cfq-iosched: log responsible 'cfqq' in idle timer arm
[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/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY          (HZ / 5)
35
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT              (2)
40
41 #define CFQ_SLICE_SCALE         (5)
42 #define CFQ_HW_QUEUE_MIN        (5)
43
44 #define RQ_CIC(rq)              \
45         ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
47
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50
51 static DEFINE_PER_CPU(unsigned long, ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
59 #define sample_valid(samples)   ((samples) > 80)
60
61 /*
62  * Most of our rbtree usage is for sorting with min extraction, so
63  * if we cache the leftmost node we don't have to walk down the tree
64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65  * move this into the elevator for the rq sorting as well.
66  */
67 struct cfq_rb_root {
68         struct rb_root rb;
69         struct rb_node *left;
70 };
71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73 /*
74  * Per block device queue structure
75  */
76 struct cfq_data {
77         struct request_queue *queue;
78
79         /*
80          * rr list of queues with requests and the count of them
81          */
82         struct cfq_rb_root service_tree;
83         unsigned int busy_queues;
84         /*
85          * Used to track any pending rt requests so we can pre-empt current
86          * non-RT cfqq in service when this value is non-zero.
87          */
88         unsigned int busy_rt_queues;
89
90         int rq_in_driver;
91         int sync_flight;
92
93         /*
94          * queue-depth detection
95          */
96         int rq_queued;
97         int hw_tag;
98         int hw_tag_samples;
99         int rq_in_driver_peak;
100
101         /*
102          * idle window management
103          */
104         struct timer_list idle_slice_timer;
105         struct work_struct unplug_work;
106
107         struct cfq_queue *active_queue;
108         struct cfq_io_context *active_cic;
109
110         /*
111          * async queue for each priority case
112          */
113         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
114         struct cfq_queue *async_idle_cfqq;
115
116         sector_t last_position;
117         unsigned long last_end_request;
118
119         /*
120          * tunables, see top of file
121          */
122         unsigned int cfq_quantum;
123         unsigned int cfq_fifo_expire[2];
124         unsigned int cfq_back_penalty;
125         unsigned int cfq_back_max;
126         unsigned int cfq_slice[2];
127         unsigned int cfq_slice_async_rq;
128         unsigned int cfq_slice_idle;
129
130         struct list_head cic_list;
131 };
132
133 /*
134  * Per process-grouping structure
135  */
136 struct cfq_queue {
137         /* reference count */
138         atomic_t ref;
139         /* various state flags, see below */
140         unsigned int flags;
141         /* parent cfq_data */
142         struct cfq_data *cfqd;
143         /* service_tree member */
144         struct rb_node rb_node;
145         /* service_tree key */
146         unsigned long rb_key;
147         /* sorted list of pending requests */
148         struct rb_root sort_list;
149         /* if fifo isn't expired, next request to serve */
150         struct request *next_rq;
151         /* requests queued in sort_list */
152         int queued[2];
153         /* currently allocated requests */
154         int allocated[2];
155         /* fifo list of requests in sort_list */
156         struct list_head fifo;
157
158         unsigned long slice_end;
159         long slice_resid;
160         unsigned int slice_dispatch;
161
162         /* pending metadata requests */
163         int meta_pending;
164         /* number of requests that are on the dispatch list or inside driver */
165         int dispatched;
166
167         /* io prio of this group */
168         unsigned short ioprio, org_ioprio;
169         unsigned short ioprio_class, org_ioprio_class;
170
171         pid_t pid;
172 };
173
174 enum cfqq_state_flags {
175         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
176         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
177         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
178         CFQ_CFQQ_FLAG_must_alloc,       /* must be allowed rq alloc */
179         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
180         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
181         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
182         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
183         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
184         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
185 };
186
187 #define CFQ_CFQQ_FNS(name)                                              \
188 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
189 {                                                                       \
190         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
191 }                                                                       \
192 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
193 {                                                                       \
194         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
195 }                                                                       \
196 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
197 {                                                                       \
198         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
199 }
200
201 CFQ_CFQQ_FNS(on_rr);
202 CFQ_CFQQ_FNS(wait_request);
203 CFQ_CFQQ_FNS(must_dispatch);
204 CFQ_CFQQ_FNS(must_alloc);
205 CFQ_CFQQ_FNS(must_alloc_slice);
206 CFQ_CFQQ_FNS(fifo_expire);
207 CFQ_CFQQ_FNS(idle_window);
208 CFQ_CFQQ_FNS(prio_changed);
209 CFQ_CFQQ_FNS(slice_new);
210 CFQ_CFQQ_FNS(sync);
211 #undef CFQ_CFQQ_FNS
212
213 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
214         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
215 #define cfq_log(cfqd, fmt, args...)     \
216         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
217
218 static void cfq_dispatch_insert(struct request_queue *, struct request *);
219 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
220                                        struct io_context *, gfp_t);
221 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
222                                                 struct io_context *);
223
224 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
225                                             int is_sync)
226 {
227         return cic->cfqq[!!is_sync];
228 }
229
230 static inline void cic_set_cfqq(struct cfq_io_context *cic,
231                                 struct cfq_queue *cfqq, int is_sync)
232 {
233         cic->cfqq[!!is_sync] = cfqq;
234 }
235
236 /*
237  * We regard a request as SYNC, if it's either a read or has the SYNC bit
238  * set (in which case it could also be direct WRITE).
239  */
240 static inline int cfq_bio_sync(struct bio *bio)
241 {
242         if (bio_data_dir(bio) == READ || bio_sync(bio))
243                 return 1;
244
245         return 0;
246 }
247
248 /*
249  * scheduler run of queue, if there are requests pending and no one in the
250  * driver that will restart queueing
251  */
252 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
253 {
254         if (cfqd->busy_queues) {
255                 cfq_log(cfqd, "schedule dispatch");
256                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
257         }
258 }
259
260 static int cfq_queue_empty(struct request_queue *q)
261 {
262         struct cfq_data *cfqd = q->elevator->elevator_data;
263
264         return !cfqd->busy_queues;
265 }
266
267 /*
268  * Scale schedule slice based on io priority. Use the sync time slice only
269  * if a queue is marked sync and has sync io queued. A sync queue with async
270  * io only, should not get full sync slice length.
271  */
272 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
273                                  unsigned short prio)
274 {
275         const int base_slice = cfqd->cfq_slice[sync];
276
277         WARN_ON(prio >= IOPRIO_BE_NR);
278
279         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
280 }
281
282 static inline int
283 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
284 {
285         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
286 }
287
288 static inline void
289 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
290 {
291         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
292         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
293 }
294
295 /*
296  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
297  * isn't valid until the first request from the dispatch is activated
298  * and the slice time set.
299  */
300 static inline int cfq_slice_used(struct cfq_queue *cfqq)
301 {
302         if (cfq_cfqq_slice_new(cfqq))
303                 return 0;
304         if (time_before(jiffies, cfqq->slice_end))
305                 return 0;
306
307         return 1;
308 }
309
310 /*
311  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
312  * We choose the request that is closest to the head right now. Distance
313  * behind the head is penalized and only allowed to a certain extent.
314  */
315 static struct request *
316 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
317 {
318         sector_t last, s1, s2, d1 = 0, d2 = 0;
319         unsigned long back_max;
320 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
321 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
322         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
323
324         if (rq1 == NULL || rq1 == rq2)
325                 return rq2;
326         if (rq2 == NULL)
327                 return rq1;
328
329         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
330                 return rq1;
331         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
332                 return rq2;
333         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
334                 return rq1;
335         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
336                 return rq2;
337
338         s1 = rq1->sector;
339         s2 = rq2->sector;
340
341         last = cfqd->last_position;
342
343         /*
344          * by definition, 1KiB is 2 sectors
345          */
346         back_max = cfqd->cfq_back_max * 2;
347
348         /*
349          * Strict one way elevator _except_ in the case where we allow
350          * short backward seeks which are biased as twice the cost of a
351          * similar forward seek.
352          */
353         if (s1 >= last)
354                 d1 = s1 - last;
355         else if (s1 + back_max >= last)
356                 d1 = (last - s1) * cfqd->cfq_back_penalty;
357         else
358                 wrap |= CFQ_RQ1_WRAP;
359
360         if (s2 >= last)
361                 d2 = s2 - last;
362         else if (s2 + back_max >= last)
363                 d2 = (last - s2) * cfqd->cfq_back_penalty;
364         else
365                 wrap |= CFQ_RQ2_WRAP;
366
367         /* Found required data */
368
369         /*
370          * By doing switch() on the bit mask "wrap" we avoid having to
371          * check two variables for all permutations: --> faster!
372          */
373         switch (wrap) {
374         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
375                 if (d1 < d2)
376                         return rq1;
377                 else if (d2 < d1)
378                         return rq2;
379                 else {
380                         if (s1 >= s2)
381                                 return rq1;
382                         else
383                                 return rq2;
384                 }
385
386         case CFQ_RQ2_WRAP:
387                 return rq1;
388         case CFQ_RQ1_WRAP:
389                 return rq2;
390         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
391         default:
392                 /*
393                  * Since both rqs are wrapped,
394                  * start with the one that's further behind head
395                  * (--> only *one* back seek required),
396                  * since back seek takes more time than forward.
