cfq: change the meaning of the cfqq_coop flag
[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, cfq_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 process-grouping structure
75  */
76 struct cfq_queue {
77         /* reference count */
78         atomic_t ref;
79         /* various state flags, see below */
80         unsigned int flags;
81         /* parent cfq_data */
82         struct cfq_data *cfqd;
83         /* service_tree member */
84         struct rb_node rb_node;
85         /* service_tree key */
86         unsigned long rb_key;
87         /* prio tree member */
88         struct rb_node p_node;
89         /* prio tree root we belong to, if any */
90         struct rb_root *p_root;
91         /* sorted list of pending requests */
92         struct rb_root sort_list;
93         /* if fifo isn't expired, next request to serve */
94         struct request *next_rq;
95         /* requests queued in sort_list */
96         int queued[2];
97         /* currently allocated requests */
98         int allocated[2];
99         /* fifo list of requests in sort_list */
100         struct list_head fifo;
101
102         unsigned long slice_end;
103         long slice_resid;
104         unsigned int slice_dispatch;
105
106         /* pending metadata requests */
107         int meta_pending;
108         /* number of requests that are on the dispatch list or inside driver */
109         int dispatched;
110
111         /* io prio of this group */
112         unsigned short ioprio, org_ioprio;
113         unsigned short ioprio_class, org_ioprio_class;
114
115         unsigned int seek_samples;
116         u64 seek_total;
117         sector_t seek_mean;
118         sector_t last_request_pos;
119
120         pid_t pid;
121
122         struct cfq_queue *new_cfqq;
123 };
124
125 /*
126  * Per block device queue structure
127  */
128 struct cfq_data {
129         struct request_queue *queue;
130
131         /*
132          * rr list of queues with requests and the count of them
133          */
134         struct cfq_rb_root service_tree;
135
136         /*
137          * Each priority tree is sorted by next_request position.  These
138          * trees are used when determining if two or more queues are
139          * interleaving requests (see cfq_close_cooperator).
140          */
141         struct rb_root prio_trees[CFQ_PRIO_LISTS];
142
143         unsigned int busy_queues;
144
145         int rq_in_driver[2];
146         int sync_flight;
147
148         /*
149          * queue-depth detection
150          */
151         int rq_queued;
152         int hw_tag;
153         int hw_tag_samples;
154         int rq_in_driver_peak;
155
156         /*
157          * idle window management
158          */
159         struct timer_list idle_slice_timer;
160         struct work_struct unplug_work;
161
162         struct cfq_queue *active_queue;
163         struct cfq_io_context *active_cic;
164
165         /*
166          * async queue for each priority case
167          */
168         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
169         struct cfq_queue *async_idle_cfqq;
170
171         sector_t last_position;
172
173         /*
174          * tunables, see top of file
175          */
176         unsigned int cfq_quantum;
177         unsigned int cfq_fifo_expire[2];
178         unsigned int cfq_back_penalty;
179         unsigned int cfq_back_max;
180         unsigned int cfq_slice[2];
181         unsigned int cfq_slice_async_rq;
182         unsigned int cfq_slice_idle;
183         unsigned int cfq_latency;
184
185         struct list_head cic_list;
186
187         /*
188          * Fallback dummy cfqq for extreme OOM conditions
189          */
190         struct cfq_queue oom_cfqq;
191
192         unsigned long last_end_sync_rq;
193 };
194
195 enum cfqq_state_flags {
196         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
197         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
198         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
199         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
200         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
201         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
202         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
203         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
204         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
205         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
206 };
207
208 #define CFQ_CFQQ_FNS(name)                                              \
209 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
210 {                                                                       \
211         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
212 }                                                                       \
213 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
214 {                                                                       \
215         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
216 }                                                                       \
217 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
218 {                                                                       \
219         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
220 }
221
222 CFQ_CFQQ_FNS(on_rr);
223 CFQ_CFQQ_FNS(wait_request);
224 CFQ_CFQQ_FNS(must_dispatch);
225 CFQ_CFQQ_FNS(must_alloc_slice);
226 CFQ_CFQQ_FNS(fifo_expire);
227 CFQ_CFQQ_FNS(idle_window);
228 CFQ_CFQQ_FNS(prio_changed);
229 CFQ_CFQQ_FNS(slice_new);
230 CFQ_CFQQ_FNS(sync);
231 CFQ_CFQQ_FNS(coop);
232 #undef CFQ_CFQQ_FNS
233
234 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
235         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
236 #define cfq_log(cfqd, fmt, args...)     \
237         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
238
239 static void cfq_dispatch_insert(struct request_queue *, struct request *);
240 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
241                                        struct io_context *, gfp_t);
242 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
243                                                 struct io_context *);
244
245 static inline int rq_in_driver(struct cfq_data *cfqd)
246 {
247         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
248 }
249
250 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
251                                             bool is_sync)
252 {
253         return cic->cfqq[is_sync];
254 }
255
256 static inline void cic_set_cfqq(struct cfq_io_context *cic,
257                                 struct cfq_queue *cfqq, bool is_sync)
258 {
259         cic->cfqq[is_sync] = cfqq;
260 }
261
262 /*
263  * We regard a request as SYNC, if it's either a read or has the SYNC bit
264  * set (in which case it could also be direct WRITE).
265  */
266 static inline bool cfq_bio_sync(struct bio *bio)
267 {
268         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
269 }
270
271 /*
272  * scheduler run of queue, if there are requests pending and no one in the
273  * driver that will restart queueing
274  */
275 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
276 {
277         if (cfqd->busy_queues) {
278                 cfq_log(cfqd, "schedule dispatch");
279                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
280         }
281 }
282
283 static int cfq_queue_empty(struct request_queue *q)
284 {
285         struct cfq_data *cfqd = q->elevator->elevator_data;
286
287         return !cfqd->busy_queues;
288 }
289
290 /*
291  * Scale schedule slice based on io priority. Use the sync time slice only
292  * if a queue is marked sync and has sync io queued. A sync queue with async
293  * io only, should not get full sync slice length.
294  */
295 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
296                                  unsigned short prio)
297 {
298         const int base_slice = cfqd->cfq_slice[sync];
299
300         WARN_ON(prio >= IOPRIO_BE_NR);
301
302         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
303 }
304
305 static inline int
306 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
307 {
308         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
309 }
310
311 static inline void
312 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
313 {
314         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
315         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
316 }
317
318 /*
319  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
320  * isn't valid until the first request from the dispatch is activated
321  * and the slice time set.
322  */
323 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
324 {
325         if (cfq_cfqq_slice_new(cfqq))
326                 return 0;
327         if (time_before(jiffies, cfqq->slice_end))
328                 return 0;
329
330         return 1;
331 }
332
333 /*
334  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
335  * We choose the request that is closest to the head right now. Distance
336  * behind the head is penalized and only allowed to a certain extent.
337  */
338 static struct request *
339 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
340 {
341         sector_t last, s1, s2, d1 = 0, d2 = 0;
342         unsigned long back_max;
343 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
344 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
345         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
346
347         if (rq1 == NULL || rq1 == rq2)
348                 return rq2;
349         if (rq2 == NULL)
350                 return rq1;
351
352         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
353                 return rq1;
354         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
355                 return rq2;
356         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
357                 return rq1;
358         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
359                 return rq2;
360
361         s1 = blk_rq_pos(rq1);
362         s2 = blk_rq_pos(rq2);
363
364         last = cfqd->last_position;
365
366         /*
367          * by definition, 1KiB is 2 sectors
368          */
369         back_max = cfqd->cfq_back_max * 2;
370
371         /*
372          * Strict one way elevator _except_ in the case where we allow
373          * short backward seeks which are biased as twice the cost of a
374          * similar forward seek.
375          */
376         if (s1 >= last)
377                 d1 = s1 - last;
378         else if (s1 + back_max >= last)
379                 d1 = (last - s1) * cfqd->cfq_back_penalty;
380         else
381                 wrap |= CFQ_RQ1_WRAP;
382
383         if (s2 >= last)
384                 d2 = s2 - last;
385         else if (s2 + back_max >= last)
386                 d2 = (last - s2) * cfqd->cfq_back_penalty;
387         else
388                 wrap |= CFQ_RQ2_WRAP;
389
390         /* Found required data */
391
392         /*
393          * By doing switch() on the bit mask "wrap" we avoid having to
394          * check two variables for all permutations: --> faster!
395          */
396         switch (wrap) {
397         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
398                 if (d1 < d2)
399                         return rq1;
400                 else if (d2 < d1)
401                         return rq2;
402                 else {
403                         if (s1 >= s2)
404                                 return rq1;
405                         else
406                                 return rq2;
407                 }
408
409         case CFQ_RQ2_WRAP:
410                 return rq1;
411         case CFQ_RQ1_WRAP:
412                 return rq2;
413         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
414         default:
415                 /*
416                  * Since both rqs are wrapped,
417                  * start with the one that's further behind head
418                  * (--> only *one* back seek required),
419                  * since back seek takes more time than forward.
420                  */
421                 if (s1 <= s2)
422                         return rq1;
423                 else
424                         return rq2;
425         }
426 }
427
428 /*
429  * The below is leftmost cache rbtree addon
430  */
431 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
432 {
433         if (!root->left)
434                 root->left = rb_first(&root->rb);
435
436         if (root->left)
437                 return rb_entry(root->left, struct cfq_queue, rb_node);
438
439         return NULL;
440 }
441
442 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
443 {
444         rb_erase(n, root);
445         RB_CLEAR_NODE(n);
446 }
447
448 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
449 {
450         if (root->left == n)
451                 root->left = NULL;
452         rb_erase_init(n, &root->rb);
453 }
454
455 /*
456  * would be nice to take fifo expire time into account as well
457  */
458 static struct request *
459 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
460                   struct request *last)
461 {
462         struct rb_node *rbnext = rb_next(&last->rb_node);
463         struct rb_node *rbprev = rb_prev(&last->rb_node);
464         struct request *next = NULL, *prev = NULL;
465
466         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
467
468         if (rbprev)
469                 prev = rb_entry_rq(rbprev);
470
471         if (rbnext)
472                 next = rb_entry_rq(rbnext);
473         else {
474                 rbnext = rb_first(&cfqq->sort_list);
475                 if (rbnext && rbnext != &last->rb_node)
476                         next = rb_entry_rq(rbnext);
477         }
478
479         return cfq_choose_req(cfqd, next, prev);
480 }
481
482 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
483                                       struct cfq_queue *cfqq)
484 {
485         /*
486          * just an approximation, should be ok.
