cfqd(cic);
2714
2715         if (cfqd) {
2716                 struct request_queue *q = cfqd->queue;
2717                 unsigned long flags;
2718
2719                 spin_lock_irqsave(q->queue_lock, flags);
2720
2721                 /*
2722                  * Ensure we get a fresh copy of the ->key to prevent
2723                  * race between exiting task and queue
2724                  */
2725                 smp_read_barrier_depends();
2726                 if (cic->key == cfqd)
2727                         __cfq_exit_single_io_context(cfqd, cic);
2728
2729                 spin_unlock_irqrestore(q->queue_lock, flags);
2730         }
2731 }
2732
2733 /*
2734  * The process that ioc belongs to has exited, we need to clean up
2735  * and put the internal structures we have that belongs to that process.
2736  */
2737 static void cfq_exit_io_context(struct io_context *ioc)
2738 {
2739         call_for_each_cic(ioc, cfq_exit_single_io_context);
2740 }
2741
2742 static struct cfq_io_context *
2743 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2744 {
2745         struct cfq_io_context *cic;
2746
2747         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2748                                                         cfqd->queue->node);
2749         if (cic) {
2750                 cic->last_end_request = jiffies;
2751                 INIT_LIST_HEAD(&cic->queue_list);
2752                 INIT_HLIST_NODE(&cic->cic_list);
2753                 cic->dtor = cfq_free_io_context;
2754                 cic->exit = cfq_exit_io_context;
2755                 elv_ioc_count_inc(cfq_ioc_count);
2756         }
2757
2758         return cic;
2759 }
2760
2761 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2762 {
2763         struct task_struct *tsk = current;
2764         int ioprio_class;
2765
2766         if (!cfq_cfqq_prio_changed(cfqq))
2767                 return;
2768
2769         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2770         switch (ioprio_class) {
2771         default:
2772                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2773         case IOPRIO_CLASS_NONE:
2774                 /*
2775                  * no prio set, inherit CPU scheduling settings
2776                  */
2777                 cfqq->ioprio = task_nice_ioprio(tsk);
2778                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2779                 break;
2780         case IOPRIO_CLASS_RT:
2781                 cfqq->ioprio = task_ioprio(ioc);
2782                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2783                 break;
2784         case IOPRIO_CLASS_BE:
2785                 cfqq->ioprio = task_ioprio(ioc);
2786                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2787                 break;
2788         case IOPRIO_CLASS_IDLE:
2789                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2790                 cfqq->ioprio = 7;
2791                 cfq_clear_cfqq_idle_window(cfqq);
2792                 break;
2793         }
2794
2795         /*
2796          * keep track of original prio settings in case we have to temporarily
2797          * elevate the priority of this queue
2798          */
2799         cfqq->org_ioprio = cfqq->ioprio;
2800         cfqq->org_ioprio_class = cfqq->ioprio_class;
2801         cfq_clear_cfqq_prio_changed(cfqq);
2802 }
2803
2804 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2805 {
2806         struct cfq_data *cfqd = cic_to_cfqd(cic);
2807         struct cfq_queue *cfqq;
2808         unsigned long flags;
2809
2810         if (unlikely(!cfqd))
2811                 return;
2812
2813         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2814
2815         cfqq = cic->cfqq[BLK_RW_ASYNC];
2816         if (cfqq) {
2817                 struct cfq_queue *new_cfqq;
2818                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2819                                                 GFP_ATOMIC);
2820                 if (new_cfqq) {
2821                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2822                         cfq_put_queue(cfqq);
2823                 }
2824         }
2825
2826         cfqq = cic->cfqq[BLK_RW_SYNC];
2827         if (cfqq)
2828                 cfq_mark_cfqq_prio_changed(cfqq);
2829
2830         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2831 }
2832
2833 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2834 {
2835         call_for_each_cic(ioc, changed_ioprio);
2836         ioc->ioprio_changed = 0;
2837 }
2838
2839 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2840                           pid_t pid, bool is_sync)
2841 {
2842         RB_CLEAR_NODE(&cfqq->rb_node);
2843         RB_CLEAR_NODE(&cfqq->p_node);
2844         INIT_LIST_HEAD(&cfqq->fifo);
2845
2846         atomic_set(&cfqq->ref, 0);
2847         cfqq->cfqd = cfqd;
2848
2849         cfq_mark_cfqq_prio_changed(cfqq);
2850
2851         if (is_sync) {
2852                 if (!cfq_class_idle(cfqq))
2853                         cfq_mark_cfqq_idle_window(cfqq);
2854                 cfq_mark_cfqq_sync(cfqq);
2855         }
2856         cfqq->pid = pid;
2857 }
2858
2859 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2860 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2861 {
2862         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2863         struct cfq_data *cfqd = cic_to_cfqd(cic);
2864         unsigned long flags;
2865         struct request_queue *q;
2866
2867         if (unlikely(!cfqd))
2868                 return;
2869
2870         q = cfqd->queue;
2871
2872         spin_lock_irqsave(q->queue_lock, flags);
2873
2874         if (sync_cfqq) {
2875                 /*
2876                  * Drop reference to sync queue. A new sync queue will be
2877                  * assigned in new group upon arrival of a fresh request.