397                  */
398                 if (s1 <= s2)
399                         return rq1;
400                 else
401                         return rq2;
402         }
403 }
404
405 /*
406  * The below is leftmost cache rbtree addon
407  */
408 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
409 {
410         if (!root->left)
411                 root->left = rb_first(&root->rb);
412
413         if (root->left)
414                 return rb_entry(root->left, struct cfq_queue, rb_node);
415
416         return NULL;
417 }
418
419 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
420 {
421         if (root->left == n)
422                 root->left = NULL;
423
424         rb_erase(n, &root->rb);
425         RB_CLEAR_NODE(n);
426 }
427
428 /*
429  * would be nice to take fifo expire time into account as well
430  */
431 static struct request *
432 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
433                   struct request *last)
434 {
435         struct rb_node *rbnext = rb_next(&last->rb_node);
436         struct rb_node *rbprev = rb_prev(&last->rb_node);
437         struct request *next = NULL, *prev = NULL;
438
439         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
440
441         if (rbprev)
442                 prev = rb_entry_rq(rbprev);
443
444         if (rbnext)
445                 next = rb_entry_rq(rbnext);
446         else {
447                 rbnext = rb_first(&cfqq->sort_list);
448                 if (rbnext && rbnext != &last->rb_node)
449                         next = rb_entry_rq(rbnext);
450         }
451
452         return cfq_choose_req(cfqd, next, prev);
453 }
454
455 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
456                                       struct cfq_queue *cfqq)
457 {
458         /*
459          * just an approximation, should be ok.
460          */
461         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
462                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
463 }
464
465 /*
466  * The cfqd->service_tree holds all pending cfq_queue's that have
467  * requests waiting to be processed. It is sorted in the order that
468  * we will service the queues.
469  */
470 static void cfq_service_tree_add(struct cfq_data *cfqd,
471                                     struct cfq_queue *cfqq, int add_front)
472 {
473         struct rb_node **p, *parent;
474         struct cfq_queue *__cfqq;
475         unsigned long rb_key;
476         int left;
477
478         if (cfq_class_idle(cfqq)) {
479                 rb_key = CFQ_IDLE_DELAY;
480                 parent = rb_last(&cfqd->service_tree.rb);
481                 if (parent && parent != &cfqq->rb_node) {
482                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
483                         rb_key += __cfqq->rb_key;
484                 } else
485                         rb_key += jiffies;
486         } else if (!add_front) {
487                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
488                 rb_key += cfqq->slice_resid;
489                 cfqq->slice_resid = 0;
490         } else
491                 rb_key = 0;
492
493         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
494                 /*
495                  * same position, nothing more to do
496                  */
497                 if (rb_key == cfqq->rb_key)
498                         return;
499
500                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
501         }
502
503         left = 1;
504         parent = NULL;
505         p = &cfqd->service_tree.rb.rb_node;
506         while (*p) {
507                 struct rb_node **n;
508
509                 parent = *p;
510                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
511
512                 /*
513                  * sort RT queues first, we always want to give
514                  * preference to them. IDLE queues goes to the back.
515                  * after that, sort on the next service time.
516                  */
517                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
518                         n = &(*p)->rb_left;
519                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
520                         n = &(*p)->rb_right;
521                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
522                         n = &(*p)->rb_left;
523                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
524                         n = &(*p)->rb_right;
525                 else if (rb_key < __cfqq->rb_key)
526                         n = &(*p)->rb_left;
527                 else
528                         n = &(*p)->rb_right;
529
530                 if (n == &(*p)->rb_right)
531                         left = 0;
532
533                 p = n;
534         }
535
536         if (left)
537                 cfqd->service_tree.left = &cfqq->rb_node;
538
539         cfqq->rb_key = rb_key;
540         rb_link_node(&cfqq->rb_node, parent, p);
541         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
542 }
543
544 /*
545  * Update cfqq's position in the service tree.
546  */
547 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
548 {
549         /*
550          * Resorting requires the cfqq to be on the RR list already.
551          */
552         if (cfq_cfqq_on_rr(cfqq))
553                 cfq_service_tree_add(cfqd, cfqq, 0);
554 }
555
556 /*
557  * add to busy list of queues for service, trying to be fair in ordering
558  * the pending list according to last request service
559  */
560 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
561 {
562         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
563         BUG_ON(cfq_cfqq_on_rr(cfqq));
564         cfq_mark_cfqq_on_rr(cfqq);
565         cfqd->busy_queues++;
566         if (cfq_class_rt(cfqq))
567                 cfqd->busy_rt_queues++;
568
569         cfq_resort_rr_list(cfqd, cfqq);
570 }
571
572 /*
573  * Called when the cfqq no longer has requests pending, remove it from
574  * the service tree.
575  */
576 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
577 {
578         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
579         BUG_ON(!cfq_cfqq_on_rr(cfqq));
580         cfq_clear_cfqq_on_rr(cfqq);
581
582         if (!RB_EMPTY_NODE(&cfqq->rb_node))
583                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
584
585         BUG_ON(!cfqd->busy_queues);
586         cfqd->busy_queues--;
587         if (cfq_class_rt(cfqq))
588                 cfqd->busy_rt_queues--;
589 }
590
591 /*
592  * rb tree support functions
593  */
594 static void cfq_del_rq_rb(struct request *rq)
595 {
596         struct cfq_queue *cfqq = RQ_CFQQ(rq);
597         struct cfq_data *cfqd = cfqq->cfqd;
598         const int sync = rq_is_sync(rq);
599
600         BUG_ON(!cfqq->queued[sync]);
601         cfqq->queued[sync]--;
602
603         elv_rb_del(&cfqq->sort_list, rq);
604
605         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
606                 cfq_del_cfqq_rr(cfqd, cfqq);
607 }
608
609 static void cfq_add_rq_rb(struct request *rq)
610 {
611         struct cfq_queue *cfqq = RQ_CFQQ(rq);
612         struct cfq_data *cfqd = cfqq->cfqd;
613         struct request *__alias;
614
615         cfqq->queued[rq_is_sync(rq)]++;
616
617         /*
618          * looks a little odd, but the first insert might return an alias.
619          * if that happens, put the alias on the dispatch list
620          */
621         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
622                 cfq_dispatch_insert(cfqd->queue, __alias);
623
624         if (!cfq_cfqq_on_rr(cfqq))
625                 cfq_add_cfqq_rr(cfqd, cfqq);
626
627         /*
628          * check if this request is a better next-serve candidate
629          */
630         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
631         BUG_ON(!cfqq->next_rq);
632 }
633
634 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
635 {
636         elv_rb_del(&cfqq->sort_list, rq);
637         cfqq->queued[rq_is_sync(rq)]--;
638         cfq_add_rq_rb(rq);
639 }
640
641 static struct request *
642 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
643 {
644         struct task_struct *tsk = current;
645         struct cfq_io_context *cic;
646         struct cfq_queue *cfqq;
647
648         cic = cfq_cic_lookup(cfqd, tsk->io_context);
649         if (!cic)
650                 return NULL;
651
652         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
653         if (cfqq) {
654                 sector_t sector = bio->bi_sector + bio_sectors(bio);
655
656                 return elv_rb_find(&cfqq->sort_list, sector);
657         }
658
659         return NULL;
660 }
661
662 static void cfq_activate_request(struct request_queue *q, struct request *rq)
663 {
664         struct cfq_data *cfqd = q->elevator->elevator_data;
665
666         cfqd->rq_in_driver++;
667         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
668                                                 cfqd->rq_in_driver);
669
670         cfqd->last_position = rq->hard_sector + rq->hard_nr_sectors;
671 }
672
673 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
674 {
675         struct cfq_data *cfqd = q->elevator->elevator_data;
676
677         WARN_ON(!cfqd->rq_in_driver);
678         cfqd->rq_in_driver--;
679         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
680                                                 cfqd->rq_in_driver);
681 }
682
683 static void cfq_remove_request(struct request *rq)
684 {
685         struct cfq_queue *cfqq = RQ_CFQQ(rq);
686
687         if (cfqq->next_rq == rq)
688                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
689
690         list_del_init(&rq->queuelist);
691         cfq_del_rq_rb(rq);
692
693         cfqq->cfqd->rq_queued--;
694         if (rq_is_meta(rq)) {
695                 WARN_ON(!cfqq->meta_pending);
696                 cfqq->meta_pending--;
697         }
698 }
699
700 static int cfq_merge(struct request_queue *q, struct request **req,
701                      struct bio *bio)
702 {
703         struct cfq_data *cfqd = q->elevator->elevator_data;
704         struct request *__rq;
705
706         __rq = cfq_find_rq_fmerge(cfqd, bio);
707         if (__rq && elv_rq_merge_ok(__rq, bio)) {
708                 *req = __rq;
709                 return ELEVATOR_FRONT_MERGE;
710         }
711
712         return ELEVATOR_NO_MERGE;
713 }
714
715 static void cfq_merged_request(struct request_queue *q, struct request *req,
716                                int type)
717 {
718         if (type == ELEVATOR_FRONT_MERGE) {
719                 struct cfq_queue *cfqq = RQ_CFQQ(req);
720
721                 cfq_reposition_rq_rb(cfqq, req);
722         }
723 }
724
725 static void
726 cfq_merged_requests(struct request_queue *q, struct request *rq,
727                     struct request *next)
728 {
729         /*
730          * reposition in fifo if next is older than rq
731          */
732         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
733             time_before(next->start_time, rq->start_time))
734                 list_move(&rq->queuelist, &next->queuelist);
735
736         cfq_remove_request(next);
737 }
738
739 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
740                            struct bio *bio)
741 {
742         struct cfq_data *cfqd = q->elevator->elevator_data;
743         struct cfq_io_context *cic;
744         struct cfq_queue *cfqq;
745
746         /*
747          * Disallow merge of a sync bio into an async request.
