block: Fix discard topology stacking and reporting
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181         unsigned int new_weight;
182         bool needs_update;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /*
188          * Per group busy queus average. Useful for workload slice calc. We
189          * create the array for each prio class but at run time it is used
190          * only for RT and BE class and slot for IDLE class remains unused.
191          * This is primarily done to avoid confusion and a gcc warning.
192          */
193         unsigned int busy_queues_avg[CFQ_PRIO_NR];
194         /*
195          * rr lists of queues with requests. We maintain service trees for
196          * RT and BE classes. These trees are subdivided in subclasses
197          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198          * class there is no subclassification and all the cfq queues go on
199          * a single tree service_tree_idle.
200          * Counts are embedded in the cfq_rb_root
201          */
202         struct cfq_rb_root service_trees[2][3];
203         struct cfq_rb_root service_tree_idle;
204
205         unsigned long saved_workload_slice;
206         enum wl_type_t saved_workload;
207         enum wl_prio_t saved_serving_prio;
208         struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210         struct hlist_node cfqd_node;
211         int ref;
212 #endif
213         /* number of requests that are on the dispatch list or inside driver */
214         int dispatched;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303         struct rcu_head rcu;
304 };
305
306 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
307
308 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
309                                             enum wl_prio_t prio,
310                                             enum wl_type_t type)
311 {
312         if (!cfqg)
313                 return NULL;
314
315         if (prio == IDLE_WORKLOAD)
316                 return &cfqg->service_tree_idle;
317
318         return &cfqg->service_trees[prio][type];
319 }
320
321 enum cfqq_state_flags {
322         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
323         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
324         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
325         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
326         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
327         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
328         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
329         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
330         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
331         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
332         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
333         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
334         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
335 };
336
337 #define CFQ_CFQQ_FNS(name)                                              \
338 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
339 {                                                                       \
340         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
341 }                                                                       \
342 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
343 {                                                                       \
344         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
345 }                                                                       \
346 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
347 {                                                                       \
348         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
349 }
350
351 CFQ_CFQQ_FNS(on_rr);
352 CFQ_CFQQ_FNS(wait_request);
353 CFQ_CFQQ_FNS(must_dispatch);
354 CFQ_CFQQ_FNS(must_alloc_slice);
355 CFQ_CFQQ_FNS(fifo_expire);
356 CFQ_CFQQ_FNS(idle_window);
357 CFQ_CFQQ_FNS(prio_changed);
358 CFQ_CFQQ_FNS(slice_new);
359 CFQ_CFQQ_FNS(sync);
360 CFQ_CFQQ_FNS(coop);
361 CFQ_CFQQ_FNS(split_coop);
362 CFQ_CFQQ_FNS(deep);
363 CFQ_CFQQ_FNS(wait_busy);
364 #undef CFQ_CFQQ_FNS
365
366 #ifdef CONFIG_CFQ_GROUP_IOSCHED
367 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
368         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
369                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
370                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
371
372 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
373         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
374                                 blkg_path(&(cfqg)->blkg), ##args);      \
375
376 #else
377 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
378         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
380 #endif
381 #define cfq_log(cfqd, fmt, args...)     \
382         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
383
384 /* Traverses through cfq group service trees */
385 #define for_each_cfqg_st(cfqg, i, j, st) \
386         for (i = 0; i <= IDLE_WORKLOAD; i++) \
387                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
388                         : &cfqg->service_tree_idle; \
389                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
390                         (i == IDLE_WORKLOAD && j == 0); \
391                         j++, st = i < IDLE_WORKLOAD ? \
392                         &cfqg->service_trees[i][j]: NULL) \
393
394
395 static inline bool iops_mode(struct cfq_data *cfqd)
396 {
397         /*
398          * If we are not idling on queues and it is a NCQ drive, parallel
399          * execution of requests is on and measuring time is not possible
400          * in most of the cases until and unless we drive shallower queue
401          * depths and that becomes a performance bottleneck. In such cases
402          * switch to start providing fairness in terms of number of IOs.
403          */
404         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
405                 return true;
406         else
407                 return false;
408 }
409
410 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
411 {
412         if (cfq_class_idle(cfqq))
413                 return IDLE_WORKLOAD;
414         if (cfq_class_rt(cfqq))
415                 return RT_WORKLOAD;
416         return BE_WORKLOAD;
417 }
418
419
420 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
421 {
422         if (!cfq_cfqq_sync(cfqq))
423                 return ASYNC_WORKLOAD;
424         if (!cfq_cfqq_idle_window(cfqq))
425                 return SYNC_NOIDLE_WORKLOAD;
426         return SYNC_WORKLOAD;
427 }
428
429 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
430                                         struct cfq_data *cfqd,
431                                         struct cfq_group *cfqg)
432 {
433         if (wl == IDLE_WORKLOAD)
434                 return cfqg->service_tree_idle.count;
435
436         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
437                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
438                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
439 }
440
441 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
442                                         struct cfq_group *cfqg)
443 {
444         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
445                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
446 }
447
448 static void cfq_dispatch_insert(struct request_queue *, struct request *);
449 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
450                                        struct io_context *, gfp_t);
451 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
452                                                 struct io_context *);
453
454 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
455                                             bool is_sync)
456 {
457         return cic->cfqq[is_sync];
458 }
459
460 static inline void cic_set_cfqq(struct cfq_io_context *cic,
461                                 struct cfq_queue *cfqq, bool is_sync)
462 {
463         cic->cfqq[is_sync] = cfqq;
464 }
465
466 #define CIC_DEAD_KEY    1ul
467 #define CIC_DEAD_INDEX_SHIFT    1
468
469 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
470 {
471         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
472 }
473
474 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
475 {
476         struct cfq_data *cfqd = cic->key;
477
478         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
479                 return NULL;
480
481         return cfqd;
482 }
483
484 /*
485  * We regard a request as SYNC, if it's either a read or has the SYNC bit
486  * set (in which case it could also be direct WRITE).
487  */
488 static inline bool cfq_bio_sync(struct bio *bio)
489 {
490         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
491 }
492
493 /*
494  * scheduler run of queue, if there are requests pending and no one in the
495  * driver that will restart queueing
496  */
497 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
498 {
499         if (cfqd->busy_queues) {
500                 cfq_log(cfqd, "schedule dispatch");
501                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
502         }
503 }
504
505 /*
506  * Scale schedule slice based on io priority. Use the sync time slice only
507  * if a queue is marked sync and has sync io queued. A sync queue with async
508  * io only, should not get full sync slice length.
509  */
510 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
511                                  unsigned short prio)
512 {
513         const int base_slice = cfqd->cfq_slice[sync];
514
515         WARN_ON(prio >= IOPRIO_BE_NR);
516
517         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
518 }
519
520 static inline int
521 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
522 {
523         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
524 }
525
526 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
527 {
528         u64 d = delta << CFQ_SERVICE_SHIFT;
529
530         d = d * BLKIO_WEIGHT_DEFAULT;
531         do_div(d, cfqg->weight);
532         return d;
533 }
534
535 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
536 {
537         s64 delta = (s64)(vdisktime - min_vdisktime);
538         if (delta > 0)
539                 min_vdisktime = vdisktime;
540
541         return min_vdisktime;
542 }
543
544 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
545 {
546         s64 delta = (s64)(vdisktime - min_vdisktime);
547         if (delta < 0)
548                 min_vdisktime = vdisktime;
549
550         return min_vdisktime;
551 }
552
553 static void update_min_vdisktime(struct cfq_rb_root *st)
554 {
555         struct cfq_group *cfqg;
556
557         if (st->left) {
558                 cfqg = rb_entry_cfqg(st->left);
559                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
560                                                   cfqg->vdisktime);
561         }
562 }
563
564 /*
565  * get averaged number of queues of RT/BE priority.
566  * average is updated, with a formula that gives more weight to higher numbers,
567  * to quickly follows sudden increases and decrease slowly
568  */
569
570 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
571                                         struct cfq_group *cfqg, bool rt)
572 {
573         unsigned min_q, max_q;
574         unsigned mult  = cfq_hist_divisor - 1;
575         unsigned round = cfq_hist_divisor / 2;
576         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
577
578         min_q = min(cfqg->busy_queues_avg[rt], busy);
579         max_q = max(cfqg->busy_queues_avg[rt], busy);
580         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
581                 cfq_hist_divisor;
582         return cfqg->busy_queues_avg[rt];
583 }
584
585 static inline unsigned
586 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
587 {
588         struct cfq_rb_root *st = &cfqd->grp_service_tree;
589
590         return cfq_target_latency * cfqg->weight / st->total_weight;
591 }
592
593 static inline unsigned
594 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
595 {
596         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
597         if (cfqd->cfq_latency) {
598                 /*
599                  * interested queues (we consider only the ones with the same
600                  * priority class in the cfq group)
601                  */
602                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
603                                                 cfq_class_rt(cfqq));
604                 unsigned sync_slice = cfqd->cfq_slice[1];
605                 unsigned expect_latency = sync_slice * iq;
606                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
607
608                 if (expect_latency > group_slice) {
609                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
610                         /* scale low_slice according to IO priority
611                          * and sync vs async */
612                         unsigned low_slice =
613                                 min(slice, base_low_slice * slice / sync_slice);
614                         /* the adapted slice value is scaled to fit all iqs
615                          * into the target latency */
616                         slice = max(slice * group_slice / expect_latency,
617                                     low_slice);
618                 }
619         }
620         return slice;
621 }
622
623 static inline void
624 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
625 {
626         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
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_update_group_weight(struct cfq_group *cfqg)
859 {
860         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
861         if (cfqg->needs_update) {
862                 cfqg->weight = cfqg->new_weight;
863                 cfqg->needs_update = false;
864         }
865 }
866
867 static void
868 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
869 {
870         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
871
872         cfq_update_group_weight(cfqg);
873         __cfq_group_service_tree_add(st, cfqg);
874         st->total_weight += cfqg->weight;
875 }
876
877 static void
878 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
879 {
880         struct cfq_rb_root *st = &cfqd->grp_service_tree;
881         struct cfq_group *__cfqg;
882         struct rb_node *n;
883
884         cfqg->nr_cfqq++;
885         if (!RB_EMPTY_NODE(&cfqg->rb_node))
886                 return;
887
888         /*
889          * Currently put the group at the end. Later implement something
890          * so that groups get lesser vtime based on their weights, so that
891          * if group does not loose all if it was not continuously backlogged.
892          */
893         n = rb_last(&st->rb);
894         if (n) {
895                 __cfqg = rb_entry_cfqg(n);
896                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
897         } else
898                 cfqg->vdisktime = st->min_vdisktime;
899         cfq_group_service_tree_add(st, cfqg);
900 }
901
902 static void
903 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
904 {
905         st->total_weight -= cfqg->weight;
906         if (!RB_EMPTY_NODE(&cfqg->rb_node))
907                 cfq_rb_erase(&cfqg->rb_node, st);
908 }
909
910 static void
911 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
912 {
913         struct cfq_rb_root *st = &cfqd->grp_service_tree;
914
915         BUG_ON(cfqg->nr_cfqq < 1);
916         cfqg->nr_cfqq--;
917
918         /* If there are other cfq queues under this group, don't delete it */
919         if (cfqg->nr_cfqq)
920                 return;
921
922         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
923         cfq_group_service_tree_del(st, cfqg);
924         cfqg->saved_workload_slice = 0;
925         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
926 }
927
928 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
929                                                 unsigned int *unaccounted_time)
930 {
931         unsigned int slice_used;
932
933         /*
934          * Queue got expired before even a single request completed or
935          * got expired immediately after first request completion.
936          */
937         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
938                 /*
939                  * Also charge the seek time incurred to the group, otherwise
940                  * if there are mutiple queues in the group, each can dispatch
941                  * a single request on seeky media and cause lots of seek time
942                  * and group will never know it.
