55d2a21f7f0678cf9404ea801cb77c88dfb0ad83
[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 /*
46  * Allow merged cfqqs to perform this amount of seeky I/O before
47  * deciding to break the queues up again.
48  */
49 #define CFQQ_COOP_TOUT          (HZ)
50
51 #define CFQ_SLICE_SCALE         (5)
52 #define CFQ_HW_QUEUE_MIN        (5)
53 #define CFQ_SERVICE_SHIFT       12
54
55 #define RQ_CIC(rq)              \
56         ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
58
59 static struct kmem_cache *cfq_pool;
60 static struct kmem_cache *cfq_ioc_pool;
61
62 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
63 static struct completion *ioc_gone;
64 static DEFINE_SPINLOCK(ioc_gone_lock);
65
66 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
67 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
68 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69
70 #define sample_valid(samples)   ((samples) > 80)
71 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
72
73 /*
74  * Most of our rbtree usage is for sorting with min extraction, so
75  * if we cache the leftmost node we don't have to walk down the tree
76  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
77  * move this into the elevator for the rq sorting as well.
78  */
79 struct cfq_rb_root {
80         struct rb_root rb;
81         struct rb_node *left;
82         unsigned count;
83         u64 min_vdisktime;
84         struct rb_node *active;
85         unsigned total_weight;
86 };
87 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
88
89 /*
90  * Per process-grouping structure
91  */
92 struct cfq_queue {
93         /* reference count */
94         atomic_t ref;
95         /* various state flags, see below */
96         unsigned int flags;
97         /* parent cfq_data */
98         struct cfq_data *cfqd;
99         /* service_tree member */
100         struct rb_node rb_node;
101         /* service_tree key */
102         unsigned long rb_key;
103         /* prio tree member */
104         struct rb_node p_node;
105         /* prio tree root we belong to, if any */
106         struct rb_root *p_root;
107         /* sorted list of pending requests */
108         struct rb_root sort_list;
109         /* if fifo isn't expired, next request to serve */
110         struct request *next_rq;
111         /* requests queued in sort_list */
112         int queued[2];
113         /* currently allocated requests */
114         int allocated[2];
115         /* fifo list of requests in sort_list */
116         struct list_head fifo;
117
118         /* time when queue got scheduled in to dispatch first request. */
119         unsigned long dispatch_start;
120         /* time when first request from queue completed and slice started. */
121         unsigned long slice_start;
122         unsigned long slice_end;
123         long slice_resid;
124         unsigned int slice_dispatch;
125
126         /* pending metadata requests */
127         int meta_pending;
128         /* number of requests that are on the dispatch list or inside driver */
129         int dispatched;
130
131         /* io prio of this group */
132         unsigned short ioprio, org_ioprio;
133         unsigned short ioprio_class, org_ioprio_class;
134
135         unsigned int seek_samples;
136         u64 seek_total;
137         sector_t seek_mean;
138         sector_t last_request_pos;
139         unsigned long seeky_start;
140
141         pid_t pid;
142
143         struct cfq_rb_root *service_tree;
144         struct cfq_queue *new_cfqq;
145         struct cfq_group *cfqg;
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 };
193
194 /*
195  * Per block device queue structure
196  */
197 struct cfq_data {
198         struct request_queue *queue;
199         /* Root service tree for cfq_groups */
200         struct cfq_rb_root grp_service_tree;
201         struct cfq_group root_group;
202         /* Number of active cfq groups on group service tree */
203         int nr_groups;
204
205         /*
206          * The priority currently being served
207          */
208         enum wl_prio_t serving_prio;
209         enum wl_type_t serving_type;
210         unsigned long workload_expires;
211         struct cfq_group *serving_group;
212         bool noidle_tree_requires_idle;
213
214         /*
215          * Each priority tree is sorted by next_request position.  These
216          * trees are used when determining if two or more queues are
217          * interleaving requests (see cfq_close_cooperator).
218          */
219         struct rb_root prio_trees[CFQ_PRIO_LISTS];
220
221         unsigned int busy_queues;
222
223         int rq_in_driver[2];
224         int sync_flight;
225
226         /*
227          * queue-depth detection
228          */
229         int rq_queued;
230         int hw_tag;
231         /*
232          * hw_tag can be
233          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
234          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
235          *  0 => no NCQ
236          */
237         int hw_tag_est_depth;
238         unsigned int hw_tag_samples;
239
240         /*
241          * idle window management
242          */
243         struct timer_list idle_slice_timer;
244         struct work_struct unplug_work;
245
246         struct cfq_queue *active_queue;
247         struct cfq_io_context *active_cic;
248
249         /*
250          * async queue for each priority case
251          */
252         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
253         struct cfq_queue *async_idle_cfqq;
254
255         sector_t last_position;
256
257         /*
258          * tunables, see top of file
259          */
260         unsigned int cfq_quantum;
261         unsigned int cfq_fifo_expire[2];
262         unsigned int cfq_back_penalty;
263         unsigned int cfq_back_max;
264         unsigned int cfq_slice[2];
265         unsigned int cfq_slice_async_rq;
266         unsigned int cfq_slice_idle;
267         unsigned int cfq_latency;
268
269         struct list_head cic_list;
270
271         /*
272          * Fallback dummy cfqq for extreme OOM conditions
273          */
274         struct cfq_queue oom_cfqq;
275
276         unsigned long last_end_sync_rq;
277 };
278
279 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
280                                             enum wl_prio_t prio,
281                                             enum wl_type_t type,
282                                             struct cfq_data *cfqd)
283 {
284         if (!cfqg)
285                 return NULL;
286
287         if (prio == IDLE_WORKLOAD)
288                 return &cfqg->service_tree_idle;
289
290         return &cfqg->service_trees[prio][type];
291 }
292
293 enum cfqq_state_flags {
294         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
295         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
296         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
297         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
298         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
299         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
300         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
301         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
302         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
303         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
304         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
305 };
306
307 #define CFQ_CFQQ_FNS(name)                                              \
308 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
309 {                                                                       \
310         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
311 }                                                                       \
312 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
313 {                                                                       \
314         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
315 }                                                                       \
316 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
317 {                                                                       \
318         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
319 }
320
321 CFQ_CFQQ_FNS(on_rr);
322 CFQ_CFQQ_FNS(wait_request);
323 CFQ_CFQQ_FNS(must_dispatch);
324 CFQ_CFQQ_FNS(must_alloc_slice);
325 CFQ_CFQQ_FNS(fifo_expire);
326 CFQ_CFQQ_FNS(idle_window);
327 CFQ_CFQQ_FNS(prio_changed);
328 CFQ_CFQQ_FNS(slice_new);
329 CFQ_CFQQ_FNS(sync);
330 CFQ_CFQQ_FNS(coop);
331 CFQ_CFQQ_FNS(deep);
332 #undef CFQ_CFQQ_FNS
333
334 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
335         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
336 #define cfq_log(cfqd, fmt, args...)     \
337         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
338
339 /* Traverses through cfq group service trees */
340 #define for_each_cfqg_st(cfqg, i, j, st) \
341         for (i = 0; i <= IDLE_WORKLOAD; i++) \
342                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
343                         : &cfqg->service_tree_idle; \
344                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
345                         (i == IDLE_WORKLOAD && j == 0); \
346                         j++, st = i < IDLE_WORKLOAD ? \
347                         &cfqg->service_trees[i][j]: NULL) \
348
349
350 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
351 {
352         if (cfq_class_idle(cfqq))
353                 return IDLE_WORKLOAD;
354         if (cfq_class_rt(cfqq))
355                 return RT_WORKLOAD;
356         return BE_WORKLOAD;
357 }
358
359
360 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
361 {
362         if (!cfq_cfqq_sync(cfqq))
363                 return ASYNC_WORKLOAD;
364         if (!cfq_cfqq_idle_window(cfqq))
365                 return SYNC_NOIDLE_WORKLOAD;
366         return SYNC_WORKLOAD;
367 }
368
369 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
370                                         struct cfq_data *cfqd,
371                                         struct cfq_group *cfqg)
372 {
373         if (wl == IDLE_WORKLOAD)
374                 return cfqg->service_tree_idle.count;
375
376         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
377                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
378                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
379 }
380
381 static void cfq_dispatch_insert(struct request_queue *, struct request *);
382 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
383                                        struct io_context *, gfp_t);
384 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
385                                                 struct io_context *);
386
387 static inline int rq_in_driver(struct cfq_data *cfqd)
388 {
389         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
390 }
391
392 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
393                                             bool is_sync)
394 {
395         return cic->cfqq[is_sync];
396 }
397
398 static inline void cic_set_cfqq(struct cfq_io_context *cic,
399                                 struct cfq_queue *cfqq, bool is_sync)
400 {
401         cic->cfqq[is_sync] = cfqq;
402 }
403
404 /*
405  * We regard a request as SYNC, if it's either a read or has the SYNC bit
406  * set (in which case it could also be direct WRITE).
407  */
408 static inline bool cfq_bio_sync(struct bio *bio)
409 {
410         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
411 }
412
413 /*
414  * scheduler run of queue, if there are requests pending and no one in the
415  * driver that will restart queueing
416  */
417 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
418 {
419         if (cfqd->busy_queues) {
420                 cfq_log(cfqd, "schedule dispatch");
421                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
422         }
423 }
424
425 static int cfq_queue_empty(struct request_queue *q)
426 {
427         struct cfq_data *cfqd = q->elevator->elevator_data;
428
429         return !cfqd->rq_queued;
430 }
431
432 /*
433  * Scale schedule slice based on io priority. Use the sync time slice only
434  * if a queue is marked sync and has sync io queued. A sync queue with async
435  * io only, should not get full sync slice length.
436  */
437 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
438                                  unsigned short prio)
439 {
440         const int base_slice = cfqd->cfq_slice[sync];
441
442         WARN_ON(prio >= IOPRIO_BE_NR);
443
444         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
445 }
446
447 static inline int
448 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
449 {
450         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
451 }
452
453 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
454 {
455         u64 d = delta << CFQ_SERVICE_SHIFT;
456
457         d = d * BLKIO_WEIGHT_DEFAULT;
458         do_div(d, cfqg->weight);
459         return d;
460 }
461
462 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
463 {
464         s64 delta = (s64)(vdisktime - min_vdisktime);
465         if (delta > 0)
466                 min_vdisktime = vdisktime;
467
468         return min_vdisktime;
469 }
470
471 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
472 {
473         s64 delta = (s64)(vdisktime - min_vdisktime);
474         if (delta < 0)
475                 min_vdisktime = vdisktime;
476
477         return min_vdisktime;
478 }
479
480 static void update_min_vdisktime(struct cfq_rb_root *st)
481 {
482         u64 vdisktime = st->min_vdisktime;
483         struct cfq_group *cfqg;
484
485         if (st->active) {
486                 cfqg = rb_entry_cfqg(st->active);
487                 vdisktime = cfqg->vdisktime;
488         }
489
490         if (st->left) {
491                 cfqg = rb_entry_cfqg(st->left);
492                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
493         }
494
495         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
496 }
497
498 /*
499  * get averaged number of queues of RT/BE priority.
500  * average is updated, with a formula that gives more weight to higher numbers,
501  * to quickly follows sudden increases and decrease slowly
502  */
503
504 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
505                                         struct cfq_group *cfqg, bool rt)
506 {
507         unsigned min_q, max_q;
508         unsigned mult  = cfq_hist_divisor - 1;
509         unsigned round = cfq_hist_divisor / 2;
510         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
511
512         min_q = min(cfqg->busy_queues_avg[rt], busy);
513         max_q = max(cfqg->busy_queues_avg[rt], busy);
514         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
515                 cfq_hist_divisor;
516         return cfqg->busy_queues_avg[rt];
517 }
518
519 static inline unsigned
520 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
521 {
522         struct cfq_rb_root *st = &cfqd->grp_service_tree;
523
524         return cfq_target_latency * cfqg->weight / st->total_weight;
525 }
526
527 static inline void
528 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
529 {
530         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
531         if (cfqd->cfq_latency) {
532                 /*
533                  * interested queues (we consider only the ones with the same
534                  * priority class in the cfq group)
535                  */
536                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
537                                                 cfq_class_rt(cfqq));
538                 unsigned sync_slice = cfqd->cfq_slice[1];
539                 unsigned expect_latency = sync_slice * iq;
540                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
541
542                 if (expect_latency > group_slice) {
543                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
544                         /* scale low_slice according to IO priority
545                          * and sync vs async */
546                         unsigned low_slice =
547                                 min(slice, base_low_slice * slice / sync_slice);
548                         /* the adapted slice value is scaled to fit all iqs
549                          * into the target latency */
550                         slice = max(slice * group_slice / expect_latency,
551                                     low_slice);
552                 }
553         }
554         cfqq->slice_start = jiffies;
555         cfqq->slice_end = jiffies + slice;
556         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
557 }
558
559 /*
560  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
561  * isn't valid until the first request from the dispatch is activated
562  * and the slice time set.
