Merge branch 'for-rmk' of git://linux-arm.org/linux-2.6
[pandora-kernel.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 work_struct 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
177         struct list_head cic_list;
178
179         /*
180          * Fallback dummy cfqq for extreme OOM conditions
181          */
182         struct cfq_queue oom_cfqq;
183 };
184
185 enum cfqq_state_flags {
186         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
187         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
188         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
189         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
190         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
191         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
192         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
193         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
194         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
195         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
196 };
197
198 #define CFQ_CFQQ_FNS(name)                                              \
199 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
200 {                                                                       \
201         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
202 }                                                                       \
203 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
204 {                                                                       \
205         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
206 }                                                                       \
207 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
208 {                                                                       \
209         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
210 }
211
212 CFQ_CFQQ_FNS(on_rr);
213 CFQ_CFQQ_FNS(wait_request);
214 CFQ_CFQQ_FNS(must_dispatch);
215 CFQ_CFQQ_FNS(must_alloc_slice);
216 CFQ_CFQQ_FNS(fifo_expire);
217 CFQ_CFQQ_FNS(idle_window);
218 CFQ_CFQQ_FNS(prio_changed);
219 CFQ_CFQQ_FNS(slice_new);
220 CFQ_CFQQ_FNS(sync);
221 CFQ_CFQQ_FNS(coop);
222 #undef CFQ_CFQQ_FNS
223
224 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
225         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
226 #define cfq_log(cfqd, fmt, args...)     \
227         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
228
229 static void cfq_dispatch_insert(struct request_queue *, struct request *);
230 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
231                                        struct io_context *, gfp_t);
232 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
233                                                 struct io_context *);
234
235 static inline int rq_in_driver(struct cfq_data *cfqd)
236 {
237         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
238 }
239
240 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
241                                             int is_sync)
242 {
243         return cic->cfqq[!!is_sync];
244 }
245
246 static inline void cic_set_cfqq(struct cfq_io_context *cic,
247                                 struct cfq_queue *cfqq, int is_sync)
248 {
249         cic->cfqq[!!is_sync] = cfqq;
250 }
251
252 /*
253  * We regard a request as SYNC, if it's either a read or has the SYNC bit
254  * set (in which case it could also be direct WRITE).
255  */
256 static inline int cfq_bio_sync(struct bio *bio)
257 {
258         if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
259                 return 1;
260
261         return 0;
262 }
263
264 /*
265  * scheduler run of queue, if there are requests pending and no one in the
266  * driver that will restart queueing
267  */
268 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
269 {
270         if (cfqd->busy_queues) {
271                 cfq_log(cfqd, "schedule dispatch");
272                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
273         }
274 }
275
276 static int cfq_queue_empty(struct request_queue *q)
277 {
278         struct cfq_data *cfqd = q->elevator->elevator_data;
279
280         return !cfqd->busy_queues;
281 }
282
283 /*
284  * Scale schedule slice based on io priority. Use the sync time slice only
285  * if a queue is marked sync and has sync io queued. A sync queue with async
286  * io only, should not get full sync slice length.
287  */
288 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
289                                  unsigned short prio)
290 {
291         const int base_slice = cfqd->cfq_slice[sync];
292
293         WARN_ON(prio >= IOPRIO_BE_NR);
294
295         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
296 }
297
298 static inline int
299 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
300 {
301         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
302 }
303
304 static inline void
305 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
306 {
307         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
308         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
309 }
310
311 /*
312  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
313  * isn't valid until the first request from the dispatch is activated
314  * and the slice time set.
315  */
316 static inline int cfq_slice_used(struct cfq_queue *cfqq)
317 {
318         if (cfq_cfqq_slice_new(cfqq))
319                 return 0;
320         if (time_before(jiffies, cfqq->slice_end))
321                 return 0;
322
323         return 1;
324 }
325
326 /*
327  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
328  * We choose the request that is closest to the head right now. Distance
329  * behind the head is penalized and only allowed to a certain extent.
330  */
331 static struct request *
332 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
333 {
334         sector_t last, s1, s2, d1 = 0, d2 = 0;
335         unsigned long back_max;
336 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
337 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
338         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
339
340         if (rq1 == NULL || rq1 == rq2)
341                 return rq2;
342         if (rq2 == NULL)
343                 return rq1;
344
345         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
346                 return rq1;
347         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
348                 return rq2;
349         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
350                 return rq1;
351         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
352                 return rq2;
353
354         s1 = blk_rq_pos(rq1);
355         s2 = blk_rq_pos(rq2);
356
357         last = cfqd->last_position;
358
359         /*
360          * by definition, 1KiB is 2 sectors
361          */
362         back_max = cfqd->cfq_back_max * 2;
363
364         /*
365          * Strict one way elevator _except_ in the case where we allow
366          * short backward seeks which are biased as twice the cost of a
367          * similar forward seek.
368          */
369         if (s1 >= last)
370                 d1 = s1 - last;
371         else if (s1 + back_max >= last)
372                 d1 = (last - s1) * cfqd->cfq_back_penalty;
373         else
374                 wrap |= CFQ_RQ1_WRAP;
375
376         if (s2 >= last)
377                 d2 = s2 - last;
378         else if (s2 + back_max >= last)
379                 d2 = (last - s2) * cfqd->cfq_back_penalty;
380         else
381                 wrap |= CFQ_RQ2_WRAP;
382
383         /* Found required data */
384
385         /*
386          * By doing switch() on the bit mask "wrap" we avoid having to
387          * check two variables for all permutations: --> faster!
388          */
389         switch (wrap) {
390         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
391                 if (d1 < d2)
392                         return rq1;
393                 else if (d2 < d1)
394                         return rq2;
395                 else {
396                         if (s1 >= s2)
397                                 return rq1;
398                         else
399                                 return rq2;
400                 }
401
402         case CFQ_RQ2_WRAP:
403                 return rq1;
404         case CFQ_RQ1_WRAP:
405                 return rq2;
406         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
407         default:
408                 /*
409                  * Since both rqs are wrapped,
410                  * start with the one that's further behind head
411                  * (--> only *one* back seek required),
412                  * since back seek takes more time than forward.
413                  */
414                 if (s1 <= s2)
415                         return rq1;
416                 else
417                         return rq2;
418         }
419 }
420
421 /*
422  * The below is leftmost cache rbtree addon
423  */
424 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
425 {
426         if (!root->left)
427                 root->left = rb_first(&root->rb);
428
429         if (root->left)
430                 return rb_entry(root->left, struct cfq_queue, rb_node);
431
432         return NULL;
433 }
434
435 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
436 {
437         rb_erase(n, root);
438         RB_CLEAR_NODE(n);
439 }
440
441 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
442 {
443         if (root->left == n)
444                 root->left = NULL;
445         rb_erase_init(n, &root->rb);
446 }
447
448 /*
449  * would be nice to take fifo expire time into account as well
450  */
451 static struct request *
452 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
453                   struct request *last)
454 {
455         struct rb_node *rbnext = rb_next(&last->rb_node);
456         struct rb_node *rbprev = rb_prev(&last->rb_node);
457         struct request *next = NULL, *prev = NULL;
458
459         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
460
461         if (rbprev)
462                 prev = rb_entry_rq(rbprev);
463
464         if (rbnext)
465                 next = rb_entry_rq(rbnext);
466         else {
467                 rbnext = rb_first(&cfqq->sort_list);
468                 if (rbnext && rbnext != &last->rb_node)
469                         next = rb_entry_rq(rbnext);
470         }
471
472         return cfq_choose_req(cfqd, next, prev);
473 }
474
475 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
476                                       struct cfq_queue *cfqq)
477 {
478         /*
479          * just an approximation, should be ok.
480          */
481         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
482                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
483 }
484
485 /*
486  * The cfqd->service_tree holds all pending cfq_queue's that have
487  * requests waiting to be processed. It is sorted in the order that
488  * we will service the queues.