748          */
749         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
750                 return 0;
751
752         /*
753          * Lookup the cfqq that this bio will be queued with. Allow
754          * merge only if rq is queued there.
755          */
756         cic = cfq_cic_lookup(cfqd, current->io_context);
757         if (!cic)
758                 return 0;
759
760         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
761         if (cfqq == RQ_CFQQ(rq))
762                 return 1;
763
764         return 0;
765 }
766
767 static void __cfq_set_active_queue(struct cfq_data *cfqd,
768                                    struct cfq_queue *cfqq)
769 {
770         if (cfqq) {
771                 cfq_log_cfqq(cfqd, cfqq, "set_active");
772                 cfqq->slice_end = 0;
773                 cfqq->slice_dispatch = 0;
774
775                 cfq_clear_cfqq_wait_request(cfqq);
776                 cfq_clear_cfqq_must_dispatch(cfqq);
777                 cfq_clear_cfqq_must_alloc_slice(cfqq);
778                 cfq_clear_cfqq_fifo_expire(cfqq);
779                 cfq_mark_cfqq_slice_new(cfqq);
780
781                 del_timer(&cfqd->idle_slice_timer);
782         }
783
784         cfqd->active_queue = cfqq;
785 }
786
787 /*
788  * current cfqq expired its slice (or was too idle), select new one
789  */
790 static void
791 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
792                     int timed_out)
793 {
794         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
795
796         if (cfq_cfqq_wait_request(cfqq))
797                 del_timer(&cfqd->idle_slice_timer);
798
799         cfq_clear_cfqq_wait_request(cfqq);
800
801         /*
802          * store what was left of this slice, if the queue idled/timed out
803          */
804         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
805                 cfqq->slice_resid = cfqq->slice_end - jiffies;
806                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
807         }
808
809         cfq_resort_rr_list(cfqd, cfqq);
810
811         if (cfqq == cfqd->active_queue)
812                 cfqd->active_queue = NULL;
813
814         if (cfqd->active_cic) {
815                 put_io_context(cfqd->active_cic->ioc);
816                 cfqd->active_cic = NULL;
817         }
818 }
819
820 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
821 {
822         struct cfq_queue *cfqq = cfqd->active_queue;
823
824         if (cfqq)
825                 __cfq_slice_expired(cfqd, cfqq, timed_out);
826 }
827
828 /*
829  * Get next queue for service. Unless we have a queue preemption,
830  * we'll simply select the first cfqq in the service tree.
831  */
832 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
833 {
834         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
835                 return NULL;
836
837         return cfq_rb_first(&cfqd->service_tree);
838 }
839
840 /*
841  * Get and set a new active queue for service.
842  */
843 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd)
844 {
845         struct cfq_queue *cfqq;
846
847         cfqq = cfq_get_next_queue(cfqd);
848         __cfq_set_active_queue(cfqd, cfqq);
849         return cfqq;
850 }
851
852 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
853                                           struct request *rq)
854 {
855         if (rq->sector >= cfqd->last_position)
856                 return rq->sector - cfqd->last_position;
857         else
858                 return cfqd->last_position - rq->sector;
859 }
860
861 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
862 {
863         struct cfq_io_context *cic = cfqd->active_cic;
864
865         if (!sample_valid(cic->seek_samples))
866                 return 0;
867
868         return cfq_dist_from_last(cfqd, rq) <= cic->seek_mean;
869 }
870
871 static int cfq_close_cooperator(struct cfq_data *cfq_data,
872                                 struct cfq_queue *cfqq)
873 {
874         /*
875          * We should notice if some of the queues are cooperating, eg
876          * working closely on the same area of the disk. In that case,
877          * we can group them together and don't waste time idling.
878          */
879         return 0;
880 }
881
882 #define CIC_SEEKY(cic) ((cic)->seek_mean > (8 * 1024))
883
884 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
885 {
886         struct cfq_queue *cfqq = cfqd->active_queue;
887         struct cfq_io_context *cic;
888         unsigned long sl;
889
890         /*
891          * SSD device without seek penalty, disable idling. But only do so
892          * for devices that support queuing, otherwise we still have a problem
893          * with sync vs async workloads.
894          */
895         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
896                 return;
897
898         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
899         WARN_ON(cfq_cfqq_slice_new(cfqq));
900
901         /*
902          * idle is disabled, either manually or by past process history
903          */
904         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
905                 return;
906
907         /*
908          * still requests with the driver, don't idle
909          */
910         if (cfqd->rq_in_driver)
911                 return;
912
913         /*
914          * task has exited, don't wait
915          */
916         cic = cfqd->active_cic;
917         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
918                 return;
919
920         /*
921          * See if this prio level has a good candidate
922          */
923         if (cfq_close_cooperator(cfqd, cfqq) &&
924             (sample_valid(cic->ttime_samples) && cic->ttime_mean > 2))
925                 return;
926
927         cfq_mark_cfqq_wait_request(cfqq);
928
929         /*
930          * we don't want to idle for seeks, but we do want to allow
931          * fair distribution of slice time for a process doing back-to-back
932          * seeks. so allow a little bit of time for him to submit a new rq
933          */
934         sl = cfqd->cfq_slice_idle;
935         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
936                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
937
938         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
939         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
940 }
941
942 /*
943  * Move request from internal lists to the request queue dispatch list.
944  */
945 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
946 {
947         struct cfq_data *cfqd = q->elevator->elevator_data;
948         struct cfq_queue *cfqq = RQ_CFQQ(rq);
949
950         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
951
952         cfq_remove_request(rq);
953         cfqq->dispatched++;
954         elv_dispatch_sort(q, rq);
955
956         if (cfq_cfqq_sync(cfqq))
957                 cfqd->sync_flight++;
958 }
959
960 /*
961  * return expired entry, or NULL to just start from scratch in rbtree
962  */
963 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
964 {
965         struct cfq_data *cfqd = cfqq->cfqd;
966         struct request *rq;
967         int fifo;
968
969         if (cfq_cfqq_fifo_expire(cfqq))
970                 return NULL;
971
972         cfq_mark_cfqq_fifo_expire(cfqq);
973
974         if (list_empty(&cfqq->fifo))
975                 return NULL;
976
977         fifo = cfq_cfqq_sync(cfqq);
978         rq = rq_entry_fifo(cfqq->fifo.next);
979
980         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
981                 rq = NULL;
982
983         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
984         return rq;
985 }
986
987 static inline int
988 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
989 {
990         const int base_rq = cfqd->cfq_slice_async_rq;
991
992         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
993
994         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
995 }
996
997 /*
998  * Select a queue for service. If we have a current active queue,
999  * check whether to continue servicing it, or retrieve and set a new one.
1000  */
1001 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1002 {
1003         struct cfq_queue *cfqq;
1004
1005         cfqq = cfqd->active_queue;
1006         if (!cfqq)
1007                 goto new_queue;
1008
1009         /*
1010          * The active queue has run out of time, expire it and select new.
1011          */
1012         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1013                 goto expire;
1014
1015         /*
1016          * If we have a RT cfqq waiting, then we pre-empt the current non-rt
1017          * cfqq.
1018          */
1019         if (!cfq_class_rt(cfqq) && cfqd->busy_rt_queues) {
1020                 /*
1021                  * We simulate this as cfqq timed out so that it gets to bank
1022                  * the remaining of its time slice.
1023                  */
1024                 cfq_log_cfqq(cfqd, cfqq, "preempt");
1025                 cfq_slice_expired(cfqd, 1);
1026                 goto new_queue;
1027         }
1028
1029         /*
1030          * The active queue has requests and isn't expired, allow it to
1031          * dispatch.
1032          */
1033         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1034                 goto keep_queue;
1035
1036         /*
1037          * No requests pending. If the active queue still has requests in
1038          * flight or is idling for a new request, allow either of these
1039          * conditions to happen (or time out) before selecting a new queue.