2878                  */
2879                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2880                 cic_set_cfqq(cic, NULL, 1);
2881                 cfq_put_queue(sync_cfqq);
2882         }
2883
2884         spin_unlock_irqrestore(q->queue_lock, flags);
2885 }
2886
2887 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2888 {
2889         call_for_each_cic(ioc, changed_cgroup);
2890         ioc->cgroup_changed = 0;
2891 }
2892 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2893
2894 static struct cfq_queue *
2895 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2896                      struct io_context *ioc, gfp_t gfp_mask)
2897 {
2898         struct cfq_queue *cfqq, *new_cfqq = NULL;
2899         struct cfq_io_context *cic;
2900         struct cfq_group *cfqg;
2901
2902 retry:
2903         cfqg = cfq_get_cfqg(cfqd, 1);
2904         cic = cfq_cic_lookup(cfqd, ioc);
2905         /* cic always exists here */
2906         cfqq = cic_to_cfqq(cic, is_sync);
2907
2908         /*
2909          * Always try a new alloc if we fell back to the OOM cfqq
2910          * originally, since it should just be a temporary situation.
2911          */
2912         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2913                 cfqq = NULL;
2914                 if (new_cfqq) {
2915                         cfqq = new_cfqq;
2916                         new_cfqq = NULL;
2917                 } else if (gfp_mask & __GFP_WAIT) {
2918                         spin_unlock_irq(cfqd->queue->queue_lock);
2919                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2920                                         gfp_mask | __GFP_ZERO,
2921                                         cfqd->queue->node);
2922                         spin_lock_irq(cfqd->queue->queue_lock);
2923                         if (new_cfqq)
2924                                 goto retry;
2925                 } else {
2926                         cfqq = kmem_cache_alloc_node(cfq_pool,
2927                                         gfp_mask | __GFP_ZERO,
2928                                         cfqd->queue->node);
2929                 }
2930
2931                 if (cfqq) {
2932                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2933                         cfq_init_prio_data(cfqq, ioc);
2934                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2935                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2936                 } else
2937                         cfqq = &cfqd->oom_cfqq;
2938         }
2939
2940         if (new_cfqq)
2941                 kmem_cache_free(cfq_pool, new_cfqq);
2942
2943         return cfqq;
2944 }
2945
2946 static struct cfq_queue **
2947 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2948 {
2949         switch (ioprio_class) {
2950         case IOPRIO_CLASS_RT:
2951                 return &cfqd->async_cfqq[0][ioprio];
2952         case IOPRIO_CLASS_BE:
2953                 return &cfqd->async_cfqq[1][ioprio];
2954         case IOPRIO_CLASS_IDLE:
2955                 return &cfqd->async_idle_cfqq;
2956         default:
2957                 BUG();
2958         }
2959 }
2960
2961 static struct cfq_queue *
2962 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2963               gfp_t gfp_mask)
2964 {
2965         const int ioprio = task_ioprio(ioc);
2966         const int ioprio_class = task_ioprio_class(ioc);
2967         struct cfq_queue **async_cfqq = NULL;
2968         struct cfq_queue *cfqq = NULL;
2969
2970         if (!is_sync) {
2971                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2972                 cfqq = *async_cfqq;
2973         }
2974
2975         if (!cfqq)
2976                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2977
2978         /*
2979          * pin the queue now that it's allocated, scheduler exit will prune it
2980          */
2981         if (!is_sync && !(*async_cfqq)) {
2982                 atomic_inc(&cfqq->ref);
2983                 *async_cfqq = cfqq;
2984         }
2985
2986         atomic_inc(&cfqq->ref);
2987         return cfqq;
2988 }
2989
2990 /*
2991  * We drop cfq io contexts lazily, so we may find a dead one.
2992  */
2993 static void
2994 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2995                   struct cfq_io_context *cic)
2996 {
2997         unsigned long flags;
2998
2999         WARN_ON(!list_empty(&cic->queue_list));
3000         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3001
3002         spin_lock_irqsave(&ioc->lock, flags);
3003
3004         BUG_ON(ioc->ioc_data == cic);
3005
3006         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3007         hlist_del_rcu(&cic->cic_list);
3008         spin_unlock_irqrestore(&ioc->lock, flags);
3009
3010         cfq_cic_free(cic);
3011 }
3012
3013 static struct cfq_io_context *
3014 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3015 {
3016         struct cfq_io_context *cic;
3017         unsigned long flags;
3018
3019         if (unlikely(!ioc))
3020                 return NULL;
3021
3022         rcu_read_lock();
3023
3024         /*
3025          * we maintain a last-hit cache, to avoid browsing over the tree
3026          */
3027         cic = rcu_dereference(ioc->ioc_data);
3028         if (cic && cic->key == cfqd) {
3029                 rcu_read_unlock();
3030                 return cic;
3031         }
3032
3033         do {
3034                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3035                 rcu_read_unlock();
3036                 if (!cic)
3037                         break;
3038                 if (unlikely(cic->key != cfqd)) {
3039                         cfq_drop_dead_cic(cfqd, ioc, cic);
3040                         rcu_read_lock();
3041                         continue;
3042                 }
3043
3044                 spin_lock_irqsave(&ioc->lock, flags);
3045                 rcu_assign_pointer(ioc->ioc_data, cic);
3046                 spin_unlock_irqrestore(&ioc->lock, flags);
3047                 break;
3048         } while (1);
3049
3050         return cic;
3051 }
3052
3053 /*
3054  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3055  * the process specific cfq io context when entered from the block layer.