943                  */
944                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
945                                         1);
946         } else {
947                 slice_used = jiffies - cfqq->slice_start;
948                 if (slice_used > cfqq->allocated_slice) {
949                         *unaccounted_time = slice_used - cfqq->allocated_slice;
950                         slice_used = cfqq->allocated_slice;
951                 }
952                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
953                         *unaccounted_time += cfqq->slice_start -
954                                         cfqq->dispatch_start;
955         }
956
957         return slice_used;
958 }
959
960 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
961                                 struct cfq_queue *cfqq)
962 {
963         struct cfq_rb_root *st = &cfqd->grp_service_tree;
964         unsigned int used_sl, charge, unaccounted_sl = 0;
965         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
966                         - cfqg->service_tree_idle.count;
967
968         BUG_ON(nr_sync < 0);
969         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
970
971         if (iops_mode(cfqd))
972                 charge = cfqq->slice_dispatch;
973         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
974                 charge = cfqq->allocated_slice;
975
976         /* Can't update vdisktime while group is on service tree */
977         cfq_group_service_tree_del(st, cfqg);
978         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
979         /* If a new weight was requested, update now, off tree */
980         cfq_group_service_tree_add(st, cfqg);
981
982         /* This group is being expired. Save the context */
983         if (time_after(cfqd->workload_expires, jiffies)) {
984                 cfqg->saved_workload_slice = cfqd->workload_expires
985                                                 - jiffies;
986                 cfqg->saved_workload = cfqd->serving_type;
987                 cfqg->saved_serving_prio = cfqd->serving_prio;
988         } else
989                 cfqg->saved_workload_slice = 0;
990
991         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
992                                         st->min_vdisktime);
993         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
994                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
995                         iops_mode(cfqd), cfqq->nr_sectors);
996         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
997                                           unaccounted_sl);
998         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
999 }
1000
1001 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1002 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1003 {
1004         if (blkg)
1005                 return container_of(blkg, struct cfq_group, blkg);
1006         return NULL;
1007 }
1008
1009 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1010                                         unsigned int weight)
1011 {
1012         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1013         cfqg->new_weight = weight;
1014         cfqg->needs_update = true;
1015 }
1016
1017 static struct cfq_group *
1018 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
1019 {
1020         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
1021         struct cfq_group *cfqg = NULL;
1022         void *key = cfqd;
1023         int i, j;
1024         struct cfq_rb_root *st;
1025         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1026         unsigned int major, minor;
1027
1028         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1029         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1030                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1031                 cfqg->blkg.dev = MKDEV(major, minor);
1032                 goto done;
1033         }
1034         if (cfqg || !create)
1035                 goto done;
1036
1037         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1038         if (!cfqg)
1039                 goto done;
1040
1041         for_each_cfqg_st(cfqg, i, j, st)
1042                 *st = CFQ_RB_ROOT;
1043         RB_CLEAR_NODE(&cfqg->rb_node);
1044
1045         /*
1046          * Take the initial reference that will be released on destroy
1047          * This can be thought of a joint reference by cgroup and
1048          * elevator which will be dropped by either elevator exit
1049          * or cgroup deletion path depending on who is exiting first.
1050          */
1051         cfqg->ref = 1;
1052
1053         /*
1054          * Add group onto cgroup list. It might happen that bdi->dev is
1055          * not initialized yet. Initialize this new group without major
1056          * and minor info and this info will be filled in once a new thread
1057          * comes for IO. See code above.
1058          */
1059         if (bdi->dev) {
1060                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1061                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1062                                         MKDEV(major, minor));
1063         } else
1064                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1065                                         0);
1066
1067         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1068
1069         /* Add group on cfqd list */
1070         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1071
1072 done:
1073         return cfqg;
1074 }
1075
1076 /*
1077  * Search for the cfq group current task belongs to. If create = 1, then also
1078  * create the cfq group if it does not exist. request_queue lock must be held.
1079  */
1080 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1081 {
1082         struct cgroup *cgroup;
1083         struct cfq_group *cfqg = NULL;
1084
1085         rcu_read_lock();
1086         cgroup = task_cgroup(current, blkio_subsys_id);
1087         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1088         if (!cfqg && create)
1089                 cfqg = &cfqd->root_group;
1090         rcu_read_unlock();
1091         return cfqg;
1092 }
1093
1094 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1095 {
1096         cfqg->ref++;
1097         return cfqg;
1098 }
1099
1100 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1101 {
1102         /* Currently, all async queues are mapped to root group */
1103         if (!cfq_cfqq_sync(cfqq))
1104                 cfqg = &cfqq->cfqd->root_group;
1105
1106         cfqq->cfqg = cfqg;
1107         /* cfqq reference on cfqg */
1108         cfqq->cfqg->ref++;
1109 }
1110
1111 static void cfq_put_cfqg(struct cfq_group *cfqg)
1112 {
1113         struct cfq_rb_root *st;
1114         int i, j;
1115
1116         BUG_ON(cfqg->ref <= 0);
1117         cfqg->ref--;
1118         if (cfqg->ref)
1119                 return;
1120         for_each_cfqg_st(cfqg, i, j, st)
1121                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1122         kfree(cfqg);
1123 }
1124
1125 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1126 {
1127         /* Something wrong if we are trying to remove same group twice */
1128         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1129
1130         hlist_del_init(&cfqg->cfqd_node);
1131
1132         /*
1133          * Put the reference taken at the time of creation so that when all
1134          * queues are gone, group can be destroyed.
1135          */
1136         cfq_put_cfqg(cfqg);
1137 }
1138
1139 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1140 {
1141         struct hlist_node *pos, *n;
1142         struct cfq_group *cfqg;
1143
1144         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1145                 /*
1146                  * If cgroup removal path got to blk_group first and removed
1147                  * it from cgroup list, then it will take care of destroying
1148                  * cfqg also.
1149                  */
1150                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1151                         cfq_destroy_cfqg(cfqd, cfqg);
1152         }
1153 }
1154
1155 /*
1156  * Blk cgroup controller notification saying that blkio_group object is being
1157  * delinked as associated cgroup object is going away. That also means that
1158  * no new IO will come in this group. So get rid of this group as soon as
1159  * any pending IO in the group is finished.
1160  *
1161  * This function is called under rcu_read_lock(). key is the rcu protected
1162  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1163  * read lock.
1164  *
1165  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1166  * it should not be NULL as even if elevator was exiting, cgroup deltion
1167  * path got to it first.
1168  */
1169 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1170 {
1171         unsigned long  flags;
1172         struct cfq_data *cfqd = key;
1173
1174         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1175         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1176         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1177 }
1178
1179 #else /* GROUP_IOSCHED */
1180 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1181 {
1182         return &cfqd->root_group;
1183 }
1184
1185 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1186 {
1187         return cfqg;
1188 }
1189
1190 static inline void
1191 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1192         cfqq->cfqg = cfqg;
1193 }
1194
1195 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1196 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1197
1198 #endif /* GROUP_IOSCHED */
1199
1200 /*
1201  * The cfqd->service_trees holds all pending cfq_queue's that have
1202  * requests waiting to be processed. It is sorted in the order that
1203  * we will service the queues.
1204  */
1205 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1206                                  bool add_front)
1207 {
1208         struct rb_node **p, *parent;
1209         struct cfq_queue *__cfqq;
1210         unsigned long rb_key;
1211         struct cfq_rb_root *service_tree;
1212         int left;
1213         int new_cfqq = 1;
1214         int group_changed = 0;
1215
1216         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1217                                                 cfqq_type(cfqq));
1218         if (cfq_class_idle(cfqq)) {
1219                 rb_key = CFQ_IDLE_DELAY;
1220                 parent = rb_last(&service_tree->rb);
1221                 if (parent && parent != &cfqq->rb_node) {
1222                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1223                         rb_key += __cfqq->rb_key;
1224                 } else
1225                         rb_key += jiffies;
1226         } else if (!add_front) {
1227                 /*
1228                  * Get our rb key offset. Subtract any residual slice
1229                  * value carried from last service. A negative resid
1230                  * count indicates slice overrun, and this should position
1231                  * the next service time further away in the tree.
1232                  */
1233                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1234                 rb_key -= cfqq->slice_resid;
1235                 cfqq->slice_resid = 0;
1236         } else {
1237                 rb_key = -HZ;
1238                 __cfqq = cfq_rb_first(service_tree);
1239                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1240         }
1241
1242         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1243                 new_cfqq = 0;
1244                 /*
1245                  * same position, nothing more to do
1246                  */
1247                 if (rb_key == cfqq->rb_key &&
1248                     cfqq->service_tree == service_tree)
1249                         return;
1250
1251                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1252                 cfqq->service_tree = NULL;
1253         }
1254
1255         left = 1;
1256         parent = NULL;
1257         cfqq->service_tree = service_tree;
1258         p = &service_tree->rb.rb_node;
1259         while (*p) {
1260                 struct rb_node **n;
1261
1262                 parent = *p;
1263                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1264
1265                 /*
1266                  * sort by key, that represents service time.
1267                  */
1268                 if (time_before(rb_key, __cfqq->rb_key))
1269                         n = &(*p)->rb_left;
1270                 else {
1271                         n = &(*p)->rb_right;
1272                         left = 0;
1273                 }
1274
1275                 p = n;
1276         }
1277
1278         if (left)
1279                 service_tree->left = &cfqq->rb_node;
1280
1281         cfqq->rb_key = rb_key;
1282         rb_link_node(&cfqq->rb_node, parent, p);
1283         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1284         service_tree->count++;
1285         if ((add_front || !new_cfqq) && !group_changed)
1286                 return;
1287         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1288 }
1289
1290 static struct cfq_queue *
1291 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1292                      sector_t sector, struct rb_node **ret_parent,
1293                      struct rb_node ***rb_link)
1294 {
1295         struct rb_node **p, *parent;
1296         struct cfq_queue *cfqq = NULL;
1297
1298         parent = NULL;
1299         p = &root->rb_node;
1300         while (*p) {
1301                 struct rb_node **n;
1302
1303                 parent = *p;
1304                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1305
1306                 /*
1307                  * Sort strictly based on sector.  Smallest to the left,
1308                  * largest to the right.
1309                  */
1310                 if (sector > blk_rq_pos(cfqq->next_rq))
1311                         n = &(*p)->rb_right;
1312                 else if (sector < blk_rq_pos(cfqq->next_rq))
1313                         n = &(*p)->rb_left;
1314                 else
1315                         break;
1316                 p = n;
1317                 cfqq = NULL;
1318         }
1319
1320         *ret_parent = parent;
1321         if (rb_link)
1322                 *rb_link = p;
1323         return cfqq;
1324 }
1325
1326 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1327 {
1328         struct rb_node **p, *parent;
1329         struct cfq_queue *__cfqq;
1330
1331         if (cfqq->p_root) {
1332                 rb_erase(&cfqq->p_node, cfqq->p_root);
1333                 cfqq->p_root = NULL;
1334         }
1335
1336         if (cfq_class_idle(cfqq))
1337                 return;
1338         if (!cfqq->next_rq)
1339                 return;
1340
1341         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1342         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1343                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1344         if (!__cfqq) {
1345                 rb_link_node(&cfqq->p_node, parent, p);
1346                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1347         } else
1348                 cfqq->p_root = NULL;
1349 }
1350
1351 /*
1352  * Update cfqq's position in the service tree.
1353  */
1354 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1355 {
1356         /*
1357          * Resorting requires the cfqq to be on the RR list already.
1358          */
1359         if (cfq_cfqq_on_rr(cfqq)) {
1360                 cfq_service_tree_add(cfqd, cfqq, 0);
1361                 cfq_prio_tree_add(cfqd, cfqq);
1362         }
1363 }
1364
1365 /*
1366  * add to busy list of queues for service, trying to be fair in ordering
1367  * the pending list according to last request service
1368  */
1369 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1370 {
1371         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1372         BUG_ON(cfq_cfqq_on_rr(cfqq));
1373         cfq_mark_cfqq_on_rr(cfqq);
1374         cfqd->busy_queues++;
1375         if (cfq_cfqq_sync(cfqq))
1376                 cfqd->busy_sync_queues++;
1377
1378         cfq_resort_rr_list(cfqd, cfqq);
1379 }
1380
1381 /*
1382  * Called when the cfqq no longer has requests pending, remove it from
1383  * the service tree.
1384  */
1385 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1386 {
1387         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1388         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1389         cfq_clear_cfqq_on_rr(cfqq);
1390
1391         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1392                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1393                 cfqq->service_tree = NULL;
1394         }
1395         if (cfqq->p_root) {
1396                 rb_erase(&cfqq->p_node, cfqq->p_root);
1397                 cfqq->p_root = NULL;
1398         }
1399
1400         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1401         BUG_ON(!cfqd->busy_queues);
1402         cfqd->busy_queues--;
1403         if (cfq_cfqq_sync(cfqq))
1404                 cfqd->busy_sync_queues--;
1405 }
1406
1407 /*
1408  * rb tree support functions
1409  */
1410 static void cfq_del_rq_rb(struct request *rq)
1411 {
1412         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1413         const int sync = rq_is_sync(rq);
1414
1415         BUG_ON(!cfqq->queued[sync]);
1416         cfqq->queued[sync]--;
1417
1418         elv_rb_del(&cfqq->sort_list, rq);
1419
1420         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1421                 /*
1422                  * Queue will be deleted from service tree when we actually
1423                  * expire it later. Right now just remove it from prio tree
1424                  * as it is empty.
1425                  */
1426                 if (cfqq->p_root) {
1427                         rb_erase(&cfqq->p_node, cfqq->p_root);
1428                         cfqq->p_root = NULL;
1429                 }
1430         }
1431 }
1432
1433 static void cfq_add_rq_rb(struct request *rq)
1434 {
1435         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1436         struct cfq_data *cfqd = cfqq->cfqd;
1437         struct request *__alias, *prev;
1438
1439         cfqq->queued[rq_is_sync(rq)]++;
1440
1441         /*
1442          * looks a little odd, but the first insert might return an alias.