1040          */
1041         if (timer_pending(&cfqd->idle_slice_timer) ||
1042             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1043                 cfqq = NULL;
1044                 goto keep_queue;
1045         }
1046
1047 expire:
1048         cfq_slice_expired(cfqd, 0);
1049 new_queue:
1050         cfqq = cfq_set_active_queue(cfqd);
1051 keep_queue:
1052         return cfqq;
1053 }
1054
1055 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1056 {
1057         int dispatched = 0;
1058
1059         while (cfqq->next_rq) {
1060                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1061                 dispatched++;
1062         }
1063
1064         BUG_ON(!list_empty(&cfqq->fifo));
1065         return dispatched;
1066 }
1067
1068 /*
1069  * Drain our current requests. Used for barriers and when switching
1070  * io schedulers on-the-fly.
1071  */
1072 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1073 {
1074         struct cfq_queue *cfqq;
1075         int dispatched = 0;
1076
1077         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1078                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1079
1080         cfq_slice_expired(cfqd, 0);
1081
1082         BUG_ON(cfqd->busy_queues);
1083
1084         cfq_log(cfqd, "forced_dispatch=%d\n", dispatched);
1085         return dispatched;
1086 }
1087
1088 /*
1089  * Dispatch a request from cfqq, moving them to the request queue
1090  * dispatch list.
1091  */
1092 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1093 {
1094         struct request *rq;
1095
1096         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1097
1098         /*
1099          * follow expired path, else get first next available
1100          */
1101         rq = cfq_check_fifo(cfqq);
1102         if (!rq)
1103                 rq = cfqq->next_rq;
1104
1105         /*
1106          * insert request into driver dispatch list
1107          */
1108         cfq_dispatch_insert(cfqd->queue, rq);
1109
1110         if (!cfqd->active_cic) {
1111                 struct cfq_io_context *cic = RQ_CIC(rq);
1112
1113                 atomic_inc(&cic->ioc->refcount);
1114                 cfqd->active_cic = cic;
1115         }
1116 }
1117
1118 /*
1119  * Find the cfqq that we need to service and move a request from that to the
1120  * dispatch list
1121  */
1122 static int cfq_dispatch_requests(struct request_queue *q, int force)
1123 {
1124         struct cfq_data *cfqd = q->elevator->elevator_data;
1125         struct cfq_queue *cfqq;
1126         unsigned int max_dispatch;
1127
1128         if (!cfqd->busy_queues)
1129                 return 0;
1130
1131         if (unlikely(force))
1132                 return cfq_forced_dispatch(cfqd);
1133
1134         cfqq = cfq_select_queue(cfqd);
1135         if (!cfqq)
1136                 return 0;
1137
1138         /*
1139          * If this is an async queue and we have sync IO in flight, let it wait
1140          */
1141         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1142                 return 0;
1143
1144         max_dispatch = cfqd->cfq_quantum;
1145         if (cfq_class_idle(cfqq))
1146                 max_dispatch = 1;
1147
1148         /*
1149          * Does this cfqq already have too much IO in flight?
1150          */
1151         if (cfqq->dispatched >= max_dispatch) {
1152                 /*
1153                  * idle queue must always only have a single IO in flight
1154                  */
1155                 if (cfq_class_idle(cfqq))
1156                         return 0;
1157
1158                 /*
1159                  * We have other queues, don't allow more IO from this one
1160                  */
1161                 if (cfqd->busy_queues > 1)
1162                         return 0;
1163
1164                 /*
1165                  * we are the only queue, allow up to 4 times of 'quantum'
1166                  */
1167                 if (cfqq->dispatched >= 4 * max_dispatch)
1168                         return 0;
1169         }
1170
1171         /*
1172          * Dispatch a request from this cfqq
1173          */
1174         cfq_dispatch_request(cfqd, cfqq);
1175         cfqq->slice_dispatch++;
1176         cfq_clear_cfqq_must_dispatch(cfqq);
1177
1178         /*
1179          * expire an async queue immediately if it has used up its slice. idle
1180          * queue always expire after 1 dispatch round.
1181          */
1182         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1183             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1184             cfq_class_idle(cfqq))) {
1185                 cfqq->slice_end = jiffies + 1;
1186                 cfq_slice_expired(cfqd, 0);
1187         }
1188
1189         cfq_log(cfqd, "dispatched a request");
1190         return 1;
1191 }
1192
1193 /*
1194  * task holds one reference to the queue, dropped when task exits. each rq
1195  * in-flight on this queue also holds a reference, dropped when rq is freed.
1196  *
1197  * queue lock must be held here.
1198  */
1199 static void cfq_put_queue(struct cfq_queue *cfqq)
1200 {
1201         struct cfq_data *cfqd = cfqq->cfqd;
1202
1203         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1204
1205         if (!atomic_dec_and_test(&cfqq->ref))
1206                 return;
1207
1208         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1209         BUG_ON(rb_first(&cfqq->sort_list));
1210         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1211         BUG_ON(cfq_cfqq_on_rr(cfqq));
1212
1213         if (unlikely(cfqd->active_queue == cfqq)) {
1214                 __cfq_slice_expired(cfqd, cfqq, 0);
1215                 cfq_schedule_dispatch(cfqd);
1216         }
1217
1218         kmem_cache_free(cfq_pool, cfqq);
1219 }
1220
1221 /*
1222  * Must always be called with the rcu_read_lock() held
1223  */
1224 static void
1225 __call_for_each_cic(struct io_context *ioc,
1226                     void (*func)(struct io_context *, struct cfq_io_context *))
1227 {
1228         struct cfq_io_context *cic;
1229         struct hlist_node *n;
1230
1231         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1232                 func(ioc, cic);
1233 }
1234
1235 /*
1236  * Call func for each cic attached to this ioc.
1237  */
1238 static void
1239 call_for_each_cic(struct io_context *ioc,
1240                   void (*func)(struct io_context *, struct cfq_io_context *))
1241 {
1242         rcu_read_lock();
1243         __call_for_each_cic(ioc, func);
1244         rcu_read_unlock();
1245 }
1246
1247 static void cfq_cic_free_rcu(struct rcu_head *head)
1248 {
1249         struct cfq_io_context *cic;
1250
1251         cic = container_of(head, struct cfq_io_context, rcu_head);
1252
1253         kmem_cache_free(cfq_ioc_pool, cic);
1254         elv_ioc_count_dec(ioc_count);
1255
1256         if (ioc_gone) {
1257                 /*
1258                  * CFQ scheduler is exiting, grab exit lock and check
1259                  * the pending io context count. If it hits zero,
1260                  * complete ioc_gone and set it back to NULL
1261                  */
1262                 spin_lock(&ioc_gone_lock);
1263                 if (ioc_gone && !elv_ioc_count_read(ioc_count)) {
1264                         complete(ioc_gone);
1265                         ioc_gone = NULL;
1266                 }
1267                 spin_unlock(&ioc_gone_lock);
1268         }
1269 }
1270
1271 static void cfq_cic_free(struct cfq_io_context *cic)
1272 {
1273         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1274 }
1275
1276 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1277 {
1278         unsigned long flags;
1279
1280         BUG_ON(!cic->dead_key);
1281
1282         spin_lock_irqsave(&ioc->lock, flags);
1283         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1284         hlist_del_rcu(&cic->cic_list);
1285         spin_unlock_irqrestore(&ioc->lock, flags);
1286
1287         cfq_cic_free(cic);
1288 }
1289
1290 /*
1291  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1292  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1293  * and ->trim() which is called with the task lock held
1294  */
1295 static void cfq_free_io_context(struct io_context *ioc)
1296 {
1297         /*
1298          * ioc->refcount is zero here, or we are called from elv_unregister(),
1299          * so no more cic's are allowed to be linked into this ioc.  So it
1300          * should be ok to iterate over the known list, we will see all cic's
1301          * since no new ones are added.