3056  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3057  */
3058 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3059                         struct cfq_io_context *cic, gfp_t gfp_mask)
3060 {
3061         unsigned long flags;
3062         int ret;
3063
3064         ret = radix_tree_preload(gfp_mask);
3065         if (!ret) {
3066                 cic->ioc = ioc;
3067                 cic->key = cfqd;
3068
3069                 spin_lock_irqsave(&ioc->lock, flags);
3070                 ret = radix_tree_insert(&ioc->radix_root,
3071                                                 cfqd->cic_index, cic);
3072                 if (!ret)
3073                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3074                 spin_unlock_irqrestore(&ioc->lock, flags);
3075
3076                 radix_tree_preload_end();
3077
3078                 if (!ret) {
3079                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3080                         list_add(&cic->queue_list, &cfqd->cic_list);
3081                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3082                 }
3083         }
3084
3085         if (ret)
3086                 printk(KERN_ERR "cfq: cic link failed!\n");
3087
3088         return ret;
3089 }
3090
3091 /*
3092  * Setup general io context and cfq io context. There can be several cfq
3093  * io contexts per general io context, if this process is doing io to more
3094  * than one device managed by cfq.
3095  */
3096 static struct cfq_io_context *
3097 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3098 {
3099         struct io_context *ioc = NULL;
3100         struct cfq_io_context *cic;
3101
3102         might_sleep_if(gfp_mask & __GFP_WAIT);
3103
3104         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3105         if (!ioc)
3106                 return NULL;
3107
3108         cic = cfq_cic_lookup(cfqd, ioc);
3109         if (cic)
3110                 goto out;
3111
3112         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3113         if (cic == NULL)
3114                 goto err;
3115
3116         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3117                 goto err_free;
3118
3119 out:
3120         smp_read_barrier_depends();
3121         if (unlikely(ioc->ioprio_changed))
3122                 cfq_ioc_set_ioprio(ioc);
3123
3124 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3125         if (unlikely(ioc->cgroup_changed))
3126                 cfq_ioc_set_cgroup(ioc);
3127 #endif
3128         return cic;
3129 err_free:
3130         cfq_cic_free(cic);
3131 err:
3132         put_io_context(ioc);
3133         return NULL;
3134 }
3135
3136 static void
3137 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3138 {
3139         unsigned long elapsed = jiffies - cic->last_end_request;
3140         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3141
3142         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3143         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3144         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3145 }
3146
3147 static void
3148 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3149                        struct request *rq)
3150 {
3151         sector_t sdist = 0;
3152         sector_t n_sec = blk_rq_sectors(rq);
3153         if (cfqq->last_request_pos) {
3154                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3155                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3156                 else
3157                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3158         }
3159
3160         cfqq->seek_history <<= 1;
3161         if (blk_queue_nonrot(cfqd->queue))
3162                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3163         else
3164                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3165 }
3166
3167 /*
3168  * Disable idle window if the process thinks too long or seeks so much that
3169  * it doesn't matter
3170  */
3171 static void
3172 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3173                        struct cfq_io_context *cic)
3174 {
3175         int old_idle, enable_idle;
3176
3177         /*
3178          * Don't idle for async or idle io prio class
3179          */
3180         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3181                 return;
3182
3183         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3184
3185         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3186                 cfq_mark_cfqq_deep(cfqq);
3187
3188         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3189                 enable_idle = 0;
3190         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3191             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3192                 enable_idle = 0;
3193         else if (sample_valid(cic->ttime_samples)) {
3194                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3195                         enable_idle = 0;
3196                 else
3197                         enable_idle = 1;
3198         }
3199
3200         if (old_idle != enable_idle) {
3201                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3202                 if (enable_idle)
3203                         cfq_mark_cfqq_idle_window(cfqq);
3204                 else
3205                         cfq_clear_cfqq_idle_window(cfqq);
3206         }
3207 }
3208
3209 /*
3210  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3211  * no or if we aren't sure, a 1 will cause a preempt.
3212  */
3213 static bool
3214 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3215                    struct request *rq)
3216 {
3217         struct cfq_queue *cfqq;
3218
3219         cfqq = cfqd->active_queue;
3220         if (!cfqq)
3221                 return false;
3222
3223         if (cfq_class_idle(new_cfqq))
3224                 return false;
3225
3226         if (cfq_class_idle(cfqq))
3227                 return true;
3228
3229         /*
3230          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3231          */
3232         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3233                 return false;
3234
3235         /*
3236          * if the new request is sync, but the currently running queue is
3237          * not, let the sync request have priority.
3238          */
3239         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3240                 return true;
3241
3242         if (new_cfqq->cfqg != cfqq->cfqg)
3243                 return false;
3244
3245         if (cfq_slice_used(cfqq))
3246                 return true;
3247
3248         /* Allow preemption only if we are idling on sync-noidle tree */
3249         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3250             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3251             new_cfqq->service_tree->count == 2 &&
3252             RB_EMPTY_ROOT(&cfqq->sort_list))
3253                 return true;
3254
3255         /*
3256          * So both queues are sync. Let the new request get disk time if
3257          * it's a metadata request and the current queue is doing regular IO.
3258          */
3259         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3260                 return true;
3261
3262         /*
3263          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3264          */
3265         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3266                 return true;
3267
3268         /* An idle queue should not be idle now for some reason */
3269         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3270                 return true;
3271
3272         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3273                 return false;
3274
3275         /*
3276          * if this request is as-good as one we would expect from the
3277          * current cfqq, let it preempt
3278          */
3279         if (cfq_rq_close(cfqd, cfqq, rq))
3280                 return true;
3281
3282         return false;
3283 }
3284
3285 /*
3286  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3287  * let it have half of its nominal slice.