489  */
490 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
491                                  int add_front)
492 {
493         struct rb_node **p, *parent;
494         struct cfq_queue *__cfqq;
495         unsigned long rb_key;
496         int left;
497
498         if (cfq_class_idle(cfqq)) {
499                 rb_key = CFQ_IDLE_DELAY;
500                 parent = rb_last(&cfqd->service_tree.rb);
501                 if (parent && parent != &cfqq->rb_node) {
502                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
503                         rb_key += __cfqq->rb_key;
504                 } else
505                         rb_key += jiffies;
506         } else if (!add_front) {
507                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
508                 rb_key += cfqq->slice_resid;
509                 cfqq->slice_resid = 0;
510         } else
511                 rb_key = 0;
512
513         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
514                 /*
515                  * same position, nothing more to do
516                  */
517                 if (rb_key == cfqq->rb_key)
518                         return;
519
520                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
521         }
522
523         left = 1;
524         parent = NULL;
525         p = &cfqd->service_tree.rb.rb_node;
526         while (*p) {
527                 struct rb_node **n;
528
529                 parent = *p;
530                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
531
532                 /*
533                  * sort RT queues first, we always want to give
534                  * preference to them. IDLE queues goes to the back.
535                  * after that, sort on the next service time.
536                  */
537                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
538                         n = &(*p)->rb_left;
539                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
540                         n = &(*p)->rb_right;
541                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
542                         n = &(*p)->rb_left;
543                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
544                         n = &(*p)->rb_right;
545                 else if (rb_key < __cfqq->rb_key)
546                         n = &(*p)->rb_left;
547                 else
548                         n = &(*p)->rb_right;
549
550                 if (n == &(*p)->rb_right)
551                         left = 0;
552
553                 p = n;
554         }
555
556         if (left)
557                 cfqd->service_tree.left = &cfqq->rb_node;
558
559         cfqq->rb_key = rb_key;
560         rb_link_node(&cfqq->rb_node, parent, p);
561         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
562 }
563
564 static struct cfq_queue *
565 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
566                      sector_t sector, struct rb_node **ret_parent,
567                      struct rb_node ***rb_link)
568 {
569         struct rb_node **p, *parent;
570         struct cfq_queue *cfqq = NULL;
571
572         parent = NULL;
573         p = &root->rb_node;
574         while (*p) {
575                 struct rb_node **n;
576
577                 parent = *p;
578                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
579
580                 /*
581                  * Sort strictly based on sector.  Smallest to the left,
582                  * largest to the right.
583                  */
584                 if (sector > blk_rq_pos(cfqq->next_rq))
585                         n = &(*p)->rb_right;
586                 else if (sector < blk_rq_pos(cfqq->next_rq))
587                         n = &(*p)->rb_left;
588                 else
589                         break;
590                 p = n;
591                 cfqq = NULL;
592         }
593
594         *ret_parent = parent;
595         if (rb_link)
596                 *rb_link = p;
597         return cfqq;
598 }
599
600 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
601 {
602         struct rb_node **p, *parent;
603         struct cfq_queue *__cfqq;
604
605         if (cfqq->p_root) {
606                 rb_erase(&cfqq->p_node, cfqq->p_root);
607                 cfqq->p_root = NULL;
608         }
609
610         if (cfq_class_idle(cfqq))
611                 return;
612         if (!cfqq->next_rq)
613                 return;
614
615         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
616         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
617                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
618         if (!__cfqq) {
619                 rb_link_node(&cfqq->p_node, parent, p);
620                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
621         } else
622                 cfqq->p_root = NULL;
623 }
624
625 /*
626  * Update cfqq's position in the service tree.
627  */
628 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
629 {
630         /*
631          * Resorting requires the cfqq to be on the RR list already.
632          */
633         if (cfq_cfqq_on_rr(cfqq)) {
634                 cfq_service_tree_add(cfqd, cfqq, 0);
635                 cfq_prio_tree_add(cfqd, cfqq);
636         }
637 }
638
639 /*
640  * add to busy list of queues for service, trying to be fair in ordering
641  * the pending list according to last request service
642  */
643 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
644 {
645         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
646         BUG_ON(cfq_cfqq_on_rr(cfqq));
647         cfq_mark_cfqq_on_rr(cfqq);
648         cfqd->busy_queues++;
649
650         cfq_resort_rr_list(cfqd, cfqq);
651 }
652
653 /*
654  * Called when the cfqq no longer has requests pending, remove it from
655  * the service tree.
656  */
657 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
658 {
659         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
660         BUG_ON(!cfq_cfqq_on_rr(cfqq));
661         cfq_clear_cfqq_on_rr(cfqq);
662
663         if (!RB_EMPTY_NODE(&cfqq->rb_node))
664                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
665         if (cfqq->p_root) {
666                 rb_erase(&cfqq->p_node, cfqq->p_root);
667                 cfqq->p_root = NULL;
668         }
669
670         BUG_ON(!cfqd->busy_queues);
671         cfqd->busy_queues--;
672 }
673
674 /*
675  * rb tree support functions
676  */
677 static void cfq_del_rq_rb(struct request *rq)
678 {
679         struct cfq_queue *cfqq = RQ_CFQQ(rq);
680         struct cfq_data *cfqd = cfqq->cfqd;
681         const int sync = rq_is_sync(rq);
682
683         BUG_ON(!cfqq->queued[sync]);
684         cfqq->queued[sync]--;
685
686         elv_rb_del(&cfqq->sort_list, rq);
687
688         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
689                 cfq_del_cfqq_rr(cfqd, cfqq);
690 }
691
692 static void cfq_add_rq_rb(struct request *rq)
693 {
694         struct cfq_queue *cfqq = RQ_CFQQ(rq);
695         struct cfq_data *cfqd = cfqq->cfqd;
696         struct request *__alias, *prev;
697
698         cfqq->queued[rq_is_sync(rq)]++;
699
700         /*
701          * looks a little odd, but the first insert might return an alias.
702          * if that happens, put the alias on the dispatch list
703          */
704         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
705                 cfq_dispatch_insert(cfqd->queue, __alias);
706
707         if (!cfq_cfqq_on_rr(cfqq))
708                 cfq_add_cfqq_rr(cfqd, cfqq);
709
710         /*
711          * check if this request is a better next-serve candidate
712          */
713         prev = cfqq->next_rq;
714         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
715
716         /*
717          * adjust priority tree position, if ->next_rq changes
718          */
719         if (prev != cfqq->next_rq)
720                 cfq_prio_tree_add(cfqd, cfqq);
721
722         BUG_ON(!cfqq->next_rq);
723 }
724
725 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
726 {
727         elv_rb_del(&cfqq->sort_list, rq);
728         cfqq->queued[rq_is_sync(rq)]--;
729         cfq_add_rq_rb(rq);
730 }
731
732 static struct request *
733 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
734 {
735         struct task_struct *tsk = current;
736         struct cfq_io_context *cic;
737         struct cfq_queue *cfqq;
738
739         cic = cfq_cic_lookup(cfqd, tsk->io_context);
740         if (!cic)
741                 return NULL;
742
743         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
744         if (cfqq) {
745                 sector_t sector = bio->bi_sector + bio_sectors(bio);
746
747                 return elv_rb_find(&cfqq->sort_list, sector);
748         }
749
750         return NULL;
751 }
752
753 static void cfq_activate_request(struct request_queue *q, struct request *rq)
754 {
755         struct cfq_data *cfqd = q->elevator->elevator_data;
756
757         cfqd->rq_in_driver[rq_is_sync(rq)]++;
758         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
759                                                 rq_in_driver(cfqd));
760
761         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
762 }
763
764 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
765 {
766         struct cfq_data *cfqd = q->elevator->elevator_data;
767         const int sync = rq_is_sync(rq);
768
769         WARN_ON(!cfqd->rq_in_driver[sync]);
770         cfqd->rq_in_driver[sync]--;
771         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
772                                                 rq_in_driver(cfqd));
773 }
774
775 static void cfq_remove_request(struct request *rq)
776 {
777         struct cfq_queue *cfqq = RQ_CFQQ(rq);
778
779         if (cfqq->next_rq == rq)
780                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
781
782         list_del_init(&rq->queuelist);
783         cfq_del_rq_rb(rq);
784
785         cfqq->cfqd->rq_queued--;
786         if (rq_is_meta(rq)) {
787                 WARN_ON(!cfqq->meta_pending);
788                 cfqq->meta_pending--;
789         }
790 }
791
792 static int cfq_merge(struct request_queue *q, struct request **req,
793                      struct bio *bio)
794 {
795         struct cfq_data *cfqd = q->elevator->elevator_data;
796         struct request *__rq;
797
798         __rq = cfq_find_rq_fmerge(cfqd, bio);
799         if (__rq && elv_rq_merge_ok(__rq, bio)) {
800                 *req = __rq;
801                 return ELEVATOR_FRONT_MERGE;
802         }
803
804         return ELEVATOR_NO_MERGE;
805 }
806
807 static void cfq_merged_request(struct request_queue *q, struct request *req,
808                                int type)
809 {
810         if (type == ELEVATOR_FRONT_MERGE) {
811                 struct cfq_queue *cfqq = RQ_CFQQ(req);
812
813                 cfq_reposition_rq_rb(cfqq, req);
814         }
815 }
816
817 static void
818 cfq_merged_requests(struct request_queue *q, struct request *rq,
819                     struct request *next)
820 {
821         /*
822          * reposition in fifo if next is older than rq
823          */
824         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
825             time_before(next->start_time, rq->start_time))
826                 list_move(&rq->queuelist, &next->queuelist);
827
828         cfq_remove_request(next);
829 }
830
831 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
832                            struct bio *bio)
833 {
834         struct cfq_data *cfqd = q->elevator->elevator_data;
835         struct cfq_io_context *cic;
836         struct cfq_queue *cfqq;
837
838         /*
839          * Disallow merge of a sync bio into an async request.