1443          * if that happens, put the alias on the dispatch list
1444          */
1445         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1446                 cfq_dispatch_insert(cfqd->queue, __alias);
1447
1448         if (!cfq_cfqq_on_rr(cfqq))
1449                 cfq_add_cfqq_rr(cfqd, cfqq);
1450
1451         /*
1452          * check if this request is a better next-serve candidate
1453          */
1454         prev = cfqq->next_rq;
1455         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1456
1457         /*
1458          * adjust priority tree position, if ->next_rq changes
1459          */
1460         if (prev != cfqq->next_rq)
1461                 cfq_prio_tree_add(cfqd, cfqq);
1462
1463         BUG_ON(!cfqq->next_rq);
1464 }
1465
1466 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1467 {
1468         elv_rb_del(&cfqq->sort_list, rq);
1469         cfqq->queued[rq_is_sync(rq)]--;
1470         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1471                                         rq_data_dir(rq), rq_is_sync(rq));
1472         cfq_add_rq_rb(rq);
1473         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1474                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1475                         rq_is_sync(rq));
1476 }
1477
1478 static struct request *
1479 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1480 {
1481         struct task_struct *tsk = current;
1482         struct cfq_io_context *cic;
1483         struct cfq_queue *cfqq;
1484
1485         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1486         if (!cic)
1487                 return NULL;
1488
1489         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1490         if (cfqq) {
1491                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1492
1493                 return elv_rb_find(&cfqq->sort_list, sector);
1494         }
1495
1496         return NULL;
1497 }
1498
1499 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1500 {
1501         struct cfq_data *cfqd = q->elevator->elevator_data;
1502
1503         cfqd->rq_in_driver++;
1504         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1505                                                 cfqd->rq_in_driver);
1506
1507         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1508 }
1509
1510 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1511 {
1512         struct cfq_data *cfqd = q->elevator->elevator_data;
1513
1514         WARN_ON(!cfqd->rq_in_driver);
1515         cfqd->rq_in_driver--;
1516         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1517                                                 cfqd->rq_in_driver);
1518 }
1519
1520 static void cfq_remove_request(struct request *rq)
1521 {
1522         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1523
1524         if (cfqq->next_rq == rq)
1525                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1526
1527         list_del_init(&rq->queuelist);
1528         cfq_del_rq_rb(rq);
1529
1530         cfqq->cfqd->rq_queued--;
1531         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1532                                         rq_data_dir(rq), rq_is_sync(rq));
1533         if (rq->cmd_flags & REQ_META) {
1534                 WARN_ON(!cfqq->meta_pending);
1535                 cfqq->meta_pending--;
1536         }
1537 }
1538
1539 static int cfq_merge(struct request_queue *q, struct request **req,
1540                      struct bio *bio)
1541 {
1542         struct cfq_data *cfqd = q->elevator->elevator_data;
1543         struct request *__rq;
1544
1545         __rq = cfq_find_rq_fmerge(cfqd, bio);
1546         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1547                 *req = __rq;
1548                 return ELEVATOR_FRONT_MERGE;
1549         }
1550
1551         return ELEVATOR_NO_MERGE;
1552 }
1553
1554 static void cfq_merged_request(struct request_queue *q, struct request *req,
1555                                int type)
1556 {
1557         if (type == ELEVATOR_FRONT_MERGE) {
1558                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1559
1560                 cfq_reposition_rq_rb(cfqq, req);
1561         }
1562 }
1563
1564 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1565                                 struct bio *bio)
1566 {
1567         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1568                                         bio_data_dir(bio), cfq_bio_sync(bio));
1569 }
1570
1571 static void
1572 cfq_merged_requests(struct request_queue *q, struct request *rq,
1573                     struct request *next)
1574 {
1575         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1576         /*
1577          * reposition in fifo if next is older than rq
1578          */
1579         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1580             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1581                 list_move(&rq->queuelist, &next->queuelist);
1582                 rq_set_fifo_time(rq, rq_fifo_time(next));
1583         }
1584
1585         if (cfqq->next_rq == next)
1586                 cfqq->next_rq = rq;
1587         cfq_remove_request(next);
1588         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1589                                         rq_data_dir(next), rq_is_sync(next));
1590 }
1591
1592 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1593                            struct bio *bio)
1594 {
1595         struct cfq_data *cfqd = q->elevator->elevator_data;
1596         struct cfq_io_context *cic;
1597         struct cfq_queue *cfqq;
1598
1599         /*
1600          * Disallow merge of a sync bio into an async request.
1601          */
1602         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1603                 return false;
1604
1605         /*
1606          * Lookup the cfqq that this bio will be queued with. Allow
1607          * merge only if rq is queued there.
1608          */
1609         cic = cfq_cic_lookup(cfqd, current->io_context);
1610         if (!cic)
1611                 return false;
1612
1613         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1614         return cfqq == RQ_CFQQ(rq);
1615 }
1616
1617 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1618 {
1619         del_timer(&cfqd->idle_slice_timer);
1620         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1621 }
1622
1623 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1624                                    struct cfq_queue *cfqq)
1625 {
1626         if (cfqq) {
1627                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1628                                 cfqd->serving_prio, cfqd->serving_type);
1629                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1630                 cfqq->slice_start = 0;
1631                 cfqq->dispatch_start = jiffies;
1632                 cfqq->allocated_slice = 0;
1633                 cfqq->slice_end = 0;
1634                 cfqq->slice_dispatch = 0;
1635                 cfqq->nr_sectors = 0;
1636
1637                 cfq_clear_cfqq_wait_request(cfqq);
1638                 cfq_clear_cfqq_must_dispatch(cfqq);
1639                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1640                 cfq_clear_cfqq_fifo_expire(cfqq);
1641                 cfq_mark_cfqq_slice_new(cfqq);
1642
1643                 cfq_del_timer(cfqd, cfqq);
1644         }
1645
1646         cfqd->active_queue = cfqq;
1647 }
1648
1649 /*
1650  * current cfqq expired its slice (or was too idle), select new one
1651  */
1652 static void
1653 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1654                     bool timed_out)
1655 {
1656         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1657
1658         if (cfq_cfqq_wait_request(cfqq))
1659                 cfq_del_timer(cfqd, cfqq);
1660
1661         cfq_clear_cfqq_wait_request(cfqq);
1662         cfq_clear_cfqq_wait_busy(cfqq);
1663
1664         /*
1665          * If this cfqq is shared between multiple processes, check to
1666          * make sure that those processes are still issuing I/Os within
1667          * the mean seek distance.  If not, it may be time to break the
1668          * queues apart again.
1669          */
1670         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1671                 cfq_mark_cfqq_split_coop(cfqq);
1672
1673         /*
1674          * store what was left of this slice, if the queue idled/timed out
1675          */
1676         if (timed_out) {
1677                 if (cfq_cfqq_slice_new(cfqq))
1678                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1679                 else
1680                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1681                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1682         }
1683
1684         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1685
1686         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1687                 cfq_del_cfqq_rr(cfqd, cfqq);
1688
1689         cfq_resort_rr_list(cfqd, cfqq);
1690
1691         if (cfqq == cfqd->active_queue)
1692                 cfqd->active_queue = NULL;
1693
1694         if (cfqd->active_cic) {
1695                 put_io_context(cfqd->active_cic->ioc);
1696                 cfqd->active_cic = NULL;
1697         }
1698 }
1699
1700 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1701 {
1702         struct cfq_queue *cfqq = cfqd->active_queue;
1703
1704         if (cfqq)
1705                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1706 }
1707
1708 /*
1709  * Get next queue for service. Unless we have a queue preemption,
1710  * we'll simply select the first cfqq in the service tree.
1711  */
1712 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1713 {
1714         struct cfq_rb_root *service_tree =
1715                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1716                                         cfqd->serving_type);
1717
1718         if (!cfqd->rq_queued)
1719                 return NULL;
1720
1721         /* There is nothing to dispatch */
1722         if (!service_tree)
1723                 return NULL;
1724         if (RB_EMPTY_ROOT(&service_tree->rb))
1725                 return NULL;
1726         return cfq_rb_first(service_tree);
1727 }
1728
1729 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1730 {
1731         struct cfq_group *cfqg;
1732         struct cfq_queue *cfqq;
1733         int i, j;
1734         struct cfq_rb_root *st;
1735
1736         if (!cfqd->rq_queued)
1737                 return NULL;
1738
1739         cfqg = cfq_get_next_cfqg(cfqd);
1740         if (!cfqg)
1741                 return NULL;
1742
1743         for_each_cfqg_st(cfqg, i, j, st)
1744                 if ((cfqq = cfq_rb_first(st)) != NULL)
1745                         return cfqq;
1746         return NULL;
1747 }
1748
1749 /*
1750  * Get and set a new active queue for service.
1751  */
1752 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1753                                               struct cfq_queue *cfqq)
1754 {
1755         if (!cfqq)
1756                 cfqq = cfq_get_next_queue(cfqd);
1757
1758         __cfq_set_active_queue(cfqd, cfqq);
1759         return cfqq;
1760 }
1761
1762 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1763                                           struct request *rq)
1764 {
1765         if (blk_rq_pos(rq) >= cfqd->last_position)
1766                 return blk_rq_pos(rq) - cfqd->last_position;
1767         else
1768                 return cfqd->last_position - blk_rq_pos(rq);
1769 }
1770
1771 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1772                                struct request *rq)
1773 {
1774         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1775 }
1776
1777 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1778                                     struct cfq_queue *cur_cfqq)
1779 {
1780         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1781         struct rb_node *parent, *node;
1782         struct cfq_queue *__cfqq;
1783         sector_t sector = cfqd->last_position;
1784
1785         if (RB_EMPTY_ROOT(root))
1786                 return NULL;
1787
1788         /*
1789          * First, if we find a request starting at the end of the last
1790          * request, choose it.
1791          */
1792         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1793         if (__cfqq)
1794                 return __cfqq;
1795
1796         /*
1797          * If the exact sector wasn't found, the parent of the NULL leaf
1798          * will contain the closest sector.
1799          */
1800         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1801         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1802                 return __cfqq;
1803
1804         if (blk_rq_pos(__cfqq->next_rq) < sector)
1805                 node = rb_next(&__cfqq->p_node);
1806         else
1807                 node = rb_prev(&__cfqq->p_node);
1808         if (!node)
1809                 return NULL;
1810
1811         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1812         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1813                 return __cfqq;
1814
1815         return NULL;
1816 }
1817
1818 /*
1819  * cfqd - obvious
1820  * cur_cfqq - passed in so that we don't decide that the current queue is
1821  *            closely cooperating with itself.
1822  *
1823  * So, basically we're assuming that that cur_cfqq has dispatched at least
1824  * one request, and that cfqd->last_position reflects a position on the disk
1825  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1826  * assumption.
1827  */
1828 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1829                                               struct cfq_queue *cur_cfqq)
1830 {
1831         struct cfq_queue *cfqq;
1832
1833         if (cfq_class_idle(cur_cfqq))
1834                 return NULL;
1835         if (!cfq_cfqq_sync(cur_cfqq))
1836                 return NULL;
1837         if (CFQQ_SEEKY(cur_cfqq))
1838                 return NULL;
1839
1840         /*
1841          * Don't search priority tree if it's the only queue in the group.
1842          */
1843         if (cur_cfqq->cfqg->nr_cfqq == 1)
1844                 return NULL;
1845
1846         /*
1847          * We should notice if some of the queues are cooperating, eg
1848          * working closely on the same area of the disk. In that case,
1849          * we can group them together and don't waste time idling.
1850          */
1851         cfqq = cfqq_close(cfqd, cur_cfqq);
1852         if (!cfqq)
1853                 return NULL;
1854
1855         /* If new queue belongs to different cfq_group, don't choose it */
1856         if (cur_cfqq->cfqg != cfqq->cfqg)
1857                 return NULL;
1858
1859         /*
1860          * It only makes sense to merge sync queues.
1861          */
1862         if (!cfq_cfqq_sync(cfqq))
1863                 return NULL;
1864         if (CFQQ_SEEKY(cfqq))
1865                 return NULL;
1866
1867         /*
1868          * Do not merge queues of different priority classes
1869          */
1870         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1871                 return NULL;
1872
1873         return cfqq;
1874 }
1875
1876 /*
1877  * Determine whether we should enforce idle window for this queue.
1878  */
1879
1880 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1881 {
1882         enum wl_prio_t prio = cfqq_prio(cfqq);
1883         struct cfq_rb_root *service_tree = cfqq->service_tree;
1884
1885         BUG_ON(!service_tree);
1886         BUG_ON(!service_tree->count);
1887
1888         if (!cfqd->cfq_slice_idle)
1889                 return false;
1890
1891         /* We never do for idle class queues. */
1892         if (prio == IDLE_WORKLOAD)
1893                 return false;
1894
1895         /* We do for queues that were marked with idle window flag. */
1896         if (cfq_cfqq_idle_window(cfqq) &&
1897            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1898                 return true;
1899
1900         /*
1901          * Otherwise, we do only if they are the last ones
1902          * in their service tree.
1903          */
1904         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1905                 return true;
1906         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1907                         service_tree->count);
1908         return false;
1909 }
1910
1911 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1912 {
1913         struct cfq_queue *cfqq = cfqd->active_queue;
1914         struct cfq_io_context *cic;
1915         unsigned long sl, group_idle = 0;
1916
1917         /*
1918          * SSD device without seek penalty, disable idling. But only do so
1919          * for devices that support queuing, otherwise we still have a problem
1920          * with sync vs async workloads.