3288  */
3289 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3290 {
3291         cfq_log_cfqq(cfqd, cfqq, "preempt");
3292         cfq_slice_expired(cfqd, 1);
3293
3294         /*
3295          * Put the new queue at the front of the of the current list,
3296          * so we know that it will be selected next.
3297          */
3298         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3299
3300         cfq_service_tree_add(cfqd, cfqq, 1);
3301
3302         cfqq->slice_end = 0;
3303         cfq_mark_cfqq_slice_new(cfqq);
3304 }
3305
3306 /*
3307  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3308  * something we should do about it
3309  */
3310 static void
3311 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3312                 struct request *rq)
3313 {
3314         struct cfq_io_context *cic = RQ_CIC(rq);
3315
3316         cfqd->rq_queued++;
3317         if (rq->cmd_flags & REQ_META)
3318                 cfqq->meta_pending++;
3319
3320         cfq_update_io_thinktime(cfqd, cic);
3321         cfq_update_io_seektime(cfqd, cfqq, rq);
3322         cfq_update_idle_window(cfqd, cfqq, cic);
3323
3324         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3325
3326         if (cfqq == cfqd->active_queue) {
3327                 /*
3328                  * Remember that we saw a request from this process, but
3329                  * don't start queuing just yet. Otherwise we risk seeing lots
3330                  * of tiny requests, because we disrupt the normal plugging
3331                  * and merging. If the request is already larger than a single
3332                  * page, let it rip immediately. For that case we assume that
3333                  * merging is already done. Ditto for a busy system that
3334                  * has other work pending, don't risk delaying until the
3335                  * idle timer unplug to continue working.
3336                  */
3337                 if (cfq_cfqq_wait_request(cfqq)) {
3338                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3339                             cfqd->busy_queues > 1) {
3340                                 cfq_del_timer(cfqd, cfqq);
3341                                 cfq_clear_cfqq_wait_request(cfqq);
3342                                 __blk_run_queue(cfqd->queue);
3343                         } else {
3344                                 cfq_blkiocg_update_idle_time_stats(
3345                                                 &cfqq->cfqg->blkg);
3346                                 cfq_mark_cfqq_must_dispatch(cfqq);
3347                         }
3348                 }
3349         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3350                 /*
3351                  * not the active queue - expire current slice if it is
3352                  * idle and has expired it's mean thinktime or this new queue
3353                  * has some old slice time left and is of higher priority or
3354                  * this new queue is RT and the current one is BE
3355                  */
3356                 cfq_preempt_queue(cfqd, cfqq);
3357                 __blk_run_queue(cfqd->queue);
3358         }
3359 }
3360
3361 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3362 {
3363         struct cfq_data *cfqd = q->elevator->elevator_data;
3364         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3365
3366         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3367         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3368
3369         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3370         list_add_tail(&rq->queuelist, &cfqq->fifo);
3371         cfq_add_rq_rb(rq);
3372         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3373                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3374                         rq_is_sync(rq));
3375         cfq_rq_enqueued(cfqd, cfqq, rq);
3376 }
3377
3378 /*
3379  * Update hw_tag based on peak queue depth over 50 samples under
3380  * sufficient load.
3381  */
3382 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3383 {
3384         struct cfq_queue *cfqq = cfqd->active_queue;
3385
3386         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3387                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3388
3389         if (cfqd->hw_tag == 1)
3390                 return;
3391
3392         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3393             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3394                 return;
3395
3396         /*
3397          * If active queue hasn't enough requests and can idle, cfq might not
3398          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3399          * case
3400          */
3401         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3402             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3403             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3404                 return;
3405
3406         if (cfqd->hw_tag_samples++ < 50)
3407                 return;
3408
3409         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3410                 cfqd->hw_tag = 1;
3411         else
3412                 cfqd->hw_tag = 0;
3413 }
3414
3415 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3416 {
3417         struct cfq_io_context *cic = cfqd->active_cic;
3418
3419         /* If there are other queues in the group, don't wait */
3420         if (cfqq->cfqg->nr_cfqq > 1)
3421                 return false;
3422
3423         if (cfq_slice_used(cfqq))
3424                 return true;
3425
3426         /* if slice left is less than think time, wait busy */
3427         if (cic && sample_valid(cic->ttime_samples)
3428             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3429                 return true;
3430
3431         /*
3432          * If think times is less than a jiffy than ttime_mean=0 and above
3433          * will not be true. It might happen that slice has not expired yet
3434          * but will expire soon (4-5 ns) during select_queue(). To cover the
3435          * case where think time is less than a jiffy, mark the queue wait
3436          * busy if only 1 jiffy is left in the slice.
3437          */
3438         if (cfqq->slice_end - jiffies == 1)
3439                 return true;
3440
3441         return false;
3442 }
3443
3444 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3445 {
3446         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3447         struct cfq_data *cfqd = cfqq->cfqd;
3448         const int sync = rq_is_sync(rq);
3449         unsigned long now;
3450
3451         now = jiffies;
3452         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3453                      !!(rq->cmd_flags & REQ_NOIDLE));
3454
3455         cfq_update_hw_tag(cfqd);
3456
3457         WARN_ON(!cfqd->rq_in_driver);
3458         WARN_ON(!cfqq->dispatched);
3459         cfqd->rq_in_driver--;
3460         cfqq->dispatched--;
3461         (RQ_CFQG(rq))->dispatched--;
3462         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3463                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3464                         rq_data_dir(rq), rq_is_sync(rq));
3465
3466         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3467
3468         if (sync) {
3469                 RQ_CIC(rq)->last_end_request = now;
3470                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3471                         cfqd->last_delayed_sync = now;
3472         }
3473
3474         /*
3475          * If this is the active queue, check if it needs to be expired,
3476          * or if we want to idle in case it has no pending requests.