840          */
841         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
842                 return 0;
843
844         /*
845          * Lookup the cfqq that this bio will be queued with. Allow
846          * merge only if rq is queued there.
847          */
848         cic = cfq_cic_lookup(cfqd, current->io_context);
849         if (!cic)
850                 return 0;
851
852         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
853         if (cfqq == RQ_CFQQ(rq))
854                 return 1;
855
856         return 0;
857 }
858
859 static void __cfq_set_active_queue(struct cfq_data *cfqd,
860                                    struct cfq_queue *cfqq)
861 {
862         if (cfqq) {
863                 cfq_log_cfqq(cfqd, cfqq, "set_active");
864                 cfqq->slice_end = 0;
865                 cfqq->slice_dispatch = 0;
866
867                 cfq_clear_cfqq_wait_request(cfqq);
868                 cfq_clear_cfqq_must_dispatch(cfqq);
869                 cfq_clear_cfqq_must_alloc_slice(cfqq);
870                 cfq_clear_cfqq_fifo_expire(cfqq);
871                 cfq_mark_cfqq_slice_new(cfqq);
872
873                 del_timer(&cfqd->idle_slice_timer);
874         }
875
876         cfqd->active_queue = cfqq;
877 }
878
879 /*
880  * current cfqq expired its slice (or was too idle), select new one
881  */
882 static void
883 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
884                     int timed_out)
885 {
886         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
887
888         if (cfq_cfqq_wait_request(cfqq))
889                 del_timer(&cfqd->idle_slice_timer);
890
891         cfq_clear_cfqq_wait_request(cfqq);
892
893         /*
894          * store what was left of this slice, if the queue idled/timed out
895          */
896         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
897                 cfqq->slice_resid = cfqq->slice_end - jiffies;
898                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
899         }
900
901         cfq_resort_rr_list(cfqd, cfqq);
902
903         if (cfqq == cfqd->active_queue)
904                 cfqd->active_queue = NULL;
905
906         if (cfqd->active_cic) {
907                 put_io_context(cfqd->active_cic->ioc);
908                 cfqd->active_cic = NULL;
909         }
910 }
911
912 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
913 {
914         struct cfq_queue *cfqq = cfqd->active_queue;
915
916         if (cfqq)
917                 __cfq_slice_expired(cfqd, cfqq, timed_out);
918 }
919
920 /*
921  * Get next queue for service. Unless we have a queue preemption,
922  * we'll simply select the first cfqq in the service tree.
923  */
924 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
925 {
926         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
927                 return NULL;
928
929         return cfq_rb_first(&cfqd->service_tree);
930 }
931
932 /*
933  * Get and set a new active queue for service.
934  */
935 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
936                                               struct cfq_queue *cfqq)
937 {
938         if (!cfqq) {
939                 cfqq = cfq_get_next_queue(cfqd);
940                 if (cfqq)
941                         cfq_clear_cfqq_coop(cfqq);
942         }
943
944         __cfq_set_active_queue(cfqd, cfqq);
945         return cfqq;
946 }
947
948 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
949                                           struct request *rq)
950 {
951         if (blk_rq_pos(rq) >= cfqd->last_position)
952                 return blk_rq_pos(rq) - cfqd->last_position;
953         else
954                 return cfqd->last_position - blk_rq_pos(rq);
955 }
956
957 #define CIC_SEEK_THR    8 * 1024
958 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
959
960 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
961 {
962         struct cfq_io_context *cic = cfqd->active_cic;
963         sector_t sdist = cic->seek_mean;
964
965         if (!sample_valid(cic->seek_samples))
966                 sdist = CIC_SEEK_THR;
967
968         return cfq_dist_from_last(cfqd, rq) <= sdist;
969 }
970
971 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
972                                     struct cfq_queue *cur_cfqq)
973 {
974         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
975         struct rb_node *parent, *node;
976         struct cfq_queue *__cfqq;
977         sector_t sector = cfqd->last_position;
978
979         if (RB_EMPTY_ROOT(root))
980                 return NULL;
981
982         /*
983          * First, if we find a request starting at the end of the last
984          * request, choose it.
985          */
986         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
987         if (__cfqq)
988                 return __cfqq;
989
990         /*
991          * If the exact sector wasn't found, the parent of the NULL leaf
992          * will contain the closest sector.
993          */
994         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
995         if (cfq_rq_close(cfqd, __cfqq->next_rq))
996                 return __cfqq;
997
998         if (blk_rq_pos(__cfqq->next_rq) < sector)
999                 node = rb_next(&__cfqq->p_node);
1000         else
1001                 node = rb_prev(&__cfqq->p_node);
1002         if (!node)
1003                 return NULL;
1004
1005         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1006         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1007                 return __cfqq;
1008
1009         return NULL;
1010 }
1011
1012 /*
1013  * cfqd - obvious
1014  * cur_cfqq - passed in so that we don't decide that the current queue is
1015  *            closely cooperating with itself.
1016  *
1017  * So, basically we're assuming that that cur_cfqq has dispatched at least
1018  * one request, and that cfqd->last_position reflects a position on the disk
1019  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1020  * assumption.
1021  */
1022 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1023                                               struct cfq_queue *cur_cfqq,
1024                                               int probe)
1025 {
1026         struct cfq_queue *cfqq;
1027
1028         /*
1029          * A valid cfq_io_context is necessary to compare requests against
1030          * the seek_mean of the current cfqq.
1031          */
1032         if (!cfqd->active_cic)
1033                 return NULL;
1034
1035         /*
1036          * We should notice if some of the queues are cooperating, eg
1037          * working closely on the same area of the disk. In that case,
1038          * we can group them together and don't waste time idling.
1039          */
1040         cfqq = cfqq_close(cfqd, cur_cfqq);
1041         if (!cfqq)
1042                 return NULL;
1043
1044         if (cfq_cfqq_coop(cfqq))
1045                 return NULL;
1046
1047         if (!probe)
1048                 cfq_mark_cfqq_coop(cfqq);
1049         return cfqq;
1050 }
1051
1052 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1053 {
1054         struct cfq_queue *cfqq = cfqd->active_queue;
1055         struct cfq_io_context *cic;
1056         unsigned long sl;
1057
1058         /*
1059          * SSD device without seek penalty, disable idling. But only do so
1060          * for devices that support queuing, otherwise we still have a problem
1061          * with sync vs async workloads.
1062          */
1063         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1064                 return;
1065
1066         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1067         WARN_ON(cfq_cfqq_slice_new(cfqq));
1068
1069         /*
1070          * idle is disabled, either manually or by past process history
1071          */
1072         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1073                 return;
1074
1075         /*
1076          * still requests with the driver, don't idle
1077          */
1078         if (rq_in_driver(cfqd))
1079                 return;
1080
1081         /*
1082          * task has exited, don't wait
1083          */
1084         cic = cfqd->active_cic;
1085         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1086                 return;
1087
1088         cfq_mark_cfqq_wait_request(cfqq);
1089
1090         /*
1091          * we don't want to idle for seeks, but we do want to allow
1092          * fair distribution of slice time for a process doing back-to-back
1093          * seeks. so allow a little bit of time for him to submit a new rq
1094          */
1095         sl = cfqd->cfq_slice_idle;
1096         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1097                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1098
1099         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1100         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1101 }
1102
1103 /*
1104  * Move request from internal lists to the request queue dispatch list.