1921          */
1922         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1923                 return;
1924
1925         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1926         WARN_ON(cfq_cfqq_slice_new(cfqq));
1927
1928         /*
1929          * idle is disabled, either manually or by past process history
1930          */
1931         if (!cfq_should_idle(cfqd, cfqq)) {
1932                 /* no queue idling. Check for group idling */
1933                 if (cfqd->cfq_group_idle)
1934                         group_idle = cfqd->cfq_group_idle;
1935                 else
1936                         return;
1937         }
1938
1939         /*
1940          * still active requests from this queue, don't idle
1941          */
1942         if (cfqq->dispatched)
1943                 return;
1944
1945         /*
1946          * task has exited, don't wait
1947          */
1948         cic = cfqd->active_cic;
1949         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1950                 return;
1951
1952         /*
1953          * If our average think time is larger than the remaining time
1954          * slice, then don't idle. This avoids overrunning the allotted
1955          * time slice.
1956          */
1957         if (sample_valid(cic->ttime_samples) &&
1958             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1959                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1960                                 cic->ttime_mean);
1961                 return;
1962         }
1963
1964         /* There are other queues in the group, don't do group idle */
1965         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
1966                 return;
1967
1968         cfq_mark_cfqq_wait_request(cfqq);
1969
1970         if (group_idle)
1971                 sl = cfqd->cfq_group_idle;
1972         else
1973                 sl = cfqd->cfq_slice_idle;
1974
1975         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1976         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1977         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
1978                         group_idle ? 1 : 0);
1979 }
1980
1981 /*
1982  * Move request from internal lists to the request queue dispatch list.
1983  */
1984 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1985 {
1986         struct cfq_data *cfqd = q->elevator->elevator_data;
1987         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1988
1989         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1990
1991         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1992         cfq_remove_request(rq);
1993         cfqq->dispatched++;
1994         (RQ_CFQG(rq))->dispatched++;
1995         elv_dispatch_sort(q, rq);
1996
1997         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1998         cfqq->nr_sectors += blk_rq_sectors(rq);
1999         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2000                                         rq_data_dir(rq), rq_is_sync(rq));
2001 }
2002
2003 /*
2004  * return expired entry, or NULL to just start from scratch in rbtree
2005  */
2006 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2007 {
2008         struct request *rq = NULL;
2009
2010         if (cfq_cfqq_fifo_expire(cfqq))
2011                 return NULL;
2012
2013         cfq_mark_cfqq_fifo_expire(cfqq);
2014
2015         if (list_empty(&cfqq->fifo))
2016                 return NULL;
2017
2018         rq = rq_entry_fifo(cfqq->fifo.next);
2019         if (time_before(jiffies, rq_fifo_time(rq)))
2020                 rq = NULL;
2021
2022         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2023         return rq;
2024 }
2025
2026 static inline int
2027 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2028 {
2029         const int base_rq = cfqd->cfq_slice_async_rq;
2030
2031         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2032
2033         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2034 }
2035
2036 /*
2037  * Must be called with the queue_lock held.
2038  */
2039 static int cfqq_process_refs(struct cfq_queue *cfqq)
2040 {
2041         int process_refs, io_refs;
2042
2043         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2044         process_refs = cfqq->ref - io_refs;
2045         BUG_ON(process_refs < 0);
2046         return process_refs;
2047 }
2048
2049 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2050 {
2051         int process_refs, new_process_refs;
2052         struct cfq_queue *__cfqq;
2053
2054         /*
2055          * If there are no process references on the new_cfqq, then it is
2056          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2057          * chain may have dropped their last reference (not just their
2058          * last process reference).
2059          */
2060         if (!cfqq_process_refs(new_cfqq))
2061                 return;
2062
2063         /* Avoid a circular list and skip interim queue merges */
2064         while ((__cfqq = new_cfqq->new_cfqq)) {
2065                 if (__cfqq == cfqq)
2066                         return;
2067                 new_cfqq = __cfqq;
2068         }
2069
2070         process_refs = cfqq_process_refs(cfqq);
2071         new_process_refs = cfqq_process_refs(new_cfqq);
2072         /*
2073          * If the process for the cfqq has gone away, there is no
2074          * sense in merging the queues.
2075          */
2076         if (process_refs == 0 || new_process_refs == 0)
2077                 return;
2078
2079         /*
2080          * Merge in the direction of the lesser amount of work.
2081          */
2082         if (new_process_refs >= process_refs) {
2083                 cfqq->new_cfqq = new_cfqq;
2084                 new_cfqq->ref += process_refs;
2085         } else {
2086                 new_cfqq->new_cfqq = cfqq;
2087                 cfqq->ref += new_process_refs;
2088         }
2089 }
2090
2091 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2092                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2093 {
2094         struct cfq_queue *queue;
2095         int i;
2096         bool key_valid = false;
2097         unsigned long lowest_key = 0;
2098         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2099
2100         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2101                 /* select the one with lowest rb_key */
2102                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2103                 if (queue &&
2104                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2105                         lowest_key = queue->rb_key;
2106                         cur_best = i;
2107                         key_valid = true;
2108                 }
2109         }
2110
2111         return cur_best;
2112 }
2113
2114 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2115 {
2116         unsigned slice;
2117         unsigned count;
2118         struct cfq_rb_root *st;
2119         unsigned group_slice;
2120         enum wl_prio_t original_prio = cfqd->serving_prio;
2121
2122         /* Choose next priority. RT > BE > IDLE */
2123         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2124                 cfqd->serving_prio = RT_WORKLOAD;
2125         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2126                 cfqd->serving_prio = BE_WORKLOAD;
2127         else {
2128                 cfqd->serving_prio = IDLE_WORKLOAD;
2129                 cfqd->workload_expires = jiffies + 1;
2130                 return;
2131         }
2132
2133         if (original_prio != cfqd->serving_prio)
2134                 goto new_workload;
2135
2136         /*
2137          * For RT and BE, we have to choose also the type
2138          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2139          * expiration time
2140          */
2141         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2142         count = st->count;
2143
2144         /*
2145          * check workload expiration, and that we still have other queues ready
2146          */
2147         if (count && !time_after(jiffies, cfqd->workload_expires))
2148                 return;
2149
2150 new_workload:
2151         /* otherwise select new workload type */
2152         cfqd->serving_type =
2153                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2154         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2155         count = st->count;
2156
2157         /*
2158          * the workload slice is computed as a fraction of target latency
2159          * proportional to the number of queues in that workload, over
2160          * all the queues in the same priority class
2161          */
2162         group_slice = cfq_group_slice(cfqd, cfqg);
2163
2164         slice = group_slice * count /
2165                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2166                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2167
2168         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2169                 unsigned int tmp;
2170
2171                 /*
2172                  * Async queues are currently system wide. Just taking
2173                  * proportion of queues with-in same group will lead to higher
2174                  * async ratio system wide as generally root group is going
2175                  * to have higher weight. A more accurate thing would be to
2176                  * calculate system wide asnc/sync ratio.
2177                  */
2178                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2179                 tmp = tmp/cfqd->busy_queues;
2180                 slice = min_t(unsigned, slice, tmp);
2181
2182                 /* async workload slice is scaled down according to
2183                  * the sync/async slice ratio. */
2184                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2185         } else
2186                 /* sync workload slice is at least 2 * cfq_slice_idle */
2187                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2188
2189         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2190         cfq_log(cfqd, "workload slice:%d", slice);
2191         cfqd->workload_expires = jiffies + slice;
2192 }
2193
2194 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2195 {
2196         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2197         struct cfq_group *cfqg;
2198
2199         if (RB_EMPTY_ROOT(&st->rb))
2200                 return NULL;
2201         cfqg = cfq_rb_first_group(st);
2202         update_min_vdisktime(st);
2203         return cfqg;
2204 }
2205
2206 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2207 {
2208         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2209
2210         cfqd->serving_group = cfqg;
2211
2212         /* Restore the workload type data */
2213         if (cfqg->saved_workload_slice) {
2214                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2215                 cfqd->serving_type = cfqg->saved_workload;
2216                 cfqd->serving_prio = cfqg->saved_serving_prio;
2217         } else
2218                 cfqd->workload_expires = jiffies - 1;
2219
2220         choose_service_tree(cfqd, cfqg);
2221 }
2222
2223 /*
2224  * Select a queue for service. If we have a current active queue,
2225  * check whether to continue servicing it, or retrieve and set a new one.
2226  */
2227 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2228 {
2229         struct cfq_queue *cfqq, *new_cfqq = NULL;
2230
2231         cfqq = cfqd->active_queue;
2232         if (!cfqq)
2233                 goto new_queue;
2234
2235         if (!cfqd->rq_queued)
2236                 return NULL;
2237
2238         /*
2239          * We were waiting for group to get backlogged. Expire the queue
2240          */
2241         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2242                 goto expire;
2243
2244         /*
2245          * The active queue has run out of time, expire it and select new.
2246          */
2247         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2248                 /*
2249                  * If slice had not expired at the completion of last request
2250                  * we might not have turned on wait_busy flag. Don't expire
2251                  * the queue yet. Allow the group to get backlogged.
2252                  *
2253                  * The very fact that we have used the slice, that means we
2254                  * have been idling all along on this queue and it should be
2255                  * ok to wait for this request to complete.
2256                  */
2257                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2258                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2259                         cfqq = NULL;
2260                         goto keep_queue;
2261                 } else
2262                         goto check_group_idle;
2263         }
2264
2265         /*
2266          * The active queue has requests and isn't expired, allow it to
2267          * dispatch.
2268          */
2269         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2270                 goto keep_queue;
2271
2272         /*
2273          * If another queue has a request waiting within our mean seek
2274          * distance, let it run.  The expire code will check for close
2275          * cooperators and put the close queue at the front of the service
2276          * tree.  If possible, merge the expiring queue with the new cfqq.
2277          */
2278         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2279         if (new_cfqq) {
2280                 if (!cfqq->new_cfqq)
2281                         cfq_setup_merge(cfqq, new_cfqq);
2282                 goto expire;
2283         }
2284
2285         /*
2286          * No requests pending. If the active queue still has requests in
2287          * flight or is idling for a new request, allow either of these
2288          * conditions to happen (or time out) before selecting a new queue.
2289          */
2290         if (timer_pending(&cfqd->idle_slice_timer)) {
2291                 cfqq = NULL;
2292                 goto keep_queue;
2293         }
2294
2295         /*
2296          * This is a deep seek queue, but the device is much faster than
2297          * the queue can deliver, don't idle
2298          **/
2299         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2300             (cfq_cfqq_slice_new(cfqq) ||
2301             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2302                 cfq_clear_cfqq_deep(cfqq);
2303                 cfq_clear_cfqq_idle_window(cfqq);
2304         }
2305
2306         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2307                 cfqq = NULL;
2308                 goto keep_queue;
2309         }
2310
2311         /*
2312          * If group idle is enabled and there are requests dispatched from
2313          * this group, wait for requests to complete.
2314          */
2315 check_group_idle:
2316         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2317             && cfqq->cfqg->dispatched) {
2318                 cfqq = NULL;
2319                 goto keep_queue;
2320         }
2321
2322 expire:
2323         cfq_slice_expired(cfqd, 0);
2324 new_queue:
2325         /*
2326          * Current queue expired. Check if we have to switch to a new
2327          * service tree
2328          */
2329         if (!new_cfqq)
2330                 cfq_choose_cfqg(cfqd);
2331
2332         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2333 keep_queue:
2334         return cfqq;
2335 }
2336
2337 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2338 {
2339         int dispatched = 0;
2340
2341         while (cfqq->next_rq) {
2342                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2343                 dispatched++;
2344         }
2345
2346         BUG_ON(!list_empty(&cfqq->fifo));
2347
2348         /* By default cfqq is not expired if it is empty. Do it explicitly */
2349         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2350         return dispatched;
2351 }
2352
2353 /*
2354  * Drain our current requests. Used for barriers and when switching
2355  * io schedulers on-the-fly.
2356  */
2357 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2358 {
2359         struct cfq_queue *cfqq;
2360         int dispatched = 0;
2361
2362         /* Expire the timeslice of the current active queue first */
2363         cfq_slice_expired(cfqd, 0);
2364         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2365                 __cfq_set_active_queue(cfqd, cfqq);
2366                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2367         }
2368
2369         BUG_ON(cfqd->busy_queues);
2370
2371         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2372         return dispatched;
2373 }
2374
2375 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2376         struct cfq_queue *cfqq)
2377 {
2378         /* the queue hasn't finished any request, can't estimate */
2379         if (cfq_cfqq_slice_new(cfqq))
2380                 return true;
2381         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2382                 cfqq->slice_end))
2383                 return true;
2384
2385         return false;
2386 }
2387
2388 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2389 {
2390         unsigned int max_dispatch;
2391
2392         /*
2393          * Drain async requests before we start sync IO
2394          */
2395         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2396                 return false;
2397
2398         /*
2399          * If this is an async queue and we have sync IO in flight, let it wait
2400          */
2401         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2402                 return false;
2403
2404         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2405         if (cfq_class_idle(cfqq))
2406                 max_dispatch = 1;
2407
2408         /*
2409          * Does this cfqq already have too much IO in flight?