3477          */
3478         if (cfqd->active_queue == cfqq) {
3479                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3480
3481                 if (cfq_cfqq_slice_new(cfqq)) {
3482                         cfq_set_prio_slice(cfqd, cfqq);
3483                         cfq_clear_cfqq_slice_new(cfqq);
3484                 }
3485
3486                 /*
3487                  * Should we wait for next request to come in before we expire
3488                  * the queue.
3489                  */
3490                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3491                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3492                         if (!cfqd->cfq_slice_idle)
3493                                 extend_sl = cfqd->cfq_group_idle;
3494                         cfqq->slice_end = jiffies + extend_sl;
3495                         cfq_mark_cfqq_wait_busy(cfqq);
3496                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3497                 }
3498
3499                 /*
3500                  * Idling is not enabled on:
3501                  * - expired queues
3502                  * - idle-priority queues
3503                  * - async queues
3504                  * - queues with still some requests queued
3505                  * - when there is a close cooperator
3506                  */
3507                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3508                         cfq_slice_expired(cfqd, 1);
3509                 else if (sync && cfqq_empty &&
3510                          !cfq_close_cooperator(cfqd, cfqq)) {
3511                         cfq_arm_slice_timer(cfqd);
3512                 }
3513         }
3514
3515         if (!cfqd->rq_in_driver)
3516                 cfq_schedule_dispatch(cfqd);
3517 }
3518
3519 /*
3520  * we temporarily boost lower priority queues if they are holding fs exclusive
3521  * resources. they are boosted to normal prio (CLASS_BE/4)
3522  */
3523 static void cfq_prio_boost(struct cfq_queue *cfqq)
3524 {
3525         if (has_fs_excl()) {
3526                 /*
3527                  * boost idle prio on transactions that would lock out other
3528                  * users of the filesystem
3529                  */
3530                 if (cfq_class_idle(cfqq))
3531                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3532                 if (cfqq->ioprio > IOPRIO_NORM)
3533                         cfqq->ioprio = IOPRIO_NORM;
3534         } else {
3535                 /*
3536                  * unboost the queue (if needed)
3537                  */
3538                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3539                 cfqq->ioprio = cfqq->org_ioprio;
3540         }
3541 }
3542
3543 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3544 {
3545         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3546                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3547                 return ELV_MQUEUE_MUST;
3548         }
3549
3550         return ELV_MQUEUE_MAY;
3551 }
3552
3553 static int cfq_may_queue(struct request_queue *q, int rw)
3554 {
3555         struct cfq_data *cfqd = q->elevator->elevator_data;
3556         struct task_struct *tsk = current;
3557         struct cfq_io_context *cic;
3558         struct cfq_queue *cfqq;
3559
3560         /*
3561          * don't force setup of a queue from here, as a call to may_queue
3562          * does not necessarily imply that a request actually will be queued.
3563          * so just lookup a possibly existing queue, or return 'may queue'
3564          * if that fails
3565          */
3566         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3567         if (!cic)
3568                 return ELV_MQUEUE_MAY;
3569
3570         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3571         if (cfqq) {
3572                 cfq_init_prio_data(cfqq, cic->ioc);
3573                 cfq_prio_boost(cfqq);
3574
3575                 return __cfq_may_queue(cfqq);
3576         }
3577
3578         return ELV_MQUEUE_MAY;
3579 }
3580
3581 /*
3582  * queue lock held here
3583  */
3584 static void cfq_put_request(struct request *rq)
3585 {
3586         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3587
3588         if (cfqq) {
3589                 const int rw = rq_data_dir(rq);
3590
3591                 BUG_ON(!cfqq->allocated[rw]);
3592                 cfqq->allocated[rw]--;
3593
3594                 put_io_context(RQ_CIC(rq)->ioc);
3595
3596                 rq->elevator_private = NULL;
3597                 rq->elevator_private2 = NULL;
3598
3599                 /* Put down rq reference on cfqg */
3600                 cfq_put_cfqg(RQ_CFQG(rq));
3601                 rq->elevator_private3 = NULL;
3602
3603                 cfq_put_queue(cfqq);
3604         }
3605 }
3606
3607 static struct cfq_queue *
3608 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3609                 struct cfq_queue *cfqq)
3610 {
3611         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3612         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3613         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3614         cfq_put_queue(cfqq);
3615         return cic_to_cfqq(cic, 1);
3616 }
3617
3618 /*
3619  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3620  * was the last process referring to said cfqq.
3621  */
3622 static struct cfq_queue *
3623 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3624 {
3625         if (cfqq_process_refs(cfqq) == 1) {
3626                 cfqq->pid = current->pid;
3627                 cfq_clear_cfqq_coop(cfqq);
3628                 cfq_clear_cfqq_split_coop(cfqq);
3629                 return cfqq;
3630         }
3631
3632         cic_set_cfqq(cic, NULL, 1);
3633
3634         cfq_put_cooperator(cfqq);
3635
3636         cfq_put_queue(cfqq);
3637         return NULL;
3638 }
3639 /*
3640  * Allocate cfq data structures associated with this request.