563  */
564 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
565 {
566         if (cfq_cfqq_slice_new(cfqq))
567                 return 0;
568         if (time_before(jiffies, cfqq->slice_end))
569                 return 0;
570
571         return 1;
572 }
573
574 /*
575  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
576  * We choose the request that is closest to the head right now. Distance
577  * behind the head is penalized and only allowed to a certain extent.
578  */
579 static struct request *
580 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
581 {
582         sector_t s1, s2, d1 = 0, d2 = 0;
583         unsigned long back_max;
584 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
585 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
586         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
587
588         if (rq1 == NULL || rq1 == rq2)
589                 return rq2;
590         if (rq2 == NULL)
591                 return rq1;
592
593         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
594                 return rq1;
595         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
596                 return rq2;
597         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
598                 return rq1;
599         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
600                 return rq2;
601
602         s1 = blk_rq_pos(rq1);
603         s2 = blk_rq_pos(rq2);
604
605         /*
606          * by definition, 1KiB is 2 sectors
607          */
608         back_max = cfqd->cfq_back_max * 2;
609
610         /*
611          * Strict one way elevator _except_ in the case where we allow
612          * short backward seeks which are biased as twice the cost of a
613          * similar forward seek.
614          */
615         if (s1 >= last)
616                 d1 = s1 - last;
617         else if (s1 + back_max >= last)
618                 d1 = (last - s1) * cfqd->cfq_back_penalty;
619         else
620                 wrap |= CFQ_RQ1_WRAP;
621
622         if (s2 >= last)
623                 d2 = s2 - last;
624         else if (s2 + back_max >= last)
625                 d2 = (last - s2) * cfqd->cfq_back_penalty;
626         else
627                 wrap |= CFQ_RQ2_WRAP;
628
629         /* Found required data */
630
631         /*
632          * By doing switch() on the bit mask "wrap" we avoid having to
633          * check two variables for all permutations: --> faster!
634          */
635         switch (wrap) {
636         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
637                 if (d1 < d2)
638                         return rq1;
639                 else if (d2 < d1)
640                         return rq2;
641                 else {
642                         if (s1 >= s2)
643                                 return rq1;
644                         else
645                                 return rq2;
646                 }
647
648         case CFQ_RQ2_WRAP:
649                 return rq1;
650         case CFQ_RQ1_WRAP:
651                 return rq2;
652         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
653         default:
654                 /*
655                  * Since both rqs are wrapped,
656                  * start with the one that's further behind head
657                  * (--> only *one* back seek required),
658                  * since back seek takes more time than forward.
659                  */
660                 if (s1 <= s2)
661                         return rq1;
662                 else
663                         return rq2;
664         }
665 }
666
667 /*
668  * The below is leftmost cache rbtree addon
669  */
670 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
671 {
672         /* Service tree is empty */
673         if (!root->count)
674                 return NULL;
675
676         if (!root->left)
677                 root->left = rb_first(&root->rb);
678
679         if (root->left)
680                 return rb_entry(root->left, struct cfq_queue, rb_node);
681
682         return NULL;
683 }
684
685 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
686 {
687         if (!root->left)
688                 root->left = rb_first(&root->rb);
689
690         if (root->left)
691                 return rb_entry_cfqg(root->left);
692
693         return NULL;
694 }
695
696 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
697 {
698         rb_erase(n, root);
699         RB_CLEAR_NODE(n);
700 }
701
702 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
703 {
704         if (root->left == n)
705                 root->left = NULL;
706         rb_erase_init(n, &root->rb);
707         --root->count;
708 }
709
710 /*
711  * would be nice to take fifo expire time into account as well
712  */
713 static struct request *
714 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
715                   struct request *last)
716 {
717         struct rb_node *rbnext = rb_next(&last->rb_node);
718         struct rb_node *rbprev = rb_prev(&last->rb_node);
719         struct request *next = NULL, *prev = NULL;
720
721         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
722
723         if (rbprev)
724                 prev = rb_entry_rq(rbprev);
725
726         if (rbnext)
727                 next = rb_entry_rq(rbnext);
728         else {
729                 rbnext = rb_first(&cfqq->sort_list);
730                 if (rbnext && rbnext != &last->rb_node)
731                         next = rb_entry_rq(rbnext);
732         }
733
734         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
735 }
736
737 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
738                                       struct cfq_queue *cfqq)
739 {
740         /*
741          * just an approximation, should be ok.
742          */
743         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
744                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
745 }
746
747 static inline s64
748 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
749 {
750         return cfqg->vdisktime - st->min_vdisktime;
751 }
752
753 static void
754 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
755 {
756         struct rb_node **node = &st->rb.rb_node;
757         struct rb_node *parent = NULL;
758         struct cfq_group *__cfqg;
759         s64 key = cfqg_key(st, cfqg);
760         int left = 1;
761
762         while (*node != NULL) {
763                 parent = *node;
764                 __cfqg = rb_entry_cfqg(parent);
765
766                 if (key < cfqg_key(st, __cfqg))
767                         node = &parent->rb_left;
768                 else {
769                         node = &parent->rb_right;
770                         left = 0;
771                 }
772         }
773
774         if (left)
775                 st->left = &cfqg->rb_node;
776
777         rb_link_node(&cfqg->rb_node, parent, node);
778         rb_insert_color(&cfqg->rb_node, &st->rb);
779 }
780
781 static void
782 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
783 {
784         struct cfq_rb_root *st = &cfqd->grp_service_tree;
785         struct cfq_group *__cfqg;
786         struct rb_node *n;
787
788         cfqg->nr_cfqq++;
789         if (cfqg->on_st)
790                 return;
791
792         /*
793          * Currently put the group at the end. Later implement something
794          * so that groups get lesser vtime based on their weights, so that
795          * if group does not loose all if it was not continously backlogged.
796          */
797         n = rb_last(&st->rb);
798         if (n) {
799                 __cfqg = rb_entry_cfqg(n);
800                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
801         } else
802                 cfqg->vdisktime = st->min_vdisktime;
803
804         __cfq_group_service_tree_add(st, cfqg);
805         cfqg->on_st = true;
806         cfqd->nr_groups++;
807         st->total_weight += cfqg->weight;
808 }
809
810 static void
811 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
812 {
813         struct cfq_rb_root *st = &cfqd->grp_service_tree;
814
815         if (st->active == &cfqg->rb_node)
816                 st->active = NULL;
817
818         BUG_ON(cfqg->nr_cfqq < 1);
819         cfqg->nr_cfqq--;
820
821         /* If there are other cfq queues under this group, don't delete it */
822         if (cfqg->nr_cfqq)
823                 return;
824
825         cfqg->on_st = false;
826         cfqd->nr_groups--;
827         st->total_weight -= cfqg->weight;
828         if (!RB_EMPTY_NODE(&cfqg->rb_node))
829                 cfq_rb_erase(&cfqg->rb_node, st);
830         cfqg->saved_workload_slice = 0;
831 }
832
833 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
834 {
835         unsigned int slice_used, allocated_slice;
836
837         /*
838          * Queue got expired before even a single request completed or
839          * got expired immediately after first request completion.
840          */
841         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
842                 /*
843                  * Also charge the seek time incurred to the group, otherwise
844                  * if there are mutiple queues in the group, each can dispatch
845                  * a single request on seeky media and cause lots of seek time
846                  * and group will never know it.
847                  */
848                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
849                                         1);
850         } else {
851                 slice_used = jiffies - cfqq->slice_start;
852                 allocated_slice = cfqq->slice_end - cfqq->slice_start;
853                 if (slice_used > allocated_slice)
854                         slice_used = allocated_slice;
855         }
856
857         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
858         return slice_used;
859 }
860
861 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
862                                 struct cfq_queue *cfqq)
863 {
864         struct cfq_rb_root *st = &cfqd->grp_service_tree;
865         unsigned int used_sl;
866
867         used_sl = cfq_cfqq_slice_usage(cfqq);
868
869         /* Can't update vdisktime while group is on service tree */
870         cfq_rb_erase(&cfqg->rb_node, st);
871         cfqg->vdisktime += cfq_scale_slice(used_sl, cfqg);
872         __cfq_group_service_tree_add(st, cfqg);
873
874         /* This group is being expired. Save the context */
875         if (time_after(cfqd->workload_expires, jiffies)) {
876                 cfqg->saved_workload_slice = cfqd->workload_expires
877                                                 - jiffies;
878                 cfqg->saved_workload = cfqd->serving_type;
879                 cfqg->saved_serving_prio = cfqd->serving_prio;
880         } else
881                 cfqg->saved_workload_slice = 0;
882 }
883
884 /*
885  * The cfqd->service_trees holds all pending cfq_queue's that have
886  * requests waiting to be processed. It is sorted in the order that
887  * we will service the queues.
888  */
889 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
890                                  bool add_front)
891 {
892         struct rb_node **p, *parent;
893         struct cfq_queue *__cfqq;
894         unsigned long rb_key;
895         struct cfq_rb_root *service_tree;
896         int left;
897         int new_cfqq = 1;
898
899         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
900                                                 cfqq_type(cfqq), cfqd);
901         if (cfq_class_idle(cfqq)) {
902                 rb_key = CFQ_IDLE_DELAY;
903                 parent = rb_last(&service_tree->rb);
904                 if (parent && parent != &cfqq->rb_node) {
905                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
906                         rb_key += __cfqq->rb_key;
907                 } else
908                         rb_key += jiffies;
909         } else if (!add_front) {
910                 /*
911                  * Get our rb key offset. Subtract any residual slice
912                  * value carried from last service. A negative resid
913                  * count indicates slice overrun, and this should position
914                  * the next service time further away in the tree.
915                  */
916                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
917                 rb_key -= cfqq->slice_resid;
918                 cfqq->slice_resid = 0;
919         } else {
920                 rb_key = -HZ;
921                 __cfqq = cfq_rb_first(service_tree);
922                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
923         }
924
925         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
926                 new_cfqq = 0;
927                 /*
928                  * same position, nothing more to do
929                  */
930                 if (rb_key == cfqq->rb_key &&
931                     cfqq->service_tree == service_tree)
932                         return;
933
934                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
935                 cfqq->service_tree = NULL;
936         }
937
938         left = 1;
939         parent = NULL;
940         cfqq->service_tree = service_tree;
941         p = &service_tree->rb.rb_node;
942         while (*p) {
943                 struct rb_node **n;
944
945                 parent = *p;
946                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
947
948                 /*
949                  * sort by key, that represents service time.
950                  */
951                 if (time_before(rb_key, __cfqq->rb_key))
952                         n = &(*p)->rb_left;
953                 else {
954                         n = &(*p)->rb_right;
955                         left = 0;
956                 }
957
958                 p = n;
959         }
960
961         if (left)
962                 service_tree->left = &cfqq->rb_node;
963
964         cfqq->rb_key = rb_key;
965         rb_link_node(&cfqq->rb_node, parent, p);
966         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
967         service_tree->count++;
968         if (add_front || !new_cfqq)
969                 return;
970         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
971 }
972
973 static struct cfq_queue *
974 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
975                      sector_t sector, struct rb_node **ret_parent,
976                      struct rb_node ***rb_link)
977 {
978         struct rb_node **p, *parent;
979         struct cfq_queue *cfqq = NULL;
980
981         parent = NULL;
982         p = &root->rb_node;
983         while (*p) {
984                 struct rb_node **n;
985
986                 parent = *p;
987                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
988
989                 /*
990                  * Sort strictly based on sector.  Smallest to the left,
991                  * largest to the right.
