ebab60c6be9d73cbe7140f3020cd0aef79cc4b3e
[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         pid_t pid;
116 };
117
118 /*
119  * Per block device queue structure
120  */
121 struct cfq_data {
122         struct request_queue *queue;
123
124         /*
125          * rr list of queues with requests and the count of them
126          */
127         struct cfq_rb_root service_tree;
128
129         /*
130          * Each priority tree is sorted by next_request position.  These
131          * trees are used when determining if two or more queues are
132          * interleaving requests (see cfq_close_cooperator).
133          */
134         struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
136         unsigned int busy_queues;
137
138         int rq_in_driver[2];
139         int sync_flight;
140
141         /*
142          * queue-depth detection
143          */
144         int rq_queued;
145         int hw_tag;
146         int hw_tag_samples;
147         int rq_in_driver_peak;
148
149         /*
150          * idle window management
151          */
152         struct timer_list idle_slice_timer;
153         struct delayed_work unplug_work;
154
155         struct cfq_queue *active_queue;
156         struct cfq_io_context *active_cic;
157
158         /*
159          * async queue for each priority case
160          */
161         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162         struct cfq_queue *async_idle_cfqq;
163
164         sector_t last_position;
165
166         /*
167          * tunables, see top of file
168          */
169         unsigned int cfq_quantum;
170         unsigned int cfq_fifo_expire[2];
171         unsigned int cfq_back_penalty;
172         unsigned int cfq_back_max;
173         unsigned int cfq_slice[2];
174         unsigned int cfq_slice_async_rq;
175         unsigned int cfq_slice_idle;
176         unsigned int cfq_latency;
177
178         struct list_head cic_list;
179
180         /*
181          * Fallback dummy cfqq for extreme OOM conditions
182          */
183         struct cfq_queue oom_cfqq;
184
185         unsigned long last_end_sync_rq;
186 };
187
188 enum cfqq_state_flags {
189         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
190         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
191         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
192         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
193         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
194         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
195         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
196         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
197         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
198         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
199 };
200
201 #define CFQ_CFQQ_FNS(name)                                              \
202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
203 {                                                                       \
204         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
205 }                                                                       \
206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
207 {                                                                       \
208         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
209 }                                                                       \
210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
211 {                                                                       \
212         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
213 }
214
215 CFQ_CFQQ_FNS(on_rr);
216 CFQ_CFQQ_FNS(wait_request);
217 CFQ_CFQQ_FNS(must_dispatch);
218 CFQ_CFQQ_FNS(must_alloc_slice);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
222 CFQ_CFQQ_FNS(slice_new);
223 CFQ_CFQQ_FNS(sync);
224 CFQ_CFQQ_FNS(coop);
225 #undef CFQ_CFQQ_FNS
226
227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
228         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229 #define cfq_log(cfqd, fmt, args...)     \
230         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231
232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
234                                        struct io_context *, gfp_t);
235 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
236                                                 struct io_context *);
237
238 static inline int rq_in_driver(struct cfq_data *cfqd)
239 {
240         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
241 }
242
243 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
244                                             int is_sync)
245 {
246         return cic->cfqq[!!is_sync];
247 }
248
249 static inline void cic_set_cfqq(struct cfq_io_context *cic,
250                                 struct cfq_queue *cfqq, int is_sync)
251 {
252         cic->cfqq[!!is_sync] = cfqq;
253 }
254
255 /*
256  * We regard a request as SYNC, if it's either a read or has the SYNC bit
257  * set (in which case it could also be direct WRITE).
258  */
259 static inline int cfq_bio_sync(struct bio *bio)
260 {
261         if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
262                 return 1;
263
264         return 0;
265 }
266
267 /*
268  * scheduler run of queue, if there are requests pending and no one in the
269  * driver that will restart queueing
270  */
271 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd,
272                                          unsigned long delay)
273 {
274         if (cfqd->busy_queues) {
275                 cfq_log(cfqd, "schedule dispatch");
276                 kblockd_schedule_delayed_work(cfqd->queue, &cfqd->unplug_work,
277                                                 delay);
278         }
279 }
280
281 static int cfq_queue_empty(struct request_queue *q)
282 {
283         struct cfq_data *cfqd = q->elevator->elevator_data;
284
285         return !cfqd->busy_queues;
286 }
287
288 /*
289  * Scale schedule slice based on io priority. Use the sync time slice only
290  * if a queue is marked sync and has sync io queued. A sync queue with async
291  * io only, should not get full sync slice length.
292  */
293 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
294                                  unsigned short prio)
295 {
296         const int base_slice = cfqd->cfq_slice[sync];
297
298         WARN_ON(prio >= IOPRIO_BE_NR);
299
300         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
301 }
302
303 static inline int
304 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
305 {
306         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
307 }
308
309 static inline void
310 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
311 {
312         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
313         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
314 }
315
316 /*
317  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
318  * isn't valid until the first request from the dispatch is activated
319  * and the slice time set.
320  */
321 static inline int cfq_slice_used(struct cfq_queue *cfqq)
322 {
323         if (cfq_cfqq_slice_new(cfqq))
324                 return 0;
325         if (time_before(jiffies, cfqq->slice_end))
326                 return 0;
327
328         return 1;
329 }
330
331 /*
332  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
333  * We choose the request that is closest to the head right now. Distance
334  * behind the head is penalized and only allowed to a certain extent.
335  */
336 static struct request *
337 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
338 {
339         sector_t last, s1, s2, d1 = 0, d2 = 0;
340         unsigned long back_max;
341 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
342 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
343         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
344
345         if (rq1 == NULL || rq1 == rq2)
346                 return rq2;
347         if (rq2 == NULL)
348                 return rq1;
349
350         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
351                 return rq1;
352         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
353                 return rq2;
354         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
355                 return rq1;
356         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
357                 return rq2;
358
359         s1 = blk_rq_pos(rq1);
360         s2 = blk_rq_pos(rq2);
361
362         last = cfqd->last_position;
363
364         /*
365          * by definition, 1KiB is 2 sectors
366          */
367         back_max = cfqd->cfq_back_max * 2;
368
369         /*
370          * Strict one way elevator _except_ in the case where we allow
371          * short backward seeks which are biased as twice the cost of a
372          * similar forward seek.
373          */
374         if (s1 >= last)
375                 d1 = s1 - last;
376         else if (s1 + back_max >= last)
377                 d1 = (last - s1) * cfqd->cfq_back_penalty;
378         else
379                 wrap |= CFQ_RQ1_WRAP;
380
381         if (s2 >= last)
382                 d2 = s2 - last;
383         else if (s2 + back_max >= last)
384                 d2 = (last - s2) * cfqd->cfq_back_penalty;
385         else
386                 wrap |= CFQ_RQ2_WRAP;
387
388         /* Found required data */
389
390         /*
391          * By doing switch() on the bit mask "wrap" we avoid having to
392          * check two variables for all permutations: --> faster!
393          */
394         switch (wrap) {
395         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
396                 if (d1 < d2)
397                         return rq1;
398                 else if (d2 < d1)
399                         return rq2;
400                 else {
401                         if (s1 >= s2)
402                                 return rq1;
403                         else
404                                 return rq2;
405                 }
406
407         case CFQ_RQ2_WRAP:
408                 return rq1;
409         case CFQ_RQ1_WRAP:
410                 return rq2;
411         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
412         default:
413                 /*
414                  * Since both rqs are wrapped,
415                  * start with the one that's further behind head
416                  * (--> only *one* back seek required),
417                  * since back seek takes more time than forward.
418                  */
419                 if (s1 <= s2)
420                         return rq1;
421                 else
422                         return rq2;
423         }
424 }
425
426 /*
427  * The below is leftmost cache rbtree addon
428  */
429 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
430 {
431         if (!root->left)
432                 root->left = rb_first(&root->rb);
433
434         if (root->left)
435                 return rb_entry(root->left, struct cfq_queue, rb_node);
436
437         return NULL;
438 }
439
440 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
441 {
442         rb_erase(n, root);
443         RB_CLEAR_NODE(n);
444 }
445
446 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
447 {
448         if (root->left == n)
449                 root->left = NULL;
450         rb_erase_init(n, &root->rb);
451 }
452
453 /*
454  * would be nice to take fifo expire time into account as well
455  */
456 static struct request *
457 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
458                   struct request *last)
459 {
460         struct rb_node *rbnext = rb_next(&last->rb_node);
461         struct rb_node *rbprev = rb_prev(&last->rb_node);
462         struct request *next = NULL, *prev = NULL;
463
464         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
465
466         if (rbprev)
467                 prev = rb_entry_rq(rbprev);
468
469         if (rbnext)
470                 next = rb_entry_rq(rbnext);
471         else {
472                 rbnext = rb_first(&cfqq->sort_list);
473                 if (rbnext && rbnext != &last->rb_node)
474                         next = rb_entry_rq(rbnext);
475         }
476
477         return cfq_choose_req(cfqd, next, prev);
478 }
479
480 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
481                                       struct cfq_queue *cfqq)
482 {
483         /*
484          * just an approximation, should be ok.