2410          */
2411         if (cfqq->dispatched >= max_dispatch) {
2412                 bool promote_sync = false;
2413                 /*
2414                  * idle queue must always only have a single IO in flight
2415                  */
2416                 if (cfq_class_idle(cfqq))
2417                         return false;
2418
2419                 /*
2420                  * If there is only one sync queue
2421                  * we can ignore async queue here and give the sync
2422                  * queue no dispatch limit. The reason is a sync queue can
2423                  * preempt async queue, limiting the sync queue doesn't make
2424                  * sense. This is useful for aiostress test.
2425                  */
2426                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2427                         promote_sync = true;
2428
2429                 /*
2430                  * We have other queues, don't allow more IO from this one
2431                  */
2432                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2433                                 !promote_sync)
2434                         return false;
2435
2436                 /*
2437                  * Sole queue user, no limit
2438                  */
2439                 if (cfqd->busy_queues == 1 || promote_sync)
2440                         max_dispatch = -1;
2441                 else
2442                         /*
2443                          * Normally we start throttling cfqq when cfq_quantum/2
2444                          * requests have been dispatched. But we can drive
2445                          * deeper queue depths at the beginning of slice
2446                          * subjected to upper limit of cfq_quantum.
2447                          * */
2448                         max_dispatch = cfqd->cfq_quantum;
2449         }
2450
2451         /*
2452          * Async queues must wait a bit before being allowed dispatch.
2453          * We also ramp up the dispatch depth gradually for async IO,
2454          * based on the last sync IO we serviced
2455          */
2456         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2457                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2458                 unsigned int depth;
2459
2460                 depth = last_sync / cfqd->cfq_slice[1];
2461                 if (!depth && !cfqq->dispatched)
2462                         depth = 1;
2463                 if (depth < max_dispatch)
2464                         max_dispatch = depth;
2465         }
2466
2467         /*
2468          * If we're below the current max, allow a dispatch
2469          */
2470         return cfqq->dispatched < max_dispatch;
2471 }
2472
2473 /*
2474  * Dispatch a request from cfqq, moving them to the request queue
2475  * dispatch list.
2476  */
2477 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2478 {
2479         struct request *rq;
2480
2481         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2482
2483         if (!cfq_may_dispatch(cfqd, cfqq))
2484                 return false;
2485
2486         /*
2487          * follow expired path, else get first next available
2488          */
2489         rq = cfq_check_fifo(cfqq);
2490         if (!rq)
2491                 rq = cfqq->next_rq;
2492
2493         /*
2494          * insert request into driver dispatch list
2495          */
2496         cfq_dispatch_insert(cfqd->queue, rq);
2497
2498         if (!cfqd->active_cic) {
2499                 struct cfq_io_context *cic = RQ_CIC(rq);
2500
2501                 atomic_long_inc(&cic->ioc->refcount);
2502                 cfqd->active_cic = cic;
2503         }
2504
2505         return true;
2506 }
2507
2508 /*
2509  * Find the cfqq that we need to service and move a request from that to the
2510  * dispatch list
2511  */
2512 static int cfq_dispatch_requests(struct request_queue *q, int force)
2513 {
2514         struct cfq_data *cfqd = q->elevator->elevator_data;
2515         struct cfq_queue *cfqq;
2516
2517         if (!cfqd->busy_queues)
2518                 return 0;
2519
2520         if (unlikely(force))
2521                 return cfq_forced_dispatch(cfqd);
2522
2523         cfqq = cfq_select_queue(cfqd);
2524         if (!cfqq)
2525                 return 0;
2526
2527         /*
2528          * Dispatch a request from this cfqq, if it is allowed
2529          */
2530         if (!cfq_dispatch_request(cfqd, cfqq))
2531                 return 0;
2532
2533         cfqq->slice_dispatch++;
2534         cfq_clear_cfqq_must_dispatch(cfqq);
2535
2536         /*
2537          * expire an async queue immediately if it has used up its slice. idle
2538          * queue always expire after 1 dispatch round.
2539          */
2540         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2541             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2542             cfq_class_idle(cfqq))) {
2543                 cfqq->slice_end = jiffies + 1;
2544                 cfq_slice_expired(cfqd, 0);
2545         }
2546
2547         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2548         return 1;
2549 }
2550
2551 /*
2552  * task holds one reference to the queue, dropped when task exits. each rq
2553  * in-flight on this queue also holds a reference, dropped when rq is freed.
2554  *
2555  * Each cfq queue took a reference on the parent group. Drop it now.
2556  * queue lock must be held here.
2557  */
2558 static void cfq_put_queue(struct cfq_queue *cfqq)
2559 {
2560         struct cfq_data *cfqd = cfqq->cfqd;
2561         struct cfq_group *cfqg;
2562
2563         BUG_ON(cfqq->ref <= 0);
2564
2565         cfqq->ref--;
2566         if (cfqq->ref)
2567                 return;
2568
2569         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2570         BUG_ON(rb_first(&cfqq->sort_list));
2571         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2572         cfqg = cfqq->cfqg;
2573
2574         if (unlikely(cfqd->active_queue == cfqq)) {
2575                 __cfq_slice_expired(cfqd, cfqq, 0);
2576                 cfq_schedule_dispatch(cfqd);
2577         }
2578
2579         BUG_ON(cfq_cfqq_on_rr(cfqq));
2580         kmem_cache_free(cfq_pool, cfqq);
2581         cfq_put_cfqg(cfqg);
2582 }
2583
2584 /*
2585  * Must always be called with the rcu_read_lock() held
2586  */
2587 static void
2588 __call_for_each_cic(struct io_context *ioc,
2589                     void (*func)(struct io_context *, struct cfq_io_context *))
2590 {
2591         struct cfq_io_context *cic;
2592         struct hlist_node *n;
2593
2594         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2595                 func(ioc, cic);
2596 }
2597
2598 /*
2599  * Call func for each cic attached to this ioc.
2600  */
2601 static void
2602 call_for_each_cic(struct io_context *ioc,
2603                   void (*func)(struct io_context *, struct cfq_io_context *))
2604 {
2605         rcu_read_lock();
2606         __call_for_each_cic(ioc, func);
2607         rcu_read_unlock();
2608 }
2609
2610 static void cfq_cic_free_rcu(struct rcu_head *head)
2611 {
2612         struct cfq_io_context *cic;
2613
2614         cic = container_of(head, struct cfq_io_context, rcu_head);
2615
2616         kmem_cache_free(cfq_ioc_pool, cic);
2617         elv_ioc_count_dec(cfq_ioc_count);
2618
2619         if (ioc_gone) {
2620                 /*
2621                  * CFQ scheduler is exiting, grab exit lock and check
2622                  * the pending io context count. If it hits zero,
2623                  * complete ioc_gone and set it back to NULL
2624                  */
2625                 spin_lock(&ioc_gone_lock);
2626                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2627                         complete(ioc_gone);
2628                         ioc_gone = NULL;
2629                 }
2630                 spin_unlock(&ioc_gone_lock);
2631         }
2632 }
2633
2634 static void cfq_cic_free(struct cfq_io_context *cic)
2635 {
2636         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2637 }
2638
2639 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2640 {
2641         unsigned long flags;
2642         unsigned long dead_key = (unsigned long) cic->key;
2643
2644         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2645
2646         spin_lock_irqsave(&ioc->lock, flags);
2647         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2648         hlist_del_rcu(&cic->cic_list);
2649         spin_unlock_irqrestore(&ioc->lock, flags);
2650
2651         cfq_cic_free(cic);
2652 }
2653
2654 /*
2655  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2656  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2657  * and ->trim() which is called with the task lock held
2658  */
2659 static void cfq_free_io_context(struct io_context *ioc)
2660 {
2661         /*
2662          * ioc->refcount is zero here, or we are called from elv_unregister(),
2663          * so no more cic's are allowed to be linked into this ioc.  So it
2664          * should be ok to iterate over the known list, we will see all cic's
2665          * since no new ones are added.
2666          */
2667         __call_for_each_cic(ioc, cic_free_func);
2668 }
2669
2670 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2671 {
2672         struct cfq_queue *__cfqq, *next;
2673
2674         /*
2675          * If this queue was scheduled to merge with another queue, be
2676          * sure to drop the reference taken on that queue (and others in
2677          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2678          */
2679         __cfqq = cfqq->new_cfqq;
2680         while (__cfqq) {
2681                 if (__cfqq == cfqq) {
2682                         WARN(1, "cfqq->new_cfqq loop detected\n");
2683                         break;
2684                 }
2685                 next = __cfqq->new_cfqq;
2686                 cfq_put_queue(__cfqq);
2687                 __cfqq = next;
2688         }
2689 }
2690
2691 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2692 {
2693         if (unlikely(cfqq == cfqd->active_queue)) {
2694                 __cfq_slice_expired(cfqd, cfqq, 0);
2695                 cfq_schedule_dispatch(cfqd);
2696         }
2697
2698         cfq_put_cooperator(cfqq);
2699
2700         cfq_put_queue(cfqq);
2701 }
2702
2703 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2704                                          struct cfq_io_context *cic)
2705 {
2706         struct io_context *ioc = cic->ioc;
2707
2708         list_del_init(&cic->queue_list);
2709
2710         /*
2711          * Make sure dead mark is seen for dead queues
2712          */
2713         smp_wmb();
2714         cic->key = cfqd_dead_key(cfqd);
2715
2716         if (ioc->ioc_data == cic)
2717                 rcu_assign_pointer(ioc->ioc_data, NULL);
2718
2719         if (cic->cfqq[BLK_RW_ASYNC]) {
2720                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2721                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2722         }
2723
2724         if (cic->cfqq[BLK_RW_SYNC]) {
2725                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2726                 cic->cfqq[BLK_RW_SYNC] = NULL;
2727         }
2728 }
2729
2730 static void cfq_exit_single_io_context(struct io_context *ioc,
2731                                        struct cfq_io_context *cic)
2732 {
2733         struct cfq_data *cfqd = cic_to_cfqd(cic);
2734
2735         if (cfqd) {
2736                 struct request_queue *q = cfqd->queue;
2737                 unsigned long flags;
2738
2739                 spin_lock_irqsave(q->queue_lock, flags);
2740
2741                 /*
2742                  * Ensure we get a fresh copy of the ->key to prevent
2743                  * race between exiting task and queue
2744                  */
2745                 smp_read_barrier_depends();
2746                 if (cic->key == cfqd)
2747                         __cfq_exit_single_io_context(cfqd, cic);
2748
2749                 spin_unlock_irqrestore(q->queue_lock, flags);
2750         }
2751 }
2752
2753 /*
2754  * The process that ioc belongs to has exited, we need to clean up
2755  * and put the internal structures we have that belongs to that process.
2756  */
2757 static void cfq_exit_io_context(struct io_context *ioc)
2758 {
2759         call_for_each_cic(ioc, cfq_exit_single_io_context);
2760 }
2761
2762 static struct cfq_io_context *
2763 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2764 {
2765         struct cfq_io_context *cic;
2766
2767         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2768                                                         cfqd->queue->node);
2769         if (cic) {
2770                 cic->last_end_request = jiffies;
2771                 INIT_LIST_HEAD(&cic->queue_list);
2772                 INIT_HLIST_NODE(&cic->cic_list);
2773                 cic->dtor = cfq_free_io_context;
2774                 cic->exit = cfq_exit_io_context;
2775                 elv_ioc_count_inc(cfq_ioc_count);
2776         }
2777
2778         return cic;
2779 }
2780
2781 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2782 {
2783         struct task_struct *tsk = current;
2784         int ioprio_class;
2785
2786         if (!cfq_cfqq_prio_changed(cfqq))
2787                 return;
2788
2789         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2790         switch (ioprio_class) {
2791         default:
2792                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2793         case IOPRIO_CLASS_NONE:
2794                 /*
2795                  * no prio set, inherit CPU scheduling settings
2796                  */
2797                 cfqq->ioprio = task_nice_ioprio(tsk);
2798                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2799                 break;
2800         case IOPRIO_CLASS_RT:
2801                 cfqq->ioprio = task_ioprio(ioc);
2802                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2803                 break;
2804         case IOPRIO_CLASS_BE:
2805                 cfqq->ioprio = task_ioprio(ioc);
2806                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2807                 break;
2808         case IOPRIO_CLASS_IDLE:
2809                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2810                 cfqq->ioprio = 7;
2811                 cfq_clear_cfqq_idle_window(cfqq);
2812                 break;
2813         }
2814
2815         /*
2816          * keep track of original prio settings in case we have to temporarily
2817          * elevate the priority of this queue
2818          */
2819         cfqq->org_ioprio = cfqq->ioprio;
2820         cfqq->org_ioprio_class = cfqq->ioprio_class;
2821         cfq_clear_cfqq_prio_changed(cfqq);
2822 }
2823
2824 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2825 {
2826         struct cfq_data *cfqd = cic_to_cfqd(cic);
2827         struct cfq_queue *cfqq;
2828         unsigned long flags;
2829
2830         if (unlikely(!cfqd))
2831                 return;
2832
2833         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2834
2835         cfqq = cic->cfqq[BLK_RW_ASYNC];
2836         if (cfqq) {
2837                 struct cfq_queue *new_cfqq;
2838                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2839                                                 GFP_ATOMIC);
2840                 if (new_cfqq) {
2841                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2842                         cfq_put_queue(cfqq);
2843                 }
2844         }
2845
2846         cfqq = cic->cfqq[BLK_RW_SYNC];
2847         if (cfqq)
2848                 cfq_mark_cfqq_prio_changed(cfqq);
2849
2850         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2851 }
2852
2853 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2854 {
2855         call_for_each_cic(ioc, changed_ioprio);
2856         ioc->ioprio_changed = 0;
2857 }
2858
2859 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2860                           pid_t pid, bool is_sync)
2861 {
2862         RB_CLEAR_NODE(&cfqq->rb_node);
2863         RB_CLEAR_NODE(&cfqq->p_node);
2864         INIT_LIST_HEAD(&cfqq->fifo);
2865
2866         cfqq->ref = 0;
2867         cfqq->cfqd = cfqd;
2868
2869         cfq_mark_cfqq_prio_changed(cfqq);
2870
2871         if (is_sync) {
2872                 if (!cfq_class_idle(cfqq))
2873                         cfq_mark_cfqq_idle_window(cfqq);
2874                 cfq_mark_cfqq_sync(cfqq);
2875         }
2876         cfqq->pid = pid;
2877 }
2878
2879 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2880 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2881 {
2882         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2883         struct cfq_data *cfqd = cic_to_cfqd(cic);
2884         unsigned long flags;
2885         struct request_queue *q;
2886
2887         if (unlikely(!cfqd))
2888                 return;
2889
2890         q = cfqd->queue;
2891
2892         spin_lock_irqsave(q->queue_lock, flags);
2893
2894         if (sync_cfqq) {
2895                 /*
2896                  * Drop reference to sync queue. A new sync queue will be
2897                  * assigned in new group upon arrival of a fresh request.