1105  */
1106 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1107 {
1108         struct cfq_data *cfqd = q->elevator->elevator_data;
1109         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1110
1111         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1112
1113         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1114         cfq_remove_request(rq);
1115         cfqq->dispatched++;
1116         elv_dispatch_sort(q, rq);
1117
1118         if (cfq_cfqq_sync(cfqq))
1119                 cfqd->sync_flight++;
1120 }
1121
1122 /*
1123  * return expired entry, or NULL to just start from scratch in rbtree
1124  */
1125 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1126 {
1127         struct cfq_data *cfqd = cfqq->cfqd;
1128         struct request *rq;
1129         int fifo;
1130
1131         if (cfq_cfqq_fifo_expire(cfqq))
1132                 return NULL;
1133
1134         cfq_mark_cfqq_fifo_expire(cfqq);
1135
1136         if (list_empty(&cfqq->fifo))
1137                 return NULL;
1138
1139         fifo = cfq_cfqq_sync(cfqq);
1140         rq = rq_entry_fifo(cfqq->fifo.next);
1141
1142         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1143                 rq = NULL;
1144
1145         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1146         return rq;
1147 }
1148
1149 static inline int
1150 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1151 {
1152         const int base_rq = cfqd->cfq_slice_async_rq;
1153
1154         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1155
1156         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1157 }
1158
1159 /*
1160  * Select a queue for service. If we have a current active queue,
1161  * check whether to continue servicing it, or retrieve and set a new one.
1162  */
1163 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1164 {
1165         struct cfq_queue *cfqq, *new_cfqq = NULL;
1166
1167         cfqq = cfqd->active_queue;
1168         if (!cfqq)
1169                 goto new_queue;
1170
1171         /*
1172          * The active queue has run out of time, expire it and select new.
1173          */
1174         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1175                 goto expire;
1176
1177         /*
1178          * The active queue has requests and isn't expired, allow it to
1179          * dispatch.
1180          */
1181         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1182                 goto keep_queue;
1183
1184         /*
1185          * If another queue has a request waiting within our mean seek
1186          * distance, let it run.  The expire code will check for close
1187          * cooperators and put the close queue at the front of the service
1188          * tree.
1189          */
1190         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1191         if (new_cfqq)
1192                 goto expire;
1193
1194         /*
1195          * No requests pending. If the active queue still has requests in
1196          * flight or is idling for a new request, allow either of these
1197          * conditions to happen (or time out) before selecting a new queue.
1198          */
1199         if (timer_pending(&cfqd->idle_slice_timer) ||
1200             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1201                 cfqq = NULL;
1202                 goto keep_queue;
1203         }
1204
1205 expire:
1206         cfq_slice_expired(cfqd, 0);
1207 new_queue:
1208         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1209 keep_queue:
1210         return cfqq;
1211 }
1212
1213 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1214 {
1215         int dispatched = 0;
1216
1217         while (cfqq->next_rq) {
1218                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1219                 dispatched++;
1220         }
1221
1222         BUG_ON(!list_empty(&cfqq->fifo));
1223         return dispatched;
1224 }
1225
1226 /*
1227  * Drain our current requests. Used for barriers and when switching
1228  * io schedulers on-the-fly.
1229  */
1230 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1231 {
1232         struct cfq_queue *cfqq;
1233         int dispatched = 0;
1234
1235         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1236                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1237
1238         cfq_slice_expired(cfqd, 0);
1239
1240         BUG_ON(cfqd->busy_queues);
1241
1242         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1243         return dispatched;
1244 }
1245
1246 /*
1247  * Dispatch a request from cfqq, moving them to the request queue
1248  * dispatch list.
1249  */
1250 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1251 {
1252         struct request *rq;
1253
1254         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1255
1256         /*
1257          * follow expired path, else get first next available
1258          */
1259         rq = cfq_check_fifo(cfqq);
1260         if (!rq)
1261                 rq = cfqq->next_rq;
1262
1263         /*
1264          * insert request into driver dispatch list
1265          */
1266         cfq_dispatch_insert(cfqd->queue, rq);
1267
1268         if (!cfqd->active_cic) {
1269                 struct cfq_io_context *cic = RQ_CIC(rq);
1270
1271                 atomic_long_inc(&cic->ioc->refcount);
1272                 cfqd->active_cic = cic;
1273         }
1274 }
1275
1276 /*
1277  * Find the cfqq that we need to service and move a request from that to the
1278  * dispatch list
1279  */
1280 static int cfq_dispatch_requests(struct request_queue *q, int force)
1281 {
1282         struct cfq_data *cfqd = q->elevator->elevator_data;
1283         struct cfq_queue *cfqq;
1284         unsigned int max_dispatch;
1285
1286         if (!cfqd->busy_queues)
1287                 return 0;
1288
1289         if (unlikely(force))
1290                 return cfq_forced_dispatch(cfqd);
1291
1292         cfqq = cfq_select_queue(cfqd);
1293         if (!cfqq)
1294                 return 0;
1295
1296         /*
1297          * Drain async requests before we start sync IO
1298          */
1299         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1300                 return 0;
1301
1302         /*
1303          * If this is an async queue and we have sync IO in flight, let it wait
1304          */
1305         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1306                 return 0;
1307
1308         max_dispatch = cfqd->cfq_quantum;
1309         if (cfq_class_idle(cfqq))
1310                 max_dispatch = 1;
1311
1312         /*
1313          * Does this cfqq already have too much IO in flight?
1314          */
1315         if (cfqq->dispatched >= max_dispatch) {
1316                 /*
1317                  * idle queue must always only have a single IO in flight
1318                  */
1319                 if (cfq_class_idle(cfqq))
1320                         return 0;
1321
1322                 /*
1323                  * We have other queues, don't allow more IO from this one
1324                  */
1325                 if (cfqd->busy_queues > 1)
1326                         return 0;
1327
1328                 /*
1329                  * we are the only queue, allow up to 4 times of 'quantum'
1330                  */
1331                 if (cfqq->dispatched >= 4 * max_dispatch)
1332                         return 0;
1333         }
1334
1335         /*
1336          * Dispatch a request from this cfqq
1337          */
1338         cfq_dispatch_request(cfqd, cfqq);
1339         cfqq->slice_dispatch++;
1340         cfq_clear_cfqq_must_dispatch(cfqq);
1341
1342         /*
1343          * expire an async queue immediately if it has used up its slice. idle
1344          * queue always expire after 1 dispatch round.
1345          */
1346         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1347             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1348             cfq_class_idle(cfqq))) {
1349                 cfqq->slice_end = jiffies + 1;
1350                 cfq_slice_expired(cfqd, 0);
1351         }
1352
1353         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1354         return 1;
1355 }
1356
1357 /*
1358  * task holds one reference to the queue, dropped when task exits. each rq
1359  * in-flight on this queue also holds a reference, dropped when rq is freed.
1360  *
1361  * queue lock must be held here.
1362  */
1363 static void cfq_put_queue(struct cfq_queue *cfqq)
1364 {
1365         struct cfq_data *cfqd = cfqq->cfqd;
1366
1367         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1368
1369         if (!atomic_dec_and_test(&cfqq->ref))
1370                 return;
1371
1372         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1373         BUG_ON(rb_first(&cfqq->sort_list));
1374         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1375         BUG_ON(cfq_cfqq_on_rr(cfqq));
1376
1377         if (unlikely(cfqd->active_queue == cfqq)) {
1378                 __cfq_slice_expired(cfqd, cfqq, 0);
1379                 cfq_schedule_dispatch(cfqd);
1380         }
1381
1382         kmem_cache_free(cfq_pool, cfqq);
1383 }
1384
1385 /*
1386  * Must always be called with the rcu_read_lock() held
1387  */
1388 static void
1389 __call_for_each_cic(struct io_context *ioc,
1390                     void (*func)(struct io_context *, struct cfq_io_context *))
1391 {
1392         struct cfq_io_context *cic;
1393         struct hlist_node *n;
1394
1395         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1396                 func(ioc, cic);
1397 }
1398
1399 /*
1400  * Call func for each cic attached to this ioc.