485          */
486         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
487                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
488 }
489
490 /*
491  * The cfqd->service_tree holds all pending cfq_queue's that have
492  * requests waiting to be processed. It is sorted in the order that
493  * we will service the queues.
494  */
495 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
496                                  int add_front)
497 {
498         struct rb_node **p, *parent;
499         struct cfq_queue *__cfqq;
500         unsigned long rb_key;
501         int left;
502
503         if (cfq_class_idle(cfqq)) {
504                 rb_key = CFQ_IDLE_DELAY;
505                 parent = rb_last(&cfqd->service_tree.rb);
506                 if (parent && parent != &cfqq->rb_node) {
507                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
508                         rb_key += __cfqq->rb_key;
509                 } else
510                         rb_key += jiffies;
511         } else if (!add_front) {
512                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
513                 rb_key += cfqq->slice_resid;
514                 cfqq->slice_resid = 0;
515         } else
516                 rb_key = 0;
517
518         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
519                 /*
520                  * same position, nothing more to do
521                  */
522                 if (rb_key == cfqq->rb_key)
523                         return;
524
525                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
526         }
527
528         left = 1;
529         parent = NULL;
530         p = &cfqd->service_tree.rb.rb_node;
531         while (*p) {
532                 struct rb_node **n;
533
534                 parent = *p;
535                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
536
537                 /*
538                  * sort RT queues first, we always want to give
539                  * preference to them. IDLE queues goes to the back.
540                  * after that, sort on the next service time.
541                  */
542                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
543                         n = &(*p)->rb_left;
544                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
545                         n = &(*p)->rb_right;
546                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
547                         n = &(*p)->rb_left;
548                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
549                         n = &(*p)->rb_right;
550                 else if (rb_key < __cfqq->rb_key)
551                         n = &(*p)->rb_left;
552                 else
553                         n = &(*p)->rb_right;
554
555                 if (n == &(*p)->rb_right)
556                         left = 0;
557
558                 p = n;
559         }
560
561         if (left)
562                 cfqd->service_tree.left = &cfqq->rb_node;
563
564         cfqq->rb_key = rb_key;
565         rb_link_node(&cfqq->rb_node, parent, p);
566         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
567 }
568
569 static struct cfq_queue *
570 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
571                      sector_t sector, struct rb_node **ret_parent,
572                      struct rb_node ***rb_link)
573 {
574         struct rb_node **p, *parent;
575         struct cfq_queue *cfqq = NULL;
576
577         parent = NULL;
578         p = &root->rb_node;
579         while (*p) {
580                 struct rb_node **n;
581
582                 parent = *p;
583                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
584
585                 /*
586                  * Sort strictly based on sector.  Smallest to the left,
587                  * largest to the right.
588                  */
589                 if (sector > blk_rq_pos(cfqq->next_rq))
590                         n = &(*p)->rb_right;
591                 else if (sector < blk_rq_pos(cfqq->next_rq))
592                         n = &(*p)->rb_left;
593                 else
594                         break;
595                 p = n;
596                 cfqq = NULL;
597         }
598
599         *ret_parent = parent;
600         if (rb_link)
601                 *rb_link = p;
602         return cfqq;
603 }
604
605 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
606 {
607         struct rb_node **p, *parent;
608         struct cfq_queue *__cfqq;
609
610         if (cfqq->p_root) {
611                 rb_erase(&cfqq->p_node, cfqq->p_root);
612                 cfqq->p_root = NULL;
613         }
614
615         if (cfq_class_idle(cfqq))
616                 return;
617         if (!cfqq->next_rq)
618                 return;
619
620         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
621         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
622                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
623         if (!__cfqq) {
624                 rb_link_node(&cfqq->p_node, parent, p);
625                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
626         } else
627                 cfqq->p_root = NULL;
628 }
629
630 /*
631  * Update cfqq's position in the service tree.
632  */
633 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
634 {
635         /*
636          * Resorting requires the cfqq to be on the RR list already.
637          */
638         if (cfq_cfqq_on_rr(cfqq)) {
639                 cfq_service_tree_add(cfqd, cfqq, 0);
640                 cfq_prio_tree_add(cfqd, cfqq);
641         }
642 }
643
644 /*
645  * add to busy list of queues for service, trying to be fair in ordering
646  * the pending list according to last request service
647  */
648 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
649 {
650         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
651         BUG_ON(cfq_cfqq_on_rr(cfqq));
652         cfq_mark_cfqq_on_rr(cfqq);
653         cfqd->busy_queues++;
654
655         cfq_resort_rr_list(cfqd, cfqq);
656 }
657
658 /*
659  * Called when the cfqq no longer has requests pending, remove it from
660  * the service tree.
661  */
662 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
663 {
664         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
665         BUG_ON(!cfq_cfqq_on_rr(cfqq));
666         cfq_clear_cfqq_on_rr(cfqq);
667
668         if (!RB_EMPTY_NODE(&cfqq->rb_node))
669                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
670         if (cfqq->p_root) {
671                 rb_erase(&cfqq->p_node, cfqq->p_root);
672                 cfqq->p_root = NULL;
673         }
674
675         BUG_ON(!cfqd->busy_queues);
676         cfqd->busy_queues--;
677 }
678
679 /*
680  * rb tree support functions
681  */
682 static void cfq_del_rq_rb(struct request *rq)
683 {
684         struct cfq_queue *cfqq = RQ_CFQQ(rq);
685         struct cfq_data *cfqd = cfqq->cfqd;
686         const int sync = rq_is_sync(rq);
687
688         BUG_ON(!cfqq->queued[sync]);
689         cfqq->queued[sync]--;
690
691         elv_rb_del(&cfqq->sort_list, rq);
692
693         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
694                 cfq_del_cfqq_rr(cfqd, cfqq);
695 }
696
697 static void cfq_add_rq_rb(struct request *rq)
698 {
699         struct cfq_queue *cfqq = RQ_CFQQ(rq);
700         struct cfq_data *cfqd = cfqq->cfqd;
701         struct request *__alias, *prev;
702
703         cfqq->queued[rq_is_sync(rq)]++;
704
705         /*
706          * looks a little odd, but the first insert might return an alias.
707          * if that happens, put the alias on the dispatch list
708          */
709         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
710                 cfq_dispatch_insert(cfqd->queue, __alias);
711
712         if (!cfq_cfqq_on_rr(cfqq))
713                 cfq_add_cfqq_rr(cfqd, cfqq);
714
715         /*
716          * check if this request is a better next-serve candidate
717          */
718         prev = cfqq->next_rq;
719         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
720
721         /*
722          * adjust priority tree position, if ->next_rq changes
723          */
724         if (prev != cfqq->next_rq)
725                 cfq_prio_tree_add(cfqd, cfqq);
726
727         BUG_ON(!cfqq->next_rq);
728 }
729
730 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
731 {
732         elv_rb_del(&cfqq->sort_list, rq);
733         cfqq->queued[rq_is_sync(rq)]--;
734         cfq_add_rq_rb(rq);
735 }
736
737 static struct request *
738 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
739 {
740         struct task_struct *tsk = current;
741         struct cfq_io_context *cic;
742         struct cfq_queue *cfqq;
743
744         cic = cfq_cic_lookup(cfqd, tsk->io_context);
745         if (!cic)
746                 return NULL;
747
748         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
749         if (cfqq) {
750                 sector_t sector = bio->bi_sector + bio_sectors(bio);
751
752                 return elv_rb_find(&cfqq->sort_list, sector);
753         }
754
755         return NULL;
756 }
757
758 static void cfq_activate_request(struct request_queue *q, struct request *rq)
759 {
760         struct cfq_data *cfqd = q->elevator->elevator_data;
761
762         cfqd->rq_in_driver[rq_is_sync(rq)]++;
763         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
764                                                 rq_in_driver(cfqd));
765
766         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
767 }
768
769 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
770 {
771         struct cfq_data *cfqd = q->elevator->elevator_data;
772         const int sync = rq_is_sync(rq);
773
774         WARN_ON(!cfqd->rq_in_driver[sync]);
775         cfqd->rq_in_driver[sync]--;
776         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
777                                                 rq_in_driver(cfqd));
778 }
779
780 static void cfq_remove_request(struct request *rq)
781 {
782         struct cfq_queue *cfqq = RQ_CFQQ(rq);
783
784         if (cfqq->next_rq == rq)
785                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
786
787         list_del_init(&rq->queuelist);
788         cfq_del_rq_rb(rq);
789
790         cfqq->cfqd->rq_queued--;
791         if (rq_is_meta(rq)) {
792                 WARN_ON(!cfqq->meta_pending);
793                 cfqq->meta_pending--;
794         }
795 }
796
797 static int cfq_merge(struct request_queue *q, struct request **req,
798                      struct bio *bio)
799 {
800         struct cfq_data *cfqd = q->elevator->elevator_data;
801         struct request *__rq;
802
803         __rq = cfq_find_rq_fmerge(cfqd, bio);
804         if (__rq && elv_rq_merge_ok(__rq, bio)) {
805                 *req = __rq;
806                 return ELEVATOR_FRONT_MERGE;
807         }
808
809         return ELEVATOR_NO_MERGE;
810 }
811
812 static void cfq_merged_request(struct request_queue *q, struct request *req,
813                                int type)
814 {
815         if (type == ELEVATOR_FRONT_MERGE) {
816                 struct cfq_queue *cfqq = RQ_CFQQ(req);
817
818                 cfq_reposition_rq_rb(cfqq, req);
819         }
820 }
821
822 static void
823 cfq_merged_requests(struct request_queue *q, struct request *rq,
824                     struct request *next)
825 {
826         /*
827          * reposition in fifo if next is older than rq
828          */
829         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
830             time_before(next->start_time, rq->start_time))
831                 list_move(&rq->queuelist, &next->queuelist);
832
833         cfq_remove_request(next);
834 }
835
836 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
837                            struct bio *bio)
838 {
839         struct cfq_data *cfqd = q->elevator->elevator_data;
840         struct cfq_io_context *cic;
841         struct cfq_queue *cfqq;
842
843         /*
844          * Disallow merge of a sync bio into an async request.