1302          */
1303         __call_for_each_cic(ioc, cic_free_func);
1304 }
1305
1306 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1307 {
1308         if (unlikely(cfqq == cfqd->active_queue)) {
1309                 __cfq_slice_expired(cfqd, cfqq, 0);
1310                 cfq_schedule_dispatch(cfqd);
1311         }
1312
1313         cfq_put_queue(cfqq);
1314 }
1315
1316 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1317                                          struct cfq_io_context *cic)
1318 {
1319         struct io_context *ioc = cic->ioc;
1320
1321         list_del_init(&cic->queue_list);
1322
1323         /*
1324          * Make sure key == NULL is seen for dead queues
1325          */
1326         smp_wmb();
1327         cic->dead_key = (unsigned long) cic->key;
1328         cic->key = NULL;
1329
1330         if (ioc->ioc_data == cic)
1331                 rcu_assign_pointer(ioc->ioc_data, NULL);
1332
1333         if (cic->cfqq[BLK_RW_ASYNC]) {
1334                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1335                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1336         }
1337
1338         if (cic->cfqq[BLK_RW_SYNC]) {
1339                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1340                 cic->cfqq[BLK_RW_SYNC] = NULL;
1341         }
1342 }
1343
1344 static void cfq_exit_single_io_context(struct io_context *ioc,
1345                                        struct cfq_io_context *cic)
1346 {
1347         struct cfq_data *cfqd = cic->key;
1348
1349         if (cfqd) {
1350                 struct request_queue *q = cfqd->queue;
1351                 unsigned long flags;
1352
1353                 spin_lock_irqsave(q->queue_lock, flags);
1354
1355                 /*
1356                  * Ensure we get a fresh copy of the ->key to prevent
1357                  * race between exiting task and queue
1358                  */
1359                 smp_read_barrier_depends();
1360                 if (cic->key)
1361                         __cfq_exit_single_io_context(cfqd, cic);
1362
1363                 spin_unlock_irqrestore(q->queue_lock, flags);
1364         }
1365 }
1366
1367 /*
1368  * The process that ioc belongs to has exited, we need to clean up
1369  * and put the internal structures we have that belongs to that process.
1370  */
1371 static void cfq_exit_io_context(struct io_context *ioc)
1372 {
1373         call_for_each_cic(ioc, cfq_exit_single_io_context);
1374 }
1375
1376 static struct cfq_io_context *
1377 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1378 {
1379         struct cfq_io_context *cic;
1380
1381         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1382                                                         cfqd->queue->node);
1383         if (cic) {
1384                 cic->last_end_request = jiffies;
1385                 INIT_LIST_HEAD(&cic->queue_list);
1386                 INIT_HLIST_NODE(&cic->cic_list);
1387                 cic->dtor = cfq_free_io_context;
1388                 cic->exit = cfq_exit_io_context;
1389                 elv_ioc_count_inc(ioc_count);
1390         }
1391
1392         return cic;
1393 }
1394
1395 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1396 {
1397         struct task_struct *tsk = current;
1398         int ioprio_class;
1399
1400         if (!cfq_cfqq_prio_changed(cfqq))
1401                 return;
1402
1403         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1404         switch (ioprio_class) {
1405         default:
1406                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1407         case IOPRIO_CLASS_NONE:
1408                 /*
1409                  * no prio set, inherit CPU scheduling settings
1410                  */
1411                 cfqq->ioprio = task_nice_ioprio(tsk);
1412                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1413                 break;
1414         case IOPRIO_CLASS_RT:
1415                 cfqq->ioprio = task_ioprio(ioc);
1416                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1417                 break;
1418         case IOPRIO_CLASS_BE:
1419                 cfqq->ioprio = task_ioprio(ioc);
1420                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1421                 break;
1422         case IOPRIO_CLASS_IDLE:
1423                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1424                 cfqq->ioprio = 7;
1425                 cfq_clear_cfqq_idle_window(cfqq);
1426                 break;
1427         }
1428
1429         /*
1430          * keep track of original prio settings in case we have to temporarily
1431          * elevate the priority of this queue
1432          */
1433         cfqq->org_ioprio = cfqq->ioprio;
1434         cfqq->org_ioprio_class = cfqq->ioprio_class;
1435         cfq_clear_cfqq_prio_changed(cfqq);
1436 }
1437
1438 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1439 {
1440         struct cfq_data *cfqd = cic->key;
1441         struct cfq_queue *cfqq;
1442         unsigned long flags;
1443
1444         if (unlikely(!cfqd))
1445                 return;
1446
1447         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1448
1449         cfqq = cic->cfqq[BLK_RW_ASYNC];
1450         if (cfqq) {
1451                 struct cfq_queue *new_cfqq;
1452                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1453                                                 GFP_ATOMIC);
1454                 if (new_cfqq) {
1455                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1456                         cfq_put_queue(cfqq);
1457                 }
1458         }
1459
1460         cfqq = cic->cfqq[BLK_RW_SYNC];
1461         if (cfqq)
1462                 cfq_mark_cfqq_prio_changed(cfqq);
1463
1464         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1465 }
1466
1467 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1468 {
1469         call_for_each_cic(ioc, changed_ioprio);
1470         ioc->ioprio_changed = 0;
1471 }
1472
1473 static struct cfq_queue *
1474 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1475                      struct io_context *ioc, gfp_t gfp_mask)
1476 {
1477         struct cfq_queue *cfqq, *new_cfqq = NULL;
1478         struct cfq_io_context *cic;
1479
1480 retry:
1481         cic = cfq_cic_lookup(cfqd, ioc);
1482         /* cic always exists here */
1483         cfqq = cic_to_cfqq(cic, is_sync);
1484
1485         if (!cfqq) {
1486                 if (new_cfqq) {
1487                         cfqq = new_cfqq;
1488                         new_cfqq = NULL;
1489                 } else if (gfp_mask & __GFP_WAIT) {
1490                         /*
1491                          * Inform the allocator of the fact that we will
1492                          * just repeat this allocation if it fails, to allow
1493                          * the allocator to do whatever it needs to attempt to
1494                          * free memory.
1495                          */
1496                         spin_unlock_irq(cfqd->queue->queue_lock);
1497                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1498                                         gfp_mask | __GFP_NOFAIL | __GFP_ZERO,
1499                                         cfqd->queue->node);
1500                         spin_lock_irq(cfqd->queue->queue_lock);
1501                         goto retry;
1502                 } else {
1503                         cfqq = kmem_cache_alloc_node(cfq_pool,
1504                                         gfp_mask | __GFP_ZERO,
1505                                         cfqd->queue->node);
1506                         if (!cfqq)
1507                                 goto out;
1508                 }
1509
1510                 RB_CLEAR_NODE(&cfqq->rb_node);
1511                 INIT_LIST_HEAD(&cfqq->fifo);
1512
1513                 atomic_set(&cfqq->ref, 0);
1514                 cfqq->cfqd = cfqd;
1515
1516                 cfq_mark_cfqq_prio_changed(cfqq);
1517
1518                 cfq_init_prio_data(cfqq, ioc);
1519
1520                 if (is_sync) {
1521                         if (!cfq_class_idle(cfqq))
1522                                 cfq_mark_cfqq_idle_window(cfqq);
1523                         cfq_mark_cfqq_sync(cfqq);
1524                 }
1525                 cfqq->pid = current->pid;
1526                 cfq_log_cfqq(cfqd, cfqq, "alloced");
1527         }
1528
1529         if (new_cfqq)
1530                 kmem_cache_free(cfq_pool, new_cfqq);
1531
1532 out:
1533         WARN_ON((gfp_mask & __GFP_WAIT) && !cfqq);
1534         return cfqq;
1535 }
1536
1537 static struct cfq_queue **
1538 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1539 {
1540         switch (ioprio_class) {
1541         case IOPRIO_CLASS_RT:
1542                 return &cfqd->async_cfqq[0][ioprio];
1543         case IOPRIO_CLASS_BE:
1544                 return &cfqd->async_cfqq[1][ioprio];
1545         case IOPRIO_CLASS_IDLE:
1546                 return &cfqd->async_idle_cfqq;
1547         default:
1548                 BUG();
1549         }
1550 }
1551
1552 static struct cfq_queue *
1553 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1554               gfp_t gfp_mask)
1555 {
1556         const int ioprio = task_ioprio(ioc);
1557         const int ioprio_class = task_ioprio_class(ioc);
1558         struct cfq_queue **async_cfqq = NULL;
1559         struct cfq_queue *cfqq = NULL;
1560
1561         if (!is_sync) {
1562                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1563                 cfqq = *async_cfqq;
1564         }
1565
1566         if (!cfqq) {
1567                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1568                 if (!cfqq)
1569                         return NULL;
1570         }
1571
1572         /*
1573          * pin the queue now that it's allocated, scheduler exit will prune it
1574          */
1575         if (!is_sync && !(*async_cfqq)) {
1576                 atomic_inc(&cfqq->ref);
1577                 *async_cfqq = cfqq;
1578         }
1579
1580         atomic_inc(&cfqq->ref);
1581         return cfqq;
1582 }
1583
1584 /*
1585  * We drop cfq io contexts lazily, so we may find a dead one.