487          */
488         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
489                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
490 }
491
492 /*
493  * The cfqd->service_tree holds all pending cfq_queue's that have
494  * requests waiting to be processed. It is sorted in the order that
495  * we will service the queues.
496  */
497 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
498                                  bool add_front)
499 {
500         struct rb_node **p, *parent;
501         struct cfq_queue *__cfqq;
502         unsigned long rb_key;
503         int left;
504
505         if (cfq_class_idle(cfqq)) {
506                 rb_key = CFQ_IDLE_DELAY;
507                 parent = rb_last(&cfqd->service_tree.rb);
508                 if (parent && parent != &cfqq->rb_node) {
509                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
510                         rb_key += __cfqq->rb_key;
511                 } else
512                         rb_key += jiffies;
513         } else if (!add_front) {
514                 /*
515                  * Get our rb key offset. Subtract any residual slice
516                  * value carried from last service. A negative resid
517                  * count indicates slice overrun, and this should position
518                  * the next service time further away in the tree.
519                  */
520                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
521                 rb_key -= cfqq->slice_resid;
522                 cfqq->slice_resid = 0;
523         } else {
524                 rb_key = -HZ;
525                 __cfqq = cfq_rb_first(&cfqd->service_tree);
526                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
527         }
528
529         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
530                 /*
531                  * same position, nothing more to do
532                  */
533                 if (rb_key == cfqq->rb_key)
534                         return;
535
536                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
537         }
538
539         left = 1;
540         parent = NULL;
541         p = &cfqd->service_tree.rb.rb_node;
542         while (*p) {
543                 struct rb_node **n;
544
545                 parent = *p;
546                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
547
548                 /*
549                  * sort RT queues first, we always want to give
550                  * preference to them. IDLE queues goes to the back.
551                  * after that, sort on the next service time.
552                  */
553                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
554                         n = &(*p)->rb_left;
555                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
556                         n = &(*p)->rb_right;
557                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
558                         n = &(*p)->rb_left;
559                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
560                         n = &(*p)->rb_right;
561                 else if (time_before(rb_key, __cfqq->rb_key))
562                         n = &(*p)->rb_left;
563                 else
564                         n = &(*p)->rb_right;
565
566                 if (n == &(*p)->rb_right)
567                         left = 0;
568
569                 p = n;
570         }
571
572         if (left)
573                 cfqd->service_tree.left = &cfqq->rb_node;
574
575         cfqq->rb_key = rb_key;
576         rb_link_node(&cfqq->rb_node, parent, p);
577         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
578 }
579
580 static struct cfq_queue *
581 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
582                      sector_t sector, struct rb_node **ret_parent,
583                      struct rb_node ***rb_link)
584 {
585         struct rb_node **p, *parent;
586         struct cfq_queue *cfqq = NULL;
587
588         parent = NULL;
589         p = &root->rb_node;
590         while (*p) {
591                 struct rb_node **n;
592
593                 parent = *p;
594                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
595
596                 /*
597                  * Sort strictly based on sector.  Smallest to the left,
598                  * largest to the right.
599                  */
600                 if (sector > blk_rq_pos(cfqq->next_rq))
601                         n = &(*p)->rb_right;
602                 else if (sector < blk_rq_pos(cfqq->next_rq))
603                         n = &(*p)->rb_left;
604                 else
605                         break;
606                 p = n;
607                 cfqq = NULL;
608         }
609
610         *ret_parent = parent;
611         if (rb_link)
612                 *rb_link = p;
613         return cfqq;
614 }
615
616 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
617 {
618         struct rb_node **p, *parent;
619         struct cfq_queue *__cfqq;
620
621         if (cfqq->p_root) {
622                 rb_erase(&cfqq->p_node, cfqq->p_root);
623                 cfqq->p_root = NULL;
624         }
625
626         if (cfq_class_idle(cfqq))
627                 return;
628         if (!cfqq->next_rq)
629                 return;
630
631         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
632         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
633                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
634         if (!__cfqq) {
635                 rb_link_node(&cfqq->p_node, parent, p);
636                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
637         } else
638                 cfqq->p_root = NULL;
639 }
640
641 /*
642  * Update cfqq's position in the service tree.
643  */
644 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
645 {
646         /*
647          * Resorting requires the cfqq to be on the RR list already.
648          */
649         if (cfq_cfqq_on_rr(cfqq)) {
650                 cfq_service_tree_add(cfqd, cfqq, 0);
651                 cfq_prio_tree_add(cfqd, cfqq);
652         }
653 }
654
655 /*
656  * add to busy list of queues for service, trying to be fair in ordering
657  * the pending list according to last request service
658  */
659 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
660 {
661         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
662         BUG_ON(cfq_cfqq_on_rr(cfqq));
663         cfq_mark_cfqq_on_rr(cfqq);
664         cfqd->busy_queues++;
665
666         cfq_resort_rr_list(cfqd, cfqq);
667 }
668
669 /*
670  * Called when the cfqq no longer has requests pending, remove it from
671  * the service tree.
672  */
673 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
674 {
675         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
676         BUG_ON(!cfq_cfqq_on_rr(cfqq));
677         cfq_clear_cfqq_on_rr(cfqq);
678
679         if (!RB_EMPTY_NODE(&cfqq->rb_node))
680                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
681         if (cfqq->p_root) {
682                 rb_erase(&cfqq->p_node, cfqq->p_root);
683                 cfqq->p_root = NULL;
684         }
685
686         BUG_ON(!cfqd->busy_queues);
687         cfqd->busy_queues--;
688 }
689
690 /*
691  * rb tree support functions
692  */
693 static void cfq_del_rq_rb(struct request *rq)
694 {
695         struct cfq_queue *cfqq = RQ_CFQQ(rq);
696         struct cfq_data *cfqd = cfqq->cfqd;
697         const int sync = rq_is_sync(rq);
698
699         BUG_ON(!cfqq->queued[sync]);
700         cfqq->queued[sync]--;
701
702         elv_rb_del(&cfqq->sort_list, rq);
703
704         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
705                 cfq_del_cfqq_rr(cfqd, cfqq);
706 }
707
708 static void cfq_add_rq_rb(struct request *rq)
709 {
710         struct cfq_queue *cfqq = RQ_CFQQ(rq);
711         struct cfq_data *cfqd = cfqq->cfqd;
712         struct request *__alias, *prev;
713
714         cfqq->queued[rq_is_sync(rq)]++;
715
716         /*
717          * looks a little odd, but the first insert might return an alias.
718          * if that happens, put the alias on the dispatch list
719          */
720         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
721                 cfq_dispatch_insert(cfqd->queue, __alias);
722
723         if (!cfq_cfqq_on_rr(cfqq))
724                 cfq_add_cfqq_rr(cfqd, cfqq);
725
726         /*
727          * check if this request is a better next-serve candidate
728          */
729         prev = cfqq->next_rq;
730         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
731
732         /*
733          * adjust priority tree position, if ->next_rq changes
734          */
735         if (prev != cfqq->next_rq)
736                 cfq_prio_tree_add(cfqd, cfqq);
737
738         BUG_ON(!cfqq->next_rq);
739 }
740
741 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
742 {
743         elv_rb_del(&cfqq->sort_list, rq);
744         cfqq->queued[rq_is_sync(rq)]--;
745         cfq_add_rq_rb(rq);
746 }
747
748 static struct request *
749 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
750 {
751         struct task_struct *tsk = current;
752         struct cfq_io_context *cic;
753         struct cfq_queue *cfqq;
754
755         cic = cfq_cic_lookup(cfqd, tsk->io_context);
756         if (!cic)
757                 return NULL;
758
759         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
760         if (cfqq) {
761                 sector_t sector = bio->bi_sector + bio_sectors(bio);
762
763                 return elv_rb_find(&cfqq->sort_list, sector);
764         }
765
766         return NULL;
767 }
768
769 static void cfq_activate_request(struct request_queue *q, struct request *rq)
770 {
771         struct cfq_data *cfqd = q->elevator->elevator_data;
772
773         cfqd->rq_in_driver[rq_is_sync(rq)]++;
774         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
775                                                 rq_in_driver(cfqd));
776
777         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
778 }
779
780 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
781 {
782         struct cfq_data *cfqd = q->elevator->elevator_data;
783         const int sync = rq_is_sync(rq);
784
785         WARN_ON(!cfqd->rq_in_driver[sync]);
786         cfqd->rq_in_driver[sync]--;
787         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
788                                                 rq_in_driver(cfqd));
789 }
790
791 static void cfq_remove_request(struct request *rq)
792 {
793         struct cfq_queue *cfqq = RQ_CFQQ(rq);
794
795         if (cfqq->next_rq == rq)
796                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
797
798         list_del_init(&rq->queuelist);
799         cfq_del_rq_rb(rq);
800
801         cfqq->cfqd->rq_queued--;
802         if (rq_is_meta(rq)) {
803                 WARN_ON(!cfqq->meta_pending);
804                 cfqq->meta_pending--;
805         }
806 }
807
808 static int cfq_merge(struct request_queue *q, struct request **req,
809                      struct bio *bio)
810 {
811         struct cfq_data *cfqd = q->elevator->elevator_data;
812         struct request *__rq;
813
814         __rq = cfq_find_rq_fmerge(cfqd, bio);
815         if (__rq && elv_rq_merge_ok(__rq, bio)) {
816                 *req = __rq;
817                 return ELEVATOR_FRONT_MERGE;
818         }
819
820         return ELEVATOR_NO_MERGE;
821 }
822
823 static void cfq_merged_request(struct request_queue *q, struct request *req,
824                                int type)
825 {
826         if (type == ELEVATOR_FRONT_MERGE) {
827                 struct cfq_queue *cfqq = RQ_CFQQ(req);
828
829                 cfq_reposition_rq_rb(cfqq, req);
830         }
831 }
832
833 static void
834 cfq_merged_requests(struct request_queue *q, struct request *rq,
835                     struct request *next)
836 {
837         /*
838          * reposition in fifo if next is older than rq
839          */
840         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
841             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
842                 list_move(&rq->queuelist, &next->queuelist);
843                 rq_set_fifo_time(rq, rq_fifo_time(next));
844         }
845
846         cfq_remove_request(next);
847 }
848
849 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
850                            struct bio *bio)
851 {
852         struct cfq_data *cfqd = q->elevator->elevator_data;
853         struct cfq_io_context *cic;
854         struct cfq_queue *cfqq;
855
856         /*
857          * Disallow merge of a sync bio into an async request.