992                  */
993                 if (sector > blk_rq_pos(cfqq->next_rq))
994                         n = &(*p)->rb_right;
995                 else if (sector < blk_rq_pos(cfqq->next_rq))
996                         n = &(*p)->rb_left;
997                 else
998                         break;
999                 p = n;
1000                 cfqq = NULL;
1001         }
1002
1003         *ret_parent = parent;
1004         if (rb_link)
1005                 *rb_link = p;
1006         return cfqq;
1007 }
1008
1009 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1010 {
1011         struct rb_node **p, *parent;
1012         struct cfq_queue *__cfqq;
1013
1014         if (cfqq->p_root) {
1015                 rb_erase(&cfqq->p_node, cfqq->p_root);
1016                 cfqq->p_root = NULL;
1017         }
1018
1019         if (cfq_class_idle(cfqq))
1020                 return;
1021         if (!cfqq->next_rq)
1022                 return;
1023
1024         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1025         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1026                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1027         if (!__cfqq) {
1028                 rb_link_node(&cfqq->p_node, parent, p);
1029                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1030         } else
1031                 cfqq->p_root = NULL;
1032 }
1033
1034 /*
1035  * Update cfqq's position in the service tree.
1036  */
1037 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1038 {
1039         /*
1040          * Resorting requires the cfqq to be on the RR list already.
1041          */
1042         if (cfq_cfqq_on_rr(cfqq)) {
1043                 cfq_service_tree_add(cfqd, cfqq, 0);
1044                 cfq_prio_tree_add(cfqd, cfqq);
1045         }
1046 }
1047
1048 /*
1049  * add to busy list of queues for service, trying to be fair in ordering
1050  * the pending list according to last request service
1051  */
1052 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1053 {
1054         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1055         BUG_ON(cfq_cfqq_on_rr(cfqq));
1056         cfq_mark_cfqq_on_rr(cfqq);
1057         cfqd->busy_queues++;
1058
1059         cfq_resort_rr_list(cfqd, cfqq);
1060 }
1061
1062 /*
1063  * Called when the cfqq no longer has requests pending, remove it from
1064  * the service tree.
1065  */
1066 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1067 {
1068         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1069         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1070         cfq_clear_cfqq_on_rr(cfqq);
1071
1072         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1073                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1074                 cfqq->service_tree = NULL;
1075         }
1076         if (cfqq->p_root) {
1077                 rb_erase(&cfqq->p_node, cfqq->p_root);
1078                 cfqq->p_root = NULL;
1079         }
1080
1081         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1082         BUG_ON(!cfqd->busy_queues);
1083         cfqd->busy_queues--;
1084 }
1085
1086 /*
1087  * rb tree support functions
1088  */
1089 static void cfq_del_rq_rb(struct request *rq)
1090 {
1091         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1092         const int sync = rq_is_sync(rq);
1093
1094         BUG_ON(!cfqq->queued[sync]);
1095         cfqq->queued[sync]--;
1096
1097         elv_rb_del(&cfqq->sort_list, rq);
1098
1099         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1100                 /*
1101                  * Queue will be deleted from service tree when we actually
1102                  * expire it later. Right now just remove it from prio tree
1103                  * as it is empty.
1104                  */
1105                 if (cfqq->p_root) {
1106                         rb_erase(&cfqq->p_node, cfqq->p_root);
1107                         cfqq->p_root = NULL;
1108                 }
1109         }
1110 }
1111
1112 static void cfq_add_rq_rb(struct request *rq)
1113 {
1114         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1115         struct cfq_data *cfqd = cfqq->cfqd;
1116         struct request *__alias, *prev;
1117
1118         cfqq->queued[rq_is_sync(rq)]++;
1119
1120         /*
1121          * looks a little odd, but the first insert might return an alias.
1122          * if that happens, put the alias on the dispatch list
1123          */
1124         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1125                 cfq_dispatch_insert(cfqd->queue, __alias);
1126
1127         if (!cfq_cfqq_on_rr(cfqq))
1128                 cfq_add_cfqq_rr(cfqd, cfqq);
1129
1130         /*
1131          * check if this request is a better next-serve candidate
1132          */
1133         prev = cfqq->next_rq;
1134         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1135
1136         /*
1137          * adjust priority tree position, if ->next_rq changes
1138          */
1139         if (prev != cfqq->next_rq)
1140                 cfq_prio_tree_add(cfqd, cfqq);
1141
1142         BUG_ON(!cfqq->next_rq);
1143 }
1144
1145 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1146 {
1147         elv_rb_del(&cfqq->sort_list, rq);
1148         cfqq->queued[rq_is_sync(rq)]--;
1149         cfq_add_rq_rb(rq);
1150 }
1151
1152 static struct request *
1153 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1154 {
1155         struct task_struct *tsk = current;
1156         struct cfq_io_context *cic;
1157         struct cfq_queue *cfqq;
1158
1159         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1160         if (!cic)
1161                 return NULL;
1162
1163         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1164         if (cfqq) {
1165                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1166
1167                 return elv_rb_find(&cfqq->sort_list, sector);
1168         }
1169
1170         return NULL;
1171 }
1172
1173 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1174 {
1175         struct cfq_data *cfqd = q->elevator->elevator_data;
1176
1177         cfqd->rq_in_driver[rq_is_sync(rq)]++;
1178         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1179                                                 rq_in_driver(cfqd));
1180
1181         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1182 }
1183
1184 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1185 {
1186         struct cfq_data *cfqd = q->elevator->elevator_data;
1187         const int sync = rq_is_sync(rq);
1188
1189         WARN_ON(!cfqd->rq_in_driver[sync]);
1190         cfqd->rq_in_driver[sync]--;
1191         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1192                                                 rq_in_driver(cfqd));
1193 }
1194
1195 static void cfq_remove_request(struct request *rq)
1196 {
1197         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1198
1199         if (cfqq->next_rq == rq)
1200                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1201
1202         list_del_init(&rq->queuelist);
1203         cfq_del_rq_rb(rq);
1204
1205         cfqq->cfqd->rq_queued--;
1206         if (rq_is_meta(rq)) {
1207                 WARN_ON(!cfqq->meta_pending);
1208                 cfqq->meta_pending--;
1209         }
1210 }
1211
1212 static int cfq_merge(struct request_queue *q, struct request **req,
1213                      struct bio *bio)
1214 {
1215         struct cfq_data *cfqd = q->elevator->elevator_data;
1216         struct request *__rq;
1217
1218         __rq = cfq_find_rq_fmerge(cfqd, bio);
1219         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1220                 *req = __rq;
1221                 return ELEVATOR_FRONT_MERGE;
1222         }
1223
1224         return ELEVATOR_NO_MERGE;
1225 }
1226
1227 static void cfq_merged_request(struct request_queue *q, struct request *req,
1228                                int type)
1229 {
1230         if (type == ELEVATOR_FRONT_MERGE) {
1231                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1232
1233                 cfq_reposition_rq_rb(cfqq, req);
1234         }
1235 }
1236
1237 static void
1238 cfq_merged_requests(struct request_queue *q, struct request *rq,
1239                     struct request *next)
1240 {
1241         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1242         /*
1243          * reposition in fifo if next is older than rq
1244          */
1245         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1246             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1247                 list_move(&rq->queuelist, &next->queuelist);
1248                 rq_set_fifo_time(rq, rq_fifo_time(next));
1249         }
1250
1251         if (cfqq->next_rq == next)
1252                 cfqq->next_rq = rq;
1253         cfq_remove_request(next);
1254 }
1255
1256 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1257                            struct bio *bio)
1258 {
1259         struct cfq_data *cfqd = q->elevator->elevator_data;
1260         struct cfq_io_context *cic;
1261         struct cfq_queue *cfqq;
1262
1263         /*
1264          * Disallow merge of a sync bio into an async request.
1265          */
1266         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1267                 return false;
1268
1269         /*
1270          * Lookup the cfqq that this bio will be queued with. Allow
1271          * merge only if rq is queued there.
1272          */
1273         cic = cfq_cic_lookup(cfqd, current->io_context);
1274         if (!cic)
1275                 return false;
1276
1277         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1278         return cfqq == RQ_CFQQ(rq);
1279 }
1280
1281 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1282                                    struct cfq_queue *cfqq)
1283 {
1284         if (cfqq) {
1285                 cfq_log_cfqq(cfqd, cfqq, "set_active");
1286                 cfqq->slice_start = 0;
1287                 cfqq->dispatch_start = jiffies;
1288                 cfqq->slice_end = 0;
1289                 cfqq->slice_dispatch = 0;
1290
1291                 cfq_clear_cfqq_wait_request(cfqq);
1292                 cfq_clear_cfqq_must_dispatch(cfqq);
1293                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1294                 cfq_clear_cfqq_fifo_expire(cfqq);
1295                 cfq_mark_cfqq_slice_new(cfqq);
1296
1297                 del_timer(&cfqd->idle_slice_timer);
1298         }
1299
1300         cfqd->active_queue = cfqq;
1301 }
1302
1303 /*
1304  * current cfqq expired its slice (or was too idle), select new one
1305  */
1306 static void
1307 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1308                     bool timed_out)
1309 {
1310         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1311
1312         if (cfq_cfqq_wait_request(cfqq))
1313                 del_timer(&cfqd->idle_slice_timer);
1314
1315         cfq_clear_cfqq_wait_request(cfqq);
1316
1317         /*
1318          * store what was left of this slice, if the queue idled/timed out
1319          */
1320         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1321                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1322                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1323         }
1324
1325         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1326
1327         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1328                 cfq_del_cfqq_rr(cfqd, cfqq);
1329
1330         cfq_resort_rr_list(cfqd, cfqq);
1331
1332         if (cfqq == cfqd->active_queue)
1333                 cfqd->active_queue = NULL;
1334
1335         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1336                 cfqd->grp_service_tree.active = NULL;
1337
1338         if (cfqd->active_cic) {
1339                 put_io_context(cfqd->active_cic->ioc);
1340                 cfqd->active_cic = NULL;
1341         }
1342 }
1343
1344 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1345 {
1346         struct cfq_queue *cfqq = cfqd->active_queue;
1347
1348         if (cfqq)
1349                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1350 }
1351
1352 /*
1353  * Get next queue for service. Unless we have a queue preemption,
1354  * we'll simply select the first cfqq in the service tree.
1355  */
1356 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1357 {
1358         struct cfq_rb_root *service_tree =
1359                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1360                                         cfqd->serving_type, cfqd);
1361
1362         if (!cfqd->rq_queued)
1363                 return NULL;
1364
1365         /* There is nothing to dispatch */
1366         if (!service_tree)
1367                 return NULL;
1368         if (RB_EMPTY_ROOT(&service_tree->rb))
1369                 return NULL;
1370         return cfq_rb_first(service_tree);
1371 }
1372
1373 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1374 {
1375         struct cfq_group *cfqg = &cfqd->root_group;
1376         struct cfq_queue *cfqq;
1377         int i, j;
1378         struct cfq_rb_root *st;
1379
1380         if (!cfqd->rq_queued)
1381                 return NULL;
1382
1383         for_each_cfqg_st(cfqg, i, j, st)
1384                 if ((cfqq = cfq_rb_first(st)) != NULL)
1385                         return cfqq;
1386         return NULL;
1387 }
1388
1389 /*
1390  * Get and set a new active queue for service.
1391  */
1392 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1393                                               struct cfq_queue *cfqq)
1394 {
1395         if (!cfqq)
1396                 cfqq = cfq_get_next_queue(cfqd);
1397
1398         __cfq_set_active_queue(cfqd, cfqq);
1399         return cfqq;
1400 }
1401
1402 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1403                                           struct request *rq)
1404 {
1405         if (blk_rq_pos(rq) >= cfqd->last_position)
1406                 return blk_rq_pos(rq) - cfqd->last_position;
1407         else
1408                 return cfqd->last_position - blk_rq_pos(rq);
1409 }
1410
1411 #define CFQQ_SEEK_THR           8 * 1024
1412 #define CFQQ_SEEKY(cfqq)        ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1413
1414 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1415                                struct request *rq)
1416 {
1417         sector_t sdist = cfqq->seek_mean;
1418
1419         if (!sample_valid(cfqq->seek_samples))
1420                 sdist = CFQQ_SEEK_THR;
1421
1422         return cfq_dist_from_last(cfqd, rq) <= sdist;
1423 }
1424
1425 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1426                                     struct cfq_queue *cur_cfqq)
1427 {
1428         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1429         struct rb_node *parent, *node;
1430         struct cfq_queue *__cfqq;
1431         sector_t sector = cfqd->last_position;
1432
1433         if (RB_EMPTY_ROOT(root))
1434                 return NULL;
1435
1436         /*
1437          * First, if we find a request starting at the end of the last
1438          * request, choose it.
