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