1401  */
1402 static void
1403 call_for_each_cic(struct io_context *ioc,
1404                   void (*func)(struct io_context *, struct cfq_io_context *))
1405 {
1406         rcu_read_lock();
1407         __call_for_each_cic(ioc, func);
1408         rcu_read_unlock();
1409 }
1410
1411 static void cfq_cic_free_rcu(struct rcu_head *head)
1412 {
1413         struct cfq_io_context *cic;
1414
1415         cic = container_of(head, struct cfq_io_context, rcu_head);
1416
1417         kmem_cache_free(cfq_ioc_pool, cic);
1418         elv_ioc_count_dec(cfq_ioc_count);
1419
1420         if (ioc_gone) {
1421                 /*
1422                  * CFQ scheduler is exiting, grab exit lock and check
1423                  * the pending io context count. If it hits zero,
1424                  * complete ioc_gone and set it back to NULL
1425                  */
1426                 spin_lock(&ioc_gone_lock);
1427                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1428                         complete(ioc_gone);
1429                         ioc_gone = NULL;
1430                 }
1431                 spin_unlock(&ioc_gone_lock);
1432         }
1433 }
1434
1435 static void cfq_cic_free(struct cfq_io_context *cic)
1436 {
1437         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1438 }
1439
1440 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1441 {
1442         unsigned long flags;
1443
1444         BUG_ON(!cic->dead_key);
1445
1446         spin_lock_irqsave(&ioc->lock, flags);
1447         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1448         hlist_del_rcu(&cic->cic_list);
1449         spin_unlock_irqrestore(&ioc->lock, flags);
1450
1451         cfq_cic_free(cic);
1452 }
1453
1454 /*
1455  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1456  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1457  * and ->trim() which is called with the task lock held
1458  */
1459 static void cfq_free_io_context(struct io_context *ioc)
1460 {
1461         /*
1462          * ioc->refcount is zero here, or we are called from elv_unregister(),
1463          * so no more cic's are allowed to be linked into this ioc.  So it
1464          * should be ok to iterate over the known list, we will see all cic's
1465          * since no new ones are added.
1466          */
1467         __call_for_each_cic(ioc, cic_free_func);
1468 }
1469
1470 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1471 {
1472         if (unlikely(cfqq == cfqd->active_queue)) {
1473                 __cfq_slice_expired(cfqd, cfqq, 0);
1474                 cfq_schedule_dispatch(cfqd);
1475         }
1476
1477         cfq_put_queue(cfqq);
1478 }
1479
1480 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1481                                          struct cfq_io_context *cic)
1482 {
1483         struct io_context *ioc = cic->ioc;
1484
1485         list_del_init(&cic->queue_list);
1486
1487         /*
1488          * Make sure key == NULL is seen for dead queues
1489          */
1490         smp_wmb();
1491         cic->dead_key = (unsigned long) cic->key;
1492         cic->key = NULL;
1493
1494         if (ioc->ioc_data == cic)
1495                 rcu_assign_pointer(ioc->ioc_data, NULL);
1496
1497         if (cic->cfqq[BLK_RW_ASYNC]) {
1498                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1499                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1500         }
1501
1502         if (cic->cfqq[BLK_RW_SYNC]) {
1503                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1504                 cic->cfqq[BLK_RW_SYNC] = NULL;
1505         }
1506 }
1507
1508 static void cfq_exit_single_io_context(struct io_context *ioc,
1509                                        struct cfq_io_context *cic)
1510 {
1511         struct cfq_data *cfqd = cic->key;
1512
1513         if (cfqd) {
1514                 struct request_queue *q = cfqd->queue;
1515                 unsigned long flags;
1516
1517                 spin_lock_irqsave(q->queue_lock, flags);
1518
1519                 /*
1520                  * Ensure we get a fresh copy of the ->key to prevent
1521                  * race between exiting task and queue
1522                  */
1523                 smp_read_barrier_depends();
1524                 if (cic->key)
1525                         __cfq_exit_single_io_context(cfqd, cic);
1526
1527                 spin_unlock_irqrestore(q->queue_lock, flags);
1528         }
1529 }
1530
1531 /*
1532  * The process that ioc belongs to has exited, we need to clean up
1533  * and put the internal structures we have that belongs to that process.
1534  */
1535 static void cfq_exit_io_context(struct io_context *ioc)
1536 {
1537         call_for_each_cic(ioc, cfq_exit_single_io_context);
1538 }
1539
1540 static struct cfq_io_context *
1541 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1542 {
1543         struct cfq_io_context *cic;
1544
1545         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1546                                                         cfqd->queue->node);
1547         if (cic) {
1548                 cic->last_end_request = jiffies;
1549                 INIT_LIST_HEAD(&cic->queue_list);
1550                 INIT_HLIST_NODE(&cic->cic_list);
1551                 cic->dtor = cfq_free_io_context;
1552                 cic->exit = cfq_exit_io_context;
1553                 elv_ioc_count_inc(cfq_ioc_count);
1554         }
1555
1556         return cic;
1557 }
1558
1559 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1560 {
1561         struct task_struct *tsk = current;
1562         int ioprio_class;
1563
1564         if (!cfq_cfqq_prio_changed(cfqq))
1565                 return;
1566
1567         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1568         switch (ioprio_class) {
1569         default:
1570                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1571         case IOPRIO_CLASS_NONE:
1572                 /*
1573                  * no prio set, inherit CPU scheduling settings
1574                  */
1575                 cfqq->ioprio = task_nice_ioprio(tsk);
1576                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1577                 break;
1578         case IOPRIO_CLASS_RT:
1579                 cfqq->ioprio = task_ioprio(ioc);
1580                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1581                 break;
1582         case IOPRIO_CLASS_BE:
1583                 cfqq->ioprio = task_ioprio(ioc);
1584                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1585                 break;
1586         case IOPRIO_CLASS_IDLE:
1587                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1588                 cfqq->ioprio = 7;
1589                 cfq_clear_cfqq_idle_window(cfqq);
1590                 break;
1591         }
1592
1593         /*
1594          * keep track of original prio settings in case we have to temporarily
1595          * elevate the priority of this queue
1596          */
1597         cfqq->org_ioprio = cfqq->ioprio;
1598         cfqq->org_ioprio_class = cfqq->ioprio_class;
1599         cfq_clear_cfqq_prio_changed(cfqq);
1600 }
1601
1602 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1603 {
1604         struct cfq_data *cfqd = cic->key;
1605         struct cfq_queue *cfqq;
1606         unsigned long flags;
1607
1608         if (unlikely(!cfqd))
1609                 return;
1610
1611         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1612
1613         cfqq = cic->cfqq[BLK_RW_ASYNC];
1614         if (cfqq) {
1615                 struct cfq_queue *new_cfqq;
1616                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1617                                                 GFP_ATOMIC);
1618                 if (new_cfqq) {
1619                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1620                         cfq_put_queue(cfqq);
1621                 }
1622         }
1623
1624         cfqq = cic->cfqq[BLK_RW_SYNC];
1625         if (cfqq)
1626                 cfq_mark_cfqq_prio_changed(cfqq);
1627
1628         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1629 }
1630
1631 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1632 {
1633         call_for_each_cic(ioc, changed_ioprio);
1634         ioc->ioprio_changed = 0;
1635 }
1636
1637 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1638                           pid_t pid, int is_sync)
1639 {
1640         RB_CLEAR_NODE(&cfqq->rb_node);
1641         RB_CLEAR_NODE(&cfqq->p_node);
1642         INIT_LIST_HEAD(&cfqq->fifo);
1643
1644         atomic_set(&cfqq->ref, 0);
1645         cfqq->cfqd = cfqd;
1646
1647         cfq_mark_cfqq_prio_changed(cfqq);
1648
1649         if (is_sync) {
1650                 if (!cfq_class_idle(cfqq))
1651                         cfq_mark_cfqq_idle_window(cfqq);
1652                 cfq_mark_cfqq_sync(cfqq);
1653         }
1654         cfqq->pid = pid;
1655 }
1656
1657 static struct cfq_queue *
1658 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1659                      struct io_context *ioc, gfp_t gfp_mask)
1660 {
1661         struct cfq_queue *cfqq, *new_cfqq = NULL;
1662         struct cfq_io_context *cic;
1663
1664 retry:
1665         cic = cfq_cic_lookup(cfqd, ioc);
1666         /* cic always exists here */
1667         cfqq = cic_to_cfqq(cic, is_sync);
1668
1669         /*
1670          * Always try a new alloc if we fell back to the OOM cfqq
1671          * originally, since it should just be a temporary situation.