1439          */
1440         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1441         if (__cfqq)
1442                 return __cfqq;
1443
1444         /*
1445          * If the exact sector wasn't found, the parent of the NULL leaf
1446          * will contain the closest sector.
1447          */
1448         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1449         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1450                 return __cfqq;
1451
1452         if (blk_rq_pos(__cfqq->next_rq) < sector)
1453                 node = rb_next(&__cfqq->p_node);
1454         else
1455                 node = rb_prev(&__cfqq->p_node);
1456         if (!node)
1457                 return NULL;
1458
1459         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1460         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1461                 return __cfqq;
1462
1463         return NULL;
1464 }
1465
1466 /*
1467  * cfqd - obvious
1468  * cur_cfqq - passed in so that we don't decide that the current queue is
1469  *            closely cooperating with itself.
1470  *
1471  * So, basically we're assuming that that cur_cfqq has dispatched at least
1472  * one request, and that cfqd->last_position reflects a position on the disk
1473  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1474  * assumption.
1475  */
1476 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1477                                               struct cfq_queue *cur_cfqq)
1478 {
1479         struct cfq_queue *cfqq;
1480
1481         if (!cfq_cfqq_sync(cur_cfqq))
1482                 return NULL;
1483         if (CFQQ_SEEKY(cur_cfqq))
1484                 return NULL;
1485
1486         /*
1487          * We should notice if some of the queues are cooperating, eg
1488          * working closely on the same area of the disk. In that case,
1489          * we can group them together and don't waste time idling.
1490          */
1491         cfqq = cfqq_close(cfqd, cur_cfqq);
1492         if (!cfqq)
1493                 return NULL;
1494
1495         /*
1496          * It only makes sense to merge sync queues.
1497          */
1498         if (!cfq_cfqq_sync(cfqq))
1499                 return NULL;
1500         if (CFQQ_SEEKY(cfqq))
1501                 return NULL;
1502
1503         /*
1504          * Do not merge queues of different priority classes
1505          */
1506         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1507                 return NULL;
1508
1509         return cfqq;
1510 }
1511
1512 /*
1513  * Determine whether we should enforce idle window for this queue.
1514  */
1515
1516 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1517 {
1518         enum wl_prio_t prio = cfqq_prio(cfqq);
1519         struct cfq_rb_root *service_tree = cfqq->service_tree;
1520
1521         BUG_ON(!service_tree);
1522         BUG_ON(!service_tree->count);
1523
1524         /* We never do for idle class queues. */
1525         if (prio == IDLE_WORKLOAD)
1526                 return false;
1527
1528         /* We do for queues that were marked with idle window flag. */
1529         if (cfq_cfqq_idle_window(cfqq))
1530                 return true;
1531
1532         /*
1533          * Otherwise, we do only if they are the last ones
1534          * in their service tree.
1535          */
1536         return service_tree->count == 1;
1537 }
1538
1539 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1540 {
1541         struct cfq_queue *cfqq = cfqd->active_queue;
1542         struct cfq_io_context *cic;
1543         unsigned long sl;
1544
1545         /*
1546          * SSD device without seek penalty, disable idling. But only do so
1547          * for devices that support queuing, otherwise we still have a problem
1548          * with sync vs async workloads.
1549          */
1550         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1551                 return;
1552
1553         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1554         WARN_ON(cfq_cfqq_slice_new(cfqq));
1555
1556         /*
1557          * idle is disabled, either manually or by past process history
1558          */
1559         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1560                 return;
1561
1562         /*
1563          * still active requests from this queue, don't idle
1564          */
1565         if (cfqq->dispatched)
1566                 return;
1567
1568         /*
1569          * task has exited, don't wait
1570          */
1571         cic = cfqd->active_cic;
1572         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1573                 return;
1574
1575         /*
1576          * If our average think time is larger than the remaining time
1577          * slice, then don't idle. This avoids overrunning the allotted
1578          * time slice.
1579          */
1580         if (sample_valid(cic->ttime_samples) &&
1581             (cfqq->slice_end - jiffies < cic->ttime_mean))
1582                 return;
1583
1584         cfq_mark_cfqq_wait_request(cfqq);
1585
1586         sl = cfqd->cfq_slice_idle;
1587
1588         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1589         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1590 }
1591
1592 /*
1593  * Move request from internal lists to the request queue dispatch list.
1594  */
1595 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1596 {
1597         struct cfq_data *cfqd = q->elevator->elevator_data;
1598         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1599
1600         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1601
1602         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1603         cfq_remove_request(rq);
1604         cfqq->dispatched++;
1605         elv_dispatch_sort(q, rq);
1606
1607         if (cfq_cfqq_sync(cfqq))
1608                 cfqd->sync_flight++;
1609 }
1610
1611 /*
1612  * return expired entry, or NULL to just start from scratch in rbtree
1613  */
1614 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1615 {
1616         struct request *rq = NULL;
1617
1618         if (cfq_cfqq_fifo_expire(cfqq))
1619                 return NULL;
1620
1621         cfq_mark_cfqq_fifo_expire(cfqq);
1622
1623         if (list_empty(&cfqq->fifo))
1624                 return NULL;
1625
1626         rq = rq_entry_fifo(cfqq->fifo.next);
1627         if (time_before(jiffies, rq_fifo_time(rq)))
1628                 rq = NULL;
1629
1630         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1631         return rq;
1632 }
1633
1634 static inline int
1635 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1636 {
1637         const int base_rq = cfqd->cfq_slice_async_rq;
1638
1639         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1640
1641         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1642 }
1643
1644 /*
1645  * Must be called with the queue_lock held.
1646  */
1647 static int cfqq_process_refs(struct cfq_queue *cfqq)
1648 {
1649         int process_refs, io_refs;
1650
1651         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1652         process_refs = atomic_read(&cfqq->ref) - io_refs;
1653         BUG_ON(process_refs < 0);
1654         return process_refs;
1655 }
1656
1657 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1658 {
1659         int process_refs, new_process_refs;
1660         struct cfq_queue *__cfqq;
1661
1662         /* Avoid a circular list and skip interim queue merges */
1663         while ((__cfqq = new_cfqq->new_cfqq)) {
1664                 if (__cfqq == cfqq)
1665                         return;
1666                 new_cfqq = __cfqq;
1667         }
1668
1669         process_refs = cfqq_process_refs(cfqq);
1670         /*
1671          * If the process for the cfqq has gone away, there is no
1672          * sense in merging the queues.
1673          */
1674         if (process_refs == 0)
1675                 return;
1676
1677         /*
1678          * Merge in the direction of the lesser amount of work.
1679          */
1680         new_process_refs = cfqq_process_refs(new_cfqq);
1681         if (new_process_refs >= process_refs) {
1682                 cfqq->new_cfqq = new_cfqq;
1683                 atomic_add(process_refs, &new_cfqq->ref);
1684         } else {
1685                 new_cfqq->new_cfqq = cfqq;
1686                 atomic_add(new_process_refs, &cfqq->ref);
1687         }
1688 }
1689
1690 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1691                                 struct cfq_group *cfqg, enum wl_prio_t prio,
1692                                 bool prio_changed)
1693 {
1694         struct cfq_queue *queue;
1695         int i;
1696         bool key_valid = false;
1697         unsigned long lowest_key = 0;
1698         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1699
1700         if (prio_changed) {
1701                 /*
1702                  * When priorities switched, we prefer starting
1703                  * from SYNC_NOIDLE (first choice), or just SYNC
1704                  * over ASYNC
1705                  */
1706                 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1707                         return cur_best;
1708                 cur_best = SYNC_WORKLOAD;
1709                 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1710                         return cur_best;
1711
1712                 return ASYNC_WORKLOAD;
1713         }
1714
1715         for (i = 0; i < 3; ++i) {
1716                 /* otherwise, select the one with lowest rb_key */
1717                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1718                 if (queue &&
1719                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1720                         lowest_key = queue->rb_key;
1721                         cur_best = i;
1722                         key_valid = true;
1723                 }
1724         }
1725
1726         return cur_best;
1727 }
1728
1729 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1730 {
1731         enum wl_prio_t previous_prio = cfqd->serving_prio;
1732         bool prio_changed;
1733         unsigned slice;
1734         unsigned count;
1735         struct cfq_rb_root *st;
1736         unsigned group_slice;
1737
1738         if (!cfqg) {
1739                 cfqd->serving_prio = IDLE_WORKLOAD;
1740                 cfqd->workload_expires = jiffies + 1;
1741                 return;
1742         }
1743
1744         /* Choose next priority. RT > BE > IDLE */
1745         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1746                 cfqd->serving_prio = RT_WORKLOAD;
1747         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1748                 cfqd->serving_prio = BE_WORKLOAD;
1749         else {
1750                 cfqd->serving_prio = IDLE_WORKLOAD;
1751                 cfqd->workload_expires = jiffies + 1;
1752                 return;
1753         }
1754
1755         /*
1756          * For RT and BE, we have to choose also the type
1757          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1758          * expiration time
1759          */
1760         prio_changed = (cfqd->serving_prio != previous_prio);
1761         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1762                                 cfqd);
1763         count = st->count;
1764
1765         /*
1766          * If priority didn't change, check workload expiration,
1767          * and that we still have other queues ready
1768          */
1769         if (!prio_changed && count &&
1770             !time_after(jiffies, cfqd->workload_expires))
1771                 return;
1772
1773         /* otherwise select new workload type */
1774         cfqd->serving_type =
1775                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
1776         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1777                                 cfqd);
1778         count = st->count;
1779
1780         /*
1781          * the workload slice is computed as a fraction of target latency
1782          * proportional to the number of queues in that workload, over
1783          * all the queues in the same priority class
1784          */
1785         group_slice = cfq_group_slice(cfqd, cfqg);
1786
1787         slice = group_slice * count /
1788                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
1789                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
1790
1791         if (cfqd->serving_type == ASYNC_WORKLOAD)
1792                 /* async workload slice is scaled down according to
1793                  * the sync/async slice ratio. */
1794                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1795         else
1796                 /* sync workload slice is at least 2 * cfq_slice_idle */
1797                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1798
1799         slice = max_t(unsigned, slice, CFQ_MIN_TT);
1800         cfqd->workload_expires = jiffies + slice;
1801         cfqd->noidle_tree_requires_idle = false;
1802 }
1803
1804 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
1805 {
1806         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1807         struct cfq_group *cfqg;
1808
1809         if (RB_EMPTY_ROOT(&st->rb))
1810                 return NULL;
1811         cfqg = cfq_rb_first_group(st);
1812         st->active = &cfqg->rb_node;
1813         update_min_vdisktime(st);
1814         return cfqg;
1815 }
1816
1817 static void cfq_choose_cfqg(struct cfq_data *cfqd)
1818 {
1819         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
1820
1821         cfqd->serving_group = cfqg;
1822
1823         /* Restore the workload type data */
1824         if (cfqg->saved_workload_slice) {
1825                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
1826                 cfqd->serving_type = cfqg->saved_workload;
1827                 cfqd->serving_prio = cfqg->saved_serving_prio;
1828         }
1829         choose_service_tree(cfqd, cfqg);
1830 }
1831
1832 /*
1833  * Select a queue for service. If we have a current active queue,
1834  * check whether to continue servicing it, or retrieve and set a new one.
1835  */
1836 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1837 {
1838         struct cfq_queue *cfqq, *new_cfqq = NULL;
1839
1840         cfqq = cfqd->active_queue;
1841         if (!cfqq)
1842                 goto new_queue;
1843
1844         if (!cfqd->rq_queued)
1845                 return NULL;
1846         /*
1847          * The active queue has run out of time, expire it and select new.
1848          */
1849         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1850                 goto expire;
1851
1852         /*
1853          * The active queue has requests and isn't expired, allow it to
1854          * dispatch.
1855          */
1856         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1857                 goto keep_queue;
1858
1859         /*
1860          * If another queue has a request waiting within our mean seek
1861          * distance, let it run.  The expire code will check for close
1862          * cooperators and put the close queue at the front of the service
1863          * tree.  If possible, merge the expiring queue with the new cfqq.
1864          */
1865         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1866         if (new_cfqq) {
1867                 if (!cfqq->new_cfqq)
1868                         cfq_setup_merge(cfqq, new_cfqq);
1869                 goto expire;
1870         }
1871
1872         /*
1873          * No requests pending. If the active queue still has requests in
1874          * flight or is idling for a new request, allow either of these
1875          * conditions to happen (or time out) before selecting a new queue.