858          */
859         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
860                 return false;
861
862         /*
863          * Lookup the cfqq that this bio will be queued with. Allow
864          * merge only if rq is queued there.
865          */
866         cic = cfq_cic_lookup(cfqd, current->io_context);
867         if (!cic)
868                 return false;
869
870         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
871         return cfqq == RQ_CFQQ(rq);
872 }
873
874 static void __cfq_set_active_queue(struct cfq_data *cfqd,
875                                    struct cfq_queue *cfqq)
876 {
877         if (cfqq) {
878                 cfq_log_cfqq(cfqd, cfqq, "set_active");
879                 cfqq->slice_end = 0;
880                 cfqq->slice_dispatch = 0;
881
882                 cfq_clear_cfqq_wait_request(cfqq);
883                 cfq_clear_cfqq_must_dispatch(cfqq);
884                 cfq_clear_cfqq_must_alloc_slice(cfqq);
885                 cfq_clear_cfqq_fifo_expire(cfqq);
886                 cfq_mark_cfqq_slice_new(cfqq);
887
888                 del_timer(&cfqd->idle_slice_timer);
889         }
890
891         cfqd->active_queue = cfqq;
892 }
893
894 /*
895  * current cfqq expired its slice (or was too idle), select new one
896  */
897 static void
898 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
899                     bool timed_out)
900 {
901         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
902
903         if (cfq_cfqq_wait_request(cfqq))
904                 del_timer(&cfqd->idle_slice_timer);
905
906         cfq_clear_cfqq_wait_request(cfqq);
907
908         /*
909          * store what was left of this slice, if the queue idled/timed out
910          */
911         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
912                 cfqq->slice_resid = cfqq->slice_end - jiffies;
913                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
914         }
915
916         cfq_resort_rr_list(cfqd, cfqq);
917
918         if (cfqq == cfqd->active_queue)
919                 cfqd->active_queue = NULL;
920
921         if (cfqd->active_cic) {
922                 put_io_context(cfqd->active_cic->ioc);
923                 cfqd->active_cic = NULL;
924         }
925 }
926
927 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
928 {
929         struct cfq_queue *cfqq = cfqd->active_queue;
930
931         if (cfqq)
932                 __cfq_slice_expired(cfqd, cfqq, timed_out);
933 }
934
935 /*
936  * Get next queue for service. Unless we have a queue preemption,
937  * we'll simply select the first cfqq in the service tree.
938  */
939 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
940 {
941         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
942                 return NULL;
943
944         return cfq_rb_first(&cfqd->service_tree);
945 }
946
947 /*
948  * Get and set a new active queue for service.
949  */
950 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
951                                               struct cfq_queue *cfqq)
952 {
953         if (!cfqq)
954                 cfqq = cfq_get_next_queue(cfqd);
955
956         __cfq_set_active_queue(cfqd, cfqq);
957         return cfqq;
958 }
959
960 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
961                                           struct request *rq)
962 {
963         if (blk_rq_pos(rq) >= cfqd->last_position)
964                 return blk_rq_pos(rq) - cfqd->last_position;
965         else
966                 return cfqd->last_position - blk_rq_pos(rq);
967 }
968
969 #define CFQQ_SEEK_THR           8 * 1024
970 #define CFQQ_SEEKY(cfqq)        ((cfqq)->seek_mean > CFQQ_SEEK_THR)
971
972 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
973                                struct request *rq)
974 {
975         sector_t sdist = cfqq->seek_mean;
976
977         if (!sample_valid(cfqq->seek_samples))
978                 sdist = CFQQ_SEEK_THR;
979
980         return cfq_dist_from_last(cfqd, rq) <= sdist;
981 }
982
983 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
984                                     struct cfq_queue *cur_cfqq)
985 {
986         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
987         struct rb_node *parent, *node;
988         struct cfq_queue *__cfqq;
989         sector_t sector = cfqd->last_position;
990
991         if (RB_EMPTY_ROOT(root))
992                 return NULL;
993
994         /*
995          * First, if we find a request starting at the end of the last
996          * request, choose it.
997          */
998         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
999         if (__cfqq)
1000                 return __cfqq;
1001
1002         /*
1003          * If the exact sector wasn't found, the parent of the NULL leaf
1004          * will contain the closest sector.
1005          */
1006         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1007         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1008                 return __cfqq;
1009
1010         if (blk_rq_pos(__cfqq->next_rq) < sector)
1011                 node = rb_next(&__cfqq->p_node);
1012         else
1013                 node = rb_prev(&__cfqq->p_node);
1014         if (!node)
1015                 return NULL;
1016
1017         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1018         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1019                 return __cfqq;
1020
1021         return NULL;
1022 }
1023
1024 /*
1025  * cfqd - obvious
1026  * cur_cfqq - passed in so that we don't decide that the current queue is
1027  *            closely cooperating with itself.
1028  *
1029  * So, basically we're assuming that that cur_cfqq has dispatched at least
1030  * one request, and that cfqd->last_position reflects a position on the disk
1031  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1032  * assumption.
1033  */
1034 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1035                                               struct cfq_queue *cur_cfqq)
1036 {
1037         struct cfq_queue *cfqq;
1038
1039         /*
1040          * We should notice if some of the queues are cooperating, eg
1041          * working closely on the same area of the disk. In that case,
1042          * we can group them together and don't waste time idling.
1043          */
1044         cfqq = cfqq_close(cfqd, cur_cfqq);
1045         if (!cfqq)
1046                 return NULL;
1047
1048         /*
1049          * It only makes sense to merge sync queues.
1050          */
1051         if (!cfq_cfqq_sync(cfqq))
1052                 return NULL;
1053
1054         return cfqq;
1055 }
1056
1057 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1058 {
1059         struct cfq_queue *cfqq = cfqd->active_queue;
1060         struct cfq_io_context *cic;
1061         unsigned long sl;
1062
1063         /*
1064          * SSD device without seek penalty, disable idling. But only do so
1065          * for devices that support queuing, otherwise we still have a problem
1066          * with sync vs async workloads.
1067          */
1068         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1069                 return;
1070
1071         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1072         WARN_ON(cfq_cfqq_slice_new(cfqq));
1073
1074         /*
1075          * idle is disabled, either manually or by past process history
1076          */
1077         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1078                 return;
1079
1080         /*
1081          * still requests with the driver, don't idle
1082          */
1083         if (rq_in_driver(cfqd))
1084                 return;
1085
1086         /*
1087          * task has exited, don't wait
1088          */
1089         cic = cfqd->active_cic;
1090         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1091                 return;
1092
1093         /*
1094          * If our average think time is larger than the remaining time
1095          * slice, then don't idle. This avoids overrunning the allotted
1096          * time slice.
1097          */
1098         if (sample_valid(cic->ttime_samples) &&
1099             (cfqq->slice_end - jiffies < cic->ttime_mean))
1100                 return;
1101
1102         cfq_mark_cfqq_wait_request(cfqq);
1103
1104         /*
1105          * we don't want to idle for seeks, but we do want to allow
1106          * fair distribution of slice time for a process doing back-to-back
1107          * seeks. so allow a little bit of time for him to submit a new rq
1108          */
1109         sl = cfqd->cfq_slice_idle;
1110         if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
1111                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1112
1113         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1114         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1115 }
1116
1117 /*
1118  * Move request from internal lists to the request queue dispatch list.
1119  */
1120 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1121 {
1122         struct cfq_data *cfqd = q->elevator->elevator_data;
1123         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1124
1125         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1126
1127         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1128         cfq_remove_request(rq);
1129         cfqq->dispatched++;
1130         elv_dispatch_sort(q, rq);
1131
1132         if (cfq_cfqq_sync(cfqq))
1133                 cfqd->sync_flight++;
1134 }
1135
1136 /*
1137  * return expired entry, or NULL to just start from scratch in rbtree
1138  */
1139 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1140 {
1141         struct request *rq = NULL;
1142
1143         if (cfq_cfqq_fifo_expire(cfqq))
1144                 return NULL;
1145
1146         cfq_mark_cfqq_fifo_expire(cfqq);
1147
1148         if (list_empty(&cfqq->fifo))
1149                 return NULL;
1150
1151         rq = rq_entry_fifo(cfqq->fifo.next);
1152         if (time_before(jiffies, rq_fifo_time(rq)))
1153                 rq = NULL;
1154
1155         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1156         return rq;
1157 }
1158
1159 static inline int
1160 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1161 {
1162         const int base_rq = cfqd->cfq_slice_async_rq;
1163
1164         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1165
1166         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1167 }
1168
1169 /*
1170  * Must be called with the queue_lock held.