3641  */
3642 static int
3643 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3644 {
3645         struct cfq_data *cfqd = q->elevator->elevator_data;
3646         struct cfq_io_context *cic;
3647         const int rw = rq_data_dir(rq);
3648         const bool is_sync = rq_is_sync(rq);
3649         struct cfq_queue *cfqq;
3650         unsigned long flags;
3651
3652         might_sleep_if(gfp_mask & __GFP_WAIT);
3653
3654         cic = cfq_get_io_context(cfqd, gfp_mask);
3655
3656         spin_lock_irqsave(q->queue_lock, flags);
3657
3658         if (!cic)
3659                 goto queue_fail;
3660
3661 new_queue:
3662         cfqq = cic_to_cfqq(cic, is_sync);
3663         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3664                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3665                 cic_set_cfqq(cic, cfqq, is_sync);
3666         } else {
3667                 /*
3668                  * If the queue was seeky for too long, break it apart.
3669                  */
3670                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3671                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3672                         cfqq = split_cfqq(cic, cfqq);
3673                         if (!cfqq)
3674                                 goto new_queue;
3675                 }
3676
3677                 /*
3678                  * Check to see if this queue is scheduled to merge with
3679                  * another, closely cooperating queue.  The merging of
3680                  * queues happens here as it must be done in process context.
3681                  * The reference on new_cfqq was taken in merge_cfqqs.
3682                  */
3683                 if (cfqq->new_cfqq)
3684                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3685         }
3686
3687         cfqq->allocated[rw]++;
3688         atomic_inc(&cfqq->ref);
3689
3690         spin_unlock_irqrestore(q->queue_lock, flags);
3691
3692         rq->elevator_private = cic;
3693         rq->elevator_private2 = cfqq;
3694         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3695         return 0;
3696
3697 queue_fail:
3698         if (cic)
3699                 put_io_context(cic->ioc);
3700
3701         cfq_schedule_dispatch(cfqd);
3702         spin_unlock_irqrestore(q->queue_lock, flags);
3703         cfq_log(cfqd, "set_request fail");
3704         return 1;
3705 }
3706
3707 static void cfq_kick_queue(struct work_struct *work)
3708 {
3709         struct cfq_data *cfqd =
3710                 container_of(work, struct cfq_data, unplug_work);
3711         struct request_queue *q = cfqd->queue;
3712
3713         spin_lock_irq(q->queue_lock);
3714         __blk_run_queue(cfqd->queue);
3715         spin_unlock_irq(q->queue_lock);
3716 }
3717
3718 /*
3719  * Timer running if the active_queue is currently idling inside its time slice
3720  */
3721 static void cfq_idle_slice_timer(unsigned long data)
3722 {
3723         struct cfq_data *cfqd = (struct cfq_data *) data;
3724         struct cfq_queue *cfqq;
3725         unsigned long flags;
3726         int timed_out = 1;
3727
3728         cfq_log(cfqd, "idle timer fired");
3729
3730         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3731
3732         cfqq = cfqd->active_queue;
3733         if (cfqq) {
3734                 timed_out = 0;
3735
3736                 /*
3737                  * We saw a request before the queue expired, let it through
3738                  */
3739                 if (cfq_cfqq_must_dispatch(cfqq))
3740                         goto out_kick;
3741
3742                 /*
3743                  * expired
3744                  */
3745                 if (cfq_slice_used(cfqq))
3746                         goto expire;
3747
3748                 /*
3749                  * only expire and reinvoke request handler, if there are
3750                  * other queues with pending requests
3751                  */
3752                 if (!cfqd->busy_queues)
3753                         goto out_cont;
3754
3755                 /*
3756                  * not expired and it has a request pending, let it dispatch
3757                  */
3758                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3759                         goto out_kick;
3760
3761                 /*
3762                  * Queue depth flag is reset only when the idle didn't succeed
3763                  */
3764                 cfq_clear_cfqq_deep(cfqq);
3765         }
3766 expire:
3767         cfq_slice_expired(cfqd, timed_out);
3768 out_kick:
3769         cfq_schedule_dispatch(cfqd);
3770 out_cont:
3771         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3772 }
3773
3774 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3775 {
3776         del_timer_sync(&cfqd->idle_slice_timer);
3777         cancel_work_sync(&cfqd->unplug_work);
3778 }
3779
3780 static void cfq_put_async_queues(struct cfq_data *cfqd)
3781 {
3782         int i;
3783
3784         for (i = 0; i < IOPRIO_BE_NR; i++) {
3785                 if (cfqd->async_cfqq[0][i])
3786                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3787                 if (cfqd->async_cfqq[1][i])
3788                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3789         }
3790
3791         if (cfqd->async_idle_cfqq)
3792                 cfq_put_queue(cfqd->async_idle_cfqq);
3793 }
3794
3795 static void cfq_cfqd_free(struct rcu_head *head)
3796 {
3797         kfree(container_of(head, struct cfq_data, rcu));
3798 }