1672          */
1673         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1674                 cfqq = NULL;
1675                 if (new_cfqq) {
1676                         cfqq = new_cfqq;
1677                         new_cfqq = NULL;
1678                 } else if (gfp_mask & __GFP_WAIT) {
1679                         spin_unlock_irq(cfqd->queue->queue_lock);
1680                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1681                                         gfp_mask | __GFP_ZERO,
1682                                         cfqd->queue->node);
1683                         spin_lock_irq(cfqd->queue->queue_lock);
1684                         if (new_cfqq)
1685                                 goto retry;
1686                 } else {
1687                         cfqq = kmem_cache_alloc_node(cfq_pool,
1688                                         gfp_mask | __GFP_ZERO,
1689                                         cfqd->queue->node);
1690                 }
1691
1692                 if (cfqq) {
1693                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1694                         cfq_init_prio_data(cfqq, ioc);
1695                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1696                 } else
1697                         cfqq = &cfqd->oom_cfqq;
1698         }
1699
1700         if (new_cfqq)
1701                 kmem_cache_free(cfq_pool, new_cfqq);
1702
1703         return cfqq;
1704 }
1705
1706 static struct cfq_queue **
1707 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1708 {
1709         switch (ioprio_class) {
1710         case IOPRIO_CLASS_RT:
1711                 return &cfqd->async_cfqq[0][ioprio];
1712         case IOPRIO_CLASS_BE:
1713                 return &cfqd->async_cfqq[1][ioprio];
1714         case IOPRIO_CLASS_IDLE:
1715                 return &cfqd->async_idle_cfqq;
1716         default:
1717                 BUG();
1718         }
1719 }
1720
1721 static struct cfq_queue *
1722 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1723               gfp_t gfp_mask)
1724 {
1725         const int ioprio = task_ioprio(ioc);
1726         const int ioprio_class = task_ioprio_class(ioc);
1727         struct cfq_queue **async_cfqq = NULL;
1728         struct cfq_queue *cfqq = NULL;
1729
1730         if (!is_sync) {
1731                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1732                 cfqq = *async_cfqq;
1733         }
1734
1735         if (!cfqq)
1736                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1737
1738         /*
1739          * pin the queue now that it's allocated, scheduler exit will prune it
1740          */
1741         if (!is_sync && !(*async_cfqq)) {
1742                 atomic_inc(&cfqq->ref);
1743                 *async_cfqq = cfqq;
1744         }
1745
1746         atomic_inc(&cfqq->ref);
1747         return cfqq;
1748 }
1749
1750 /*
1751  * We drop cfq io contexts lazily, so we may find a dead one.
1752  */
1753 static void
1754 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1755                   struct cfq_io_context *cic)
1756 {
1757         unsigned long flags;
1758
1759         WARN_ON(!list_empty(&cic->queue_list));
1760
1761         spin_lock_irqsave(&ioc->lock, flags);
1762
1763         BUG_ON(ioc->ioc_data == cic);
1764
1765         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1766         hlist_del_rcu(&cic->cic_list);
1767         spin_unlock_irqrestore(&ioc->lock, flags);
1768
1769         cfq_cic_free(cic);
1770 }
1771
1772 static struct cfq_io_context *
1773 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1774 {
1775         struct cfq_io_context *cic;
1776         unsigned long flags;
1777         void *k;
1778
1779         if (unlikely(!ioc))
1780                 return NULL;
1781
1782         rcu_read_lock();
1783
1784         /*
1785          * we maintain a last-hit cache, to avoid browsing over the tree
1786          */
1787         cic = rcu_dereference(ioc->ioc_data);
1788         if (cic && cic->key == cfqd) {
1789                 rcu_read_unlock();
1790                 return cic;
1791         }
1792
1793         do {
1794                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1795                 rcu_read_unlock();
1796                 if (!cic)
1797                         break;
1798                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1799                 k = cic->key;
1800                 if (unlikely(!k)) {
1801                         cfq_drop_dead_cic(cfqd, ioc, cic);
1802                         rcu_read_lock();
1803                         continue;
1804                 }
1805
1806                 spin_lock_irqsave(&ioc->lock, flags);
1807                 rcu_assign_pointer(ioc->ioc_data, cic);
1808                 spin_unlock_irqrestore(&ioc->lock, flags);
1809                 break;
1810         } while (1);
1811
1812         return cic;
1813 }
1814
1815 /*
1816  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1817  * the process specific cfq io context when entered from the block layer.
1818  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1819  */
1820 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1821                         struct cfq_io_context *cic, gfp_t gfp_mask)
1822 {
1823         unsigned long flags;
1824         int ret;
1825
1826         ret = radix_tree_preload(gfp_mask);
1827         if (!ret) {
1828                 cic->ioc = ioc;
1829                 cic->key = cfqd;
1830
1831                 spin_lock_irqsave(&ioc->lock, flags);
1832                 ret = radix_tree_insert(&ioc->radix_root,
1833                                                 (unsigned long) cfqd, cic);
1834                 if (!ret)
1835                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1836                 spin_unlock_irqrestore(&ioc->lock, flags);
1837
1838                 radix_tree_preload_end();
1839
1840                 if (!ret) {
1841                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1842                         list_add(&cic->queue_list, &cfqd->cic_list);
1843                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1844                 }
1845         }
1846
1847         if (ret)
1848                 printk(KERN_ERR "cfq: cic link failed!\n");
1849
1850         return ret;
1851 }
1852
1853 /*
1854  * Setup general io context and cfq io context. There can be several cfq
1855  * io contexts per general io context, if this process is doing io to more
1856  * than one device managed by cfq.
1857  */
1858 static struct cfq_io_context *
1859 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1860 {
1861         struct io_context *ioc = NULL;
1862         struct cfq_io_context *cic;
1863
1864         might_sleep_if(gfp_mask & __GFP_WAIT);
1865
1866         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1867         if (!ioc)
1868                 return NULL;
1869
1870         cic = cfq_cic_lookup(cfqd, ioc);
1871         if (cic)
1872                 goto out;
1873
1874         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1875         if (cic == NULL)
1876                 goto err;
1877
1878         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1879                 goto err_free;
1880
1881 out:
1882         smp_read_barrier_depends();
1883         if (unlikely(ioc->ioprio_changed))
1884                 cfq_ioc_set_ioprio(ioc);
1885
1886         return cic;
1887 err_free:
1888         cfq_cic_free(cic);
1889 err:
1890         put_io_context(ioc);
1891         return NULL;
1892 }
1893
1894 static void
1895 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1896 {
1897         unsigned long elapsed = jiffies - cic->last_end_request;
1898         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1899
1900         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1901         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1902         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1903 }
1904
1905 static void
1906 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1907                        struct request *rq)
1908 {
1909         sector_t sdist;
1910         u64 total;
1911
1912         if (!cic->last_request_pos)
1913                 sdist = 0;
1914         else if (cic->last_request_pos < blk_rq_pos(rq))
1915                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1916         else
1917                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1918
1919         /*
1920          * Don't allow the seek distance to get too large from the
1921          * odd fragment, pagein, etc
1922          */
1923         if (cic->seek_samples <= 60) /* second&third seek */
1924                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1925         else
1926                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1927
1928         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1929         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1930         total = cic->seek_total + (cic->seek_samples/2);
1931         do_div(total, cic->seek_samples);
1932         cic->seek_mean = (sector_t)total;
1933 }
1934
1935 /*
1936  * Disable idle window if the process thinks too long or seeks so much that
1937  * it doesn't matter
1938  */
1939 static void
1940 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1941                        struct cfq_io_context *cic)
1942 {
1943         int old_idle, enable_idle;
1944
1945         /*
1946          * Don't idle for async or idle io prio class
1947          */
1948         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1949                 return;
1950
1951         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1952
1953         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1954             (cfqd->hw_tag && CIC_SEEKY(cic)))
1955                 enable_idle = 0;
1956         else if (sample_valid(cic->ttime_samples)) {
1957                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1958                         enable_idle = 0;
1959                 else
1960                         enable_idle = 1;
1961         }
1962
1963         if (old_idle != enable_idle) {
1964                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1965                 if (enable_idle)
1966                         cfq_mark_cfqq_idle_window(cfqq);
1967                 else
1968                         cfq_clear_cfqq_idle_window(cfqq);
1969         }
1970 }
1971
1972 /*
1973  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1974  * no or if we aren't sure, a 1 will cause a preempt.
1975  */
1976 static int
1977 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1978                    struct request *rq)
1979 {
1980         struct cfq_queue *cfqq;
1981
1982         cfqq = cfqd->active_queue;
1983         if (!cfqq)
1984                 return 0;
1985
1986         if (cfq_slice_used(cfqq))
1987                 return 1;
1988
1989         if (cfq_class_idle(new_cfqq))
1990                 return 0;
1991
1992         if (cfq_class_idle(cfqq))
1993                 return 1;
1994
1995         /*
1996          * if the new request is sync, but the currently running queue is
1997          * not, let the sync request have priority.