1171  */
1172 static int cfqq_process_refs(struct cfq_queue *cfqq)
1173 {
1174         int process_refs, io_refs;
1175
1176         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1177         process_refs = atomic_read(&cfqq->ref) - io_refs;
1178         BUG_ON(process_refs < 0);
1179         return process_refs;
1180 }
1181
1182 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1183 {
1184         int process_refs;
1185         struct cfq_queue *__cfqq;
1186
1187         /* Avoid a circular list and skip interim queue merges */
1188         while ((__cfqq = new_cfqq->new_cfqq)) {
1189                 if (__cfqq == cfqq)
1190                         return;
1191                 new_cfqq = __cfqq;
1192         }
1193
1194         process_refs = cfqq_process_refs(cfqq);
1195         /*
1196          * If the process for the cfqq has gone away, there is no
1197          * sense in merging the queues.
1198          */
1199         if (process_refs == 0)
1200                 return;
1201
1202         cfqq->new_cfqq = new_cfqq;
1203         atomic_add(process_refs, &new_cfqq->ref);
1204 }
1205
1206 /*
1207  * Select a queue for service. If we have a current active queue,
1208  * check whether to continue servicing it, or retrieve and set a new one.
1209  */
1210 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1211 {
1212         struct cfq_queue *cfqq, *new_cfqq = NULL;
1213
1214         cfqq = cfqd->active_queue;
1215         if (!cfqq)
1216                 goto new_queue;
1217
1218         /*
1219          * The active queue has run out of time, expire it and select new.
1220          */
1221         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1222                 goto expire;
1223
1224         /*
1225          * The active queue has requests and isn't expired, allow it to
1226          * dispatch.
1227          */
1228         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1229                 goto keep_queue;
1230
1231         /*
1232          * If another queue has a request waiting within our mean seek
1233          * distance, let it run.  The expire code will check for close
1234          * cooperators and put the close queue at the front of the service
1235          * tree.  If possible, merge the expiring queue with the new cfqq.
1236          */
1237         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1238         if (new_cfqq) {
1239                 if (!cfqq->new_cfqq)
1240                         cfq_setup_merge(cfqq, new_cfqq);
1241                 goto expire;
1242         }
1243
1244         /*
1245          * No requests pending. If the active queue still has requests in
1246          * flight or is idling for a new request, allow either of these
1247          * conditions to happen (or time out) before selecting a new queue.
1248          */
1249         if (timer_pending(&cfqd->idle_slice_timer) ||
1250             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1251                 cfqq = NULL;
1252                 goto keep_queue;
1253         }
1254
1255 expire:
1256         cfq_slice_expired(cfqd, 0);
1257 new_queue:
1258         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1259 keep_queue:
1260         return cfqq;
1261 }
1262
1263 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1264 {
1265         int dispatched = 0;
1266
1267         while (cfqq->next_rq) {
1268                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1269                 dispatched++;
1270         }
1271
1272         BUG_ON(!list_empty(&cfqq->fifo));
1273         return dispatched;
1274 }
1275
1276 /*
1277  * Drain our current requests. Used for barriers and when switching
1278  * io schedulers on-the-fly.
1279  */
1280 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1281 {
1282         struct cfq_queue *cfqq;
1283         int dispatched = 0;
1284
1285         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1286                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1287
1288         cfq_slice_expired(cfqd, 0);
1289
1290         BUG_ON(cfqd->busy_queues);
1291
1292         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1293         return dispatched;
1294 }
1295
1296 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1297 {
1298         unsigned int max_dispatch;
1299
1300         /*
1301          * Drain async requests before we start sync IO
1302          */
1303         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1304                 return false;
1305
1306         /*
1307          * If this is an async queue and we have sync IO in flight, let it wait
1308          */
1309         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1310                 return false;
1311
1312         max_dispatch = cfqd->cfq_quantum;
1313         if (cfq_class_idle(cfqq))
1314                 max_dispatch = 1;
1315
1316         /*
1317          * Does this cfqq already have too much IO in flight?
1318          */
1319         if (cfqq->dispatched >= max_dispatch) {
1320                 /*
1321                  * idle queue must always only have a single IO in flight
1322                  */
1323                 if (cfq_class_idle(cfqq))
1324                         return false;
1325
1326                 /*
1327                  * We have other queues, don't allow more IO from this one
1328                  */
1329                 if (cfqd->busy_queues > 1)
1330                         return false;
1331
1332                 /*
1333                  * Sole queue user, allow bigger slice
1334                  */
1335                 max_dispatch *= 4;
1336         }
1337
1338         /*
1339          * Async queues must wait a bit before being allowed dispatch.
1340          * We also ramp up the dispatch depth gradually for async IO,
1341          * based on the last sync IO we serviced
1342          */
1343         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1344                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1345                 unsigned int depth;
1346
1347                 depth = last_sync / cfqd->cfq_slice[1];
1348                 if (!depth && !cfqq->dispatched)
1349                         depth = 1;
1350                 if (depth < max_dispatch)
1351                         max_dispatch = depth;
1352         }
1353
1354         /*
1355          * If we're below the current max, allow a dispatch
1356          */
1357         return cfqq->dispatched < max_dispatch;
1358 }
1359
1360 /*
1361  * Dispatch a request from cfqq, moving them to the request queue
1362  * dispatch list.
1363  */
1364 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1365 {
1366         struct request *rq;
1367
1368         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1369
1370         if (!cfq_may_dispatch(cfqd, cfqq))
1371                 return false;
1372
1373         /*
1374          * follow expired path, else get first next available
1375          */
1376         rq = cfq_check_fifo(cfqq);
1377         if (!rq)
1378                 rq = cfqq->next_rq;
1379
1380         /*
1381          * insert request into driver dispatch list
1382          */
1383         cfq_dispatch_insert(cfqd->queue, rq);
1384
1385         if (!cfqd->active_cic) {
1386                 struct cfq_io_context *cic = RQ_CIC(rq);
1387
1388                 atomic_long_inc(&cic->ioc->refcount);
1389                 cfqd->active_cic = cic;
1390         }
1391
1392         return true;
1393 }
1394
1395 /*
1396  * Find the cfqq that we need to service and move a request from that to the
1397  * dispatch list
1398  */
1399 static int cfq_dispatch_requests(struct request_queue *q, int force)
1400 {
1401         struct cfq_data *cfqd = q->elevator->elevator_data;
1402         struct cfq_queue *cfqq;
1403
1404         if (!cfqd->busy_queues)
1405                 return 0;
1406
1407         if (unlikely(force))
1408                 return cfq_forced_dispatch(cfqd);
1409
1410         cfqq = cfq_select_queue(cfqd);
1411         if (!cfqq)
1412                 return 0;
1413
1414         /*
1415          * Dispatch a request from this cfqq, if it is allowed
1416          */
1417         if (!cfq_dispatch_request(cfqd, cfqq))
1418                 return 0;
1419
1420         cfqq->slice_dispatch++;
1421         cfq_clear_cfqq_must_dispatch(cfqq);
1422
1423         /*
1424          * expire an async queue immediately if it has used up its slice. idle
1425          * queue always expire after 1 dispatch round.
1426          */
1427         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1428             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1429             cfq_class_idle(cfqq))) {
1430                 cfqq->slice_end = jiffies + 1;
1431                 cfq_slice_expired(cfqd, 0);
1432         }
1433
1434         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1435         return 1;
1436 }
1437
1438 /*
1439  * task holds one reference to the queue, dropped when task exits. each rq
1440  * in-flight on this queue also holds a reference, dropped when rq is freed.
1441  *
1442  * queue lock must be held here.
1443  */
1444 static void cfq_put_queue(struct cfq_queue *cfqq)
1445 {
1446         struct cfq_data *cfqd = cfqq->cfqd;
1447
1448         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1449
1450         if (!atomic_dec_and_test(&cfqq->ref))
1451                 return;
1452
1453         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1454         BUG_ON(rb_first(&cfqq->sort_list));
1455         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1456         BUG_ON(cfq_cfqq_on_rr(cfqq));
1457
1458         if (unlikely(cfqd->active_queue == cfqq)) {
1459                 __cfq_slice_expired(cfqd, cfqq, 0);
1460                 cfq_schedule_dispatch(cfqd);
1461         }
1462
1463         kmem_cache_free(cfq_pool, cfqq);
1464 }
1465
1466 /*
1467  * Must always be called with the rcu_read_lock() held
1468  */
1469 static void
1470 __call_for_each_cic(struct io_context *ioc,
1471                     void (*func)(struct io_context *, struct cfq_io_context *))
1472 {
1473         struct cfq_io_context *cic;
1474         struct hlist_node *n;
1475
1476         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1477                 func(ioc, cic);
1478 }
1479
1480 /*
1481  * Call func for each cic attached to this ioc.