3799
3800 static void cfq_exit_queue(struct elevator_queue *e)
3801 {
3802         struct cfq_data *cfqd = e->elevator_data;
3803         struct request_queue *q = cfqd->queue;
3804
3805         cfq_shutdown_timer_wq(cfqd);
3806
3807         spin_lock_irq(q->queue_lock);
3808
3809         if (cfqd->active_queue)
3810                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3811
3812         while (!list_empty(&cfqd->cic_list)) {
3813                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3814                                                         struct cfq_io_context,
3815                                                         queue_list);
3816
3817                 __cfq_exit_single_io_context(cfqd, cic);
3818         }
3819
3820         cfq_put_async_queues(cfqd);
3821         cfq_release_cfq_groups(cfqd);
3822         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3823
3824         spin_unlock_irq(q->queue_lock);
3825
3826         cfq_shutdown_timer_wq(cfqd);
3827
3828         spin_lock(&cic_index_lock);
3829         ida_remove(&cic_index_ida, cfqd->cic_index);
3830         spin_unlock(&cic_index_lock);
3831
3832         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3833         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3834 }
3835
3836 static int cfq_alloc_cic_index(void)
3837 {
3838         int index, error;
3839
3840         do {
3841                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3842                         return -ENOMEM;
3843
3844                 spin_lock(&cic_index_lock);
3845                 error = ida_get_new(&cic_index_ida, &index);
3846                 spin_unlock(&cic_index_lock);
3847                 if (error && error != -EAGAIN)
3848                         return error;
3849         } while (error);
3850
3851         return index;
3852 }
3853
3854 static void *cfq_init_queue(struct request_queue *q)
3855 {
3856         struct cfq_data *cfqd;
3857         int i, j;
3858         struct cfq_group *cfqg;
3859         struct cfq_rb_root *st;
3860
3861         i = cfq_alloc_cic_index();
3862         if (i < 0)
3863                 return NULL;
3864
3865         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3866         if (!cfqd)
3867                 return NULL;
3868
3869         cfqd->cic_index = i;
3870
3871         /* Init root service tree */
3872         cfqd->grp_service_tree = CFQ_RB_ROOT;
3873
3874         /* Init root group */
3875         cfqg = &cfqd->root_group;
3876         for_each_cfqg_st(cfqg, i, j, st)
3877                 *st = CFQ_RB_ROOT;
3878         RB_CLEAR_NODE(&cfqg->rb_node);
3879
3880         /* Give preference to root group over other groups */
3881         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3882
3883 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3884         /*
3885          * Take a reference to root group which we never drop. This is just
3886          * to make sure that cfq_put_cfqg() does not try to kfree root group
3887          */
3888         atomic_set(&cfqg->ref, 1);
3889         rcu_read_lock();
3890         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3891                                         (void *)cfqd, 0);
3892         rcu_read_unlock();
3893 #endif
3894         /*
3895          * Not strictly needed (since RB_ROOT just clears the node and we
3896          * zeroed cfqd on alloc), but better be safe in case someone decides
3897          * to add magic to the rb code
3898          */
3899         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3900                 cfqd->prio_trees[i] = RB_ROOT;
3901
3902         /*
3903          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3904          * Grab a permanent reference to it, so that the normal code flow
3905          * will not attempt to free it.
3906          */
3907         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3908         atomic_inc(&cfqd->oom_cfqq.ref);
3909         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3910
3911         INIT_LIST_HEAD(&cfqd->cic_list);
3912
3913         cfqd->queue = q;
3914
3915         init_timer(&cfqd->idle_slice_timer);
3916         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3917         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3918
3919         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3920
3921         cfqd->cfq_quantum = cfq_quantum;
3922         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3923         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3924         cfqd->cfq_back_max = cfq_back_max;
3925         cfqd->cfq_back_penalty = cfq_back_penalty;
3926         cfqd->cfq_slice[0] = cfq_slice_async;
3927         cfqd->cfq_slice[1] = cfq_slice_sync;
3928         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3929         cfqd->cfq_slice_idle = cfq_slice_idle;
3930         cfqd->cfq_group_idle = cfq_group_idle;
3931         cfqd->cfq_latency = 1;
3932         cfqd->cfq_group_isolation = 0;
3933         cfqd->hw_tag = -1;
3934         /*
3935          * we optimistically start assuming sync ops weren't delayed in last
3936          * second, in order to have larger depth for async operations.
3937          */
3938         cfqd->last_delayed_sync = jiffies - HZ;
3939         return cfqd;
3940 }
3941
3942 static void cfq_slab_kill(void)
3943 {
3944         /*
3945          * Caller already ensured that pending RCU callbacks are completed,
3946          * so we should have no busy allocations at this point.