1876          */
1877         if (timer_pending(&cfqd->idle_slice_timer) ||
1878             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1879                 cfqq = NULL;
1880                 goto keep_queue;
1881         }
1882
1883 expire:
1884         cfq_slice_expired(cfqd, 0);
1885 new_queue:
1886         /*
1887          * Current queue expired. Check if we have to switch to a new
1888          * service tree
1889          */
1890         if (!new_cfqq)
1891                 cfq_choose_cfqg(cfqd);
1892
1893         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1894 keep_queue:
1895         return cfqq;
1896 }
1897
1898 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1899 {
1900         int dispatched = 0;
1901
1902         while (cfqq->next_rq) {
1903                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1904                 dispatched++;
1905         }
1906
1907         BUG_ON(!list_empty(&cfqq->fifo));
1908
1909         /* By default cfqq is not expired if it is empty. Do it explicitly */
1910         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
1911         return dispatched;
1912 }
1913
1914 /*
1915  * Drain our current requests. Used for barriers and when switching
1916  * io schedulers on-the-fly.
1917  */
1918 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1919 {
1920         struct cfq_queue *cfqq;
1921         int dispatched = 0;
1922
1923         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
1924                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1925
1926         cfq_slice_expired(cfqd, 0);
1927         BUG_ON(cfqd->busy_queues);
1928
1929         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1930         return dispatched;
1931 }
1932
1933 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1934 {
1935         unsigned int max_dispatch;
1936
1937         /*
1938          * Drain async requests before we start sync IO
1939          */
1940         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1941                 return false;
1942
1943         /*
1944          * If this is an async queue and we have sync IO in flight, let it wait
1945          */
1946         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1947                 return false;
1948
1949         max_dispatch = cfqd->cfq_quantum;
1950         if (cfq_class_idle(cfqq))
1951                 max_dispatch = 1;
1952
1953         /*
1954          * Does this cfqq already have too much IO in flight?
1955          */
1956         if (cfqq->dispatched >= max_dispatch) {
1957                 /*
1958                  * idle queue must always only have a single IO in flight
1959                  */
1960                 if (cfq_class_idle(cfqq))
1961                         return false;
1962
1963                 /*
1964                  * We have other queues, don't allow more IO from this one
1965                  */
1966                 if (cfqd->busy_queues > 1)
1967                         return false;
1968
1969                 /*
1970                  * Sole queue user, no limit
1971                  */
1972                 max_dispatch = -1;
1973         }
1974
1975         /*
1976          * Async queues must wait a bit before being allowed dispatch.
1977          * We also ramp up the dispatch depth gradually for async IO,
1978          * based on the last sync IO we serviced
1979          */
1980         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1981                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1982                 unsigned int depth;
1983
1984                 depth = last_sync / cfqd->cfq_slice[1];
1985                 if (!depth && !cfqq->dispatched)
1986                         depth = 1;
1987                 if (depth < max_dispatch)
1988                         max_dispatch = depth;
1989         }
1990
1991         /*
1992          * If we're below the current max, allow a dispatch
1993          */
1994         return cfqq->dispatched < max_dispatch;
1995 }
1996
1997 /*
1998  * Dispatch a request from cfqq, moving them to the request queue
1999  * dispatch list.
2000  */
2001 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2002 {
2003         struct request *rq;
2004
2005         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2006
2007         if (!cfq_may_dispatch(cfqd, cfqq))
2008                 return false;
2009
2010         /*
2011          * follow expired path, else get first next available
2012          */
2013         rq = cfq_check_fifo(cfqq);
2014         if (!rq)
2015                 rq = cfqq->next_rq;
2016
2017         /*
2018          * insert request into driver dispatch list
2019          */
2020         cfq_dispatch_insert(cfqd->queue, rq);
2021
2022         if (!cfqd->active_cic) {
2023                 struct cfq_io_context *cic = RQ_CIC(rq);
2024
2025                 atomic_long_inc(&cic->ioc->refcount);
2026                 cfqd->active_cic = cic;
2027         }
2028
2029         return true;
2030 }
2031
2032 /*
2033  * Find the cfqq that we need to service and move a request from that to the
2034  * dispatch list
2035  */
2036 static int cfq_dispatch_requests(struct request_queue *q, int force)
2037 {
2038         struct cfq_data *cfqd = q->elevator->elevator_data;
2039         struct cfq_queue *cfqq;
2040
2041         if (!cfqd->busy_queues)
2042                 return 0;
2043
2044         if (unlikely(force))
2045                 return cfq_forced_dispatch(cfqd);
2046
2047         cfqq = cfq_select_queue(cfqd);
2048         if (!cfqq)
2049                 return 0;
2050
2051         /*
2052          * Dispatch a request from this cfqq, if it is allowed
2053          */
2054         if (!cfq_dispatch_request(cfqd, cfqq))
2055                 return 0;
2056
2057         cfqq->slice_dispatch++;
2058         cfq_clear_cfqq_must_dispatch(cfqq);
2059
2060         /*
2061          * expire an async queue immediately if it has used up its slice. idle
2062          * queue always expire after 1 dispatch round.
2063          */
2064         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2065             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2066             cfq_class_idle(cfqq))) {
2067                 cfqq->slice_end = jiffies + 1;
2068                 cfq_slice_expired(cfqd, 0);
2069         }
2070
2071         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2072         return 1;
2073 }
2074
2075 /*
2076  * task holds one reference to the queue, dropped when task exits. each rq
2077  * in-flight on this queue also holds a reference, dropped when rq is freed.
2078  *
2079  * queue lock must be held here.
2080  */
2081 static void cfq_put_queue(struct cfq_queue *cfqq)
2082 {
2083         struct cfq_data *cfqd = cfqq->cfqd;
2084
2085         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2086
2087         if (!atomic_dec_and_test(&cfqq->ref))
2088                 return;
2089
2090         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2091         BUG_ON(rb_first(&cfqq->sort_list));
2092         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2093
2094         if (unlikely(cfqd->active_queue == cfqq)) {
2095                 __cfq_slice_expired(cfqd, cfqq, 0);
2096                 cfq_schedule_dispatch(cfqd);
2097         }
2098
2099         BUG_ON(cfq_cfqq_on_rr(cfqq));
2100         kmem_cache_free(cfq_pool, cfqq);
2101 }
2102
2103 /*
2104  * Must always be called with the rcu_read_lock() held
2105  */
2106 static void
2107 __call_for_each_cic(struct io_context *ioc,
2108                     void (*func)(struct io_context *, struct cfq_io_context *))
2109 {
2110         struct cfq_io_context *cic;
2111         struct hlist_node *n;
2112
2113         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2114                 func(ioc, cic);
2115 }
2116
2117 /*
2118  * Call func for each cic attached to this ioc.
2119  */
2120 static void
2121 call_for_each_cic(struct io_context *ioc,
2122                   void (*func)(struct io_context *, struct cfq_io_context *))
2123 {
2124         rcu_read_lock();
2125         __call_for_each_cic(ioc, func);
2126         rcu_read_unlock();
2127 }
2128
2129 static void cfq_cic_free_rcu(struct rcu_head *head)
2130 {
2131         struct cfq_io_context *cic;
2132
2133         cic = container_of(head, struct cfq_io_context, rcu_head);
2134
2135         kmem_cache_free(cfq_ioc_pool, cic);
2136         elv_ioc_count_dec(cfq_ioc_count);
2137
2138         if (ioc_gone) {
2139                 /*
2140                  * CFQ scheduler is exiting, grab exit lock and check
2141                  * the pending io context count. If it hits zero,
2142                  * complete ioc_gone and set it back to NULL
2143                  */
2144                 spin_lock(&ioc_gone_lock);
2145                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2146                         complete(ioc_gone);
2147                         ioc_gone = NULL;
2148                 }
2149                 spin_unlock(&ioc_gone_lock);
2150         }
2151 }
2152
2153 static void cfq_cic_free(struct cfq_io_context *cic)
2154 {
2155         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2156 }
2157
2158 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2159 {
2160         unsigned long flags;
2161
2162         BUG_ON(!cic->dead_key);
2163
2164         spin_lock_irqsave(&ioc->lock, flags);
2165         radix_tree_delete(&ioc->radix_root, cic->dead_key);
2166         hlist_del_rcu(&cic->cic_list);
2167         spin_unlock_irqrestore(&ioc->lock, flags);
2168
2169         cfq_cic_free(cic);
2170 }
2171
2172 /*
2173  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2174  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2175  * and ->trim() which is called with the task lock held
2176  */
2177 static void cfq_free_io_context(struct io_context *ioc)
2178 {
2179         /*
2180          * ioc->refcount is zero here, or we are called from elv_unregister(),
2181          * so no more cic's are allowed to be linked into this ioc.  So it
2182          * should be ok to iterate over the known list, we will see all cic's
2183          * since no new ones are added.
2184          */
2185         __call_for_each_cic(ioc, cic_free_func);
2186 }
2187
2188 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2189 {
2190         struct cfq_queue *__cfqq, *next;
2191
2192         if (unlikely(cfqq == cfqd->active_queue)) {
2193                 __cfq_slice_expired(cfqd, cfqq, 0);
2194                 cfq_schedule_dispatch(cfqd);
2195         }
2196
2197         /*
2198          * If this queue was scheduled to merge with another queue, be
2199          * sure to drop the reference taken on that queue (and others in
2200          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2201          */
2202         __cfqq = cfqq->new_cfqq;
2203         while (__cfqq) {
2204                 if (__cfqq == cfqq) {
2205                         WARN(1, "cfqq->new_cfqq loop detected\n");
2206                         break;
2207                 }
2208                 next = __cfqq->new_cfqq;
2209                 cfq_put_queue(__cfqq);
2210                 __cfqq = next;
2211         }
2212
2213         cfq_put_queue(cfqq);
2214 }
2215
2216 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2217                                          struct cfq_io_context *cic)
2218 {
2219         struct io_context *ioc = cic->ioc;
2220
2221         list_del_init(&cic->queue_list);
2222
2223         /*
2224          * Make sure key == NULL is seen for dead queues
2225          */
2226         smp_wmb();
2227         cic->dead_key = (unsigned long) cic->key;
2228         cic->key = NULL;
2229
2230         if (ioc->ioc_data == cic)
2231                 rcu_assign_pointer(ioc->ioc_data, NULL);
2232
2233         if (cic->cfqq[BLK_RW_ASYNC]) {
2234                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2235                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2236         }
2237
2238         if (cic->cfqq[BLK_RW_SYNC]) {
2239                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2240                 cic->cfqq[BLK_RW_SYNC] = NULL;
2241         }
2242 }
2243
2244 static void cfq_exit_single_io_context(struct io_context *ioc,
2245                                        struct cfq_io_context *cic)
2246 {
2247         struct cfq_data *cfqd = cic->key;
2248
2249         if (cfqd) {
2250                 struct request_queue *q = cfqd->queue;
2251                 unsigned long flags;
2252
2253                 spin_lock_irqsave(q->queue_lock, flags);
2254
2255                 /*
2256                  * Ensure we get a fresh copy of the ->key to prevent
2257                  * race between exiting task and queue
2258                  */
2259                 smp_read_barrier_depends();
2260                 if (cic->key)
2261                         __cfq_exit_single_io_context(cfqd, cic);
2262
2263                 spin_unlock_irqrestore(q->queue_lock, flags);
2264         }
2265 }
2266
2267 /*
2268  * The process that ioc belongs to has exited, we need to clean up
2269  * and put the internal structures we have that belongs to that process.