1482  */
1483 static void
1484 call_for_each_cic(struct io_context *ioc,
1485                   void (*func)(struct io_context *, struct cfq_io_context *))
1486 {
1487         rcu_read_lock();
1488         __call_for_each_cic(ioc, func);
1489         rcu_read_unlock();
1490 }
1491
1492 static void cfq_cic_free_rcu(struct rcu_head *head)
1493 {
1494         struct cfq_io_context *cic;
1495
1496         cic = container_of(head, struct cfq_io_context, rcu_head);
1497
1498         kmem_cache_free(cfq_ioc_pool, cic);
1499         elv_ioc_count_dec(cfq_ioc_count);
1500
1501         if (ioc_gone) {
1502                 /*
1503                  * CFQ scheduler is exiting, grab exit lock and check
1504                  * the pending io context count. If it hits zero,
1505                  * complete ioc_gone and set it back to NULL
1506                  */
1507                 spin_lock(&ioc_gone_lock);
1508                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1509                         complete(ioc_gone);
1510                         ioc_gone = NULL;
1511                 }
1512                 spin_unlock(&ioc_gone_lock);
1513         }
1514 }
1515
1516 static void cfq_cic_free(struct cfq_io_context *cic)
1517 {
1518         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1519 }
1520
1521 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1522 {
1523         unsigned long flags;
1524
1525         BUG_ON(!cic->dead_key);
1526
1527         spin_lock_irqsave(&ioc->lock, flags);
1528         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1529         hlist_del_rcu(&cic->cic_list);
1530         spin_unlock_irqrestore(&ioc->lock, flags);
1531
1532         cfq_cic_free(cic);
1533 }
1534
1535 /*
1536  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1537  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1538  * and ->trim() which is called with the task lock held
1539  */
1540 static void cfq_free_io_context(struct io_context *ioc)
1541 {
1542         /*
1543          * ioc->refcount is zero here, or we are called from elv_unregister(),
1544          * so no more cic's are allowed to be linked into this ioc.  So it
1545          * should be ok to iterate over the known list, we will see all cic's
1546          * since no new ones are added.
1547          */
1548         __call_for_each_cic(ioc, cic_free_func);
1549 }
1550
1551 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1552 {
1553         struct cfq_queue *__cfqq, *next;
1554
1555         if (unlikely(cfqq == cfqd->active_queue)) {
1556                 __cfq_slice_expired(cfqd, cfqq, 0);
1557                 cfq_schedule_dispatch(cfqd);
1558         }
1559
1560         /*
1561          * If this queue was scheduled to merge with another queue, be
1562          * sure to drop the reference taken on that queue (and others in
1563          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
1564          */
1565         __cfqq = cfqq->new_cfqq;
1566         while (__cfqq) {
1567                 if (__cfqq == cfqq) {
1568                         WARN(1, "cfqq->new_cfqq loop detected\n");
1569                         break;
1570                 }
1571                 next = __cfqq->new_cfqq;
1572                 cfq_put_queue(__cfqq);
1573                 __cfqq = next;
1574         }
1575
1576         cfq_put_queue(cfqq);
1577 }
1578
1579 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1580                                          struct cfq_io_context *cic)
1581 {
1582         struct io_context *ioc = cic->ioc;
1583
1584         list_del_init(&cic->queue_list);
1585
1586         /*
1587          * Make sure key == NULL is seen for dead queues
1588          */
1589         smp_wmb();
1590         cic->dead_key = (unsigned long) cic->key;
1591         cic->key = NULL;
1592
1593         if (ioc->ioc_data == cic)
1594                 rcu_assign_pointer(ioc->ioc_data, NULL);
1595
1596         if (cic->cfqq[BLK_RW_ASYNC]) {
1597                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1598                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1599         }
1600
1601         if (cic->cfqq[BLK_RW_SYNC]) {
1602                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1603                 cic->cfqq[BLK_RW_SYNC] = NULL;
1604         }
1605 }
1606
1607 static void cfq_exit_single_io_context(struct io_context *ioc,
1608                                        struct cfq_io_context *cic)
1609 {
1610         struct cfq_data *cfqd = cic->key;
1611
1612         if (cfqd) {
1613                 struct request_queue *q = cfqd->queue;
1614                 unsigned long flags;
1615
1616                 spin_lock_irqsave(q->queue_lock, flags);
1617
1618                 /*
1619                  * Ensure we get a fresh copy of the ->key to prevent
1620                  * race between exiting task and queue
1621                  */
1622                 smp_read_barrier_depends();
1623                 if (cic->key)
1624                         __cfq_exit_single_io_context(cfqd, cic);
1625
1626                 spin_unlock_irqrestore(q->queue_lock, flags);
1627         }
1628 }
1629
1630 /*
1631  * The process that ioc belongs to has exited, we need to clean up
1632  * and put the internal structures we have that belongs to that process.
1633  */
1634 static void cfq_exit_io_context(struct io_context *ioc)
1635 {
1636         call_for_each_cic(ioc, cfq_exit_single_io_context);
1637 }
1638
1639 static struct cfq_io_context *
1640 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1641 {
1642         struct cfq_io_context *cic;
1643
1644         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1645                                                         cfqd->queue->node);
1646         if (cic) {
1647                 cic->last_end_request = jiffies;
1648                 INIT_LIST_HEAD(&cic->queue_list);
1649                 INIT_HLIST_NODE(&cic->cic_list);
1650                 cic->dtor = cfq_free_io_context;
1651                 cic->exit = cfq_exit_io_context;
1652                 elv_ioc_count_inc(cfq_ioc_count);
1653         }
1654
1655         return cic;
1656 }
1657
1658 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1659 {
1660         struct task_struct *tsk = current;
1661         int ioprio_class;
1662
1663         if (!cfq_cfqq_prio_changed(cfqq))
1664                 return;
1665
1666         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1667         switch (ioprio_class) {
1668         default:
1669                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1670         case IOPRIO_CLASS_NONE:
1671                 /*
1672                  * no prio set, inherit CPU scheduling settings
1673                  */
1674                 cfqq->ioprio = task_nice_ioprio(tsk);
1675                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1676                 break;
1677         case IOPRIO_CLASS_RT:
1678                 cfqq->ioprio = task_ioprio(ioc);
1679                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1680                 break;
1681         case IOPRIO_CLASS_BE:
1682                 cfqq->ioprio = task_ioprio(ioc);
1683                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1684                 break;
1685         case IOPRIO_CLASS_IDLE:
1686                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1687                 cfqq->ioprio = 7;
1688                 cfq_clear_cfqq_idle_window(cfqq);
1689                 break;
1690         }
1691
1692         /*
1693          * keep track of original prio settings in case we have to temporarily
1694          * elevate the priority of this queue
1695          */
1696         cfqq->org_ioprio = cfqq->ioprio;
1697         cfqq->org_ioprio_class = cfqq->ioprio_class;
1698         cfq_clear_cfqq_prio_changed(cfqq);
1699 }
1700
1701 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1702 {
1703         struct cfq_data *cfqd = cic->key;
1704         struct cfq_queue *cfqq;
1705         unsigned long flags;
1706
1707         if (unlikely(!cfqd))
1708                 return;
1709
1710         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1711
1712         cfqq = cic->cfqq[BLK_RW_ASYNC];
1713         if (cfqq) {
1714                 struct cfq_queue *new_cfqq;
1715                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1716                                                 GFP_ATOMIC);
1717                 if (new_cfqq) {
1718                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1719                         cfq_put_queue(cfqq);
1720                 }
1721         }
1722
1723         cfqq = cic->cfqq[BLK_RW_SYNC];
1724         if (cfqq)
1725                 cfq_mark_cfqq_prio_changed(cfqq);
1726
1727         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1728 }
1729
1730 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1731 {
1732         call_for_each_cic(ioc, changed_ioprio);
1733         ioc->ioprio_changed = 0;
1734 }
1735
1736 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1737                           pid_t pid, bool is_sync)
1738 {
1739         RB_CLEAR_NODE(&cfqq->rb_node);
1740         RB_CLEAR_NODE(&cfqq->p_node);
1741         INIT_LIST_HEAD(&cfqq->fifo);
1742
1743         atomic_set(&cfqq->ref, 0);
1744         cfqq->cfqd = cfqd;
1745
1746         cfq_mark_cfqq_prio_changed(cfqq);
1747
1748         if (is_sync) {
1749                 if (!cfq_class_idle(cfqq))
1750                         cfq_mark_cfqq_idle_window(cfqq);
1751                 cfq_mark_cfqq_sync(cfqq);
1752         }
1753         cfqq->pid = pid;
1754 }
1755
1756 static struct cfq_queue *
1757 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1758                      struct io_context *ioc, gfp_t gfp_mask)
1759 {
1760         struct cfq_queue *cfqq, *new_cfqq = NULL;
1761         struct cfq_io_context *cic;
1762
1763 retry:
1764         cic = cfq_cic_lookup(cfqd, ioc);
1765         /* cic always exists here */
1766         cfqq = cic_to_cfqq(cic, is_sync);
1767
1768         /*
1769          * Always try a new alloc if we fell back to the OOM cfqq
1770          * originally, since it should just be a temporary situation.
1771          */
1772         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1773                 cfqq = NULL;
1774                 if (new_cfqq) {
1775                         cfqq = new_cfqq;
1776                         new_cfqq = NULL;
1777                 } else if (gfp_mask & __GFP_WAIT) {
1778                         spin_unlock_irq(cfqd->queue->queue_lock);
1779                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1780                                         gfp_mask | __GFP_ZERO,
1781                                         cfqd->queue->node);
1782                         spin_lock_irq(cfqd->queue->queue_lock);
1783                         if (new_cfqq)
1784                                 goto retry;
1785                 } else {
1786                         cfqq = kmem_cache_alloc_node(cfq_pool,
1787                                         gfp_mask | __GFP_ZERO,
1788                                         cfqd->queue->node);
1789                 }
1790
1791                 if (cfqq) {
1792                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1793                         cfq_init_prio_data(cfqq, ioc);
1794                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1795                 } else
1796                         cfqq = &cfqd->oom_cfqq;
1797         }
1798
1799         if (new_cfqq)
1800                 kmem_cache_free(cfq_pool, new_cfqq);
1801
1802         return cfqq;
1803 }
1804
1805 static struct cfq_queue **
1806 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1807 {
1808         switch (ioprio_class) {
1809         case IOPRIO_CLASS_RT:
1810                 return &cfqd->async_cfqq[0][ioprio];
1811         case IOPRIO_CLASS_BE:
1812                 return &cfqd->async_cfqq[1][ioprio];
1813         case IOPRIO_CLASS_IDLE:
1814                 return &cfqd->async_idle_cfqq;
1815         default:
1816                 BUG();
1817         }
1818 }
1819
1820 static struct cfq_queue *
1821 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1822               gfp_t gfp_mask)
1823 {
1824         const int ioprio = task_ioprio(ioc);
1825         const int ioprio_class = task_ioprio_class(ioc);
1826         struct cfq_queue **async_cfqq = NULL;
1827         struct cfq_queue *cfqq = NULL;
1828
1829         if (!is_sync) {
1830                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1831                 cfqq = *async_cfqq;
1832         }
1833
1834         if (!cfqq)
1835                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1836
1837         /*
1838          * pin the queue now that it's allocated, scheduler exit will prune it
1839          */
1840         if (!is_sync && !(*async_cfqq)) {
1841                 atomic_inc(&cfqq->ref);
1842                 *async_cfqq = cfqq;
1843         }
1844
1845         atomic_inc(&cfqq->ref);
1846         return cfqq;
1847 }
1848
1849 /*
1850  * We drop cfq io contexts lazily, so we may find a dead one.