845          */
846         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
847                 return 0;
848
849         /*
850          * Lookup the cfqq that this bio will be queued with. Allow
851          * merge only if rq is queued there.
852          */
853         cic = cfq_cic_lookup(cfqd, current->io_context);
854         if (!cic)
855                 return 0;
856
857         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
858         if (cfqq == RQ_CFQQ(rq))
859                 return 1;
860
861         return 0;
862 }
863
864 static void __cfq_set_active_queue(struct cfq_data *cfqd,
865                                    struct cfq_queue *cfqq)
866 {
867         if (cfqq) {
868                 cfq_log_cfqq(cfqd, cfqq, "set_active");
869                 cfqq->slice_end = 0;
870                 cfqq->slice_dispatch = 0;
871
872                 cfq_clear_cfqq_wait_request(cfqq);
873                 cfq_clear_cfqq_must_dispatch(cfqq);
874                 cfq_clear_cfqq_must_alloc_slice(cfqq);
875                 cfq_clear_cfqq_fifo_expire(cfqq);
876                 cfq_mark_cfqq_slice_new(cfqq);
877
878                 del_timer(&cfqd->idle_slice_timer);
879         }
880
881         cfqd->active_queue = cfqq;
882 }
883
884 /*
885  * current cfqq expired its slice (or was too idle), select new one
886  */
887 static void
888 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
889                     int timed_out)
890 {
891         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
892
893         if (cfq_cfqq_wait_request(cfqq))
894                 del_timer(&cfqd->idle_slice_timer);
895
896         cfq_clear_cfqq_wait_request(cfqq);
897
898         /*
899          * store what was left of this slice, if the queue idled/timed out
900          */
901         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
902                 cfqq->slice_resid = cfqq->slice_end - jiffies;
903                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
904         }
905
906         cfq_resort_rr_list(cfqd, cfqq);
907
908         if (cfqq == cfqd->active_queue)
909                 cfqd->active_queue = NULL;
910
911         if (cfqd->active_cic) {
912                 put_io_context(cfqd->active_cic->ioc);
913                 cfqd->active_cic = NULL;
914         }
915 }
916
917 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
918 {
919         struct cfq_queue *cfqq = cfqd->active_queue;
920
921         if (cfqq)
922                 __cfq_slice_expired(cfqd, cfqq, timed_out);
923 }
924
925 /*
926  * Get next queue for service. Unless we have a queue preemption,
927  * we'll simply select the first cfqq in the service tree.
928  */
929 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
930 {
931         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
932                 return NULL;
933
934         return cfq_rb_first(&cfqd->service_tree);
935 }
936
937 /*
938  * Get and set a new active queue for service.
939  */
940 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
941                                               struct cfq_queue *cfqq)
942 {
943         if (!cfqq) {
944                 cfqq = cfq_get_next_queue(cfqd);
945                 if (cfqq)
946                         cfq_clear_cfqq_coop(cfqq);
947         }
948
949         __cfq_set_active_queue(cfqd, cfqq);
950         return cfqq;
951 }
952
953 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
954                                           struct request *rq)
955 {
956         if (blk_rq_pos(rq) >= cfqd->last_position)
957                 return blk_rq_pos(rq) - cfqd->last_position;
958         else
959                 return cfqd->last_position - blk_rq_pos(rq);
960 }
961
962 #define CIC_SEEK_THR    8 * 1024
963 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
964
965 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
966 {
967         struct cfq_io_context *cic = cfqd->active_cic;
968         sector_t sdist = cic->seek_mean;
969
970         if (!sample_valid(cic->seek_samples))
971                 sdist = CIC_SEEK_THR;
972
973         return cfq_dist_from_last(cfqd, rq) <= sdist;
974 }
975
976 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
977                                     struct cfq_queue *cur_cfqq)
978 {
979         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
980         struct rb_node *parent, *node;
981         struct cfq_queue *__cfqq;
982         sector_t sector = cfqd->last_position;
983
984         if (RB_EMPTY_ROOT(root))
985                 return NULL;
986
987         /*
988          * First, if we find a request starting at the end of the last
989          * request, choose it.
990          */
991         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
992         if (__cfqq)
993                 return __cfqq;
994
995         /*
996          * If the exact sector wasn't found, the parent of the NULL leaf
997          * will contain the closest sector.
998          */
999         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1000         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1001                 return __cfqq;
1002
1003         if (blk_rq_pos(__cfqq->next_rq) < sector)
1004                 node = rb_next(&__cfqq->p_node);
1005         else
1006                 node = rb_prev(&__cfqq->p_node);
1007         if (!node)
1008                 return NULL;
1009
1010         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1011         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1012                 return __cfqq;
1013
1014         return NULL;
1015 }
1016
1017 /*
1018  * cfqd - obvious
1019  * cur_cfqq - passed in so that we don't decide that the current queue is
1020  *            closely cooperating with itself.
1021  *
1022  * So, basically we're assuming that that cur_cfqq has dispatched at least
1023  * one request, and that cfqd->last_position reflects a position on the disk
1024  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1025  * assumption.
1026  */
1027 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1028                                               struct cfq_queue *cur_cfqq,
1029                                               int probe)
1030 {
1031         struct cfq_queue *cfqq;
1032
1033         /*
1034          * A valid cfq_io_context is necessary to compare requests against
1035          * the seek_mean of the current cfqq.
1036          */
1037         if (!cfqd->active_cic)
1038                 return NULL;
1039
1040         /*
1041          * We should notice if some of the queues are cooperating, eg
1042          * working closely on the same area of the disk. In that case,
1043          * we can group them together and don't waste time idling.
1044          */
1045         cfqq = cfqq_close(cfqd, cur_cfqq);
1046         if (!cfqq)
1047                 return NULL;
1048
1049         if (cfq_cfqq_coop(cfqq))
1050                 return NULL;
1051
1052         if (!probe)
1053                 cfq_mark_cfqq_coop(cfqq);
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         cfq_mark_cfqq_wait_request(cfqq);
1094
1095         /*
1096          * we don't want to idle for seeks, but we do want to allow
1097          * fair distribution of slice time for a process doing back-to-back
1098          * seeks. so allow a little bit of time for him to submit a new rq
1099          */
1100         sl = cfqd->cfq_slice_idle;
1101         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1102                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1103
1104         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1105         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1106 }
1107
1108 /*
1109  * Move request from internal lists to the request queue dispatch list.
1110  */
1111 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1112 {
1113         struct cfq_data *cfqd = q->elevator->elevator_data;
1114         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1115
1116         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1117
1118         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1119         cfq_remove_request(rq);
1120         cfqq->dispatched++;
1121         elv_dispatch_sort(q, rq);
1122
1123         if (cfq_cfqq_sync(cfqq))
1124                 cfqd->sync_flight++;
1125 }
1126
1127 /*
1128  * return expired entry, or NULL to just start from scratch in rbtree
1129  */
1130 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1131 {
1132         struct cfq_data *cfqd = cfqq->cfqd;
1133         struct request *rq;
1134         int fifo;
1135
1136         if (cfq_cfqq_fifo_expire(cfqq))
1137                 return NULL;
1138
1139         cfq_mark_cfqq_fifo_expire(cfqq);
1140
1141         if (list_empty(&cfqq->fifo))
1142                 return NULL;
1143
1144         fifo = cfq_cfqq_sync(cfqq);
1145         rq = rq_entry_fifo(cfqq->fifo.next);
1146
1147         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1148                 rq = NULL;
1149
1150         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1151         return rq;
1152 }
1153
1154 static inline int
1155 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1156 {
1157         const int base_rq = cfqd->cfq_slice_async_rq;
1158
1159         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1160
1161         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1162 }
1163
1164 /*
1165  * Select a queue for service. If we have a current active queue,
1166  * check whether to continue servicing it, or retrieve and set a new one.
1167  */
1168 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1169 {
1170         struct cfq_queue *cfqq, *new_cfqq = NULL;
1171
1172         cfqq = cfqd->active_queue;
1173         if (!cfqq)
1174                 goto new_queue;
1175
1176         /*
1177          * The active queue has run out of time, expire it and select new.