1586  */
1587 static void
1588 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1589                   struct cfq_io_context *cic)
1590 {
1591         unsigned long flags;
1592
1593         WARN_ON(!list_empty(&cic->queue_list));
1594
1595         spin_lock_irqsave(&ioc->lock, flags);
1596
1597         BUG_ON(ioc->ioc_data == cic);
1598
1599         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1600         hlist_del_rcu(&cic->cic_list);
1601         spin_unlock_irqrestore(&ioc->lock, flags);
1602
1603         cfq_cic_free(cic);
1604 }
1605
1606 static struct cfq_io_context *
1607 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1608 {
1609         struct cfq_io_context *cic;
1610         unsigned long flags;
1611         void *k;
1612
1613         if (unlikely(!ioc))
1614                 return NULL;
1615
1616         rcu_read_lock();
1617
1618         /*
1619          * we maintain a last-hit cache, to avoid browsing over the tree
1620          */
1621         cic = rcu_dereference(ioc->ioc_data);
1622         if (cic && cic->key == cfqd) {
1623                 rcu_read_unlock();
1624                 return cic;
1625         }
1626
1627         do {
1628                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1629                 rcu_read_unlock();
1630                 if (!cic)
1631                         break;
1632                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1633                 k = cic->key;
1634                 if (unlikely(!k)) {
1635                         cfq_drop_dead_cic(cfqd, ioc, cic);
1636                         rcu_read_lock();
1637                         continue;
1638                 }
1639
1640                 spin_lock_irqsave(&ioc->lock, flags);
1641                 rcu_assign_pointer(ioc->ioc_data, cic);
1642                 spin_unlock_irqrestore(&ioc->lock, flags);
1643                 break;
1644         } while (1);
1645
1646         return cic;
1647 }
1648
1649 /*
1650  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1651  * the process specific cfq io context when entered from the block layer.
1652  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1653  */
1654 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1655                         struct cfq_io_context *cic, gfp_t gfp_mask)
1656 {
1657         unsigned long flags;
1658         int ret;
1659
1660         ret = radix_tree_preload(gfp_mask);
1661         if (!ret) {
1662                 cic->ioc = ioc;
1663                 cic->key = cfqd;
1664
1665                 spin_lock_irqsave(&ioc->lock, flags);
1666                 ret = radix_tree_insert(&ioc->radix_root,
1667                                                 (unsigned long) cfqd, cic);
1668                 if (!ret)
1669                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1670                 spin_unlock_irqrestore(&ioc->lock, flags);
1671
1672                 radix_tree_preload_end();
1673
1674                 if (!ret) {
1675                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1676                         list_add(&cic->queue_list, &cfqd->cic_list);
1677                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1678                 }
1679         }
1680
1681         if (ret)
1682                 printk(KERN_ERR "cfq: cic link failed!\n");
1683
1684         return ret;
1685 }
1686
1687 /*
1688  * Setup general io context and cfq io context. There can be several cfq
1689  * io contexts per general io context, if this process is doing io to more
1690  * than one device managed by cfq.
1691  */
1692 static struct cfq_io_context *
1693 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1694 {
1695         struct io_context *ioc = NULL;
1696         struct cfq_io_context *cic;
1697
1698         might_sleep_if(gfp_mask & __GFP_WAIT);
1699
1700         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1701         if (!ioc)
1702                 return NULL;
1703
1704         cic = cfq_cic_lookup(cfqd, ioc);
1705         if (cic)
1706                 goto out;
1707
1708         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1709         if (cic == NULL)
1710                 goto err;
1711
1712         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1713                 goto err_free;
1714
1715 out:
1716         smp_read_barrier_depends();
1717         if (unlikely(ioc->ioprio_changed))
1718                 cfq_ioc_set_ioprio(ioc);
1719
1720         return cic;
1721 err_free:
1722         cfq_cic_free(cic);
1723 err:
1724         put_io_context(ioc);
1725         return NULL;
1726 }
1727
1728 static void
1729 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1730 {
1731         unsigned long elapsed = jiffies - cic->last_end_request;
1732         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1733
1734         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1735         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1736         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1737 }
1738
1739 static void
1740 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1741                        struct request *rq)
1742 {
1743         sector_t sdist;
1744         u64 total;
1745
1746         if (cic->last_request_pos < rq->sector)
1747                 sdist = rq->sector - cic->last_request_pos;
1748         else
1749                 sdist = cic->last_request_pos - rq->sector;
1750
1751         /*
1752          * Don't allow the seek distance to get too large from the
1753          * odd fragment, pagein, etc
1754          */
1755         if (cic->seek_samples <= 60) /* second&third seek */
1756                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1757         else
1758                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1759
1760         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1761         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1762         total = cic->seek_total + (cic->seek_samples/2);
1763         do_div(total, cic->seek_samples);
1764         cic->seek_mean = (sector_t)total;
1765 }
1766
1767 /*
1768  * Disable idle window if the process thinks too long or seeks so much that
1769  * it doesn't matter
1770  */
1771 static void
1772 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1773                        struct cfq_io_context *cic)
1774 {
1775         int old_idle, enable_idle;
1776
1777         /*
1778          * Don't idle for async or idle io prio class
1779          */
1780         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1781                 return;
1782
1783         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1784
1785         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1786             (cfqd->hw_tag && CIC_SEEKY(cic)))
1787                 enable_idle = 0;
1788         else if (sample_valid(cic->ttime_samples)) {
1789                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1790                         enable_idle = 0;
1791                 else
1792                         enable_idle = 1;
1793         }
1794
1795         if (old_idle != enable_idle) {
1796                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1797                 if (enable_idle)
1798                         cfq_mark_cfqq_idle_window(cfqq);
1799                 else
1800                         cfq_clear_cfqq_idle_window(cfqq);
1801         }
1802 }
1803
1804 /*
1805  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1806  * no or if we aren't sure, a 1 will cause a preempt.
1807  */
1808 static int
1809 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1810                    struct request *rq)
1811 {
1812         struct cfq_queue *cfqq;
1813
1814         cfqq = cfqd->active_queue;
1815         if (!cfqq)
1816                 return 0;
1817
1818         if (cfq_slice_used(cfqq))
1819                 return 1;
1820
1821         if (cfq_class_idle(new_cfqq))
1822                 return 0;
1823
1824         if (cfq_class_idle(cfqq))
1825                 return 1;
1826
1827         /*
1828          * if the new request is sync, but the currently running queue is
1829          * not, let the sync request have priority.
1830          */
1831         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
1832                 return 1;
1833
1834         /*
1835          * So both queues are sync. Let the new request get disk time if
1836          * it's a metadata request and the current queue is doing regular IO.
1837          */
1838         if (rq_is_meta(rq) && !cfqq->meta_pending)
1839                 return 1;
1840
1841         /*
1842          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
1843          */
1844         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
1845                 return 1;
1846
1847         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
1848                 return 0;
1849
1850         /*
1851          * if this request is as-good as one we would expect from the
1852          * current cfqq, let it preempt
1853          */
1854         if (cfq_rq_close(cfqd, rq))
1855                 return 1;
1856
1857         return 0;
1858 }
1859
1860 /*
1861  * cfqq preempts the active queue. if we allowed preempt with no slice left,
1862  * let it have half of its nominal slice.
1863  */
1864 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1865 {
1866         cfq_log_cfqq(cfqd, cfqq, "preempt");
1867         cfq_slice_expired(cfqd, 1);
1868
1869         /*
1870          * Put the new queue at the front of the of the current list,
1871          * so we know that it will be selected next.
1872          */
1873         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1874
1875         cfq_service_tree_add(cfqd, cfqq, 1);
1876
1877         cfqq->slice_end = 0;
1878         cfq_mark_cfqq_slice_new(cfqq);
1879 }
1880
1881 /*
1882  * Called when a new fs request (rq) is added (to cfqq). Check if there's
1883  * something we should do about it
1884  */
1885 static void
1886 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1887                 struct request *rq)
1888 {
1889         struct cfq_io_context *cic = RQ_CIC(rq);
1890
1891         cfqd->rq_queued++;
1892         if (rq_is_meta(rq))
1893                 cfqq->meta_pending++;
1894
1895         cfq_update_io_thinktime(cfqd, cic);
1896         cfq_update_io_seektime(cfqd, cic, rq);
1897         cfq_update_idle_window(cfqd, cfqq, cic);
1898
1899         cic->last_request_pos = rq->sector + rq->nr_sectors;
1900
1901         if (cfqq == cfqd->active_queue) {
1902                 /*
1903                  * Remember that we saw a request from this process, but
1904                  * don't start queuing just yet. Otherwise we risk seeing lots
1905                  * of tiny requests, because we disrupt the normal plugging
1906                  * and merging. If the request is already larger than a single
1907                  * page, let it rip immediately. For that case we assume that
1908                  * merging is already done. Ditto for a busy system that
1909                  * has other work pending, don't risk delaying until the
1910                  * idle timer unplug to continue working.
1911                  */
1912                 if (cfq_cfqq_wait_request(cfqq)) {
1913                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
1914                             cfqd->busy_queues > 1) {
1915                                 del_timer(&cfqd->idle_slice_timer);
1916                                 blk_start_queueing(cfqd->queue);
1917                         }
1918                         cfq_mark_cfqq_must_dispatch(cfqq);
1919                 }
1920         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
1921                 /*
1922                  * not the active queue - expire current slice if it is
1923                  * idle and has expired it's mean thinktime or this new queue
1924                  * has some old slice time left and is of higher priority or
1925                  * this new queue is RT and the current one is BE
1926                  */
1927                 cfq_preempt_queue(cfqd, cfqq);
1928                 blk_start_queueing(cfqd->queue);
1929         }
1930 }
1931
1932 static void cfq_insert_request(struct request_queue *q, struct request *rq)
1933 {
1934         struct cfq_data *cfqd = q->elevator->elevator_data;
1935         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1936
1937         cfq_log_cfqq(cfqd, cfqq, "insert_request");
1938         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
1939
1940         cfq_add_rq_rb(rq);
1941
1942         list_add_tail(&rq->queuelist, &cfqq->fifo);
1943
1944         cfq_rq_enqueued(cfqd, cfqq, rq);
1945 }
1946
1947 /*
1948  * Update hw_tag based on peak queue depth over 50 samples under
1949  * sufficient load.