1851  */
1852 static void
1853 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1854                   struct cfq_io_context *cic)
1855 {
1856         unsigned long flags;
1857
1858         WARN_ON(!list_empty(&cic->queue_list));
1859
1860         spin_lock_irqsave(&ioc->lock, flags);
1861
1862         BUG_ON(ioc->ioc_data == cic);
1863
1864         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1865         hlist_del_rcu(&cic->cic_list);
1866         spin_unlock_irqrestore(&ioc->lock, flags);
1867
1868         cfq_cic_free(cic);
1869 }
1870
1871 static struct cfq_io_context *
1872 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1873 {
1874         struct cfq_io_context *cic;
1875         unsigned long flags;
1876         void *k;
1877
1878         if (unlikely(!ioc))
1879                 return NULL;
1880
1881         rcu_read_lock();
1882
1883         /*
1884          * we maintain a last-hit cache, to avoid browsing over the tree
1885          */
1886         cic = rcu_dereference(ioc->ioc_data);
1887         if (cic && cic->key == cfqd) {
1888                 rcu_read_unlock();
1889                 return cic;
1890         }
1891
1892         do {
1893                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1894                 rcu_read_unlock();
1895                 if (!cic)
1896                         break;
1897                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1898                 k = cic->key;
1899                 if (unlikely(!k)) {
1900                         cfq_drop_dead_cic(cfqd, ioc, cic);
1901                         rcu_read_lock();
1902                         continue;
1903                 }
1904
1905                 spin_lock_irqsave(&ioc->lock, flags);
1906                 rcu_assign_pointer(ioc->ioc_data, cic);
1907                 spin_unlock_irqrestore(&ioc->lock, flags);
1908                 break;
1909         } while (1);
1910
1911         return cic;
1912 }
1913
1914 /*
1915  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1916  * the process specific cfq io context when entered from the block layer.
1917  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1918  */
1919 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1920                         struct cfq_io_context *cic, gfp_t gfp_mask)
1921 {
1922         unsigned long flags;
1923         int ret;
1924
1925         ret = radix_tree_preload(gfp_mask);
1926         if (!ret) {
1927                 cic->ioc = ioc;
1928                 cic->key = cfqd;
1929
1930                 spin_lock_irqsave(&ioc->lock, flags);
1931                 ret = radix_tree_insert(&ioc->radix_root,
1932                                                 (unsigned long) cfqd, cic);
1933                 if (!ret)
1934                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1935                 spin_unlock_irqrestore(&ioc->lock, flags);
1936
1937                 radix_tree_preload_end();
1938
1939                 if (!ret) {
1940                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1941                         list_add(&cic->queue_list, &cfqd->cic_list);
1942                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1943                 }
1944         }
1945
1946         if (ret)
1947                 printk(KERN_ERR "cfq: cic link failed!\n");
1948
1949         return ret;
1950 }
1951
1952 /*
1953  * Setup general io context and cfq io context. There can be several cfq
1954  * io contexts per general io context, if this process is doing io to more
1955  * than one device managed by cfq.
1956  */
1957 static struct cfq_io_context *
1958 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1959 {
1960         struct io_context *ioc = NULL;
1961         struct cfq_io_context *cic;
1962
1963         might_sleep_if(gfp_mask & __GFP_WAIT);
1964
1965         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1966         if (!ioc)
1967                 return NULL;
1968
1969         cic = cfq_cic_lookup(cfqd, ioc);
1970         if (cic)
1971                 goto out;
1972
1973         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1974         if (cic == NULL)
1975                 goto err;
1976
1977         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1978                 goto err_free;
1979
1980 out:
1981         smp_read_barrier_depends();
1982         if (unlikely(ioc->ioprio_changed))
1983                 cfq_ioc_set_ioprio(ioc);
1984
1985         return cic;
1986 err_free:
1987         cfq_cic_free(cic);
1988 err:
1989         put_io_context(ioc);
1990         return NULL;
1991 }
1992
1993 static void
1994 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1995 {
1996         unsigned long elapsed = jiffies - cic->last_end_request;
1997         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1998
1999         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2000         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2001         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2002 }
2003
2004 static void
2005 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2006                        struct request *rq)
2007 {
2008         sector_t sdist;
2009         u64 total;
2010
2011         if (!cfqq->last_request_pos)
2012                 sdist = 0;
2013         else if (cfqq->last_request_pos < blk_rq_pos(rq))
2014                 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2015         else
2016                 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2017
2018         /*
2019          * Don't allow the seek distance to get too large from the
2020          * odd fragment, pagein, etc
2021          */
2022         if (cfqq->seek_samples <= 60) /* second&third seek */
2023                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2024         else
2025                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2026
2027         cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2028         cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2029         total = cfqq->seek_total + (cfqq->seek_samples/2);
2030         do_div(total, cfqq->seek_samples);
2031         cfqq->seek_mean = (sector_t)total;
2032 }
2033
2034 /*
2035  * Disable idle window if the process thinks too long or seeks so much that
2036  * it doesn't matter
2037  */
2038 static void
2039 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2040                        struct cfq_io_context *cic)
2041 {
2042         int old_idle, enable_idle;
2043
2044         /*
2045          * Don't idle for async or idle io prio class
2046          */
2047         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2048                 return;
2049
2050         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2051
2052         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2053             (!cfqd->cfq_latency && cfqd->hw_tag && CFQQ_SEEKY(cfqq)))
2054                 enable_idle = 0;
2055         else if (sample_valid(cic->ttime_samples)) {
2056                 unsigned int slice_idle = cfqd->cfq_slice_idle;
2057                 if (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq))
2058                         slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
2059                 if (cic->ttime_mean > slice_idle)
2060                         enable_idle = 0;
2061                 else
2062                         enable_idle = 1;
2063         }
2064
2065         if (old_idle != enable_idle) {
2066                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2067                 if (enable_idle)
2068                         cfq_mark_cfqq_idle_window(cfqq);
2069                 else
2070                         cfq_clear_cfqq_idle_window(cfqq);
2071         }
2072 }
2073
2074 /*
2075  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2076  * no or if we aren't sure, a 1 will cause a preempt.
2077  */
2078 static bool
2079 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2080                    struct request *rq)
2081 {
2082         struct cfq_queue *cfqq;
2083
2084         cfqq = cfqd->active_queue;
2085         if (!cfqq)
2086                 return false;
2087
2088         if (cfq_slice_used(cfqq))
2089                 return true;
2090
2091         if (cfq_class_idle(new_cfqq))
2092                 return false;
2093
2094         if (cfq_class_idle(cfqq))
2095                 return true;
2096
2097         /*
2098          * if the new request is sync, but the currently running queue is
2099          * not, let the sync request have priority.
2100          */
2101         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2102                 return true;
2103
2104         /*
2105          * So both queues are sync. Let the new request get disk time if
2106          * it's a metadata request and the current queue is doing regular IO.
2107          */
2108         if (rq_is_meta(rq) && !cfqq->meta_pending)
2109                 return false;
2110
2111         /*
2112          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2113          */
2114         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2115                 return true;
2116
2117         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2118                 return false;
2119
2120         /*
2121          * if this request is as-good as one we would expect from the
2122          * current cfqq, let it preempt
2123          */
2124         if (cfq_rq_close(cfqd, cfqq, rq))
2125                 return true;
2126
2127         return false;
2128 }
2129
2130 /*
2131  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2132  * let it have half of its nominal slice.
2133  */
2134 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2135 {
2136         cfq_log_cfqq(cfqd, cfqq, "preempt");
2137         cfq_slice_expired(cfqd, 1);
2138
2139         /*
2140          * Put the new queue at the front of the of the current list,
2141          * so we know that it will be selected next.
2142          */
2143         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2144
2145         cfq_service_tree_add(cfqd, cfqq, 1);
2146
2147         cfqq->slice_end = 0;
2148         cfq_mark_cfqq_slice_new(cfqq);
2149 }
2150
2151 /*
2152  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2153  * something we should do about it
2154  */
2155 static void
2156 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2157                 struct request *rq)
2158 {
2159         struct cfq_io_context *cic = RQ_CIC(rq);
2160
2161         cfqd->rq_queued++;
2162         if (rq_is_meta(rq))
2163                 cfqq->meta_pending++;
2164
2165         cfq_update_io_thinktime(cfqd, cic);
2166         cfq_update_io_seektime(cfqd, cfqq, rq);
2167         cfq_update_idle_window(cfqd, cfqq, cic);
2168
2169         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2170
2171         if (cfqq == cfqd->active_queue) {
2172                 /*
2173                  * Remember that we saw a request from this process, but
2174                  * don't start queuing just yet. Otherwise we risk seeing lots
2175                  * of tiny requests, because we disrupt the normal plugging
2176                  * and merging. If the request is already larger than a single
2177                  * page, let it rip immediately. For that case we assume that
2178                  * merging is already done. Ditto for a busy system that
2179                  * has other work pending, don't risk delaying until the
2180                  * idle timer unplug to continue working.