1998          */
1999         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2000                 return 1;
2001
2002         /*
2003          * So both queues are sync. Let the new request get disk time if
2004          * it's a metadata request and the current queue is doing regular IO.
2005          */
2006         if (rq_is_meta(rq) && !cfqq->meta_pending)
2007                 return 1;
2008
2009         /*
2010          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2011          */
2012         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2013                 return 1;
2014
2015         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2016                 return 0;
2017
2018         /*
2019          * if this request is as-good as one we would expect from the
2020          * current cfqq, let it preempt
2021          */
2022         if (cfq_rq_close(cfqd, rq))
2023                 return 1;
2024
2025         return 0;
2026 }
2027
2028 /*
2029  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2030  * let it have half of its nominal slice.
2031  */
2032 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2033 {
2034         cfq_log_cfqq(cfqd, cfqq, "preempt");
2035         cfq_slice_expired(cfqd, 1);
2036
2037         /*
2038          * Put the new queue at the front of the of the current list,
2039          * so we know that it will be selected next.
2040          */
2041         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2042
2043         cfq_service_tree_add(cfqd, cfqq, 1);
2044
2045         cfqq->slice_end = 0;
2046         cfq_mark_cfqq_slice_new(cfqq);
2047 }
2048
2049 /*
2050  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2051  * something we should do about it
2052  */
2053 static void
2054 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2055                 struct request *rq)
2056 {
2057         struct cfq_io_context *cic = RQ_CIC(rq);
2058
2059         cfqd->rq_queued++;
2060         if (rq_is_meta(rq))
2061                 cfqq->meta_pending++;
2062
2063         cfq_update_io_thinktime(cfqd, cic);
2064         cfq_update_io_seektime(cfqd, cic, rq);
2065         cfq_update_idle_window(cfqd, cfqq, cic);
2066
2067         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2068
2069         if (cfqq == cfqd->active_queue) {
2070                 /*
2071                  * Remember that we saw a request from this process, but
2072                  * don't start queuing just yet. Otherwise we risk seeing lots
2073                  * of tiny requests, because we disrupt the normal plugging
2074                  * and merging. If the request is already larger than a single
2075                  * page, let it rip immediately. For that case we assume that
2076                  * merging is already done. Ditto for a busy system that
2077                  * has other work pending, don't risk delaying until the
2078                  * idle timer unplug to continue working.
2079                  */
2080                 if (cfq_cfqq_wait_request(cfqq)) {
2081                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2082                             cfqd->busy_queues > 1) {
2083                                 del_timer(&cfqd->idle_slice_timer);
2084                         __blk_run_queue(cfqd->queue);
2085                         }
2086                         cfq_mark_cfqq_must_dispatch(cfqq);
2087                 }
2088         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2089                 /*
2090                  * not the active queue - expire current slice if it is
2091                  * idle and has expired it's mean thinktime or this new queue
2092                  * has some old slice time left and is of higher priority or
2093                  * this new queue is RT and the current one is BE
2094                  */
2095                 cfq_preempt_queue(cfqd, cfqq);
2096                 __blk_run_queue(cfqd->queue);
2097         }
2098 }
2099
2100 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2101 {
2102         struct cfq_data *cfqd = q->elevator->elevator_data;
2103         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2104
2105         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2106         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2107
2108         cfq_add_rq_rb(rq);
2109
2110         list_add_tail(&rq->queuelist, &cfqq->fifo);
2111
2112         cfq_rq_enqueued(cfqd, cfqq, rq);
2113 }
2114
2115 /*
2116  * Update hw_tag based on peak queue depth over 50 samples under
2117  * sufficient load.
2118  */
2119 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2120 {
2121         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2122                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2123
2124         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2125             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2126                 return;
2127
2128         if (cfqd->hw_tag_samples++ < 50)
2129                 return;
2130
2131         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2132                 cfqd->hw_tag = 1;
2133         else
2134                 cfqd->hw_tag = 0;
2135
2136         cfqd->hw_tag_samples = 0;
2137         cfqd->rq_in_driver_peak = 0;
2138 }
2139
2140 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2141 {
2142         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2143         struct cfq_data *cfqd = cfqq->cfqd;
2144         const int sync = rq_is_sync(rq);
2145         unsigned long now;
2146
2147         now = jiffies;
2148         cfq_log_cfqq(cfqd, cfqq, "complete");
2149
2150         cfq_update_hw_tag(cfqd);
2151
2152         WARN_ON(!cfqd->rq_in_driver[sync]);
2153         WARN_ON(!cfqq->dispatched);
2154         cfqd->rq_in_driver[sync]--;
2155         cfqq->dispatched--;
2156
2157         if (cfq_cfqq_sync(cfqq))
2158                 cfqd->sync_flight--;
2159
2160         if (sync)
2161                 RQ_CIC(rq)->last_end_request = now;
2162
2163         /*
2164          * If this is the active queue, check if it needs to be expired,
2165          * or if we want to idle in case it has no pending requests.
2166          */
2167         if (cfqd->active_queue == cfqq) {
2168                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2169
2170                 if (cfq_cfqq_slice_new(cfqq)) {
2171                         cfq_set_prio_slice(cfqd, cfqq);
2172                         cfq_clear_cfqq_slice_new(cfqq);
2173                 }
2174                 /*
2175                  * If there are no requests waiting in this queue, and
2176                  * there are other queues ready to issue requests, AND
2177                  * those other queues are issuing requests within our
2178                  * mean seek distance, give them a chance to run instead
2179                  * of idling.
2180                  */
2181                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2182                         cfq_slice_expired(cfqd, 1);
2183                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2184                          sync && !rq_noidle(rq))
2185                         cfq_arm_slice_timer(cfqd);
2186         }
2187
2188         if (!rq_in_driver(cfqd))
2189                 cfq_schedule_dispatch(cfqd);
2190 }
2191
2192 /*
2193  * we temporarily boost lower priority queues if they are holding fs exclusive
2194  * resources. they are boosted to normal prio (CLASS_BE/4)
2195  */
2196 static void cfq_prio_boost(struct cfq_queue *cfqq)
2197 {
2198         if (has_fs_excl()) {
2199                 /*
2200                  * boost idle prio on transactions that would lock out other
2201                  * users of the filesystem
2202                  */
2203                 if (cfq_class_idle(cfqq))
2204                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2205                 if (cfqq->ioprio > IOPRIO_NORM)
2206                         cfqq->ioprio = IOPRIO_NORM;
2207         } else {
2208                 /*
2209                  * check if we need to unboost the queue
2210                  */
2211                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2212                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2213                 if (cfqq->ioprio != cfqq->org_ioprio)
2214                         cfqq->ioprio = cfqq->org_ioprio;
2215         }
2216 }
2217
2218 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2219 {
2220         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2221                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2222                 return ELV_MQUEUE_MUST;
2223         }
2224
2225         return ELV_MQUEUE_MAY;
2226 }
2227
2228 static int cfq_may_queue(struct request_queue *q, int rw)
2229 {
2230         struct cfq_data *cfqd = q->elevator->elevator_data;
2231         struct task_struct *tsk = current;
2232         struct cfq_io_context *cic;
2233         struct cfq_queue *cfqq;
2234
2235         /*
2236          * don't force setup of a queue from here, as a call to may_queue
2237          * does not necessarily imply that a request actually will be queued.
2238          * so just lookup a possibly existing queue, or return 'may queue'
2239          * if that fails
2240          */
2241         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2242         if (!cic)
2243                 return ELV_MQUEUE_MAY;
2244
2245         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2246         if (cfqq) {
2247                 cfq_init_prio_data(cfqq, cic->ioc);
2248                 cfq_prio_boost(cfqq);
2249
2250                 return __cfq_may_queue(cfqq);
2251         }
2252
2253         return ELV_MQUEUE_MAY;
2254 }
2255
2256 /*
2257  * queue lock held here
2258  */
2259 static void cfq_put_request(struct request *rq)
2260 {
2261         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2262
2263         if (cfqq) {
2264                 const int rw = rq_data_dir(rq);
2265
2266                 BUG_ON(!cfqq->allocated[rw]);
2267                 cfqq->allocated[rw]--;
2268
2269                 put_io_context(RQ_CIC(rq)->ioc);
2270
2271                 rq->elevator_private = NULL;
2272                 rq->elevator_private2 = NULL;
2273
2274                 cfq_put_queue(cfqq);
2275         }
2276 }
2277
2278 /*
2279  * Allocate cfq data structures associated with this request.