2270  */
2271 static void cfq_exit_io_context(struct io_context *ioc)
2272 {
2273         call_for_each_cic(ioc, cfq_exit_single_io_context);
2274 }
2275
2276 static struct cfq_io_context *
2277 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2278 {
2279         struct cfq_io_context *cic;
2280
2281         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2282                                                         cfqd->queue->node);
2283         if (cic) {
2284                 cic->last_end_request = jiffies;
2285                 INIT_LIST_HEAD(&cic->queue_list);
2286                 INIT_HLIST_NODE(&cic->cic_list);
2287                 cic->dtor = cfq_free_io_context;
2288                 cic->exit = cfq_exit_io_context;
2289                 elv_ioc_count_inc(cfq_ioc_count);
2290         }
2291
2292         return cic;
2293 }
2294
2295 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2296 {
2297         struct task_struct *tsk = current;
2298         int ioprio_class;
2299
2300         if (!cfq_cfqq_prio_changed(cfqq))
2301                 return;
2302
2303         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2304         switch (ioprio_class) {
2305         default:
2306                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2307         case IOPRIO_CLASS_NONE:
2308                 /*
2309                  * no prio set, inherit CPU scheduling settings
2310                  */
2311                 cfqq->ioprio = task_nice_ioprio(tsk);
2312                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2313                 break;
2314         case IOPRIO_CLASS_RT:
2315                 cfqq->ioprio = task_ioprio(ioc);
2316                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2317                 break;
2318         case IOPRIO_CLASS_BE:
2319                 cfqq->ioprio = task_ioprio(ioc);
2320                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2321                 break;
2322         case IOPRIO_CLASS_IDLE:
2323                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2324                 cfqq->ioprio = 7;
2325                 cfq_clear_cfqq_idle_window(cfqq);
2326                 break;
2327         }
2328
2329         /*
2330          * keep track of original prio settings in case we have to temporarily
2331          * elevate the priority of this queue
2332          */
2333         cfqq->org_ioprio = cfqq->ioprio;
2334         cfqq->org_ioprio_class = cfqq->ioprio_class;
2335         cfq_clear_cfqq_prio_changed(cfqq);
2336 }
2337
2338 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2339 {
2340         struct cfq_data *cfqd = cic->key;
2341         struct cfq_queue *cfqq;
2342         unsigned long flags;
2343
2344         if (unlikely(!cfqd))
2345                 return;
2346
2347         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2348
2349         cfqq = cic->cfqq[BLK_RW_ASYNC];
2350         if (cfqq) {
2351                 struct cfq_queue *new_cfqq;
2352                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2353                                                 GFP_ATOMIC);
2354                 if (new_cfqq) {
2355                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2356                         cfq_put_queue(cfqq);
2357                 }
2358         }
2359
2360         cfqq = cic->cfqq[BLK_RW_SYNC];
2361         if (cfqq)
2362                 cfq_mark_cfqq_prio_changed(cfqq);
2363
2364         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2365 }
2366
2367 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2368 {
2369         call_for_each_cic(ioc, changed_ioprio);
2370         ioc->ioprio_changed = 0;
2371 }
2372
2373 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2374                           pid_t pid, bool is_sync)
2375 {
2376         RB_CLEAR_NODE(&cfqq->rb_node);
2377         RB_CLEAR_NODE(&cfqq->p_node);
2378         INIT_LIST_HEAD(&cfqq->fifo);
2379
2380         atomic_set(&cfqq->ref, 0);
2381         cfqq->cfqd = cfqd;
2382
2383         cfq_mark_cfqq_prio_changed(cfqq);
2384
2385         if (is_sync) {
2386                 if (!cfq_class_idle(cfqq))
2387                         cfq_mark_cfqq_idle_window(cfqq);
2388                 cfq_mark_cfqq_sync(cfqq);
2389         }
2390         cfqq->pid = pid;
2391 }
2392
2393 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
2394 {
2395         cfqq->cfqg = cfqg;
2396 }
2397
2398 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
2399 {
2400         return &cfqd->root_group;
2401 }
2402
2403 static struct cfq_queue *
2404 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2405                      struct io_context *ioc, gfp_t gfp_mask)
2406 {
2407         struct cfq_queue *cfqq, *new_cfqq = NULL;
2408         struct cfq_io_context *cic;
2409         struct cfq_group *cfqg;
2410
2411 retry:
2412         cfqg = cfq_get_cfqg(cfqd, 1);
2413         cic = cfq_cic_lookup(cfqd, ioc);
2414         /* cic always exists here */
2415         cfqq = cic_to_cfqq(cic, is_sync);
2416
2417         /*
2418          * Always try a new alloc if we fell back to the OOM cfqq
2419          * originally, since it should just be a temporary situation.
2420          */
2421         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2422                 cfqq = NULL;
2423                 if (new_cfqq) {
2424                         cfqq = new_cfqq;
2425                         new_cfqq = NULL;
2426                 } else if (gfp_mask & __GFP_WAIT) {
2427                         spin_unlock_irq(cfqd->queue->queue_lock);
2428                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2429                                         gfp_mask | __GFP_ZERO,
2430                                         cfqd->queue->node);
2431                         spin_lock_irq(cfqd->queue->queue_lock);
2432                         if (new_cfqq)
2433                                 goto retry;
2434                 } else {
2435                         cfqq = kmem_cache_alloc_node(cfq_pool,
2436                                         gfp_mask | __GFP_ZERO,
2437                                         cfqd->queue->node);
2438                 }
2439
2440                 if (cfqq) {
2441                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2442                         cfq_init_prio_data(cfqq, ioc);
2443                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2444                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2445                 } else
2446                         cfqq = &cfqd->oom_cfqq;
2447         }
2448
2449         if (new_cfqq)
2450                 kmem_cache_free(cfq_pool, new_cfqq);
2451
2452         return cfqq;
2453 }
2454
2455 static struct cfq_queue **
2456 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2457 {
2458         switch (ioprio_class) {
2459         case IOPRIO_CLASS_RT:
2460                 return &cfqd->async_cfqq[0][ioprio];
2461         case IOPRIO_CLASS_BE:
2462                 return &cfqd->async_cfqq[1][ioprio];
2463         case IOPRIO_CLASS_IDLE:
2464                 return &cfqd->async_idle_cfqq;
2465         default:
2466                 BUG();
2467         }
2468 }
2469
2470 static struct cfq_queue *
2471 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2472               gfp_t gfp_mask)
2473 {
2474         const int ioprio = task_ioprio(ioc);
2475         const int ioprio_class = task_ioprio_class(ioc);
2476         struct cfq_queue **async_cfqq = NULL;
2477         struct cfq_queue *cfqq = NULL;
2478
2479         if (!is_sync) {
2480                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2481                 cfqq = *async_cfqq;
2482         }
2483
2484         if (!cfqq)
2485                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2486
2487         /*
2488          * pin the queue now that it's allocated, scheduler exit will prune it
2489          */
2490         if (!is_sync && !(*async_cfqq)) {
2491                 atomic_inc(&cfqq->ref);
2492                 *async_cfqq = cfqq;
2493         }
2494
2495         atomic_inc(&cfqq->ref);
2496         return cfqq;
2497 }
2498
2499 /*
2500  * We drop cfq io contexts lazily, so we may find a dead one.
2501  */
2502 static void
2503 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2504                   struct cfq_io_context *cic)
2505 {
2506         unsigned long flags;
2507
2508         WARN_ON(!list_empty(&cic->queue_list));
2509
2510         spin_lock_irqsave(&ioc->lock, flags);
2511
2512         BUG_ON(ioc->ioc_data == cic);
2513
2514         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2515         hlist_del_rcu(&cic->cic_list);
2516         spin_unlock_irqrestore(&ioc->lock, flags);
2517
2518         cfq_cic_free(cic);
2519 }
2520
2521 static struct cfq_io_context *
2522 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2523 {
2524         struct cfq_io_context *cic;
2525         unsigned long flags;
2526         void *k;
2527
2528         if (unlikely(!ioc))
2529                 return NULL;
2530
2531         rcu_read_lock();
2532
2533         /*
2534          * we maintain a last-hit cache, to avoid browsing over the tree
2535          */
2536         cic = rcu_dereference(ioc->ioc_data);
2537         if (cic && cic->key == cfqd) {
2538                 rcu_read_unlock();
2539                 return cic;
2540         }
2541
2542         do {
2543                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2544                 rcu_read_unlock();
2545                 if (!cic)
2546                         break;
2547                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2548                 k = cic->key;
2549                 if (unlikely(!k)) {
2550                         cfq_drop_dead_cic(cfqd, ioc, cic);
2551                         rcu_read_lock();
2552                         continue;
2553                 }
2554
2555                 spin_lock_irqsave(&ioc->lock, flags);
2556                 rcu_assign_pointer(ioc->ioc_data, cic);
2557                 spin_unlock_irqrestore(&ioc->lock, flags);
2558                 break;
2559         } while (1);
2560
2561         return cic;
2562 }
2563
2564 /*
2565  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2566  * the process specific cfq io context when entered from the block layer.
2567  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2568  */
2569 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2570                         struct cfq_io_context *cic, gfp_t gfp_mask)
2571 {
2572         unsigned long flags;
2573         int ret;
2574
2575         ret = radix_tree_preload(gfp_mask);
2576         if (!ret) {
2577                 cic->ioc = ioc;
2578                 cic->key = cfqd;
2579
2580                 spin_lock_irqsave(&ioc->lock, flags);
2581                 ret = radix_tree_insert(&ioc->radix_root,
2582                                                 (unsigned long) cfqd, cic);
2583                 if (!ret)
2584                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2585                 spin_unlock_irqrestore(&ioc->lock, flags);
2586
2587                 radix_tree_preload_end();
2588
2589                 if (!ret) {
2590                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2591                         list_add(&cic->queue_list, &cfqd->cic_list);
2592                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2593                 }
2594         }
2595
2596         if (ret)
2597                 printk(KERN_ERR "cfq: cic link failed!\n");
2598
2599         return ret;
2600 }
2601
2602 /*
2603  * Setup general io context and cfq io context. There can be several cfq
2604  * io contexts per general io context, if this process is doing io to more
2605  * than one device managed by cfq.
2606  */
2607 static struct cfq_io_context *
2608 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2609 {
2610         struct io_context *ioc = NULL;
2611         struct cfq_io_context *cic;
2612
2613         might_sleep_if(gfp_mask & __GFP_WAIT);
2614
2615         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2616         if (!ioc)
2617                 return NULL;
2618
2619         cic = cfq_cic_lookup(cfqd, ioc);
2620         if (cic)
2621                 goto out;
2622
2623         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2624         if (cic == NULL)
2625                 goto err;
2626
2627         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2628                 goto err_free;
2629
2630 out:
2631         smp_read_barrier_depends();
2632         if (unlikely(ioc->ioprio_changed))
2633                 cfq_ioc_set_ioprio(ioc);
2634
2635         return cic;
2636 err_free:
2637         cfq_cic_free(cic);
2638 err:
2639         put_io_context(ioc);
2640         return NULL;
2641 }
2642
2643 static void
2644 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2645 {
2646         unsigned long elapsed = jiffies - cic->last_end_request;
2647         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2648
2649         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2650         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2651         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2652 }
2653
2654 static void
2655 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2656                        struct request *rq)
2657 {
2658         sector_t sdist;
2659         u64 total;
2660
2661         if (!cfqq->last_request_pos)
2662                 sdist = 0;
2663         else if (cfqq->last_request_pos < blk_rq_pos(rq))
2664                 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2665         else
2666                 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2667
2668         /*
2669          * Don't allow the seek distance to get too large from the
2670          * odd fragment, pagein, etc
2671          */
2672         if (cfqq->seek_samples <= 60) /* second&third seek */
2673                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2674         else
2675                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2676
2677         cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2678         cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2679         total = cfqq->seek_total + (cfqq->seek_samples/2);
2680         do_div(total, cfqq->seek_samples);
2681         cfqq->seek_mean = (sector_t)total;
2682
2683         /*
2684          * If this cfqq is shared between multiple processes, check to
2685          * make sure that those processes are still issuing I/Os within
2686          * the mean seek distance.  If not, it may be time to break the
2687          * queues apart again.
2688          */
2689         if (cfq_cfqq_coop(cfqq)) {
2690                 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2691                         cfqq->seeky_start = jiffies;
2692                 else if (!CFQQ_SEEKY(cfqq))
2693                         cfqq->seeky_start = 0;
2694         }
2695 }
2696
2697 /*
2698  * Disable idle window if the process thinks too long or seeks so much that
2699  * it doesn't matter
2700  */
2701 static void
2702 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2703                        struct cfq_io_context *cic)
2704 {
2705         int old_idle, enable_idle;
2706
2707         /*
2708          * Don't idle for async or idle io prio class
2709          */
2710         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2711                 return;
2712
2713         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2714
2715         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2716                 cfq_mark_cfqq_deep(cfqq);
2717
2718         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2719             (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2720              && CFQQ_SEEKY(cfqq)))
2721                 enable_idle = 0;
2722         else if (sample_valid(cic->ttime_samples)) {
2723                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2724                         enable_idle = 0;
2725                 else
2726                         enable_idle = 1;
2727         }
2728
2729         if (old_idle != enable_idle) {
2730                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2731                 if (enable_idle)
2732                         cfq_mark_cfqq_idle_window(cfqq);
2733                 else
2734                         cfq_clear_cfqq_idle_window(cfqq);
2735         }
2736 }
2737
2738 /*
2739  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2740  * no or if we aren't sure, a 1 will cause a preempt.