1178          */
1179         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1180                 goto expire;
1181
1182         /*
1183          * The active queue has requests and isn't expired, allow it to
1184          * dispatch.
1185          */
1186         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1187                 goto keep_queue;
1188
1189         /*
1190          * If another queue has a request waiting within our mean seek
1191          * distance, let it run.  The expire code will check for close
1192          * cooperators and put the close queue at the front of the service
1193          * tree.
1194          */
1195         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1196         if (new_cfqq)
1197                 goto expire;
1198
1199         /*
1200          * No requests pending. If the active queue still has requests in
1201          * flight or is idling for a new request, allow either of these
1202          * conditions to happen (or time out) before selecting a new queue.
1203          */
1204         if (timer_pending(&cfqd->idle_slice_timer) ||
1205             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1206                 cfqq = NULL;
1207                 goto keep_queue;
1208         }
1209
1210 expire:
1211         cfq_slice_expired(cfqd, 0);
1212 new_queue:
1213         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1214 keep_queue:
1215         return cfqq;
1216 }
1217
1218 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1219 {
1220         int dispatched = 0;
1221
1222         while (cfqq->next_rq) {
1223                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1224                 dispatched++;
1225         }
1226
1227         BUG_ON(!list_empty(&cfqq->fifo));
1228         return dispatched;
1229 }
1230
1231 /*
1232  * Drain our current requests. Used for barriers and when switching
1233  * io schedulers on-the-fly.
1234  */
1235 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1236 {
1237         struct cfq_queue *cfqq;
1238         int dispatched = 0;
1239
1240         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1241                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1242
1243         cfq_slice_expired(cfqd, 0);
1244
1245         BUG_ON(cfqd->busy_queues);
1246
1247         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1248         return dispatched;
1249 }
1250
1251 /*
1252  * Dispatch a request from cfqq, moving them to the request queue
1253  * dispatch list.
1254  */
1255 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1256 {
1257         struct request *rq;
1258
1259         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1260
1261         /*
1262          * follow expired path, else get first next available
1263          */
1264         rq = cfq_check_fifo(cfqq);
1265         if (!rq)
1266                 rq = cfqq->next_rq;
1267
1268         /*
1269          * insert request into driver dispatch list
1270          */
1271         cfq_dispatch_insert(cfqd->queue, rq);
1272
1273         if (!cfqd->active_cic) {
1274                 struct cfq_io_context *cic = RQ_CIC(rq);
1275
1276                 atomic_long_inc(&cic->ioc->refcount);
1277                 cfqd->active_cic = cic;
1278         }
1279 }
1280
1281 /*
1282  * Find the cfqq that we need to service and move a request from that to the
1283  * dispatch list
1284  */
1285 static int cfq_dispatch_requests(struct request_queue *q, int force)
1286 {
1287         struct cfq_data *cfqd = q->elevator->elevator_data;
1288         struct cfq_queue *cfqq;
1289         unsigned int max_dispatch;
1290
1291         if (!cfqd->busy_queues)
1292                 return 0;
1293
1294         if (unlikely(force))
1295                 return cfq_forced_dispatch(cfqd);
1296
1297         cfqq = cfq_select_queue(cfqd);
1298         if (!cfqq)
1299                 return 0;
1300
1301         /*
1302          * Drain async requests before we start sync IO
1303          */
1304         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1305                 return 0;
1306
1307         /*
1308          * If this is an async queue and we have sync IO in flight, let it wait
1309          */
1310         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1311                 return 0;
1312
1313         max_dispatch = cfqd->cfq_quantum;
1314         if (cfq_class_idle(cfqq))
1315                 max_dispatch = 1;
1316
1317         /*
1318          * Does this cfqq already have too much IO in flight?
1319          */
1320         if (cfqq->dispatched >= max_dispatch) {
1321                 /*
1322                  * idle queue must always only have a single IO in flight
1323                  */
1324                 if (cfq_class_idle(cfqq))
1325                         return 0;
1326
1327                 /*
1328                  * We have other queues, don't allow more IO from this one
1329                  */
1330                 if (cfqd->busy_queues > 1)
1331                         return 0;
1332
1333                 /*
1334                  * Sole queue user, allow bigger slice
1335                  */
1336                 max_dispatch *= 4;
1337         }
1338
1339         /*
1340          * Async queues must wait a bit before being allowed dispatch.
1341          * We also ramp up the dispatch depth gradually for async IO,
1342          * based on the last sync IO we serviced
1343          */
1344         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1345                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1346                 unsigned int depth;
1347
1348                 /*
1349                  * must wait a bit longer
1350                  */
1351                 if (last_sync < cfqd->cfq_slice[1]) {
1352                         cfq_schedule_dispatch(cfqd,
1353                                                 cfqd->cfq_slice[1] - last_sync);
1354                         return 0;
1355                 }
1356
1357                 depth = last_sync / cfqd->cfq_slice[1];
1358                 if (depth < max_dispatch)
1359                         max_dispatch = depth;
1360         }
1361
1362         if (cfqq->dispatched >= max_dispatch)
1363                 return 0;
1364
1365         /*
1366          * Dispatch a request from this cfqq
1367          */
1368         cfq_dispatch_request(cfqd, cfqq);
1369         cfqq->slice_dispatch++;
1370         cfq_clear_cfqq_must_dispatch(cfqq);
1371
1372         /*
1373          * expire an async queue immediately if it has used up its slice. idle
1374          * queue always expire after 1 dispatch round.
1375          */
1376         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1377             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1378             cfq_class_idle(cfqq))) {
1379                 cfqq->slice_end = jiffies + 1;
1380                 cfq_slice_expired(cfqd, 0);
1381         }
1382
1383         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1384         return 1;
1385 }
1386
1387 /*
1388  * task holds one reference to the queue, dropped when task exits. each rq
1389  * in-flight on this queue also holds a reference, dropped when rq is freed.
1390  *
1391  * queue lock must be held here.
1392  */
1393 static void cfq_put_queue(struct cfq_queue *cfqq)
1394 {
1395         struct cfq_data *cfqd = cfqq->cfqd;
1396
1397         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1398
1399         if (!atomic_dec_and_test(&cfqq->ref))
1400                 return;
1401
1402         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1403         BUG_ON(rb_first(&cfqq->sort_list));
1404         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1405         BUG_ON(cfq_cfqq_on_rr(cfqq));
1406
1407         if (unlikely(cfqd->active_queue == cfqq)) {
1408                 __cfq_slice_expired(cfqd, cfqq, 0);
1409                 cfq_schedule_dispatch(cfqd, 0);
1410         }
1411
1412         kmem_cache_free(cfq_pool, cfqq);
1413 }
1414
1415 /*
1416  * Must always be called with the rcu_read_lock() held
1417  */
1418 static void
1419 __call_for_each_cic(struct io_context *ioc,
1420                     void (*func)(struct io_context *, struct cfq_io_context *))
1421 {
1422         struct cfq_io_context *cic;
1423         struct hlist_node *n;
1424
1425         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1426                 func(ioc, cic);
1427 }
1428
1429 /*
1430  * Call func for each cic attached to this ioc.
1431  */
1432 static void
1433 call_for_each_cic(struct io_context *ioc,
1434                   void (*func)(struct io_context *, struct cfq_io_context *))
1435 {
1436         rcu_read_lock();
1437         __call_for_each_cic(ioc, func);
1438         rcu_read_unlock();
1439 }
1440
1441 static void cfq_cic_free_rcu(struct rcu_head *head)
1442 {
1443         struct cfq_io_context *cic;
1444
1445         cic = container_of(head, struct cfq_io_context, rcu_head);
1446
1447         kmem_cache_free(cfq_ioc_pool, cic);
1448         elv_ioc_count_dec(cfq_ioc_count);
1449
1450         if (ioc_gone) {
1451                 /*
1452                  * CFQ scheduler is exiting, grab exit lock and check
1453                  * the pending io context count. If it hits zero,
1454                  * complete ioc_gone and set it back to NULL
1455                  */
1456                 spin_lock(&ioc_gone_lock);
1457                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1458                         complete(ioc_gone);
1459                         ioc_gone = NULL;
1460                 }
1461                 spin_unlock(&ioc_gone_lock);
1462         }
1463 }
1464
1465 static void cfq_cic_free(struct cfq_io_context *cic)
1466 {
1467         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1468 }
1469
1470 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1471 {
1472         unsigned long flags;
1473
1474         BUG_ON(!cic->dead_key);
1475
1476         spin_lock_irqsave(&ioc->lock, flags);
1477         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1478         hlist_del_rcu(&cic->cic_list);
1479         spin_unlock_irqrestore(&ioc->lock, flags);
1480
1481         cfq_cic_free(cic);
1482 }
1483
1484 /*
1485  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1486  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1487  * and ->trim() which is called with the task lock held
1488  */
1489 static void cfq_free_io_context(struct io_context *ioc)
1490 {
1491         /*
1492          * ioc->refcount is zero here, or we are called from elv_unregister(),
1493          * so no more cic's are allowed to be linked into this ioc.  So it
1494          * should be ok to iterate over the known list, we will see all cic's
1495          * since no new ones are added.