2181                  */
2182                 if (cfq_cfqq_wait_request(cfqq)) {
2183                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2184                             cfqd->busy_queues > 1) {
2185                                 del_timer(&cfqd->idle_slice_timer);
2186                         __blk_run_queue(cfqd->queue);
2187                         }
2188                         cfq_mark_cfqq_must_dispatch(cfqq);
2189                 }
2190         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2191                 /*
2192                  * not the active queue - expire current slice if it is
2193                  * idle and has expired it's mean thinktime or this new queue
2194                  * has some old slice time left and is of higher priority or
2195                  * this new queue is RT and the current one is BE
2196                  */
2197                 cfq_preempt_queue(cfqd, cfqq);
2198                 __blk_run_queue(cfqd->queue);
2199         }
2200 }
2201
2202 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2203 {
2204         struct cfq_data *cfqd = q->elevator->elevator_data;
2205         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2206
2207         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2208         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2209
2210         cfq_add_rq_rb(rq);
2211
2212         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2213         list_add_tail(&rq->queuelist, &cfqq->fifo);
2214
2215         cfq_rq_enqueued(cfqd, cfqq, rq);
2216 }
2217
2218 /*
2219  * Update hw_tag based on peak queue depth over 50 samples under
2220  * sufficient load.
2221  */
2222 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2223 {
2224         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2225                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2226
2227         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2228             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2229                 return;
2230
2231         if (cfqd->hw_tag_samples++ < 50)
2232                 return;
2233
2234         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2235                 cfqd->hw_tag = 1;
2236         else
2237                 cfqd->hw_tag = 0;
2238
2239         cfqd->hw_tag_samples = 0;
2240         cfqd->rq_in_driver_peak = 0;
2241 }
2242
2243 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2244 {
2245         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2246         struct cfq_data *cfqd = cfqq->cfqd;
2247         const int sync = rq_is_sync(rq);
2248         unsigned long now;
2249
2250         now = jiffies;
2251         cfq_log_cfqq(cfqd, cfqq, "complete");
2252
2253         cfq_update_hw_tag(cfqd);
2254
2255         WARN_ON(!cfqd->rq_in_driver[sync]);
2256         WARN_ON(!cfqq->dispatched);
2257         cfqd->rq_in_driver[sync]--;
2258         cfqq->dispatched--;
2259
2260         if (cfq_cfqq_sync(cfqq))
2261                 cfqd->sync_flight--;
2262
2263         if (sync) {
2264                 RQ_CIC(rq)->last_end_request = now;
2265                 cfqd->last_end_sync_rq = now;
2266         }
2267
2268         /*
2269          * If this is the active queue, check if it needs to be expired,
2270          * or if we want to idle in case it has no pending requests.
2271          */
2272         if (cfqd->active_queue == cfqq) {
2273                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2274
2275                 if (cfq_cfqq_slice_new(cfqq)) {
2276                         cfq_set_prio_slice(cfqd, cfqq);
2277                         cfq_clear_cfqq_slice_new(cfqq);
2278                 }
2279                 /*
2280                  * If there are no requests waiting in this queue, and
2281                  * there are other queues ready to issue requests, AND
2282                  * those other queues are issuing requests within our
2283                  * mean seek distance, give them a chance to run instead
2284                  * of idling.
2285                  */
2286                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2287                         cfq_slice_expired(cfqd, 1);
2288                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2289                          sync && !rq_noidle(rq))
2290                         cfq_arm_slice_timer(cfqd);
2291         }
2292
2293         if (!rq_in_driver(cfqd))
2294                 cfq_schedule_dispatch(cfqd);
2295 }
2296
2297 /*
2298  * we temporarily boost lower priority queues if they are holding fs exclusive
2299  * resources. they are boosted to normal prio (CLASS_BE/4)
2300  */
2301 static void cfq_prio_boost(struct cfq_queue *cfqq)
2302 {
2303         if (has_fs_excl()) {
2304                 /*
2305                  * boost idle prio on transactions that would lock out other
2306                  * users of the filesystem
2307                  */
2308                 if (cfq_class_idle(cfqq))
2309                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2310                 if (cfqq->ioprio > IOPRIO_NORM)
2311                         cfqq->ioprio = IOPRIO_NORM;
2312         } else {
2313                 /*
2314                  * check if we need to unboost the queue
2315                  */
2316                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2317                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2318                 if (cfqq->ioprio != cfqq->org_ioprio)
2319                         cfqq->ioprio = cfqq->org_ioprio;
2320         }
2321 }
2322
2323 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2324 {
2325         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2326                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2327                 return ELV_MQUEUE_MUST;
2328         }
2329
2330         return ELV_MQUEUE_MAY;
2331 }
2332
2333 static int cfq_may_queue(struct request_queue *q, int rw)
2334 {
2335         struct cfq_data *cfqd = q->elevator->elevator_data;
2336         struct task_struct *tsk = current;
2337         struct cfq_io_context *cic;
2338         struct cfq_queue *cfqq;
2339
2340         /*
2341          * don't force setup of a queue from here, as a call to may_queue
2342          * does not necessarily imply that a request actually will be queued.
2343          * so just lookup a possibly existing queue, or return 'may queue'
2344          * if that fails
2345          */
2346         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2347         if (!cic)
2348                 return ELV_MQUEUE_MAY;
2349
2350         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2351         if (cfqq) {
2352                 cfq_init_prio_data(cfqq, cic->ioc);
2353                 cfq_prio_boost(cfqq);
2354
2355                 return __cfq_may_queue(cfqq);
2356         }
2357
2358         return ELV_MQUEUE_MAY;
2359 }
2360
2361 /*
2362  * queue lock held here
2363  */
2364 static void cfq_put_request(struct request *rq)
2365 {
2366         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2367
2368         if (cfqq) {
2369                 const int rw = rq_data_dir(rq);
2370
2371                 BUG_ON(!cfqq->allocated[rw]);
2372                 cfqq->allocated[rw]--;
2373
2374                 put_io_context(RQ_CIC(rq)->ioc);
2375
2376                 rq->elevator_private = NULL;
2377                 rq->elevator_private2 = NULL;
2378
2379                 cfq_put_queue(cfqq);
2380         }
2381 }
2382
2383 static struct cfq_queue *
2384 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2385                 struct cfq_queue *cfqq)
2386 {
2387         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2388         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2389         cfq_mark_cfqq_coop(cfqq->new_cfqq);
2390         cfq_put_queue(cfqq);
2391         return cic_to_cfqq(cic, 1);
2392 }
2393
2394 /*
2395  * Allocate cfq data structures associated with this request.
2396  */
2397 static int
2398 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2399 {
2400         struct cfq_data *cfqd = q->elevator->elevator_data;
2401         struct cfq_io_context *cic;
2402         const int rw = rq_data_dir(rq);
2403         const bool is_sync = rq_is_sync(rq);
2404         struct cfq_queue *cfqq;
2405         unsigned long flags;
2406
2407         might_sleep_if(gfp_mask & __GFP_WAIT);
2408
2409         cic = cfq_get_io_context(cfqd, gfp_mask);
2410
2411         spin_lock_irqsave(q->queue_lock, flags);
2412
2413         if (!cic)
2414                 goto queue_fail;
2415
2416         cfqq = cic_to_cfqq(cic, is_sync);
2417         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2418                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2419                 cic_set_cfqq(cic, cfqq, is_sync);
2420         } else {
2421                 /*
2422                  * Check to see if this queue is scheduled to merge with
2423                  * another, closely cooperating queue.  The merging of
2424                  * queues happens here as it must be done in process context.
2425                  * The reference on new_cfqq was taken in merge_cfqqs.