2741  */
2742 static bool
2743 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2744                    struct request *rq)
2745 {
2746         struct cfq_queue *cfqq;
2747
2748         cfqq = cfqd->active_queue;
2749         if (!cfqq)
2750                 return false;
2751
2752         if (cfq_slice_used(cfqq))
2753                 return true;
2754
2755         if (cfq_class_idle(new_cfqq))
2756                 return false;
2757
2758         if (cfq_class_idle(cfqq))
2759                 return true;
2760
2761         /* Allow preemption only if we are idling on sync-noidle tree */
2762         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2763             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2764             new_cfqq->service_tree->count == 2 &&
2765             RB_EMPTY_ROOT(&cfqq->sort_list))
2766                 return true;
2767
2768         /*
2769          * if the new request is sync, but the currently running queue is
2770          * not, let the sync request have priority.
2771          */
2772         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2773                 return true;
2774
2775         /*
2776          * So both queues are sync. Let the new request get disk time if
2777          * it's a metadata request and the current queue is doing regular IO.
2778          */
2779         if (rq_is_meta(rq) && !cfqq->meta_pending)
2780                 return true;
2781
2782         /*
2783          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2784          */
2785         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2786                 return true;
2787
2788         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2789                 return false;
2790
2791         /*
2792          * if this request is as-good as one we would expect from the
2793          * current cfqq, let it preempt
2794          */
2795         if (cfq_rq_close(cfqd, cfqq, rq))
2796                 return true;
2797
2798         return false;
2799 }
2800
2801 /*
2802  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2803  * let it have half of its nominal slice.
2804  */
2805 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2806 {
2807         cfq_log_cfqq(cfqd, cfqq, "preempt");
2808         cfq_slice_expired(cfqd, 1);
2809
2810         /*
2811          * Put the new queue at the front of the of the current list,
2812          * so we know that it will be selected next.
2813          */
2814         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2815
2816         cfq_service_tree_add(cfqd, cfqq, 1);
2817
2818         cfqq->slice_end = 0;
2819         cfq_mark_cfqq_slice_new(cfqq);
2820 }
2821
2822 /*
2823  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2824  * something we should do about it
2825  */
2826 static void
2827 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2828                 struct request *rq)
2829 {
2830         struct cfq_io_context *cic = RQ_CIC(rq);
2831
2832         cfqd->rq_queued++;
2833         if (rq_is_meta(rq))
2834                 cfqq->meta_pending++;
2835
2836         cfq_update_io_thinktime(cfqd, cic);
2837         cfq_update_io_seektime(cfqd, cfqq, rq);
2838         cfq_update_idle_window(cfqd, cfqq, cic);
2839
2840         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2841
2842         if (cfqq == cfqd->active_queue) {
2843                 /*
2844                  * Remember that we saw a request from this process, but
2845                  * don't start queuing just yet. Otherwise we risk seeing lots
2846                  * of tiny requests, because we disrupt the normal plugging
2847                  * and merging. If the request is already larger than a single
2848                  * page, let it rip immediately. For that case we assume that
2849                  * merging is already done. Ditto for a busy system that
2850                  * has other work pending, don't risk delaying until the
2851                  * idle timer unplug to continue working.
2852                  */
2853                 if (cfq_cfqq_wait_request(cfqq)) {
2854                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2855                             cfqd->busy_queues > 1) {
2856                                 del_timer(&cfqd->idle_slice_timer);
2857                                 __blk_run_queue(cfqd->queue);
2858                         } else
2859                                 cfq_mark_cfqq_must_dispatch(cfqq);
2860                 }
2861         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2862                 /*
2863                  * not the active queue - expire current slice if it is
2864                  * idle and has expired it's mean thinktime or this new queue
2865                  * has some old slice time left and is of higher priority or
2866                  * this new queue is RT and the current one is BE
2867                  */
2868                 cfq_preempt_queue(cfqd, cfqq);
2869                 __blk_run_queue(cfqd->queue);
2870         }
2871 }
2872
2873 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2874 {
2875         struct cfq_data *cfqd = q->elevator->elevator_data;
2876         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2877
2878         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2879         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2880
2881         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2882         list_add_tail(&rq->queuelist, &cfqq->fifo);
2883         cfq_add_rq_rb(rq);
2884
2885         cfq_rq_enqueued(cfqd, cfqq, rq);
2886 }
2887
2888 /*
2889  * Update hw_tag based on peak queue depth over 50 samples under
2890  * sufficient load.
2891  */
2892 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2893 {
2894         struct cfq_queue *cfqq = cfqd->active_queue;
2895
2896         if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2897                 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2898
2899         if (cfqd->hw_tag == 1)
2900                 return;
2901
2902         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2903             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2904                 return;
2905
2906         /*
2907          * If active queue hasn't enough requests and can idle, cfq might not
2908          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2909          * case
2910          */
2911         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2912             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2913             CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2914                 return;
2915
2916         if (cfqd->hw_tag_samples++ < 50)
2917                 return;
2918
2919         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2920                 cfqd->hw_tag = 1;
2921         else
2922                 cfqd->hw_tag = 0;
2923 }
2924
2925 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2926 {
2927         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2928         struct cfq_data *cfqd = cfqq->cfqd;
2929         const int sync = rq_is_sync(rq);
2930         unsigned long now;
2931
2932         now = jiffies;
2933         cfq_log_cfqq(cfqd, cfqq, "complete");
2934
2935         cfq_update_hw_tag(cfqd);
2936
2937         WARN_ON(!cfqd->rq_in_driver[sync]);
2938         WARN_ON(!cfqq->dispatched);
2939         cfqd->rq_in_driver[sync]--;
2940         cfqq->dispatched--;
2941
2942         if (cfq_cfqq_sync(cfqq))
2943                 cfqd->sync_flight--;
2944
2945         if (sync) {
2946                 RQ_CIC(rq)->last_end_request = now;
2947                 cfqd->last_end_sync_rq = now;
2948         }
2949
2950         /*
2951          * If this is the active queue, check if it needs to be expired,
2952          * or if we want to idle in case it has no pending requests.
2953          */
2954         if (cfqd->active_queue == cfqq) {
2955                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2956
2957                 if (cfq_cfqq_slice_new(cfqq)) {
2958                         cfq_set_prio_slice(cfqd, cfqq);
2959                         cfq_clear_cfqq_slice_new(cfqq);
2960                 }
2961                 /*
2962                  * Idling is not enabled on:
2963                  * - expired queues
2964                  * - idle-priority queues
2965                  * - async queues
2966                  * - queues with still some requests queued
2967                  * - when there is a close cooperator
2968                  */
2969                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2970                         cfq_slice_expired(cfqd, 1);
2971                 else if (sync && cfqq_empty &&
2972                          !cfq_close_cooperator(cfqd, cfqq)) {
2973                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
2974                         /*
2975                          * Idling is enabled for SYNC_WORKLOAD.
2976                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2977                          * only if we processed at least one !rq_noidle request
2978                          */
2979                         if (cfqd->serving_type == SYNC_WORKLOAD
2980                             || cfqd->noidle_tree_requires_idle)
2981                                 cfq_arm_slice_timer(cfqd);
2982                 }
2983         }
2984
2985         if (!rq_in_driver(cfqd))
2986                 cfq_schedule_dispatch(cfqd);
2987 }
2988
2989 /*
2990  * we temporarily boost lower priority queues if they are holding fs exclusive
2991  * resources. they are boosted to normal prio (CLASS_BE/4)
2992  */
2993 static void cfq_prio_boost(struct cfq_queue *cfqq)
2994 {
2995         if (has_fs_excl()) {
2996                 /*
2997                  * boost idle prio on transactions that would lock out other
2998                  * users of the filesystem
2999                  */
3000                 if (cfq_class_idle(cfqq))
3001                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3002                 if (cfqq->ioprio > IOPRIO_NORM)
3003                         cfqq->ioprio = IOPRIO_NORM;
3004         } else {
3005                 /*
3006                  * unboost the queue (if needed)
3007                  */
3008                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3009                 cfqq->ioprio = cfqq->org_ioprio;
3010         }
3011 }
3012
3013 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3014 {
3015         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3016                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3017                 return ELV_MQUEUE_MUST;
3018         }
3019
3020         return ELV_MQUEUE_MAY;
3021 }
3022
3023 static int cfq_may_queue(struct request_queue *q, int rw)
3024 {
3025         struct cfq_data *cfqd = q->elevator->elevator_data;
3026         struct task_struct *tsk = current;
3027         struct cfq_io_context *cic;
3028         struct cfq_queue *cfqq;
3029
3030         /*
3031          * don't force setup of a queue from here, as a call to may_queue
3032          * does not necessarily imply that a request actually will be queued.
3033          * so just lookup a possibly existing queue, or return 'may queue'
3034          * if that fails
3035          */
3036         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3037         if (!cic)
3038                 return ELV_MQUEUE_MAY;
3039
3040         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3041         if (cfqq) {
3042                 cfq_init_prio_data(cfqq, cic->ioc);
3043                 cfq_prio_boost(cfqq);
3044
3045                 return __cfq_may_queue(cfqq);
3046         }
3047
3048         return ELV_MQUEUE_MAY;
3049 }
3050
3051 /*
3052  * queue lock held here
3053  */
3054 static void cfq_put_request(struct request *rq)
3055 {
3056         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3057
3058         if (cfqq) {
3059                 const int rw = rq_data_dir(rq);
3060
3061                 BUG_ON(!cfqq->allocated[rw]);
3062                 cfqq->allocated[rw]--;
3063
3064                 put_io_context(RQ_CIC(rq)->ioc);
3065
3066                 rq->elevator_private = NULL;
3067                 rq->elevator_private2 = NULL;
3068
3069                 cfq_put_queue(cfqq);
3070         }
3071 }
3072
3073 static struct cfq_queue *
3074 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3075                 struct cfq_queue *cfqq)
3076 {
3077         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3078         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3079         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3080         cfq_put_queue(cfqq);
3081         return cic_to_cfqq(cic, 1);
3082 }
3083
3084 static int should_split_cfqq(struct cfq_queue *cfqq)
3085 {
3086         if (cfqq->seeky_start &&
3087             time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
3088                 return 1;
3089         return 0;
3090 }
3091
3092 /*
3093  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3094  * was the last process referring to said cfqq.
3095  */
3096 static struct cfq_queue *
3097 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3098 {
3099         if (cfqq_process_refs(cfqq) == 1) {
3100                 cfqq->seeky_start = 0;
3101                 cfqq->pid = current->pid;
3102                 cfq_clear_cfqq_coop(cfqq);
3103                 return cfqq;
3104         }
3105
3106         cic_set_cfqq(cic, NULL, 1);
3107         cfq_put_queue(cfqq);
3108         return NULL;
3109 }
3110 /*
3111  * Allocate cfq data structures associated with this request.
3112  */
3113 static int
3114 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3115 {
3116         struct cfq_data *cfqd = q->elevator->elevator_data;
3117         struct cfq_io_context *cic;
3118         const int rw = rq_data_dir(rq);
3119         const bool is_sync = rq_is_sync(rq);
3120         struct cfq_queue *cfqq;
3121         unsigned long flags;
3122
3123         might_sleep_if(gfp_mask & __GFP_WAIT);
3124
3125         cic = cfq_get_io_context(cfqd, gfp_mask);
3126
3127         spin_lock_irqsave(q->queue_lock, flags);
3128
3129         if (!cic)
3130                 goto queue_fail;
3131
3132 new_queue:
3133         cfqq = cic_to_cfqq(cic, is_sync);
3134         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3135                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3136                 cic_set_cfqq(cic, cfqq, is_sync);
3137         } else {
3138                 /*
3139                  * If the queue was seeky for too long, break it apart.
3140                  */
3141                 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
3142                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3143                         cfqq = split_cfqq(cic, cfqq);
3144                         if (!cfqq)
3145                                 goto new_queue;
3146                 }
3147
3148                 /*
3149                  * Check to see if this queue is scheduled to merge with
3150                  * another, closely cooperating queue.  The merging of
3151                  * queues happens here as it must be done in process context.