1496          */
1497         __call_for_each_cic(ioc, cic_free_func);
1498 }
1499
1500 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1501 {
1502         if (unlikely(cfqq == cfqd->active_queue)) {
1503                 __cfq_slice_expired(cfqd, cfqq, 0);
1504                 cfq_schedule_dispatch(cfqd, 0);
1505         }
1506
1507         cfq_put_queue(cfqq);
1508 }
1509
1510 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1511                                          struct cfq_io_context *cic)
1512 {
1513         struct io_context *ioc = cic->ioc;
1514
1515         list_del_init(&cic->queue_list);
1516
1517         /*
1518          * Make sure key == NULL is seen for dead queues
1519          */
1520         smp_wmb();
1521         cic->dead_key = (unsigned long) cic->key;
1522         cic->key = NULL;
1523
1524         if (ioc->ioc_data == cic)
1525                 rcu_assign_pointer(ioc->ioc_data, NULL);
1526
1527         if (cic->cfqq[BLK_RW_ASYNC]) {
1528                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1529                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1530         }
1531
1532         if (cic->cfqq[BLK_RW_SYNC]) {
1533                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1534                 cic->cfqq[BLK_RW_SYNC] = NULL;
1535         }
1536 }
1537
1538 static void cfq_exit_single_io_context(struct io_context *ioc,
1539                                        struct cfq_io_context *cic)
1540 {
1541         struct cfq_data *cfqd = cic->key;
1542
1543         if (cfqd) {
1544                 struct request_queue *q = cfqd->queue;
1545                 unsigned long flags;
1546
1547                 spin_lock_irqsave(q->queue_lock, flags);
1548
1549                 /*
1550                  * Ensure we get a fresh copy of the ->key to prevent
1551                  * race between exiting task and queue
1552                  */
1553                 smp_read_barrier_depends();
1554                 if (cic->key)
1555                         __cfq_exit_single_io_context(cfqd, cic);
1556
1557                 spin_unlock_irqrestore(q->queue_lock, flags);
1558         }
1559 }
1560
1561 /*
1562  * The process that ioc belongs to has exited, we need to clean up
1563  * and put the internal structures we have that belongs to that process.
1564  */
1565 static void cfq_exit_io_context(struct io_context *ioc)
1566 {
1567         call_for_each_cic(ioc, cfq_exit_single_io_context);
1568 }
1569
1570 static struct cfq_io_context *
1571 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1572 {
1573         struct cfq_io_context *cic;
1574
1575         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1576                                                         cfqd->queue->node);
1577         if (cic) {
1578                 cic->last_end_request = jiffies;
1579                 INIT_LIST_HEAD(&cic->queue_list);
1580                 INIT_HLIST_NODE(&cic->cic_list);
1581                 cic->dtor = cfq_free_io_context;
1582                 cic->exit = cfq_exit_io_context;
1583                 elv_ioc_count_inc(cfq_ioc_count);
1584         }
1585
1586         return cic;
1587 }
1588
1589 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1590 {
1591         struct task_struct *tsk = current;
1592         int ioprio_class;
1593
1594         if (!cfq_cfqq_prio_changed(cfqq))
1595                 return;
1596
1597         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1598         switch (ioprio_class) {
1599         default:
1600                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1601         case IOPRIO_CLASS_NONE:
1602                 /*
1603                  * no prio set, inherit CPU scheduling settings
1604                  */
1605                 cfqq->ioprio = task_nice_ioprio(tsk);
1606                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1607                 break;
1608         case IOPRIO_CLASS_RT:
1609                 cfqq->ioprio = task_ioprio(ioc);
1610                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1611                 break;
1612         case IOPRIO_CLASS_BE:
1613                 cfqq->ioprio = task_ioprio(ioc);
1614                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1615                 break;
1616         case IOPRIO_CLASS_IDLE:
1617                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1618                 cfqq->ioprio = 7;
1619                 cfq_clear_cfqq_idle_window(cfqq);
1620                 break;
1621         }
1622
1623         /*
1624          * keep track of original prio settings in case we have to temporarily
1625          * elevate the priority of this queue
1626          */
1627         cfqq->org_ioprio = cfqq->ioprio;
1628         cfqq->org_ioprio_class = cfqq->ioprio_class;
1629         cfq_clear_cfqq_prio_changed(cfqq);
1630 }
1631
1632 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1633 {
1634         struct cfq_data *cfqd = cic->key;
1635         struct cfq_queue *cfqq;
1636         unsigned long flags;
1637
1638         if (unlikely(!cfqd))
1639                 return;
1640
1641         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1642
1643         cfqq = cic->cfqq[BLK_RW_ASYNC];
1644         if (cfqq) {
1645                 struct cfq_queue *new_cfqq;
1646                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1647                                                 GFP_ATOMIC);
1648                 if (new_cfqq) {
1649                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1650                         cfq_put_queue(cfqq);
1651                 }
1652         }
1653
1654         cfqq = cic->cfqq[BLK_RW_SYNC];
1655         if (cfqq)
1656                 cfq_mark_cfqq_prio_changed(cfqq);
1657
1658         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1659 }
1660
1661 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1662 {
1663         call_for_each_cic(ioc, changed_ioprio);
1664         ioc->ioprio_changed = 0;
1665 }
1666
1667 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1668                           pid_t pid, int is_sync)
1669 {
1670         RB_CLEAR_NODE(&cfqq->rb_node);
1671         RB_CLEAR_NODE(&cfqq->p_node);
1672         INIT_LIST_HEAD(&cfqq->fifo);
1673
1674         atomic_set(&cfqq->ref, 0);
1675         cfqq->cfqd = cfqd;
1676
1677         cfq_mark_cfqq_prio_changed(cfqq);
1678
1679         if (is_sync) {
1680                 if (!cfq_class_idle(cfqq))
1681                         cfq_mark_cfqq_idle_window(cfqq);
1682                 cfq_mark_cfqq_sync(cfqq);
1683         }
1684         cfqq->pid = pid;
1685 }
1686
1687 static struct cfq_queue *
1688 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1689                      struct io_context *ioc, gfp_t gfp_mask)
1690 {
1691         struct cfq_queue *cfqq, *new_cfqq = NULL;
1692         struct cfq_io_context *cic;
1693
1694 retry:
1695         cic = cfq_cic_lookup(cfqd, ioc);
1696         /* cic always exists here */
1697         cfqq = cic_to_cfqq(cic, is_sync);
1698
1699         /*
1700          * Always try a new alloc if we fell back to the OOM cfqq
1701          * originally, since it should just be a temporary situation.
1702          */
1703         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1704                 cfqq = NULL;
1705                 if (new_cfqq) {
1706                         cfqq = new_cfqq;
1707                         new_cfqq = NULL;
1708                 } else if (gfp_mask & __GFP_WAIT) {
1709                         spin_unlock_irq(cfqd->queue->queue_lock);
1710                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1711                                         gfp_mask | __GFP_ZERO,
1712                                         cfqd->queue->node);
1713                         spin_lock_irq(cfqd->queue->queue_lock);
1714                         if (new_cfqq)
1715                                 goto retry;
1716                 } else {
1717                         cfqq = kmem_cache_alloc_node(cfq_pool,
1718                                         gfp_mask | __GFP_ZERO,
1719                                         cfqd->queue->node);
1720                 }
1721
1722                 if (cfqq) {
1723                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1724                         cfq_init_prio_data(cfqq, ioc);
1725                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1726                 } else
1727                         cfqq = &cfqd->oom_cfqq;
1728         }
1729
1730         if (new_cfqq)
1731                 kmem_cache_free(cfq_pool, new_cfqq);
1732
1733         return cfqq;
1734 }
1735
1736 static struct cfq_queue **
1737 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1738 {
1739         switch (ioprio_class) {
1740         case IOPRIO_CLASS_RT:
1741                 return &cfqd->async_cfqq[0][ioprio];
1742         case IOPRIO_CLASS_BE:
1743                 return &cfqd->async_cfqq[1][ioprio];
1744         case IOPRIO_CLASS_IDLE:
1745                 return &cfqd->async_idle_cfqq;
1746         default:
1747                 BUG();
1748         }
1749 }
1750
1751 static struct cfq_queue *
1752 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1753               gfp_t gfp_mask)
1754 {
1755         const int ioprio = task_ioprio(ioc);
1756         const int ioprio_class = task_ioprio_class(ioc);
1757         struct cfq_queue **async_cfqq = NULL;
1758         struct cfq_queue *cfqq = NULL;
1759
1760         if (!is_sync) {
1761                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1762                 cfqq = *async_cfqq;
1763         }
1764
1765         if (!cfqq)
1766                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1767
1768         /*
1769          * pin the queue now that it's allocated, scheduler exit will prune it
1770          */
1771         if (!is_sync && !(*async_cfqq)) {
1772                 atomic_inc(&cfqq->ref);
1773                 *async_cfqq = cfqq;
1774         }
1775
1776         atomic_inc(&cfqq->ref);
1777         return cfqq;
1778 }
1779
1780 /*
1781  * We drop cfq io contexts lazily, so we may find a dead one.