2426                  */
2427                 if (cfqq->new_cfqq)
2428                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2429         }
2430
2431         cfqq->allocated[rw]++;
2432         atomic_inc(&cfqq->ref);
2433
2434         spin_unlock_irqrestore(q->queue_lock, flags);
2435
2436         rq->elevator_private = cic;
2437         rq->elevator_private2 = cfqq;
2438         return 0;
2439
2440 queue_fail:
2441         if (cic)
2442                 put_io_context(cic->ioc);
2443
2444         cfq_schedule_dispatch(cfqd);
2445         spin_unlock_irqrestore(q->queue_lock, flags);
2446         cfq_log(cfqd, "set_request fail");
2447         return 1;
2448 }
2449
2450 static void cfq_kick_queue(struct work_struct *work)
2451 {
2452         struct cfq_data *cfqd =
2453                 container_of(work, struct cfq_data, unplug_work);
2454         struct request_queue *q = cfqd->queue;
2455
2456         spin_lock_irq(q->queue_lock);
2457         __blk_run_queue(cfqd->queue);
2458         spin_unlock_irq(q->queue_lock);
2459 }
2460
2461 /*
2462  * Timer running if the active_queue is currently idling inside its time slice
2463  */
2464 static void cfq_idle_slice_timer(unsigned long data)
2465 {
2466         struct cfq_data *cfqd = (struct cfq_data *) data;
2467         struct cfq_queue *cfqq;
2468         unsigned long flags;
2469         int timed_out = 1;
2470
2471         cfq_log(cfqd, "idle timer fired");
2472
2473         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2474
2475         cfqq = cfqd->active_queue;
2476         if (cfqq) {
2477                 timed_out = 0;
2478
2479                 /*
2480                  * We saw a request before the queue expired, let it through
2481                  */
2482                 if (cfq_cfqq_must_dispatch(cfqq))
2483                         goto out_kick;
2484
2485                 /*
2486                  * expired
2487                  */
2488                 if (cfq_slice_used(cfqq))
2489                         goto expire;
2490
2491                 /*
2492                  * only expire and reinvoke request handler, if there are
2493                  * other queues with pending requests
2494                  */
2495                 if (!cfqd->busy_queues)
2496                         goto out_cont;
2497
2498                 /*
2499                  * not expired and it has a request pending, let it dispatch
2500                  */
2501                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2502                         goto out_kick;
2503         }
2504 expire:
2505         cfq_slice_expired(cfqd, timed_out);
2506 out_kick:
2507         cfq_schedule_dispatch(cfqd);
2508 out_cont:
2509         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2510 }
2511
2512 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2513 {
2514         del_timer_sync(&cfqd->idle_slice_timer);
2515         cancel_work_sync(&cfqd->unplug_work);
2516 }
2517
2518 static void cfq_put_async_queues(struct cfq_data *cfqd)
2519 {
2520         int i;
2521
2522         for (i = 0; i < IOPRIO_BE_NR; i++) {
2523                 if (cfqd->async_cfqq[0][i])
2524                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2525                 if (cfqd->async_cfqq[1][i])
2526                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2527         }
2528
2529         if (cfqd->async_idle_cfqq)
2530                 cfq_put_queue(cfqd->async_idle_cfqq);
2531 }
2532
2533 static void cfq_exit_queue(struct elevator_queue *e)
2534 {
2535         struct cfq_data *cfqd = e->elevator_data;
2536         struct request_queue *q = cfqd->queue;
2537
2538         cfq_shutdown_timer_wq(cfqd);
2539
2540         spin_lock_irq(q->queue_lock);
2541
2542         if (cfqd->active_queue)
2543                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2544
2545         while (!list_empty(&cfqd->cic_list)) {
2546                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2547                                                         struct cfq_io_context,
2548                                                         queue_list);
2549
2550                 __cfq_exit_single_io_context(cfqd, cic);
2551         }
2552
2553         cfq_put_async_queues(cfqd);
2554
2555         spin_unlock_irq(q->queue_lock);
2556
2557         cfq_shutdown_timer_wq(cfqd);
2558
2559         kfree(cfqd);
2560 }
2561
2562 static void *cfq_init_queue(struct request_queue *q)
2563 {
2564         struct cfq_data *cfqd;
2565         int i;
2566
2567         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2568         if (!cfqd)
2569                 return NULL;
2570
2571         cfqd->service_tree = CFQ_RB_ROOT;
2572
2573         /*
2574          * Not strictly needed (since RB_ROOT just clears the node and we
2575          * zeroed cfqd on alloc), but better be safe in case someone decides
2576          * to add magic to the rb code
2577          */
2578         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2579                 cfqd->prio_trees[i] = RB_ROOT;
2580
2581         /*
2582          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2583          * Grab a permanent reference to it, so that the normal code flow
2584          * will not attempt to free it.
2585          */
2586         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2587         atomic_inc(&cfqd->oom_cfqq.ref);
2588
2589         INIT_LIST_HEAD(&cfqd->cic_list);
2590
2591         cfqd->queue = q;
2592
2593         init_timer(&cfqd->idle_slice_timer);
2594         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2595         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2596
2597         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2598
2599         cfqd->cfq_quantum = cfq_quantum;
2600         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2601         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2602         cfqd->cfq_back_max = cfq_back_max;
2603         cfqd->cfq_back_penalty = cfq_back_penalty;
2604         cfqd->cfq_slice[0] = cfq_slice_async;
2605         cfqd->cfq_slice[1] = cfq_slice_sync;
2606         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2607         cfqd->cfq_slice_idle = cfq_slice_idle;
2608         cfqd->cfq_latency = 1;
2609         cfqd->hw_tag = 1;
2610         cfqd->last_end_sync_rq = jiffies;
2611         return cfqd;
2612 }
2613
2614 static void cfq_slab_kill(void)
2615 {
2616         /*
2617          * Caller already ensured that pending RCU callbacks are completed,
2618          * so we should have no busy allocations at this point.
2619          */
2620         if (cfq_pool)
2621                 kmem_cache_destroy(cfq_pool);
2622         if (cfq_ioc_pool)
2623                 kmem_cache_destroy(cfq_ioc_pool);
2624 }
2625
2626 static int __init cfq_slab_setup(void)
2627 {
2628         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2629         if (!cfq_pool)
2630                 goto fail;
2631
2632         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2633         if (!cfq_ioc_pool)
2634                 goto fail;
2635
2636         return 0;
2637 fail:
2638         cfq_slab_kill();
2639         return -ENOMEM;
2640 }
2641
2642 /*
2643  * sysfs parts below -->
2644  */
2645 static ssize_t
2646 cfq_var_show(unsigned int var, char *page)
2647 {
2648         return sprintf(page, "%d\n", var);
2649 }
2650
2651 static ssize_t
2652 cfq_var_store(unsigned int *var, const char *page, size_t count)
2653 {
2654         char *p = (char *) page;
2655
2656         *var = simple_strtoul(p, &p, 10);
2657         return count;
2658 }
2659
2660 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2661 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2662 {                                                                       \
2663         struct cfq_data *cfqd = e->elevator_data;                       \
2664         unsigned int __data = __VAR;                                    \
2665         if (__CONV)                                                     \
2666                 __data = jiffies_to_msecs(__data);                      \
2667         return cfq_var_show(__data, (page));                            \
2668 }
2669 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2670 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2671 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2672 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2673 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2674 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2675 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2676 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2677 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2678 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2679 #undef SHOW_FUNCTION
2680
2681 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2682 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2683 {                                                                       \
2684         struct cfq_data *cfqd = e->elevator_data;                       \
2685         unsigned int __data;                                            \
2686         int ret = cfq_var_store(&__data, (page), count);                \
2687         if (__data < (MIN))                                             \
2688                 __data = (MIN);                                         \
2689         else if (__data > (MAX))                                        \
2690                 __data = (MAX);                                         \
2691         if (__CONV)                                                     \
2692                 *(__PTR) = msecs_to_jiffies(__data);                    \
2693         else                                                            \
2694                 *(__PTR) = __data;                                      \
2695         return ret;                                                     \
2696 }
2697 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2698 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2699                 UINT_MAX, 1);
2700 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2701                 UINT_MAX, 1);
2702 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2703 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2704                 UINT_MAX, 0);
2705 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2706 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2707 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2708 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2709                 UINT_MAX, 0);
2710 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2711 #undef STORE_FUNCTION
2712
2713 #define CFQ_ATTR(name) \
2714         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2715
2716 static struct elv_fs_entry cfq_attrs[] = {
2717         CFQ_ATTR(quantum),
2718         CFQ_ATTR(fifo_expire_sync),
2719         CFQ_ATTR(fifo_expire_async),
2720         CFQ_ATTR(back_seek_max),
2721         CFQ_ATTR(back_seek_penalty),
2722         CFQ_ATTR(slice_sync),
2723         CFQ_ATTR(slice_async),
2724         CFQ_ATTR(slice_async_rq),
2725         CFQ_ATTR(slice_idle),
2726         CFQ_ATTR(low_latency),
2727         __ATTR_NULL
2728 };
2729
2730 static struct elevator_type iosched_cfq = {
2731         .ops = {
2732                 .elevator_merge_fn =            cfq_merge,
2733                 .elevator_merged_fn =           cfq_merged_request,
2734                 .elevator_merge_req_fn =        cfq_merged_requests,
2735                 .elevator_allow_merge_fn =      cfq_allow_merge,
2736                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2737                 .elevator_add_req_fn =          cfq_insert_request,
2738                 .elevator_activate_req_fn =     cfq_activate_request,
2739                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2740                 .elevator_queue_empty_fn =      cfq_queue_empty,
2741                 .elevator_completed_req_fn =    cfq_completed_request,
2742                 .elevator_former_req_fn =       elv_rb_former_request,
2743                 .elevator_latter_req_fn =       elv_rb_latter_request,
2744                 .elevator_set_req_fn =          cfq_set_request,
2745                 .elevator_put_req_fn =          cfq_put_request,
2746                 .elevator_may_queue_fn =        cfq_may_queue,
2747                 .elevator_init_fn =             cfq_init_queue,
2748                 .elevator_exit_fn =             cfq_exit_queue,
2749                 .trim =                         cfq_free_io_context,
2750         },
2751         .elevator_attrs =       cfq_attrs,
2752         .elevator_name =        "cfq",
2753         .elevator_owner =       THIS_MODULE,
2754 };
2755
2756 static int __init cfq_init(void)
2757 {
2758         /*
2759          * could be 0 on HZ < 1000 setups
2760          */
2761         if (!cfq_slice_async)
2762                 cfq_slice_async = 1;
2763         if (!cfq_slice_idle)
2764                 cfq_slice_idle = 1;
2765
2766         if (cfq_slab_setup())
2767                 return -ENOMEM;
2768
2769         elv_register(&iosched_cfq);
2770
2771         return 0;
2772 }
2773
2774 static void __exit cfq_exit(void)
2775 {
2776         DECLARE_COMPLETION_ONSTACK(all_gone);
2777         elv_unregister(&iosched_cfq);
2778         ioc_gone = &all_gone;
2779         /* ioc_gone's update must be visible before reading ioc_count */
2780         smp_wmb();
2781
2782         /*
2783          * this also protects us from entering cfq_slab_kill() with
2784          * pending RCU callbacks
2785          */
2786         if (elv_ioc_count_read(cfq_ioc_count))
2787                 wait_for_completion(&all_gone);
2788         cfq_slab_kill();
2789 }
2790
2791 module_init(cfq_init);
2792 module_exit(cfq_exit);
2793
2794 MODULE_AUTHOR("Jens Axboe");
2795 MODULE_LICENSE("GPL");
2796 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");