2898                  */
2899                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2900                 cic_set_cfqq(cic, NULL, 1);
2901                 cfq_put_queue(sync_cfqq);
2902         }
2903
2904         spin_unlock_irqrestore(q->queue_lock, flags);
2905 }
2906
2907 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2908 {
2909         call_for_each_cic(ioc, changed_cgroup);
2910         ioc->cgroup_changed = 0;
2911 }
2912 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2913
2914 static struct cfq_queue *
2915 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2916                      struct io_context *ioc, gfp_t gfp_mask)
2917 {
2918         struct cfq_queue *cfqq, *new_cfqq = NULL;
2919         struct cfq_io_context *cic;
2920         struct cfq_group *cfqg;
2921
2922 retry:
2923         cfqg = cfq_get_cfqg(cfqd, 1);
2924         cic = cfq_cic_lookup(cfqd, ioc);
2925         /* cic always exists here */
2926         cfqq = cic_to_cfqq(cic, is_sync);
2927
2928         /*
2929          * Always try a new alloc if we fell back to the OOM cfqq
2930          * originally, since it should just be a temporary situation.
2931          */
2932         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2933                 cfqq = NULL;
2934                 if (new_cfqq) {
2935                         cfqq = new_cfqq;
2936                         new_cfqq = NULL;
2937                 } else if (gfp_mask & __GFP_WAIT) {
2938                         spin_unlock_irq(cfqd->queue->queue_lock);
2939                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2940                                         gfp_mask | __GFP_ZERO,
2941                                         cfqd->queue->node);
2942                         spin_lock_irq(cfqd->queue->queue_lock);
2943                         if (new_cfqq)
2944                                 goto retry;
2945                 } else {
2946                         cfqq = kmem_cache_alloc_node(cfq_pool,
2947                                         gfp_mask | __GFP_ZERO,
2948                                         cfqd->queue->node);
2949                 }
2950
2951                 if (cfqq) {
2952                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2953                         cfq_init_prio_data(cfqq, ioc);
2954                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2955                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2956                 } else
2957                         cfqq = &cfqd->oom_cfqq;
2958         }
2959
2960         if (new_cfqq)
2961                 kmem_cache_free(cfq_pool, new_cfqq);
2962
2963         return cfqq;
2964 }
2965
2966 static struct cfq_queue **
2967 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2968 {
2969         switch (ioprio_class) {
2970         case IOPRIO_CLASS_RT:
2971                 return &cfqd->async_cfqq[0][ioprio];
2972         case IOPRIO_CLASS_BE:
2973                 return &cfqd->async_cfqq[1][ioprio];
2974         case IOPRIO_CLASS_IDLE:
2975                 return &cfqd->async_idle_cfqq;
2976         default:
2977                 BUG();
2978         }
2979 }
2980
2981 static struct cfq_queue *
2982 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2983               gfp_t gfp_mask)
2984 {
2985         const int ioprio = task_ioprio(ioc);
2986         const int ioprio_class = task_ioprio_class(ioc);
2987         struct cfq_queue **async_cfqq = NULL;
2988         struct cfq_queue *cfqq = NULL;
2989
2990         if (!is_sync) {
2991                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2992                 cfqq = *async_cfqq;
2993         }
2994
2995         if (!cfqq)
2996                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2997
2998         /*
2999          * pin the queue now that it's allocated, scheduler exit will prune it
3000          */
3001         if (!is_sync && !(*async_cfqq)) {
3002                 cfqq->ref++;
3003                 *async_cfqq = cfqq;
3004         }
3005
3006         cfqq->ref++;
3007         return cfqq;
3008 }
3009
3010 /*
3011  * We drop cfq io contexts lazily, so we may find a dead one.
3012  */
3013 static void
3014 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3015                   struct cfq_io_context *cic)
3016 {
3017         unsigned long flags;
3018
3019         WARN_ON(!list_empty(&cic->queue_list));
3020         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3021
3022         spin_lock_irqsave(&ioc->lock, flags);
3023
3024         BUG_ON(ioc->ioc_data == cic);
3025
3026         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3027         hlist_del_rcu(&cic->cic_list);
3028         spin_unlock_irqrestore(&ioc->lock, flags);
3029
3030         cfq_cic_free(cic);
3031 }
3032
3033 static struct cfq_io_context *
3034 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3035 {
3036         struct cfq_io_context *cic;
3037         unsigned long flags;
3038
3039         if (unlikely(!ioc))
3040                 return NULL;
3041
3042         rcu_read_lock();
3043
3044         /*
3045          * we maintain a last-hit cache, to avoid browsing over the tree
3046          */
3047         cic = rcu_dereference(ioc->ioc_data);
3048         if (cic && cic->key == cfqd) {
3049                 rcu_read_unlock();
3050                 return cic;
3051         }
3052
3053         do {
3054                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3055                 rcu_read_unlock();
3056                 if (!cic)
3057                         break;
3058                 if (unlikely(cic->key != cfqd)) {
3059                         cfq_drop_dead_cic(cfqd, ioc, cic);
3060                         rcu_read_lock();
3061                         continue;
3062                 }
3063
3064                 spin_lock_irqsave(&ioc->lock, flags);
3065                 rcu_assign_pointer(ioc->ioc_data, cic);
3066                 spin_unlock_irqrestore(&ioc->lock, flags);
3067                 break;
3068         } while (1);
3069
3070         return cic;
3071 }
3072
3073 /*
3074  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3075  * the process specific cfq io context when entered from the block layer.
3076  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3077  */
3078 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3079                         struct cfq_io_context *cic, gfp_t gfp_mask)
3080 {
3081         unsigned long flags;
3082         int ret;
3083
3084         ret = radix_tree_preload(gfp_mask);
3085         if (!ret) {
3086                 cic->ioc = ioc;
3087                 cic->key = cfqd;
3088
3089                 spin_lock_irqsave(&ioc->lock, flags);
3090                 ret = radix_tree_insert(&ioc->radix_root,
3091                                                 cfqd->cic_index, cic);
3092                 if (!ret)
3093                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3094                 spin_unlock_irqrestore(&ioc->lock, flags);
3095
3096                 radix_tree_preload_end();
3097
3098                 if (!ret) {
3099                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3100                         list_add(&cic->queue_list, &cfqd->cic_list);
3101                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3102                 }
3103         }
3104
3105         if (ret)
3106                 printk(KERN_ERR "cfq: cic link failed!\n");
3107
3108         return ret;
3109 }
3110
3111 /*
3112  * Setup general io context and cfq io context. There can be several cfq
3113  * io contexts per general io context, if this process is doing io to more
3114  * than one device managed by cfq.
3115  */
3116 static struct cfq_io_context *
3117 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3118 {
3119         struct io_context *ioc = NULL;
3120         struct cfq_io_context *cic;
3121
3122         might_sleep_if(gfp_mask & __GFP_WAIT);
3123
3124         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3125         if (!ioc)
3126                 return NULL;
3127
3128         cic = cfq_cic_lookup(cfqd, ioc);
3129         if (cic)
3130                 goto out;
3131
3132         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3133         if (cic == NULL)
3134                 goto err;
3135
3136         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3137                 goto err_free;
3138
3139 out:
3140         smp_read_barrier_depends();
3141         if (unlikely(ioc->ioprio_changed))
3142                 cfq_ioc_set_ioprio(ioc);
3143
3144 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3145         if (unlikely(ioc->cgroup_changed))
3146                 cfq_ioc_set_cgroup(ioc);
3147 #endif
3148         return cic;
3149 err_free:
3150         cfq_cic_free(cic);
3151 err:
3152         put_io_context(ioc);
3153         return NULL;
3154 }
3155
3156 static void
3157 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3158 {
3159         unsigned long elapsed = jiffies - cic->last_end_request;
3160         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3161
3162         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3163         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3164         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3165 }
3166
3167 static void
3168 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3169                        struct request *rq)
3170 {
3171         sector_t sdist = 0;
3172         sector_t n_sec = blk_rq_sectors(rq);
3173         if (cfqq->last_request_pos) {
3174                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3175                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3176                 else
3177                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3178         }
3179
3180         cfqq->seek_history <<= 1;
3181         if (blk_queue_nonrot(cfqd->queue))
3182                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3183         else
3184                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3185 }
3186
3187 /*
3188  * Disable idle window if the process thinks too long or seeks so much that
3189  * it doesn't matter
3190  */
3191 static void
3192 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3193                        struct cfq_io_context *cic)
3194 {
3195         int old_idle, enable_idle;
3196
3197         /*
3198          * Don't idle for async or idle io prio class
3199          */
3200         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3201                 return;
3202
3203         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3204
3205         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3206                 cfq_mark_cfqq_deep(cfqq);
3207
3208         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3209                 enable_idle = 0;
3210         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3211             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3212                 enable_idle = 0;
3213         else if (sample_valid(cic->ttime_samples)) {
3214                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3215                         enable_idle = 0;
3216                 else
3217                         enable_idle = 1;
3218         }
3219
3220         if (old_idle != enable_idle) {
3221                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3222                 if (enable_idle)
3223                         cfq_mark_cfqq_idle_window(cfqq);
3224                 else
3225                         cfq_clear_cfqq_idle_window(cfqq);
3226         }
3227 }
3228
3229 /*
3230  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3231  * no or if we aren't sure, a 1 will cause a preempt.
3232  */
3233 static bool
3234 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3235                    struct request *rq)
3236 {
3237         struct cfq_queue *cfqq;
3238
3239         cfqq = cfqd->active_queue;
3240         if (!cfqq)
3241                 return false;
3242
3243         if (cfq_class_idle(new_cfqq))
3244                 return false;
3245
3246         if (cfq_class_idle(cfqq))
3247                 return true;
3248
3249         /*
3250          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3251          */
3252         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3253                 return false;
3254
3255         /*
3256          * if the new request is sync, but the currently running queue is
3257          * not, let the sync request have priority.
3258          */
3259         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3260                 return true;
3261
3262         if (new_cfqq->cfqg != cfqq->cfqg)
3263                 return false;
3264
3265         if (cfq_slice_used(cfqq))
3266                 return true;
3267
3268         /* Allow preemption only if we are idling on sync-noidle tree */
3269         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3270             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3271             new_cfqq->service_tree->count == 2 &&
3272             RB_EMPTY_ROOT(&cfqq->sort_list))
3273                 return true;
3274
3275         /*
3276          * So both queues are sync. Let the new request get disk time if
3277          * it's a metadata request and the current queue is doing regular IO.
3278          */
3279         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3280                 return true;
3281
3282         /*
3283          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3284          */
3285         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3286                 return true;
3287
3288         /* An idle queue should not be idle now for some reason */
3289         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3290                 return true;
3291
3292         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3293                 return false;
3294
3295         /*
3296          * if this request is as-good as one we would expect from the
3297          * current cfqq, let it preempt
3298          */
3299         if (cfq_rq_close(cfqd, cfqq, rq))
3300                 return true;
3301
3302         return false;
3303 }
3304
3305 /*
3306  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3307  * let it have half of its nominal slice.
3308  */
3309 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3310 {
3311         struct cfq_queue *old_cfqq = cfqd->active_queue;
3312
3313         cfq_log_cfqq(cfqd, cfqq, "preempt");
3314         cfq_slice_expired(cfqd, 1);
3315
3316         /*
3317          * workload type is changed, don't save slice, otherwise preempt
3318          * doesn't happen
3319          */
3320         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3321                 cfqq->cfqg->saved_workload_slice = 0;
3322
3323         /*
3324          * Put the new queue at the front of the of the current list,
3325          * so we know that it will be selected next.