1782  */
1783 static void
1784 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1785                   struct cfq_io_context *cic)
1786 {
1787         unsigned long flags;
1788
1789         WARN_ON(!list_empty(&cic->queue_list));
1790
1791         spin_lock_irqsave(&ioc->lock, flags);
1792
1793         BUG_ON(ioc->ioc_data == cic);
1794
1795         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1796         hlist_del_rcu(&cic->cic_list);
1797         spin_unlock_irqrestore(&ioc->lock, flags);
1798
1799         cfq_cic_free(cic);
1800 }
1801
1802 static struct cfq_io_context *
1803 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1804 {
1805         struct cfq_io_context *cic;
1806         unsigned long flags;
1807         void *k;
1808
1809         if (unlikely(!ioc))
1810                 return NULL;
1811
1812         rcu_read_lock();
1813
1814         /*
1815          * we maintain a last-hit cache, to avoid browsing over the tree
1816          */
1817         cic = rcu_dereference(ioc->ioc_data);
1818         if (cic && cic->key == cfqd) {
1819                 rcu_read_unlock();
1820                 return cic;
1821         }
1822
1823         do {
1824                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1825                 rcu_read_unlock();
1826                 if (!cic)
1827                         break;
1828                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1829                 k = cic->key;
1830                 if (unlikely(!k)) {
1831                         cfq_drop_dead_cic(cfqd, ioc, cic);
1832                         rcu_read_lock();
1833                         continue;
1834                 }
1835
1836                 spin_lock_irqsave(&ioc->lock, flags);
1837                 rcu_assign_pointer(ioc->ioc_data, cic);
1838                 spin_unlock_irqrestore(&ioc->lock, flags);
1839                 break;
1840         } while (1);
1841
1842         return cic;
1843 }
1844
1845 /*
1846  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1847  * the process specific cfq io context when entered from the block layer.
1848  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1849  */
1850 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1851                         struct cfq_io_context *cic, gfp_t gfp_mask)
1852 {
1853         unsigned long flags;
1854         int ret;
1855
1856         ret = radix_tree_preload(gfp_mask);
1857         if (!ret) {
1858                 cic->ioc = ioc;
1859                 cic->key = cfqd;
1860
1861                 spin_lock_irqsave(&ioc->lock, flags);
1862                 ret = radix_tree_insert(&ioc->radix_root,
1863                                                 (unsigned long) cfqd, cic);
1864                 if (!ret)
1865                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1866                 spin_unlock_irqrestore(&ioc->lock, flags);
1867
1868                 radix_tree_preload_end();
1869
1870                 if (!ret) {
1871                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1872                         list_add(&cic->queue_list, &cfqd->cic_list);
1873                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1874                 }
1875         }
1876
1877         if (ret)
1878                 printk(KERN_ERR "cfq: cic link failed!\n");
1879
1880         return ret;
1881 }
1882
1883 /*
1884  * Setup general io context and cfq io context. There can be several cfq
1885  * io contexts per general io context, if this process is doing io to more
1886  * than one device managed by cfq.
1887  */
1888 static struct cfq_io_context *
1889 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1890 {
1891         struct io_context *ioc = NULL;
1892         struct cfq_io_context *cic;
1893
1894         might_sleep_if(gfp_mask & __GFP_WAIT);
1895
1896         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1897         if (!ioc)
1898                 return NULL;
1899
1900         cic = cfq_cic_lookup(cfqd, ioc);
1901         if (cic)
1902                 goto out;
1903
1904         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1905         if (cic == NULL)
1906                 goto err;
1907
1908         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1909                 goto err_free;
1910
1911 out:
1912         smp_read_barrier_depends();
1913         if (unlikely(ioc->ioprio_changed))
1914                 cfq_ioc_set_ioprio(ioc);
1915
1916         return cic;
1917 err_free:
1918         cfq_cic_free(cic);
1919 err:
1920         put_io_context(ioc);
1921         return NULL;
1922 }
1923
1924 static void
1925 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1926 {
1927         unsigned long elapsed = jiffies - cic->last_end_request;
1928         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1929
1930         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1931         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1932         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1933 }
1934
1935 static void
1936 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1937                        struct request *rq)
1938 {
1939         sector_t sdist;
1940         u64 total;
1941
1942         if (!cic->last_request_pos)
1943                 sdist = 0;
1944         else if (cic->last_request_pos < blk_rq_pos(rq))
1945                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1946         else
1947                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1948
1949         /*
1950          * Don't allow the seek distance to get too large from the
1951          * odd fragment, pagein, etc
1952          */
1953         if (cic->seek_samples <= 60) /* second&third seek */
1954                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1955         else
1956                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1957
1958         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1959         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1960         total = cic->seek_total + (cic->seek_samples/2);
1961         do_div(total, cic->seek_samples);
1962         cic->seek_mean = (sector_t)total;
1963 }
1964
1965 /*
1966  * Disable idle window if the process thinks too long or seeks so much that
1967  * it doesn't matter
1968  */
1969 static void
1970 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1971                        struct cfq_io_context *cic)
1972 {
1973         int old_idle, enable_idle;
1974
1975         /*
1976          * Don't idle for async or idle io prio class
1977          */
1978         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1979                 return;
1980
1981         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1982
1983         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1984             (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1985                 enable_idle = 0;
1986         else if (sample_valid(cic->ttime_samples)) {
1987                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1988                         enable_idle = 0;
1989                 else
1990                         enable_idle = 1;
1991         }
1992
1993         if (old_idle != enable_idle) {
1994                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1995                 if (enable_idle)
1996                         cfq_mark_cfqq_idle_window(cfqq);
1997                 else
1998                         cfq_clear_cfqq_idle_window(cfqq);
1999         }
2000 }
2001
2002 /*
2003  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2004  * no or if we aren't sure, a 1 will cause a preempt.
2005  */
2006 static int
2007 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2008                    struct request *rq)
2009 {
2010         struct cfq_queue *cfqq;
2011
2012         cfqq = cfqd->active_queue;
2013         if (!cfqq)
2014                 return 0;
2015
2016         if (cfq_slice_used(cfqq))
2017                 return 1;
2018
2019         if (cfq_class_idle(new_cfqq))
2020                 return 0;
2021
2022         if (cfq_class_idle(cfqq))
2023                 return 1;
2024
2025         /*
2026          * if the new request is sync, but the currently running queue is
2027          * not, let the sync request have priority.
2028          */
2029         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2030                 return 1;
2031
2032         /*
2033          * So both queues are sync. Let the new request get disk time if
2034          * it's a metadata request and the current queue is doing regular IO.
2035          */
2036         if (rq_is_meta(rq) && !cfqq->meta_pending)
2037                 return 1;
2038
2039         /*
2040          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2041          */
2042         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2043                 return 1;
2044
2045         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2046                 return 0;
2047
2048         /*
2049          * if this request is as-good as one we would expect from the
2050          * current cfqq, let it preempt
2051          */
2052         if (cfq_rq_close(cfqd, rq))
2053                 return 1;
2054
2055         return 0;
2056 }
2057
2058 /*
2059  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2060  * let it have half of its nominal slice.
2061  */
2062 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2063 {
2064         cfq_log_cfqq(cfqd, cfqq, "preempt");
2065         cfq_slice_expired(cfqd, 1);
2066
2067         /*
2068          * Put the new queue at the front of the of the current list,
2069          * so we know that it will be selected next.
2070          */
2071         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2072
2073         cfq_service_tree_add(cfqd, cfqq, 1);
2074
2075         cfqq->slice_end = 0;
2076         cfq_mark_cfqq_slice_new(cfqq);
2077 }
2078
2079 /*
2080  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2081  * something we should do about it
2082  */
2083 static void
2084 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2085                 struct request *rq)
2086 {
2087         struct cfq_io_context *cic = RQ_CIC(rq);
2088
2089         cfqd->rq_queued++;
2090         if (rq_is_meta(rq))
2091                 cfqq->meta_pending++;
2092
2093         cfq_update_io_thinktime(cfqd, cic);
2094         cfq_update_io_seektime(cfqd, cic, rq);
2095         cfq_update_idle_window(cfqd, cfqq, cic);
2096
2097         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2098
2099         if (cfqq == cfqd->active_queue) {
2100                 /*
2101                  * Remember that we saw a request from this process, but
2102                  * don't start queuing just yet. Otherwise we risk seeing lots
2103                  * of tiny requests, because we disrupt the normal plugging
2104                  * and merging. If the request is already larger than a single
2105                  * page, let it rip immediately. For that case we assume that
2106                  * merging is already done. Ditto for a busy system that
2107                  * has other work pending, don't risk delaying until the
2108                  * idle timer unplug to continue working.