1950  */
1951 static void cfq_update_hw_tag(struct cfq_data *cfqd)
1952 {
1953         if (cfqd->rq_in_driver > cfqd->rq_in_driver_peak)
1954                 cfqd->rq_in_driver_peak = cfqd->rq_in_driver;
1955
1956         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
1957             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
1958                 return;
1959
1960         if (cfqd->hw_tag_samples++ < 50)
1961                 return;
1962
1963         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
1964                 cfqd->hw_tag = 1;
1965         else
1966                 cfqd->hw_tag = 0;
1967
1968         cfqd->hw_tag_samples = 0;
1969         cfqd->rq_in_driver_peak = 0;
1970 }
1971
1972 static void cfq_completed_request(struct request_queue *q, struct request *rq)
1973 {
1974         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1975         struct cfq_data *cfqd = cfqq->cfqd;
1976         const int sync = rq_is_sync(rq);
1977         unsigned long now;
1978
1979         now = jiffies;
1980         cfq_log_cfqq(cfqd, cfqq, "complete");
1981
1982         cfq_update_hw_tag(cfqd);
1983
1984         WARN_ON(!cfqd->rq_in_driver);
1985         WARN_ON(!cfqq->dispatched);
1986         cfqd->rq_in_driver--;
1987         cfqq->dispatched--;
1988
1989         if (cfq_cfqq_sync(cfqq))
1990                 cfqd->sync_flight--;
1991
1992         if (!cfq_class_idle(cfqq))
1993                 cfqd->last_end_request = now;
1994
1995         if (sync)
1996                 RQ_CIC(rq)->last_end_request = now;
1997
1998         /*
1999          * If this is the active queue, check if it needs to be expired,
2000          * or if we want to idle in case it has no pending requests.
2001          */
2002         if (cfqd->active_queue == cfqq) {
2003                 if (cfq_cfqq_slice_new(cfqq)) {
2004                         cfq_set_prio_slice(cfqd, cfqq);
2005                         cfq_clear_cfqq_slice_new(cfqq);
2006                 }
2007                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2008                         cfq_slice_expired(cfqd, 1);
2009                 else if (sync && !rq_noidle(rq) &&
2010                          RB_EMPTY_ROOT(&cfqq->sort_list)) {
2011                         cfq_arm_slice_timer(cfqd);
2012                 }
2013         }
2014
2015         if (!cfqd->rq_in_driver)
2016                 cfq_schedule_dispatch(cfqd);
2017 }
2018
2019 /*
2020  * we temporarily boost lower priority queues if they are holding fs exclusive
2021  * resources. they are boosted to normal prio (CLASS_BE/4)
2022  */
2023 static void cfq_prio_boost(struct cfq_queue *cfqq)
2024 {
2025         if (has_fs_excl()) {
2026                 /*
2027                  * boost idle prio on transactions that would lock out other
2028                  * users of the filesystem
2029                  */
2030                 if (cfq_class_idle(cfqq))
2031                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2032                 if (cfqq->ioprio > IOPRIO_NORM)
2033                         cfqq->ioprio = IOPRIO_NORM;
2034         } else {
2035                 /*
2036                  * check if we need to unboost the queue
2037                  */
2038                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2039                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2040                 if (cfqq->ioprio != cfqq->org_ioprio)
2041                         cfqq->ioprio = cfqq->org_ioprio;
2042         }
2043 }
2044
2045 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2046 {
2047         if ((cfq_cfqq_wait_request(cfqq) || cfq_cfqq_must_alloc(cfqq)) &&
2048             !cfq_cfqq_must_alloc_slice(cfqq)) {
2049                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2050                 return ELV_MQUEUE_MUST;
2051         }
2052
2053         return ELV_MQUEUE_MAY;
2054 }
2055
2056 static int cfq_may_queue(struct request_queue *q, int rw)
2057 {
2058         struct cfq_data *cfqd = q->elevator->elevator_data;
2059         struct task_struct *tsk = current;
2060         struct cfq_io_context *cic;
2061         struct cfq_queue *cfqq;
2062
2063         /*
2064          * don't force setup of a queue from here, as a call to may_queue
2065          * does not necessarily imply that a request actually will be queued.
2066          * so just lookup a possibly existing queue, or return 'may queue'
2067          * if that fails
2068          */
2069         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2070         if (!cic)
2071                 return ELV_MQUEUE_MAY;
2072
2073         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2074         if (cfqq) {
2075                 cfq_init_prio_data(cfqq, cic->ioc);
2076                 cfq_prio_boost(cfqq);
2077
2078                 return __cfq_may_queue(cfqq);
2079         }
2080
2081         return ELV_MQUEUE_MAY;
2082 }
2083
2084 /*
2085  * queue lock held here
2086  */
2087 static void cfq_put_request(struct request *rq)
2088 {
2089         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2090
2091         if (cfqq) {
2092                 const int rw = rq_data_dir(rq);
2093
2094                 BUG_ON(!cfqq->allocated[rw]);
2095                 cfqq->allocated[rw]--;
2096
2097                 put_io_context(RQ_CIC(rq)->ioc);
2098
2099                 rq->elevator_private = NULL;
2100                 rq->elevator_private2 = NULL;
2101
2102                 cfq_put_queue(cfqq);
2103         }
2104 }
2105
2106 /*
2107  * Allocate cfq data structures associated with this request.
2108  */
2109 static int
2110 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2111 {
2112         struct cfq_data *cfqd = q->elevator->elevator_data;
2113         struct cfq_io_context *cic;
2114         const int rw = rq_data_dir(rq);
2115         const int is_sync = rq_is_sync(rq);
2116         struct cfq_queue *cfqq;
2117         unsigned long flags;
2118
2119         might_sleep_if(gfp_mask & __GFP_WAIT);
2120
2121         cic = cfq_get_io_context(cfqd, gfp_mask);
2122
2123         spin_lock_irqsave(q->queue_lock, flags);
2124
2125         if (!cic)
2126                 goto queue_fail;
2127
2128         cfqq = cic_to_cfqq(cic, is_sync);
2129         if (!cfqq) {
2130                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2131
2132                 if (!cfqq)
2133                         goto queue_fail;
2134
2135                 cic_set_cfqq(cic, cfqq, is_sync);
2136         }
2137
2138         cfqq->allocated[rw]++;
2139         cfq_clear_cfqq_must_alloc(cfqq);
2140         atomic_inc(&cfqq->ref);
2141
2142         spin_unlock_irqrestore(q->queue_lock, flags);
2143
2144         rq->elevator_private = cic;
2145         rq->elevator_private2 = cfqq;
2146         return 0;
2147
2148 queue_fail:
2149         if (cic)
2150                 put_io_context(cic->ioc);
2151
2152         cfq_schedule_dispatch(cfqd);
2153         spin_unlock_irqrestore(q->queue_lock, flags);
2154         cfq_log(cfqd, "set_request fail");
2155         return 1;
2156 }
2157
2158 static void cfq_kick_queue(struct work_struct *work)
2159 {
2160         struct cfq_data *cfqd =
2161                 container_of(work, struct cfq_data, unplug_work);
2162         struct request_queue *q = cfqd->queue;
2163
2164         spin_lock_irq(q->queue_lock);
2165         blk_start_queueing(q);
2166         spin_unlock_irq(q->queue_lock);
2167 }
2168
2169 /*
2170  * Timer running if the active_queue is currently idling inside its time slice
2171  */
2172 static void cfq_idle_slice_timer(unsigned long data)
2173 {
2174         struct cfq_data *cfqd = (struct cfq_data *) data;
2175         struct cfq_queue *cfqq;
2176         unsigned long flags;
2177         int timed_out = 1;
2178
2179         cfq_log(cfqd, "idle timer fired");
2180
2181         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2182
2183         cfqq = cfqd->active_queue;
2184         if (cfqq) {
2185                 timed_out = 0;
2186
2187                 /*
2188                  * We saw a request before the queue expired, let it through
2189                  */
2190                 if (cfq_cfqq_must_dispatch(cfqq))
2191                         goto out_kick;
2192
2193                 /*
2194                  * expired
2195                  */
2196                 if (cfq_slice_used(cfqq))
2197                         goto expire;
2198
2199                 /*
2200                  * only expire and reinvoke request handler, if there are
2201                  * other queues with pending requests
2202                  */
2203                 if (!cfqd->busy_queues)
2204                         goto out_cont;
2205
2206                 /*
2207                  * not expired and it has a request pending, let it dispatch
2208                  */
2209                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2210                         goto out_kick;
2211         }
2212 expire:
2213         cfq_slice_expired(cfqd, timed_out);
2214 out_kick:
2215         cfq_schedule_dispatch(cfqd);
2216 out_cont:
2217         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2218 }
2219
2220 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2221 {
2222         del_timer_sync(&cfqd->idle_slice_timer);
2223         cancel_work_sync(&cfqd->unplug_work);
2224 }
2225
2226 static void cfq_put_async_queues(struct cfq_data *cfqd)
2227 {
2228         int i;
2229
2230         for (i = 0; i < IOPRIO_BE_NR; i++) {
2231                 if (cfqd->async_cfqq[0][i])
2232                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2233                 if (cfqd->async_cfqq[1][i])
2234                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2235         }