3326          */
3327         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3328
3329         cfq_service_tree_add(cfqd, cfqq, 1);
3330
3331         cfqq->slice_end = 0;
3332         cfq_mark_cfqq_slice_new(cfqq);
3333 }
3334
3335 /*
3336  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3337  * something we should do about it
3338  */
3339 static void
3340 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3341                 struct request *rq)
3342 {
3343         struct cfq_io_context *cic = RQ_CIC(rq);
3344
3345         cfqd->rq_queued++;
3346         if (rq->cmd_flags & REQ_META)
3347                 cfqq->meta_pending++;
3348
3349         cfq_update_io_thinktime(cfqd, cic);
3350         cfq_update_io_seektime(cfqd, cfqq, rq);
3351         cfq_update_idle_window(cfqd, cfqq, cic);
3352
3353         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3354
3355         if (cfqq == cfqd->active_queue) {
3356                 /*
3357                  * Remember that we saw a request from this process, but
3358                  * don't start queuing just yet. Otherwise we risk seeing lots
3359                  * of tiny requests, because we disrupt the normal plugging
3360                  * and merging. If the request is already larger than a single
3361                  * page, let it rip immediately. For that case we assume that
3362                  * merging is already done. Ditto for a busy system that
3363                  * has other work pending, don't risk delaying until the
3364                  * idle timer unplug to continue working.
3365                  */
3366                 if (cfq_cfqq_wait_request(cfqq)) {
3367                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3368                             cfqd->busy_queues > 1) {
3369                                 cfq_del_timer(cfqd, cfqq);
3370                                 cfq_clear_cfqq_wait_request(cfqq);
3371                                 __blk_run_queue(cfqd->queue);
3372                         } else {
3373                                 cfq_blkiocg_update_idle_time_stats(
3374                                                 &cfqq->cfqg->blkg);
3375                                 cfq_mark_cfqq_must_dispatch(cfqq);
3376                         }
3377                 }
3378         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3379                 /*
3380                  * not the active queue - expire current slice if it is
3381                  * idle and has expired it's mean thinktime or this new queue
3382                  * has some old slice time left and is of higher priority or
3383                  * this new queue is RT and the current one is BE
3384                  */
3385                 cfq_preempt_queue(cfqd, cfqq);
3386                 __blk_run_queue(cfqd->queue);
3387         }
3388 }
3389
3390 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3391 {
3392         struct cfq_data *cfqd = q->elevator->elevator_data;
3393         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3394
3395         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3396         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3397
3398         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3399         list_add_tail(&rq->queuelist, &cfqq->fifo);
3400         cfq_add_rq_rb(rq);
3401         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3402                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3403                         rq_is_sync(rq));
3404         cfq_rq_enqueued(cfqd, cfqq, rq);
3405 }
3406
3407 /*
3408  * Update hw_tag based on peak queue depth over 50 samples under
3409  * sufficient load.
3410  */
3411 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3412 {
3413         struct cfq_queue *cfqq = cfqd->active_queue;
3414
3415         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3416                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3417
3418         if (cfqd->hw_tag == 1)
3419                 return;
3420
3421         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3422             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3423                 return;
3424
3425         /*
3426          * If active queue hasn't enough requests and can idle, cfq might not
3427          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3428          * case
3429          */
3430         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3431             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3432             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3433                 return;
3434
3435         if (cfqd->hw_tag_samples++ < 50)
3436                 return;
3437
3438         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3439                 cfqd->hw_tag = 1;
3440         else
3441                 cfqd->hw_tag = 0;
3442 }
3443
3444 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3445 {
3446         struct cfq_io_context *cic = cfqd->active_cic;
3447
3448         /* If the queue already has requests, don't wait */
3449         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3450                 return false;
3451
3452         /* If there are other queues in the group, don't wait */
3453         if (cfqq->cfqg->nr_cfqq > 1)
3454                 return false;
3455
3456         if (cfq_slice_used(cfqq))
3457                 return true;
3458
3459         /* if slice left is less than think time, wait busy */
3460         if (cic && sample_valid(cic->ttime_samples)
3461             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3462                 return true;
3463
3464         /*
3465          * If think times is less than a jiffy than ttime_mean=0 and above
3466          * will not be true. It might happen that slice has not expired yet
3467          * but will expire soon (4-5 ns) during select_queue(). To cover the
3468          * case where think time is less than a jiffy, mark the queue wait
3469          * busy if only 1 jiffy is left in the slice.
3470          */
3471         if (cfqq->slice_end - jiffies == 1)
3472                 return true;
3473
3474         return false;
3475 }
3476
3477 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3478 {
3479         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3480         struct cfq_data *cfqd = cfqq->cfqd;
3481         const int sync = rq_is_sync(rq);
3482         unsigned long now;
3483
3484         now = jiffies;
3485         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3486                      !!(rq->cmd_flags & REQ_NOIDLE));
3487
3488         cfq_update_hw_tag(cfqd);
3489
3490         WARN_ON(!cfqd->rq_in_driver);
3491         WARN_ON(!cfqq->dispatched);
3492         cfqd->rq_in_driver--;
3493         cfqq->dispatched--;
3494         (RQ_CFQG(rq))->dispatched--;
3495         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3496                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3497                         rq_data_dir(rq), rq_is_sync(rq));
3498
3499         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3500
3501         if (sync) {
3502                 RQ_CIC(rq)->last_end_request = now;
3503                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3504                         cfqd->last_delayed_sync = now;
3505         }
3506
3507         /*
3508          * If this is the active queue, check if it needs to be expired,
3509          * or if we want to idle in case it has no pending requests.
3510          */
3511         if (cfqd->active_queue == cfqq) {
3512                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3513
3514                 if (cfq_cfqq_slice_new(cfqq)) {
3515                         cfq_set_prio_slice(cfqd, cfqq);
3516                         cfq_clear_cfqq_slice_new(cfqq);
3517                 }
3518
3519                 /*
3520                  * Should we wait for next request to come in before we expire
3521                  * the queue.
3522                  */
3523                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3524                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3525                         if (!cfqd->cfq_slice_idle)
3526                                 extend_sl = cfqd->cfq_group_idle;
3527                         cfqq->slice_end = jiffies + extend_sl;
3528                         cfq_mark_cfqq_wait_busy(cfqq);
3529                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3530                 }
3531
3532                 /*
3533                  * Idling is not enabled on:
3534                  * - expired queues
3535                  * - idle-priority queues
3536                  * - async queues
3537                  * - queues with still some requests queued
3538                  * - when there is a close cooperator
3539                  */
3540                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3541                         cfq_slice_expired(cfqd, 1);
3542                 else if (sync && cfqq_empty &&
3543                          !cfq_close_cooperator(cfqd, cfqq)) {
3544                         cfq_arm_slice_timer(cfqd);
3545                 }
3546         }
3547
3548         if (!cfqd->rq_in_driver)
3549                 cfq_schedule_dispatch(cfqd);
3550 }
3551
3552 /*
3553  * we temporarily boost lower priority queues if they are holding fs exclusive
3554  * resources. they are boosted to normal prio (CLASS_BE/4)
3555  */
3556 static void cfq_prio_boost(struct cfq_queue *cfqq)
3557 {
3558         if (has_fs_excl()) {
3559                 /*
3560                  * boost idle prio on transactions that would lock out other
3561                  * users of the filesystem
3562                  */
3563                 if (cfq_class_idle(cfqq))
3564                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3565                 if (cfqq->ioprio > IOPRIO_NORM)
3566                         cfqq->ioprio = IOPRIO_NORM;
3567         } else {
3568                 /*
3569                  * unboost the queue (if needed)
3570                  */
3571                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3572                 cfqq->ioprio = cfqq->org_ioprio;
3573         }
3574 }
3575
3576 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3577 {
3578         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3579                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3580                 return ELV_MQUEUE_MUST;
3581         }
3582
3583         return ELV_MQUEUE_MAY;
3584 }
3585
3586 static int cfq_may_queue(struct request_queue *q, int rw)
3587 {
3588         struct cfq_data *cfqd = q->elevator->elevator_data;
3589         struct task_struct *tsk = current;
3590         struct cfq_io_context *cic;
3591         struct cfq_queue *cfqq;
3592
3593         /*
3594          * don't force setup of a queue from here, as a call to may_queue
3595          * does not necessarily imply that a request actually will be queued.
3596          * so just lookup a possibly existing queue, or return 'may queue'
3597          * if that fails
3598          */
3599         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3600         if (!cic)
3601                 return ELV_MQUEUE_MAY;
3602
3603         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3604         if (cfqq) {
3605                 cfq_init_prio_data(cfqq, cic->ioc);
3606                 cfq_prio_boost(cfqq);
3607
3608                 return __cfq_may_queue(cfqq);
3609         }
3610
3611         return ELV_MQUEUE_MAY;
3612 }
3613
3614 /*
3615  * queue lock held here
3616  */
3617 static void cfq_put_request(struct request *rq)
3618 {
3619         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3620
3621         if (cfqq) {
3622                 const int rw = rq_data_dir(rq);
3623
3624                 BUG_ON(!cfqq->allocated[rw]);
3625                 cfqq->allocated[rw]--;
3626
3627                 put_io_context(RQ_CIC(rq)->ioc);
3628
3629                 rq->elevator_private[0] = NULL;
3630                 rq->elevator_private[1] = NULL;
3631
3632                 /* Put down rq reference on cfqg */
3633                 cfq_put_cfqg(RQ_CFQG(rq));
3634                 rq->elevator_private[2] = NULL;
3635
3636                 cfq_put_queue(cfqq);
3637         }
3638 }
3639
3640 static struct cfq_queue *
3641 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3642                 struct cfq_queue *cfqq)
3643 {
3644         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3645         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3646         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3647         cfq_put_queue(cfqq);
3648         return cic_to_cfqq(cic, 1);
3649 }
3650
3651 /*
3652  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3653  * was the last process referring to said cfqq.
3654  */
3655 static struct cfq_queue *
3656 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3657 {
3658         if (cfqq_process_refs(cfqq) == 1) {
3659                 cfqq->pid = current->pid;
3660                 cfq_clear_cfqq_coop(cfqq);
3661                 cfq_clear_cfqq_split_coop(cfqq);
3662                 return cfqq;
3663         }
3664
3665         cic_set_cfqq(cic, NULL, 1);
3666
3667         cfq_put_cooperator(cfqq);
3668
3669         cfq_put_queue(cfqq);
3670         return NULL;
3671 }
3672 /*
3673  * Allocate cfq data structures associated with this request.
3674  */
3675 static int
3676 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3677 {
3678         struct cfq_data *cfqd = q->elevator->elevator_data;
3679         struct cfq_io_context *cic;
3680         const int rw = rq_data_dir(rq);
3681         const bool is_sync = rq_is_sync(rq);
3682         struct cfq_queue *cfqq;
3683         unsigned long flags;
3684
3685         might_sleep_if(gfp_mask & __GFP_WAIT);
3686
3687         cic = cfq_get_io_context(cfqd, gfp_mask);
3688
3689         spin_lock_irqsave(q->queue_lock, flags);
3690
3691         if (!cic)
3692                 goto queue_fail;
3693
3694 new_queue:
3695         cfqq = cic_to_cfqq(cic, is_sync);
3696         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3697                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3698                 cic_set_cfqq(cic, cfqq, is_sync);
3699         } else {
3700                 /*
3701                  * If the queue was seeky for too long, break it apart.
3702                  */
3703                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3704                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3705                         cfqq = split_cfqq(cic, cfqq);
3706                         if (!cfqq)
3707                                 goto new_queue;
3708                 }
3709
3710                 /*
3711                  * Check to see if this queue is scheduled to merge with
3712                  * another, closely cooperating queue.  The merging of
3713                  * queues happens here as it must be done in process context.
3714                  * The reference on new_cfqq was taken in merge_cfqqs.