3947          */
3948         if (cfq_pool)
3949                 kmem_cache_destroy(cfq_pool);
3950         if (cfq_ioc_pool)
3951                 kmem_cache_destroy(cfq_ioc_pool);
3952 }
3953
3954 static int __init cfq_slab_setup(void)
3955 {
3956         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3957         if (!cfq_pool)
3958                 goto fail;
3959
3960         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3961         if (!cfq_ioc_pool)
3962                 goto fail;
3963
3964         return 0;
3965 fail:
3966         cfq_slab_kill();
3967         return -ENOMEM;
3968 }
3969
3970 /*
3971  * sysfs parts below -->
3972  */
3973 static ssize_t
3974 cfq_var_show(unsigned int var, char *page)
3975 {
3976         return sprintf(page, "%d\n", var);
3977 }
3978
3979 static ssize_t
3980 cfq_var_store(unsigned int *var, const char *page, size_t count)
3981 {
3982         char *p = (char *) page;
3983
3984         *var = simple_strtoul(p, &p, 10);
3985         return count;
3986 }
3987
3988 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3989 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3990 {                                                                       \
3991         struct cfq_data *cfqd = e->elevator_data;                       \
3992         unsigned int __data = __VAR;                                    \
3993         if (__CONV)                                                     \
3994                 __data = jiffies_to_msecs(__data);                      \
3995         return cfq_var_show(__data, (page));                            \
3996 }
3997 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3998 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3999 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4000 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4001 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4002 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4003 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4004 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4005 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4006 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4007 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4008 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
4009 #undef SHOW_FUNCTION
4010
4011 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4012 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4013 {                                                                       \
4014         struct cfq_data *cfqd = e->elevator_data;                       \
4015         unsigned int __data;                                            \
4016         int ret = cfq_var_store(&__data, (page), count);                \
4017         if (__data < (MIN))                                             \
4018                 __data = (MIN);                                         \
4019         else if (__data > (MAX))                                        \
4020                 __data = (MAX);                                         \
4021         if (__CONV)                                                     \
4022                 *(__PTR) = msecs_to_jiffies(__data);                    \
4023         else                                                            \
4024                 *(__PTR) = __data;                                      \
4025         return ret;                                                     \
4026 }
4027 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4028 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4029                 UINT_MAX, 1);
4030 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4031                 UINT_MAX, 1);
4032 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4033 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4034                 UINT_MAX, 0);
4035 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4036 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4037 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4038 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4039 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4040                 UINT_MAX, 0);
4041 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4042 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
4043 #undef STORE_FUNCTION
4044
4045 #define CFQ_ATTR(name) \
4046         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4047
4048 static struct elv_fs_entry cfq_attrs[] = {
4049         CFQ_ATTR(quantum),
4050         CFQ_ATTR(fifo_expire_sync),
4051         CFQ_ATTR(fifo_expire_async),
4052         CFQ_ATTR(back_seek_max),
4053         CFQ_ATTR(back_seek_penalty),
4054         CFQ_ATTR(slice_sync),
4055         CFQ_ATTR(slice_async),
4056         CFQ_ATTR(slice_async_rq),
4057         CFQ_ATTR(slice_idle),
4058         CFQ_ATTR(group_idle),
4059         CFQ_ATTR(low_latency),
4060         CFQ_ATTR(group_isolation),
4061         __ATTR_NULL
4062 };
4063
4064 static struct elevator_type iosched_cfq = {
4065         .ops = {
4066                 .elevator_merge_fn =            cfq_merge,
4067                 .elevator_merged_fn =           cfq_merged_request,
4068                 .elevator_merge_req_fn =        cfq_merged_requests,
4069                 .elevator_allow_merge_fn =      cfq_allow_merge,
4070                 .elevator_bio_merged_fn =       cfq_bio_merged,
4071                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4072                 .elevator_add_req_fn =          cfq_insert_request,
4073                 .elevator_activate_req_fn =     cfq_activate_request,
4074                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4075                 .elevator_queue_empty_fn =      cfq_queue_empty,
4076                 .elevator_completed_req_fn =    cfq_completed_request,
4077                 .elevator_former_req_fn =       elv_rb_former_request,
4078                 .elevator_latter_req_fn =       elv_rb_latter_request,
4079                 .elevator_set_req_fn =          cfq_set_request,
4080                 .elevator_put_req_fn =          cfq_put_request,
4081                 .elevator_may_queue_fn =        cfq_may_queue,
4082                 .elevator_init_fn =             cfq_init_queue,
4083                 .elevator_exit_fn =             cfq_exit_queue,
4084                 .trim =                         cfq_free_io_context,
4085         },
4086         .elevator_attrs =       cfq_attrs,
4087         .elevator_name =        "cfq",
4088         .elevator_owner =       THIS_MODULE,
4089 };
4090
4091 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4092 static struct blkio_policy_type blkio_policy_cfq = {
4093         .ops = {
4094                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4095                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4096         },
4097         .plid = BLKIO_POLICY_PROP,
4098 };
4099 #else
4100 static struct blkio_policy_type blkio_policy_cfq;
4101 #endif
4102
4103 static int __init cfq_init(void)
4104 {
4105         /*
4106          * could be 0 on HZ < 1000 setups
4107          */
4108         if (!cfq_slice_async)
4109                 cfq_slice_async = 1;
4110         if (!cfq_slice_idle)
4111                 cfq_slice_idle = 1;
4112
4113 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4114         if (!cfq_group_idle)
4115                 cfq_group_idle = 1;
4116 #else
4117                 cfq_group_idle = 0;
4118 #endif
4119         if (cfq_slab_setup())
4120                 return -ENOMEM;
4121
4122         elv_register(&iosched_cfq);
4123         blkio_policy_register(&blkio_policy_cfq);
4124
4125         return 0;
4126 }
4127
4128 static void __exit cfq_exit(void)
4129 {
4130         DECLARE_COMPLETION_ONSTACK(all_gone);
4131         blkio_policy_unregister(&blkio_policy_cfq);
4132         elv_unregister(&iosched_cfq);
4133         ioc_gone = &all_gone;
4134         /* ioc_gone's update must be visible before reading ioc_count */
4135         smp_wmb();
4136
4137         /*
4138          * this also protects us from entering cfq_slab_kill() with
4139          * pending RCU callbacks
4140          */
4141         if (elv_ioc_count_read(cfq_ioc_count))
4142                 wait_for_completion(&all_gone);
4143         ida_destroy(&cic_index_ida);
4144         cfq_slab_kill();
4145 }
4146
4147 module_init(cfq_init);
4148 module_exit(cfq_exit);
4149
4150 MODULE_AUTHOR("Jens Axboe");
4151 MODULE_LICENSE("GPL");
4152 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");