2109                  */
2110                 if (cfq_cfqq_wait_request(cfqq)) {
2111                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2112                             cfqd->busy_queues > 1) {
2113                                 del_timer(&cfqd->idle_slice_timer);
2114                         __blk_run_queue(cfqd->queue);
2115                         }
2116                         cfq_mark_cfqq_must_dispatch(cfqq);
2117                 }
2118         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2119                 /*
2120                  * not the active queue - expire current slice if it is
2121                  * idle and has expired it's mean thinktime or this new queue
2122                  * has some old slice time left and is of higher priority or
2123                  * this new queue is RT and the current one is BE
2124                  */
2125                 cfq_preempt_queue(cfqd, cfqq);
2126                 __blk_run_queue(cfqd->queue);
2127         }
2128 }
2129
2130 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2131 {
2132         struct cfq_data *cfqd = q->elevator->elevator_data;
2133         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2134
2135         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2136         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2137
2138         cfq_add_rq_rb(rq);
2139
2140         list_add_tail(&rq->queuelist, &cfqq->fifo);
2141
2142         cfq_rq_enqueued(cfqd, cfqq, rq);
2143 }
2144
2145 /*
2146  * Update hw_tag based on peak queue depth over 50 samples under
2147  * sufficient load.
2148  */
2149 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2150 {
2151         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2152                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2153
2154         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2155             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2156                 return;
2157
2158         if (cfqd->hw_tag_samples++ < 50)
2159                 return;
2160
2161         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2162                 cfqd->hw_tag = 1;
2163         else
2164                 cfqd->hw_tag = 0;
2165
2166         cfqd->hw_tag_samples = 0;
2167         cfqd->rq_in_driver_peak = 0;
2168 }
2169
2170 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2171 {
2172         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2173         struct cfq_data *cfqd = cfqq->cfqd;
2174         const int sync = rq_is_sync(rq);
2175         unsigned long now;
2176
2177         now = jiffies;
2178         cfq_log_cfqq(cfqd, cfqq, "complete");
2179
2180         cfq_update_hw_tag(cfqd);
2181
2182         WARN_ON(!cfqd->rq_in_driver[sync]);
2183         WARN_ON(!cfqq->dispatched);
2184         cfqd->rq_in_driver[sync]--;
2185         cfqq->dispatched--;
2186
2187         if (cfq_cfqq_sync(cfqq))
2188                 cfqd->sync_flight--;
2189
2190         if (sync) {
2191                 RQ_CIC(rq)->last_end_request = now;
2192                 cfqd->last_end_sync_rq = now;
2193         }
2194
2195         /*
2196          * If this is the active queue, check if it needs to be expired,
2197          * or if we want to idle in case it has no pending requests.
2198          */
2199         if (cfqd->active_queue == cfqq) {
2200                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2201
2202                 if (cfq_cfqq_slice_new(cfqq)) {
2203                         cfq_set_prio_slice(cfqd, cfqq);
2204                         cfq_clear_cfqq_slice_new(cfqq);
2205                 }
2206                 /*
2207                  * If there are no requests waiting in this queue, and
2208                  * there are other queues ready to issue requests, AND
2209                  * those other queues are issuing requests within our
2210                  * mean seek distance, give them a chance to run instead
2211                  * of idling.
2212                  */
2213                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2214                         cfq_slice_expired(cfqd, 1);
2215                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2216                          sync && !rq_noidle(rq))
2217                         cfq_arm_slice_timer(cfqd);
2218         }
2219
2220         if (!rq_in_driver(cfqd))
2221                 cfq_schedule_dispatch(cfqd, 0);
2222 }
2223
2224 /*
2225  * we temporarily boost lower priority queues if they are holding fs exclusive
2226  * resources. they are boosted to normal prio (CLASS_BE/4)
2227  */
2228 static void cfq_prio_boost(struct cfq_queue *cfqq)
2229 {
2230         if (has_fs_excl()) {
2231                 /*
2232                  * boost idle prio on transactions that would lock out other
2233                  * users of the filesystem
2234                  */
2235                 if (cfq_class_idle(cfqq))
2236                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2237                 if (cfqq->ioprio > IOPRIO_NORM)
2238                         cfqq->ioprio = IOPRIO_NORM;
2239         } else {
2240                 /*
2241                  * check if we need to unboost the queue
2242                  */
2243                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2244                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2245                 if (cfqq->ioprio != cfqq->org_ioprio)
2246                         cfqq->ioprio = cfqq->org_ioprio;
2247         }
2248 }
2249
2250 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2251 {
2252         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2253                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2254                 return ELV_MQUEUE_MUST;
2255         }
2256
2257         return ELV_MQUEUE_MAY;
2258 }
2259
2260 static int cfq_may_queue(struct request_queue *q, int rw)
2261 {
2262         struct cfq_data *cfqd = q->elevator->elevator_data;
2263         struct task_struct *tsk = current;
2264         struct cfq_io_context *cic;
2265         struct cfq_queue *cfqq;
2266
2267         /*
2268          * don't force setup of a queue from here, as a call to may_queue
2269          * does not necessarily imply that a request actually will be queued.
2270          * so just lookup a possibly existing queue, or return 'may queue'
2271          * if that fails
2272          */
2273         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2274         if (!cic)
2275                 return ELV_MQUEUE_MAY;
2276
2277         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2278         if (cfqq) {
2279                 cfq_init_prio_data(cfqq, cic->ioc);
2280                 cfq_prio_boost(cfqq);
2281
2282                 return __cfq_may_queue(cfqq);
2283         }
2284
2285         return ELV_MQUEUE_MAY;
2286 }
2287
2288 /*
2289  * queue lock held here
2290  */
2291 static void cfq_put_request(struct request *rq)
2292 {
2293         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2294
2295         if (cfqq) {
2296                 const int rw = rq_data_dir(rq);
2297
2298                 BUG_ON(!cfqq->allocated[rw]);
2299                 cfqq->allocated[rw]--;
2300
2301                 put_io_context(RQ_CIC(rq)->ioc);
2302
2303                 rq->elevator_private = NULL;
2304                 rq->elevator_private2 = NULL;
2305
2306                 cfq_put_queue(cfqq);
2307         }
2308 }
2309
2310 /*
2311  * Allocate cfq data structures associated with this request.
2312  */
2313 static int
2314 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2315 {
2316         struct cfq_data *cfqd = q->elevator->elevator_data;
2317         struct cfq_io_context *cic;
2318         const int rw = rq_data_dir(rq);
2319         const int is_sync = rq_is_sync(rq);
2320         struct cfq_queue *cfqq;
2321         unsigned long flags;
2322
2323         might_sleep_if(gfp_mask & __GFP_WAIT);
2324
2325         cic = cfq_get_io_context(cfqd, gfp_mask);
2326
2327         spin_lock_irqsave(q->queue_lock, flags);
2328
2329         if (!cic)
2330                 goto queue_fail;
2331
2332         cfqq = cic_to_cfqq(cic, is_sync);
2333         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2334                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2335                 cic_set_cfqq(cic, cfqq, is_sync);
2336         }
2337
2338         cfqq->allocated[rw]++;
2339         atomic_inc(&cfqq->ref);
2340
2341         spin_unlock_irqrestore(q->queue_lock, flags);
2342
2343         rq->elevator_private = cic;
2344         rq->elevator_private2 = cfqq;
2345         return 0;
2346
2347 queue_fail:
2348         if (cic)
2349                 put_io_context(cic->ioc);
2350
2351         cfq_schedule_dispatch(cfqd, 0);
2352         spin_unlock_irqrestore(q->queue_lock, flags);
2353         cfq_log(cfqd, "set_request fail");
2354         return 1;
2355 }
2356
2357 static void cfq_kick_queue(struct work_struct *work)
2358 {
2359         struct cfq_data *cfqd =
2360                 container_of(work, struct cfq_data, unplug_work.work);
2361         struct request_queue *q = cfqd->queue;
2362
2363         spin_lock_irq(q->queue_lock);
2364         __blk_run_queue(cfqd->queue);
2365         spin_unlock_irq(q->queue_lock);
2366 }
2367
2368 /*
2369  * Timer running if the active_queue is currently idling inside its time slice
2370  */
2371 static void cfq_idle_slice_timer(unsigned long data)
2372 {
2373         struct cfq_data *cfqd = (struct cfq_data *) data;
2374         struct cfq_queue *cfqq;
2375         unsigned long flags;
2376         int timed_out = 1;
2377
2378         cfq_log(cfqd, "idle timer fired");
2379
2380         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2381
2382         cfqq = cfqd->active_queue;
2383         if (cfqq) {
2384                 timed_out = 0;
2385
2386                 /*
2387                  * We saw a request before the queue expired, let it through
2388                  */
2389                 if (cfq_cfqq_must_dispatch(cfqq))
2390                         goto out_kick;
2391
2392                 /*
2393                  * expired
2394                  */
2395                 if (cfq_slice_used(cfqq))
2396                         goto expire;
2397
2398                 /*
2399                  * only expire and reinvoke request handler, if there are
2400                  * other queues with pending requests
2401                  */
2402                 if (!cfqd->busy_queues)
2403                         goto out_cont;
2404
2405                 /*
2406                  * not expired and it has a request pending, let it dispatch
2407                  */
2408                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2409                         goto out_kick;
2410         }
2411 expire:
2412         cfq_slice_expired(cfqd, timed_out);
2413 out_kick:
2414         cfq_schedule_dispatch(cfqd, 0);
2415 out_cont:
2416         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2417 }
2418
2419 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2420 {
2421         del_timer_sync(&cfqd->idle_slice_timer);
2422         cancel_delayed_work_sync(&cfqd->unplug_work);
2423 }
2424
2425 static void cfq_put_async_queues(struct cfq_data *cfqd)
2426 {
2427         int i;
2428
2429         for (i = 0; i < IOPRIO_BE_NR; i++) {
2430                 if (cfqd->async_cfqq[0][i])
2431                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2432                 if (cfqd->async_cfqq[1][i])
2433                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2434         }
2435
2436         if (cfqd->async_idle_cfqq)
2437                 cfq_put_queue(cfqd->async_idle_cfqq);
2438 }
2439
2440 static void cfq_exit_queue(struct elevator_queue *e)
2441 {
2442         struct cfq_data *cfqd = e->elevator_data;
2443         struct request_queue *q = cfqd->queue;
2444
2445         cfq_shutdown_timer_wq(cfqd);
2446
2447         spin_lock_irq(q->queue_lock);
2448
2449         if (cfqd->active_queue)
2450                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2451
2452         while (!list_empty(&cfqd->cic_list)) {
2453                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2454                                                         struct cfq_io_context,
2455                                                         queue_list);
2456
2457                 __cfq_exit_single_io_context(cfqd, cic);
2458         }
2459
2460         cfq_put_async_queues(cfqd);
2461
2462         spin_unlock_irq(q->queue_lock);
2463
2464         cfq_shutdown_timer_wq(cfqd);
2465
2466         kfree(cfqd);
2467 }
2468
2469 static void *cfq_init_queue(struct request_queue *q)
2470 {
2471         struct cfq_data *cfqd;
2472         int i;
2473
2474         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2475         if (!cfqd)
2476                 return NULL;
2477
2478         cfqd->service_tree = CFQ_RB_ROOT;
2479
2480         /*
2481          * Not strictly needed (since RB_ROOT just clears the node and we
2482          * zeroed cfqd on alloc), but better be safe in case someone decides
2483          * to add magic to the rb code
2484          */
2485         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2486                 cfqd->prio_trees[i] = RB_ROOT;
2487
2488         /*
2489          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2490          * Grab a permanent reference to it, so that the normal code flow
2491          * will not attempt to free it.
