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