2236
2237         if (cfqd->async_idle_cfqq)
2238                 cfq_put_queue(cfqd->async_idle_cfqq);
2239 }
2240
2241 static void cfq_exit_queue(struct elevator_queue *e)
2242 {
2243         struct cfq_data *cfqd = e->elevator_data;
2244         struct request_queue *q = cfqd->queue;
2245
2246         cfq_shutdown_timer_wq(cfqd);
2247
2248         spin_lock_irq(q->queue_lock);
2249
2250         if (cfqd->active_queue)
2251                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2252
2253         while (!list_empty(&cfqd->cic_list)) {
2254                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2255                                                         struct cfq_io_context,
2256                                                         queue_list);
2257
2258                 __cfq_exit_single_io_context(cfqd, cic);
2259         }
2260
2261         cfq_put_async_queues(cfqd);
2262
2263         spin_unlock_irq(q->queue_lock);
2264
2265         cfq_shutdown_timer_wq(cfqd);
2266
2267         kfree(cfqd);
2268 }
2269
2270 static void *cfq_init_queue(struct request_queue *q)
2271 {
2272         struct cfq_data *cfqd;
2273
2274         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2275         if (!cfqd)
2276                 return NULL;
2277
2278         cfqd->service_tree = CFQ_RB_ROOT;
2279         INIT_LIST_HEAD(&cfqd->cic_list);
2280
2281         cfqd->queue = q;
2282
2283         init_timer(&cfqd->idle_slice_timer);
2284         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2285         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2286
2287         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2288
2289         cfqd->last_end_request = jiffies;
2290         cfqd->cfq_quantum = cfq_quantum;
2291         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2292         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2293         cfqd->cfq_back_max = cfq_back_max;
2294         cfqd->cfq_back_penalty = cfq_back_penalty;
2295         cfqd->cfq_slice[0] = cfq_slice_async;
2296         cfqd->cfq_slice[1] = cfq_slice_sync;
2297         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2298         cfqd->cfq_slice_idle = cfq_slice_idle;
2299         cfqd->hw_tag = 1;
2300
2301         return cfqd;
2302 }
2303
2304 static void cfq_slab_kill(void)
2305 {
2306         /*
2307          * Caller already ensured that pending RCU callbacks are completed,
2308          * so we should have no busy allocations at this point.
2309          */
2310         if (cfq_pool)
2311                 kmem_cache_destroy(cfq_pool);
2312         if (cfq_ioc_pool)
2313                 kmem_cache_destroy(cfq_ioc_pool);
2314 }
2315
2316 static int __init cfq_slab_setup(void)
2317 {
2318         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2319         if (!cfq_pool)
2320                 goto fail;
2321
2322         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2323         if (!cfq_ioc_pool)
2324                 goto fail;
2325
2326         return 0;
2327 fail:
2328         cfq_slab_kill();
2329         return -ENOMEM;
2330 }
2331
2332 /*
2333  * sysfs parts below -->
2334  */
2335 static ssize_t
2336 cfq_var_show(unsigned int var, char *page)
2337 {
2338         return sprintf(page, "%d\n", var);
2339 }
2340
2341 static ssize_t
2342 cfq_var_store(unsigned int *var, const char *page, size_t count)
2343 {
2344         char *p = (char *) page;
2345
2346         *var = simple_strtoul(p, &p, 10);
2347         return count;
2348 }
2349
2350 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2351 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2352 {                                                                       \
2353         struct cfq_data *cfqd = e->elevator_data;                       \
2354         unsigned int __data = __VAR;                                    \
2355         if (__CONV)                                                     \
2356                 __data = jiffies_to_msecs(__data);                      \
2357         return cfq_var_show(__data, (page));                            \
2358 }
2359 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2360 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2361 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2362 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2363 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2364 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2365 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2366 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2367 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2368 #undef SHOW_FUNCTION
2369
2370 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2371 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2372 {                                                                       \
2373         struct cfq_data *cfqd = e->elevator_data;                       \
2374         unsigned int __data;                                            \
2375         int ret = cfq_var_store(&__data, (page), count);                \
2376         if (__data < (MIN))                                             \
2377                 __data = (MIN);                                         \
2378         else if (__data > (MAX))                                        \
2379                 __data = (MAX);                                         \
2380         if (__CONV)                                                     \
2381                 *(__PTR) = msecs_to_jiffies(__data);                    \
2382         else                                                            \
2383                 *(__PTR) = __data;                                      \
2384         return ret;                                                     \
2385 }
2386 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2387 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2388                 UINT_MAX, 1);
2389 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2390                 UINT_MAX, 1);
2391 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2392 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2393                 UINT_MAX, 0);
2394 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2395 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2396 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2397 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2398                 UINT_MAX, 0);
2399 #undef STORE_FUNCTION
2400
2401 #define CFQ_ATTR(name) \
2402         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2403
2404 static struct elv_fs_entry cfq_attrs[] = {
2405         CFQ_ATTR(quantum),
2406         CFQ_ATTR(fifo_expire_sync),
2407         CFQ_ATTR(fifo_expire_async),
2408         CFQ_ATTR(back_seek_max),
2409         CFQ_ATTR(back_seek_penalty),
2410         CFQ_ATTR(slice_sync),
2411         CFQ_ATTR(slice_async),
2412         CFQ_ATTR(slice_async_rq),
2413         CFQ_ATTR(slice_idle),
2414         __ATTR_NULL
2415 };
2416
2417 static struct elevator_type iosched_cfq = {
2418         .ops = {
2419                 .elevator_merge_fn =            cfq_merge,
2420                 .elevator_merged_fn =           cfq_merged_request,
2421                 .elevator_merge_req_fn =        cfq_merged_requests,
2422                 .elevator_allow_merge_fn =      cfq_allow_merge,
2423                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2424                 .elevator_add_req_fn =          cfq_insert_request,
2425                 .elevator_activate_req_fn =     cfq_activate_request,
2426                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2427                 .elevator_queue_empty_fn =      cfq_queue_empty,
2428                 .elevator_completed_req_fn =    cfq_completed_request,
2429                 .elevator_former_req_fn =       elv_rb_former_request,
2430                 .elevator_latter_req_fn =       elv_rb_latter_request,
2431                 .elevator_set_req_fn =          cfq_set_request,
2432                 .elevator_put_req_fn =          cfq_put_request,
2433                 .elevator_may_queue_fn =        cfq_may_queue,
2434                 .elevator_init_fn =             cfq_init_queue,
2435                 .elevator_exit_fn =             cfq_exit_queue,
2436                 .trim =                         cfq_free_io_context,
2437         },
2438         .elevator_attrs =       cfq_attrs,
2439         .elevator_name =        "cfq",
2440         .elevator_owner =       THIS_MODULE,
2441 };
2442
2443 static int __init cfq_init(void)
2444 {
2445         /*
2446          * could be 0 on HZ < 1000 setups
2447          */
2448         if (!cfq_slice_async)
2449                 cfq_slice_async = 1;
2450         if (!cfq_slice_idle)
2451                 cfq_slice_idle = 1;
2452
2453         if (cfq_slab_setup())
2454                 return -ENOMEM;
2455
2456         elv_register(&iosched_cfq);
2457
2458         return 0;
2459 }
2460
2461 static void __exit cfq_exit(void)
2462 {
2463         DECLARE_COMPLETION_ONSTACK(all_gone);
2464         elv_unregister(&iosched_cfq);
2465         ioc_gone = &all_gone;
2466         /* ioc_gone's update must be visible before reading ioc_count */
2467         smp_wmb();
2468
2469         /*
2470          * this also protects us from entering cfq_slab_kill() with
2471          * pending RCU callbacks
2472          */
2473         if (elv_ioc_count_read(ioc_count))
2474                 wait_for_completion(&all_gone);
2475         cfq_slab_kill();
2476 }
2477
2478 module_init(cfq_init);
2479 module_exit(cfq_exit);
2480
2481 MODULE_AUTHOR("Jens Axboe");
2482 MODULE_LICENSE("GPL");
2483 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");