2492          */
2493         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2494         atomic_inc(&cfqd->oom_cfqq.ref);
2495
2496         INIT_LIST_HEAD(&cfqd->cic_list);
2497
2498         cfqd->queue = q;
2499
2500         init_timer(&cfqd->idle_slice_timer);
2501         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2502         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2503
2504         INIT_DELAYED_WORK(&cfqd->unplug_work, cfq_kick_queue);
2505
2506         cfqd->cfq_quantum = cfq_quantum;
2507         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2508         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2509         cfqd->cfq_back_max = cfq_back_max;
2510         cfqd->cfq_back_penalty = cfq_back_penalty;
2511         cfqd->cfq_slice[0] = cfq_slice_async;
2512         cfqd->cfq_slice[1] = cfq_slice_sync;
2513         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2514         cfqd->cfq_slice_idle = cfq_slice_idle;
2515         cfqd->cfq_latency = 1;
2516         cfqd->hw_tag = 1;
2517         cfqd->last_end_sync_rq = jiffies;
2518         return cfqd;
2519 }
2520
2521 static void cfq_slab_kill(void)
2522 {
2523         /*
2524          * Caller already ensured that pending RCU callbacks are completed,
2525          * so we should have no busy allocations at this point.
2526          */
2527         if (cfq_pool)
2528                 kmem_cache_destroy(cfq_pool);
2529         if (cfq_ioc_pool)
2530                 kmem_cache_destroy(cfq_ioc_pool);
2531 }
2532
2533 static int __init cfq_slab_setup(void)
2534 {
2535         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2536         if (!cfq_pool)
2537                 goto fail;
2538
2539         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2540         if (!cfq_ioc_pool)
2541                 goto fail;
2542
2543         return 0;
2544 fail:
2545         cfq_slab_kill();
2546         return -ENOMEM;
2547 }
2548
2549 /*
2550  * sysfs parts below -->
2551  */
2552 static ssize_t
2553 cfq_var_show(unsigned int var, char *page)
2554 {
2555         return sprintf(page, "%d\n", var);
2556 }
2557
2558 static ssize_t
2559 cfq_var_store(unsigned int *var, const char *page, size_t count)
2560 {
2561         char *p = (char *) page;
2562
2563         *var = simple_strtoul(p, &p, 10);
2564         return count;
2565 }
2566
2567 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2568 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2569 {                                                                       \
2570         struct cfq_data *cfqd = e->elevator_data;                       \
2571         unsigned int __data = __VAR;                                    \
2572         if (__CONV)                                                     \
2573                 __data = jiffies_to_msecs(__data);                      \
2574         return cfq_var_show(__data, (page));                            \
2575 }
2576 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2577 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2578 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2579 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2580 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2581 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2582 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2583 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2584 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2585 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2586 #undef SHOW_FUNCTION
2587
2588 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2589 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2590 {                                                                       \
2591         struct cfq_data *cfqd = e->elevator_data;                       \
2592         unsigned int __data;                                            \
2593         int ret = cfq_var_store(&__data, (page), count);                \
2594         if (__data < (MIN))                                             \
2595                 __data = (MIN);                                         \
2596         else if (__data > (MAX))                                        \
2597                 __data = (MAX);                                         \
2598         if (__CONV)                                                     \
2599                 *(__PTR) = msecs_to_jiffies(__data);                    \
2600         else                                                            \
2601                 *(__PTR) = __data;                                      \
2602         return ret;                                                     \
2603 }
2604 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2605 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2606                 UINT_MAX, 1);
2607 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2608                 UINT_MAX, 1);
2609 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2610 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2611                 UINT_MAX, 0);
2612 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2613 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2614 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2615 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2616                 UINT_MAX, 0);
2617 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2618 #undef STORE_FUNCTION
2619
2620 #define CFQ_ATTR(name) \
2621         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2622
2623 static struct elv_fs_entry cfq_attrs[] = {
2624         CFQ_ATTR(quantum),
2625         CFQ_ATTR(fifo_expire_sync),
2626         CFQ_ATTR(fifo_expire_async),
2627         CFQ_ATTR(back_seek_max),
2628         CFQ_ATTR(back_seek_penalty),
2629         CFQ_ATTR(slice_sync),
2630         CFQ_ATTR(slice_async),
2631         CFQ_ATTR(slice_async_rq),
2632         CFQ_ATTR(slice_idle),
2633         CFQ_ATTR(low_latency),
2634         __ATTR_NULL
2635 };
2636
2637 static struct elevator_type iosched_cfq = {
2638         .ops = {
2639                 .elevator_merge_fn =            cfq_merge,
2640                 .elevator_merged_fn =           cfq_merged_request,
2641                 .elevator_merge_req_fn =        cfq_merged_requests,
2642                 .elevator_allow_merge_fn =      cfq_allow_merge,
2643                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2644                 .elevator_add_req_fn =          cfq_insert_request,
2645                 .elevator_activate_req_fn =     cfq_activate_request,
2646                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2647                 .elevator_queue_empty_fn =      cfq_queue_empty,
2648                 .elevator_completed_req_fn =    cfq_completed_request,
2649                 .elevator_former_req_fn =       elv_rb_former_request,
2650                 .elevator_latter_req_fn =       elv_rb_latter_request,
2651                 .elevator_set_req_fn =          cfq_set_request,
2652                 .elevator_put_req_fn =          cfq_put_request,
2653                 .elevator_may_queue_fn =        cfq_may_queue,
2654                 .elevator_init_fn =             cfq_init_queue,
2655                 .elevator_exit_fn =             cfq_exit_queue,
2656                 .trim =                         cfq_free_io_context,
2657         },
2658         .elevator_attrs =       cfq_attrs,
2659         .elevator_name =        "cfq",
2660         .elevator_owner =       THIS_MODULE,
2661 };
2662
2663 static int __init cfq_init(void)
2664 {
2665         /*
2666          * could be 0 on HZ < 1000 setups
2667          */
2668         if (!cfq_slice_async)
2669                 cfq_slice_async = 1;
2670         if (!cfq_slice_idle)
2671                 cfq_slice_idle = 1;
2672
2673         if (cfq_slab_setup())
2674                 return -ENOMEM;
2675
2676         elv_register(&iosched_cfq);
2677
2678         return 0;
2679 }
2680
2681 static void __exit cfq_exit(void)
2682 {
2683         DECLARE_COMPLETION_ONSTACK(all_gone);
2684         elv_unregister(&iosched_cfq);
2685         ioc_gone = &all_gone;
2686         /* ioc_gone's update must be visible before reading ioc_count */
2687         smp_wmb();
2688
2689         /*
2690          * this also protects us from entering cfq_slab_kill() with
2691          * pending RCU callbacks
2692          */
2693         if (elv_ioc_count_read(cfq_ioc_count))
2694                 wait_for_completion(&all_gone);
2695         cfq_slab_kill();
2696 }
2697
2698 module_init(cfq_init);
2699 module_exit(cfq_exit);
2700
2701 MODULE_AUTHOR("Jens Axboe");
2702 MODULE_LICENSE("GPL");
2703 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");