3715                  */
3716                 if (cfqq->new_cfqq)
3717                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3718         }
3719
3720         cfqq->allocated[rw]++;
3721
3722         cfqq->ref++;
3723         rq->elevator_private[0] = cic;
3724         rq->elevator_private[1] = cfqq;
3725         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3726         spin_unlock_irqrestore(q->queue_lock, flags);
3727         return 0;
3728
3729 queue_fail:
3730         if (cic)
3731                 put_io_context(cic->ioc);
3732
3733         cfq_schedule_dispatch(cfqd);
3734         spin_unlock_irqrestore(q->queue_lock, flags);
3735         cfq_log(cfqd, "set_request fail");
3736         return 1;
3737 }
3738
3739 static void cfq_kick_queue(struct work_struct *work)
3740 {
3741         struct cfq_data *cfqd =
3742                 container_of(work, struct cfq_data, unplug_work);
3743         struct request_queue *q = cfqd->queue;
3744
3745         spin_lock_irq(q->queue_lock);
3746         __blk_run_queue(cfqd->queue);
3747         spin_unlock_irq(q->queue_lock);
3748 }
3749
3750 /*
3751  * Timer running if the active_queue is currently idling inside its time slice
3752  */
3753 static void cfq_idle_slice_timer(unsigned long data)
3754 {
3755         struct cfq_data *cfqd = (struct cfq_data *) data;
3756         struct cfq_queue *cfqq;
3757         unsigned long flags;
3758         int timed_out = 1;
3759
3760         cfq_log(cfqd, "idle timer fired");
3761
3762         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3763
3764         cfqq = cfqd->active_queue;
3765         if (cfqq) {
3766                 timed_out = 0;
3767
3768                 /*
3769                  * We saw a request before the queue expired, let it through
3770                  */
3771                 if (cfq_cfqq_must_dispatch(cfqq))
3772                         goto out_kick;
3773
3774                 /*
3775                  * expired
3776                  */
3777                 if (cfq_slice_used(cfqq))
3778                         goto expire;
3779
3780                 /*
3781                  * only expire and reinvoke request handler, if there are
3782                  * other queues with pending requests
3783                  */
3784                 if (!cfqd->busy_queues)
3785                         goto out_cont;
3786
3787                 /*
3788                  * not expired and it has a request pending, let it dispatch
3789                  */
3790                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3791                         goto out_kick;
3792
3793                 /*
3794                  * Queue depth flag is reset only when the idle didn't succeed
3795                  */
3796                 cfq_clear_cfqq_deep(cfqq);
3797         }
3798 expire:
3799         cfq_slice_expired(cfqd, timed_out);
3800 out_kick:
3801         cfq_schedule_dispatch(cfqd);
3802 out_cont:
3803         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3804 }
3805
3806 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3807 {
3808         del_timer_sync(&cfqd->idle_slice_timer);
3809         cancel_work_sync(&cfqd->unplug_work);
3810 }
3811
3812 static void cfq_put_async_queues(struct cfq_data *cfqd)
3813 {
3814         int i;
3815
3816         for (i = 0; i < IOPRIO_BE_NR; i++) {
3817                 if (cfqd->async_cfqq[0][i])
3818                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3819                 if (cfqd->async_cfqq[1][i])
3820                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3821         }
3822
3823         if (cfqd->async_idle_cfqq)
3824                 cfq_put_queue(cfqd->async_idle_cfqq);
3825 }
3826
3827 static void cfq_cfqd_free(struct rcu_head *head)
3828 {
3829         kfree(container_of(head, struct cfq_data, rcu));
3830 }
3831
3832 static void cfq_exit_queue(struct elevator_queue *e)
3833 {
3834         struct cfq_data *cfqd = e->elevator_data;
3835         struct request_queue *q = cfqd->queue;
3836
3837         cfq_shutdown_timer_wq(cfqd);
3838
3839         spin_lock_irq(q->queue_lock);
3840
3841         if (cfqd->active_queue)
3842                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3843
3844         while (!list_empty(&cfqd->cic_list)) {
3845                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3846                                                         struct cfq_io_context,
3847                                                         queue_list);
3848
3849                 __cfq_exit_single_io_context(cfqd, cic);
3850         }
3851
3852         cfq_put_async_queues(cfqd);
3853         cfq_release_cfq_groups(cfqd);
3854         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3855
3856         spin_unlock_irq(q->queue_lock);
3857
3858         cfq_shutdown_timer_wq(cfqd);
3859
3860         spin_lock(&cic_index_lock);
3861         ida_remove(&cic_index_ida, cfqd->cic_index);
3862         spin_unlock(&cic_index_lock);
3863
3864         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3865         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3866 }
3867
3868 static int cfq_alloc_cic_index(void)
3869 {
3870         int index, error;
3871
3872         do {
3873                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3874                         return -ENOMEM;
3875
3876                 spin_lock(&cic_index_lock);
3877                 error = ida_get_new(&cic_index_ida, &index);
3878                 spin_unlock(&cic_index_lock);
3879                 if (error && error != -EAGAIN)
3880                         return error;
3881         } while (error);
3882
3883         return index;
3884 }
3885
3886 static void *cfq_init_queue(struct request_queue *q)
3887 {
3888         struct cfq_data *cfqd;
3889         int i, j;
3890         struct cfq_group *cfqg;
3891         struct cfq_rb_root *st;
3892
3893         i = cfq_alloc_cic_index();
3894         if (i < 0)
3895                 return NULL;
3896
3897         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3898         if (!cfqd)
3899                 return NULL;
3900
3901         /*
3902          * Don't need take queue_lock in the routine, since we are
3903          * initializing the ioscheduler, and nobody is using cfqd
3904          */
3905         cfqd->cic_index = i;
3906
3907         /* Init root service tree */
3908         cfqd->grp_service_tree = CFQ_RB_ROOT;
3909
3910         /* Init root group */
3911         cfqg = &cfqd->root_group;
3912         for_each_cfqg_st(cfqg, i, j, st)
3913                 *st = CFQ_RB_ROOT;
3914         RB_CLEAR_NODE(&cfqg->rb_node);
3915
3916         /* Give preference to root group over other groups */
3917         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3918
3919 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3920         /*
3921          * Take a reference to root group which we never drop. This is just
3922          * to make sure that cfq_put_cfqg() does not try to kfree root group
3923          */
3924         cfqg->ref = 1;
3925         rcu_read_lock();
3926         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3927                                         (void *)cfqd, 0);
3928         rcu_read_unlock();
3929 #endif
3930         /*
3931          * Not strictly needed (since RB_ROOT just clears the node and we
3932          * zeroed cfqd on alloc), but better be safe in case someone decides
3933          * to add magic to the rb code
3934          */
3935         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3936                 cfqd->prio_trees[i] = RB_ROOT;
3937
3938         /*
3939          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3940          * Grab a permanent reference to it, so that the normal code flow
3941          * will not attempt to free it.
3942          */
3943         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3944         cfqd->oom_cfqq.ref++;
3945         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3946
3947         INIT_LIST_HEAD(&cfqd->cic_list);
3948
3949         cfqd->queue = q;
3950
3951         init_timer(&cfqd->idle_slice_timer);
3952         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3953         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3954
3955         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3956
3957         cfqd->cfq_quantum = cfq_quantum;
3958         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3959         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3960         cfqd->cfq_back_max = cfq_back_max;
3961         cfqd->cfq_back_penalty = cfq_back_penalty;
3962         cfqd->cfq_slice[0] = cfq_slice_async;
3963         cfqd->cfq_slice[1] = cfq_slice_sync;
3964         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3965         cfqd->cfq_slice_idle = cfq_slice_idle;
3966         cfqd->cfq_group_idle = cfq_group_idle;
3967         cfqd->cfq_latency = 1;
3968         cfqd->hw_tag = -1;
3969         /*
3970          * we optimistically start assuming sync ops weren't delayed in last
3971          * second, in order to have larger depth for async operations.
3972          */
3973         cfqd->last_delayed_sync = jiffies - HZ;
3974         return cfqd;
3975 }
3976
3977 static void cfq_slab_kill(void)
3978 {
3979         /*
3980          * Caller already ensured that pending RCU callbacks are completed,
3981          * so we should have no busy allocations at this point.
3982          */
3983         if (cfq_pool)
3984                 kmem_cache_destroy(cfq_pool);
3985         if (cfq_ioc_pool)
3986                 kmem_cache_destroy(cfq_ioc_pool);
3987 }
3988
3989 static int __init cfq_slab_setup(void)
3990 {
3991         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3992         if (!cfq_pool)
3993                 goto fail;
3994
3995         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3996         if (!cfq_ioc_pool)
3997                 goto fail;
3998
3999         return 0;
4000 fail:
4001         cfq_slab_kill();
4002         return -ENOMEM;
4003 }
4004
4005 /*
4006  * sysfs parts below -->
4007  */
4008 static ssize_t
4009 cfq_var_show(unsigned int var, char *page)
4010 {
4011         return sprintf(page, "%d\n", var);
4012 }
4013
4014 static ssize_t
4015 cfq_var_store(unsigned int *var, const char *page, size_t count)
4016 {
4017         char *p = (char *) page;
4018
4019         *var = simple_strtoul(p, &p, 10);
4020         return count;
4021 }
4022
4023 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4024 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4025 {                                                                       \
4026         struct cfq_data *cfqd = e->elevator_data;                       \
4027         unsigned int __data = __VAR;                                    \
4028         if (__CONV)                                                     \
4029                 __data = jiffies_to_msecs(__data);                      \
4030         return cfq_var_show(__data, (page));                            \
4031 }
4032 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4033 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4034 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4035 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4036 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4037 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4038 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4039 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4040 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4041 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4042 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4043 #undef SHOW_FUNCTION
4044
4045 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4046 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4047 {                                                                       \
4048         struct cfq_data *cfqd = e->elevator_data;                       \
4049         unsigned int __data;                                            \
4050         int ret = cfq_var_store(&__data, (page), count);                \
4051         if (__data < (MIN))                                             \
4052                 __data = (MIN);                                         \
4053         else if (__data > (MAX))                                        \
4054                 __data = (MAX);                                         \
4055         if (__CONV)                                                     \
4056                 *(__PTR) = msecs_to_jiffies(__data);                    \
4057         else                                                            \
4058                 *(__PTR) = __data;                                      \
4059         return ret;                                                     \
4060 }
4061 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4062 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4063                 UINT_MAX, 1);
4064 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4065                 UINT_MAX, 1);
4066 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4067 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4068                 UINT_MAX, 0);
4069 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4070 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4071 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4072 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4073 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4074                 UINT_MAX, 0);
4075 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4076 #undef STORE_FUNCTION
4077
4078 #define CFQ_ATTR(name) \
4079         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4080
4081 static struct elv_fs_entry cfq_attrs[] = {
4082         CFQ_ATTR(quantum),
4083         CFQ_ATTR(fifo_expire_sync),
4084         CFQ_ATTR(fifo_expire_async),
4085         CFQ_ATTR(back_seek_max),
4086         CFQ_ATTR(back_seek_penalty),
4087         CFQ_ATTR(slice_sync),
4088         CFQ_ATTR(slice_async),
4089         CFQ_ATTR(slice_async_rq),
4090         CFQ_ATTR(slice_idle),
4091         CFQ_ATTR(group_idle),
4092         CFQ_ATTR(low_latency),
4093         __ATTR_NULL
4094 };
4095
4096 static struct elevator_type iosched_cfq = {
4097         .ops = {
4098                 .elevator_merge_fn =            cfq_merge,
4099                 .elevator_merged_fn =           cfq_merged_request,
4100                 .elevator_merge_req_fn =        cfq_merged_requests,
4101                 .elevator_allow_merge_fn =      cfq_allow_merge,
4102                 .elevator_bio_merged_fn =       cfq_bio_merged,
4103                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4104                 .elevator_add_req_fn =          cfq_insert_request,
4105                 .elevator_activate_req_fn =     cfq_activate_request,
4106                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4107                 .elevator_completed_req_fn =    cfq_completed_request,
4108                 .elevator_former_req_fn =       elv_rb_former_request,
4109                 .elevator_latter_req_fn =       elv_rb_latter_request,
4110                 .elevator_set_req_fn =          cfq_set_request,
4111                 .elevator_put_req_fn =          cfq_put_request,
4112                 .elevator_may_queue_fn =        cfq_may_queue,
4113                 .elevator_init_fn =             cfq_init_queue,
4114                 .elevator_exit_fn =             cfq_exit_queue,
4115                 .trim =                         cfq_free_io_context,
4116         },
4117         .elevator_attrs =       cfq_attrs,
4118         .elevator_name =        "cfq",
4119         .elevator_owner =       THIS_MODULE,
4120 };
4121
4122 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4123 static struct blkio_policy_type blkio_policy_cfq = {
4124         .ops = {
4125                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4126                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4127         },
4128         .plid = BLKIO_POLICY_PROP,
4129 };
4130 #else
4131 static struct blkio_policy_type blkio_policy_cfq;
4132 #endif
4133
4134 static int __init cfq_init(void)
4135 {
4136         /*
4137          * could be 0 on HZ < 1000 setups
4138          */
4139         if (!cfq_slice_async)
4140                 cfq_slice_async = 1;
4141         if (!cfq_slice_idle)
4142                 cfq_slice_idle = 1;
4143
4144 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4145         if (!cfq_group_idle)
4146                 cfq_group_idle = 1;
4147 #else
4148                 cfq_group_idle = 0;
4149 #endif
4150         if (cfq_slab_setup())
4151                 return -ENOMEM;
4152
4153         elv_register(&iosched_cfq);
4154         blkio_policy_register(&blkio_policy_cfq);
4155
4156         return 0;
4157 }
4158
4159 static void __exit cfq_exit(void)
4160 {
4161         DECLARE_COMPLETION_ONSTACK(all_gone);
4162         blkio_policy_unregister(&blkio_policy_cfq);
4163         elv_unregister(&iosched_cfq);
4164         ioc_gone = &all_gone;
4165         /* ioc_gone's update must be visible before reading ioc_count */
4166         smp_wmb();
4167
4168         /*
4169          * this also protects us from entering cfq_slab_kill() with
4170          * pending RCU callbacks
4171          */
4172         if (elv_ioc_count_read(cfq_ioc_count))
4173                 wait_for_completion(&all_gone);
4174         ida_destroy(&cic_index_ida);
4175         cfq_slab_kill();
4176 }
4177
4178 module_init(cfq_init);
4179 module_exit(cfq_exit);
4180
4181 MODULE_AUTHOR("Jens Axboe");
4182 MODULE_LICENSE("GPL");
4183 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");