3152                  * The reference on new_cfqq was taken in merge_cfqqs.
3153                  */
3154                 if (cfqq->new_cfqq)
3155                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3156         }
3157
3158         cfqq->allocated[rw]++;
3159         atomic_inc(&cfqq->ref);
3160
3161         spin_unlock_irqrestore(q->queue_lock, flags);
3162
3163         rq->elevator_private = cic;
3164         rq->elevator_private2 = cfqq;
3165         return 0;
3166
3167 queue_fail:
3168         if (cic)
3169                 put_io_context(cic->ioc);
3170
3171         cfq_schedule_dispatch(cfqd);
3172         spin_unlock_irqrestore(q->queue_lock, flags);
3173         cfq_log(cfqd, "set_request fail");
3174         return 1;
3175 }
3176
3177 static void cfq_kick_queue(struct work_struct *work)
3178 {
3179         struct cfq_data *cfqd =
3180                 container_of(work, struct cfq_data, unplug_work);
3181         struct request_queue *q = cfqd->queue;
3182
3183         spin_lock_irq(q->queue_lock);
3184         __blk_run_queue(cfqd->queue);
3185         spin_unlock_irq(q->queue_lock);
3186 }
3187
3188 /*
3189  * Timer running if the active_queue is currently idling inside its time slice
3190  */
3191 static void cfq_idle_slice_timer(unsigned long data)
3192 {
3193         struct cfq_data *cfqd = (struct cfq_data *) data;
3194         struct cfq_queue *cfqq;
3195         unsigned long flags;
3196         int timed_out = 1;
3197
3198         cfq_log(cfqd, "idle timer fired");
3199
3200         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3201
3202         cfqq = cfqd->active_queue;
3203         if (cfqq) {
3204                 timed_out = 0;
3205
3206                 /*
3207                  * We saw a request before the queue expired, let it through
3208                  */
3209                 if (cfq_cfqq_must_dispatch(cfqq))
3210                         goto out_kick;
3211
3212                 /*
3213                  * expired
3214                  */
3215                 if (cfq_slice_used(cfqq))
3216                         goto expire;
3217
3218                 /*
3219                  * only expire and reinvoke request handler, if there are
3220                  * other queues with pending requests
3221                  */
3222                 if (!cfqd->busy_queues)
3223                         goto out_cont;
3224
3225                 /*
3226                  * not expired and it has a request pending, let it dispatch
3227                  */
3228                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3229                         goto out_kick;
3230
3231                 /*
3232                  * Queue depth flag is reset only when the idle didn't succeed
3233                  */
3234                 cfq_clear_cfqq_deep(cfqq);
3235         }
3236 expire:
3237         cfq_slice_expired(cfqd, timed_out);
3238 out_kick:
3239         cfq_schedule_dispatch(cfqd);
3240 out_cont:
3241         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3242 }
3243
3244 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3245 {
3246         del_timer_sync(&cfqd->idle_slice_timer);
3247         cancel_work_sync(&cfqd->unplug_work);
3248 }
3249
3250 static void cfq_put_async_queues(struct cfq_data *cfqd)
3251 {
3252         int i;
3253
3254         for (i = 0; i < IOPRIO_BE_NR; i++) {
3255                 if (cfqd->async_cfqq[0][i])
3256                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3257                 if (cfqd->async_cfqq[1][i])
3258                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3259         }
3260
3261         if (cfqd->async_idle_cfqq)
3262                 cfq_put_queue(cfqd->async_idle_cfqq);
3263 }
3264
3265 static void cfq_exit_queue(struct elevator_queue *e)
3266 {
3267         struct cfq_data *cfqd = e->elevator_data;
3268         struct request_queue *q = cfqd->queue;
3269
3270         cfq_shutdown_timer_wq(cfqd);
3271
3272         spin_lock_irq(q->queue_lock);
3273
3274         if (cfqd->active_queue)
3275                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3276
3277         while (!list_empty(&cfqd->cic_list)) {
3278                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3279                                                         struct cfq_io_context,
3280                                                         queue_list);
3281
3282                 __cfq_exit_single_io_context(cfqd, cic);
3283         }
3284
3285         cfq_put_async_queues(cfqd);
3286
3287         spin_unlock_irq(q->queue_lock);
3288
3289         cfq_shutdown_timer_wq(cfqd);
3290
3291         kfree(cfqd);
3292 }
3293
3294 static void *cfq_init_queue(struct request_queue *q)
3295 {
3296         struct cfq_data *cfqd;
3297         int i, j;
3298         struct cfq_group *cfqg;
3299         struct cfq_rb_root *st;
3300
3301         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3302         if (!cfqd)
3303                 return NULL;
3304
3305         /* Init root service tree */
3306         cfqd->grp_service_tree = CFQ_RB_ROOT;
3307
3308         /* Init root group */
3309         cfqg = &cfqd->root_group;
3310         for_each_cfqg_st(cfqg, i, j, st)
3311                 *st = CFQ_RB_ROOT;
3312         RB_CLEAR_NODE(&cfqg->rb_node);
3313
3314         /* Give preference to root group over other groups */
3315         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3316
3317         /*
3318          * Not strictly needed (since RB_ROOT just clears the node and we
3319          * zeroed cfqd on alloc), but better be safe in case someone decides
3320          * to add magic to the rb code
3321          */
3322         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3323                 cfqd->prio_trees[i] = RB_ROOT;
3324
3325         /*
3326          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3327          * Grab a permanent reference to it, so that the normal code flow
3328          * will not attempt to free it.
3329          */
3330         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3331         atomic_inc(&cfqd->oom_cfqq.ref);
3332         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3333
3334         INIT_LIST_HEAD(&cfqd->cic_list);
3335
3336         cfqd->queue = q;
3337
3338         init_timer(&cfqd->idle_slice_timer);
3339         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3340         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3341
3342         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3343
3344         cfqd->cfq_quantum = cfq_quantum;
3345         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3346         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3347         cfqd->cfq_back_max = cfq_back_max;
3348         cfqd->cfq_back_penalty = cfq_back_penalty;
3349         cfqd->cfq_slice[0] = cfq_slice_async;
3350         cfqd->cfq_slice[1] = cfq_slice_sync;
3351         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3352         cfqd->cfq_slice_idle = cfq_slice_idle;
3353         cfqd->cfq_latency = 1;
3354         cfqd->hw_tag = -1;
3355         cfqd->last_end_sync_rq = jiffies;
3356         return cfqd;
3357 }
3358
3359 static void cfq_slab_kill(void)
3360 {
3361         /*
3362          * Caller already ensured that pending RCU callbacks are completed,
3363          * so we should have no busy allocations at this point.
3364          */
3365         if (cfq_pool)
3366                 kmem_cache_destroy(cfq_pool);
3367         if (cfq_ioc_pool)
3368                 kmem_cache_destroy(cfq_ioc_pool);
3369 }
3370
3371 static int __init cfq_slab_setup(void)
3372 {
3373         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3374         if (!cfq_pool)
3375                 goto fail;
3376
3377         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3378         if (!cfq_ioc_pool)
3379                 goto fail;
3380
3381         return 0;
3382 fail:
3383         cfq_slab_kill();
3384         return -ENOMEM;
3385 }
3386
3387 /*
3388  * sysfs parts below -->
3389  */
3390 static ssize_t
3391 cfq_var_show(unsigned int var, char *page)
3392 {
3393         return sprintf(page, "%d\n", var);
3394 }
3395
3396 static ssize_t
3397 cfq_var_store(unsigned int *var, const char *page, size_t count)
3398 {
3399         char *p = (char *) page;
3400
3401         *var = simple_strtoul(p, &p, 10);
3402         return count;
3403 }
3404
3405 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3406 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3407 {                                                                       \
3408         struct cfq_data *cfqd = e->elevator_data;                       \
3409         unsigned int __data = __VAR;                                    \
3410         if (__CONV)                                                     \
3411                 __data = jiffies_to_msecs(__data);                      \
3412         return cfq_var_show(__data, (page));                            \
3413 }
3414 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3415 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3416 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3417 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3418 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3419 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3420 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3421 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3422 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3423 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3424 #undef SHOW_FUNCTION
3425
3426 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3427 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3428 {                                                                       \
3429         struct cfq_data *cfqd = e->elevator_data;                       \
3430         unsigned int __data;                                            \
3431         int ret = cfq_var_store(&__data, (page), count);                \
3432         if (__data < (MIN))                                             \
3433                 __data = (MIN);                                         \
3434         else if (__data > (MAX))                                        \
3435                 __data = (MAX);                                         \
3436         if (__CONV)                                                     \
3437                 *(__PTR) = msecs_to_jiffies(__data);                    \
3438         else                                                            \
3439                 *(__PTR) = __data;                                      \
3440         return ret;                                                     \
3441 }
3442 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3443 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3444                 UINT_MAX, 1);
3445 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3446                 UINT_MAX, 1);
3447 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3448 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3449                 UINT_MAX, 0);
3450 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3451 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3452 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3453 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3454                 UINT_MAX, 0);
3455 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3456 #undef STORE_FUNCTION
3457
3458 #define CFQ_ATTR(name) \
3459         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3460
3461 static struct elv_fs_entry cfq_attrs[] = {
3462         CFQ_ATTR(quantum),
3463         CFQ_ATTR(fifo_expire_sync),
3464         CFQ_ATTR(fifo_expire_async),
3465         CFQ_ATTR(back_seek_max),
3466         CFQ_ATTR(back_seek_penalty),
3467         CFQ_ATTR(slice_sync),
3468         CFQ_ATTR(slice_async),
3469         CFQ_ATTR(slice_async_rq),
3470         CFQ_ATTR(slice_idle),
3471         CFQ_ATTR(low_latency),
3472         __ATTR_NULL
3473 };
3474
3475 static struct elevator_type iosched_cfq = {
3476         .ops = {
3477                 .elevator_merge_fn =            cfq_merge,
3478                 .elevator_merged_fn =           cfq_merged_request,
3479                 .elevator_merge_req_fn =        cfq_merged_requests,
3480                 .elevator_allow_merge_fn =      cfq_allow_merge,
3481                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3482                 .elevator_add_req_fn =          cfq_insert_request,
3483                 .elevator_activate_req_fn =     cfq_activate_request,
3484                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3485                 .elevator_queue_empty_fn =      cfq_queue_empty,
3486                 .elevator_completed_req_fn =    cfq_completed_request,
3487                 .elevator_former_req_fn =       elv_rb_former_request,
3488                 .elevator_latter_req_fn =       elv_rb_latter_request,
3489                 .elevator_set_req_fn =          cfq_set_request,
3490                 .elevator_put_req_fn =          cfq_put_request,
3491                 .elevator_may_queue_fn =        cfq_may_queue,
3492                 .elevator_init_fn =             cfq_init_queue,
3493                 .elevator_exit_fn =             cfq_exit_queue,
3494                 .trim =                         cfq_free_io_context,
3495         },
3496         .elevator_attrs =       cfq_attrs,
3497         .elevator_name =        "cfq",
3498         .elevator_owner =       THIS_MODULE,
3499 };
3500
3501 static int __init cfq_init(void)
3502 {
3503         /*
3504          * could be 0 on HZ < 1000 setups
3505          */
3506         if (!cfq_slice_async)
3507                 cfq_slice_async = 1;
3508         if (!cfq_slice_idle)
3509                 cfq_slice_idle = 1;
3510
3511         if (cfq_slab_setup())
3512                 return -ENOMEM;
3513
3514         elv_register(&iosched_cfq);
3515
3516         return 0;
3517 }
3518
3519 static void __exit cfq_exit(void)
3520 {
3521         DECLARE_COMPLETION_ONSTACK(all_gone);
3522         elv_unregister(&iosched_cfq);
3523         ioc_gone = &all_gone;
3524         /* ioc_gone's update must be visible before reading ioc_count */
3525         smp_wmb();
3526
3527         /*
3528          * this also protects us from entering cfq_slab_kill() with
3529          * pending RCU callbacks
3530          */
3531         if (elv_ioc_count_read(cfq_ioc_count))
3532                 wait_for_completion(&all_gone);
3533         cfq_slab_kill();
3534 }
3535
3536 module_init(cfq_init);
3537 module_exit(cfq_exit);
3538
3539 MODULE_AUTHOR("Jens Axboe");
3540 MODULE_LICENSE("GPL");
3541 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");