2280  */
2281 static int
2282 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2283 {
2284         struct cfq_data *cfqd = q->elevator->elevator_data;
2285         struct cfq_io_context *cic;
2286         const int rw = rq_data_dir(rq);
2287         const int is_sync = rq_is_sync(rq);
2288         struct cfq_queue *cfqq;
2289         unsigned long flags;
2290
2291         might_sleep_if(gfp_mask & __GFP_WAIT);
2292
2293         cic = cfq_get_io_context(cfqd, gfp_mask);
2294
2295         spin_lock_irqsave(q->queue_lock, flags);
2296
2297         if (!cic)
2298                 goto queue_fail;
2299
2300         cfqq = cic_to_cfqq(cic, is_sync);
2301         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2302                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2303                 cic_set_cfqq(cic, cfqq, is_sync);
2304         }
2305
2306         cfqq->allocated[rw]++;
2307         atomic_inc(&cfqq->ref);
2308
2309         spin_unlock_irqrestore(q->queue_lock, flags);
2310
2311         rq->elevator_private = cic;
2312         rq->elevator_private2 = cfqq;
2313         return 0;
2314
2315 queue_fail:
2316         if (cic)
2317                 put_io_context(cic->ioc);
2318
2319         cfq_schedule_dispatch(cfqd);
2320         spin_unlock_irqrestore(q->queue_lock, flags);
2321         cfq_log(cfqd, "set_request fail");
2322         return 1;
2323 }
2324
2325 static void cfq_kick_queue(struct work_struct *work)
2326 {
2327         struct cfq_data *cfqd =
2328                 container_of(work, struct cfq_data, unplug_work);
2329         struct request_queue *q = cfqd->queue;
2330
2331         spin_lock_irq(q->queue_lock);
2332         __blk_run_queue(cfqd->queue);
2333         spin_unlock_irq(q->queue_lock);
2334 }
2335
2336 /*
2337  * Timer running if the active_queue is currently idling inside its time slice
2338  */
2339 static void cfq_idle_slice_timer(unsigned long data)
2340 {
2341         struct cfq_data *cfqd = (struct cfq_data *) data;
2342         struct cfq_queue *cfqq;
2343         unsigned long flags;
2344         int timed_out = 1;
2345
2346         cfq_log(cfqd, "idle timer fired");
2347
2348         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2349
2350         cfqq = cfqd->active_queue;
2351         if (cfqq) {
2352                 timed_out = 0;
2353
2354                 /*
2355                  * We saw a request before the queue expired, let it through
2356                  */
2357                 if (cfq_cfqq_must_dispatch(cfqq))
2358                         goto out_kick;
2359
2360                 /*
2361                  * expired
2362                  */
2363                 if (cfq_slice_used(cfqq))
2364                         goto expire;
2365
2366                 /*
2367                  * only expire and reinvoke request handler, if there are
2368                  * other queues with pending requests
2369                  */
2370                 if (!cfqd->busy_queues)
2371                         goto out_cont;
2372
2373                 /*
2374                  * not expired and it has a request pending, let it dispatch
2375                  */
2376                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2377                         goto out_kick;
2378         }
2379 expire:
2380         cfq_slice_expired(cfqd, timed_out);
2381 out_kick:
2382         cfq_schedule_dispatch(cfqd);
2383 out_cont:
2384         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2385 }
2386
2387 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2388 {
2389         del_timer_sync(&cfqd->idle_slice_timer);
2390         cancel_work_sync(&cfqd->unplug_work);
2391 }
2392
2393 static void cfq_put_async_queues(struct cfq_data *cfqd)
2394 {
2395         int i;
2396
2397         for (i = 0; i < IOPRIO_BE_NR; i++) {
2398                 if (cfqd->async_cfqq[0][i])
2399                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2400                 if (cfqd->async_cfqq[1][i])
2401                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2402         }
2403
2404         if (cfqd->async_idle_cfqq)
2405                 cfq_put_queue(cfqd->async_idle_cfqq);
2406 }
2407
2408 static void cfq_exit_queue(struct elevator_queue *e)
2409 {
2410         struct cfq_data *cfqd = e->elevator_data;
2411         struct request_queue *q = cfqd->queue;
2412
2413         cfq_shutdown_timer_wq(cfqd);
2414
2415         spin_lock_irq(q->queue_lock);
2416
2417         if (cfqd->active_queue)
2418                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2419
2420         while (!list_empty(&cfqd->cic_list)) {
2421                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2422                                                         struct cfq_io_context,
2423                                                         queue_list);
2424
2425                 __cfq_exit_single_io_context(cfqd, cic);
2426         }
2427
2428         cfq_put_async_queues(cfqd);
2429
2430         spin_unlock_irq(q->queue_lock);
2431
2432         cfq_shutdown_timer_wq(cfqd);
2433
2434         kfree(cfqd);
2435 }
2436
2437 static void *cfq_init_queue(struct request_queue *q)
2438 {
2439         struct cfq_data *cfqd;
2440         int i;
2441
2442         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2443         if (!cfqd)
2444                 return NULL;
2445
2446         cfqd->service_tree = CFQ_RB_ROOT;
2447
2448         /*
2449          * Not strictly needed (since RB_ROOT just clears the node and we
2450          * zeroed cfqd on alloc), but better be safe in case someone decides
2451          * to add magic to the rb code
2452          */
2453         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2454                 cfqd->prio_trees[i] = RB_ROOT;
2455
2456         /*
2457          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2458          * Grab a permanent reference to it, so that the normal code flow
2459          * will not attempt to free it.
2460          */
2461         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2462         atomic_inc(&cfqd->oom_cfqq.ref);
2463
2464         INIT_LIST_HEAD(&cfqd->cic_list);
2465
2466         cfqd->queue = q;
2467
2468         init_timer(&cfqd->idle_slice_timer);
2469         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2470         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2471
2472         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2473
2474         cfqd->cfq_quantum = cfq_quantum;
2475         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2476         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2477         cfqd->cfq_back_max = cfq_back_max;
2478         cfqd->cfq_back_penalty = cfq_back_penalty;
2479         cfqd->cfq_slice[0] = cfq_slice_async;
2480         cfqd->cfq_slice[1] = cfq_slice_sync;
2481         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2482         cfqd->cfq_slice_idle = cfq_slice_idle;
2483         cfqd->hw_tag = 1;
2484
2485         return cfqd;
2486 }
2487
2488 static void cfq_slab_kill(void)
2489 {
2490         /*
2491          * Caller already ensured that pending RCU callbacks are completed,
2492          * so we should have no busy allocations at this point.
2493          */
2494         if (cfq_pool)
2495                 kmem_cache_destroy(cfq_pool);
2496         if (cfq_ioc_pool)
2497                 kmem_cache_destroy(cfq_ioc_pool);
2498 }
2499
2500 static int __init cfq_slab_setup(void)
2501 {
2502         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2503         if (!cfq_pool)
2504                 goto fail;
2505
2506         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2507         if (!cfq_ioc_pool)
2508                 goto fail;
2509
2510         return 0;
2511 fail:
2512         cfq_slab_kill();
2513         return -ENOMEM;
2514 }
2515
2516 /*
2517  * sysfs parts below -->
2518  */
2519 static ssize_t
2520 cfq_var_show(unsigned int var, char *page)
2521 {
2522         return sprintf(page, "%d\n", var);
2523 }
2524
2525 static ssize_t
2526 cfq_var_store(unsigned int *var, const char *page, size_t count)
2527 {
2528         char *p = (char *) page;
2529
2530         *var = simple_strtoul(p, &p, 10);
2531         return count;
2532 }
2533
2534 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2535 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2536 {                                                                       \
2537         struct cfq_data *cfqd = e->elevator_data;                       \
2538         unsigned int __data = __VAR;                                    \
2539         if (__CONV)                                                     \
2540                 __data = jiffies_to_msecs(__data);                      \
2541         return cfq_var_show(__data, (page));                            \
2542 }
2543 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2544 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2545 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2546 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2547 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2548 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2549 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2550 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2551 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2552 #undef SHOW_FUNCTION
2553
2554 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2555 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2556 {                                                                       \
2557         struct cfq_data *cfqd = e->elevator_data;                       \
2558         unsigned int __data;                                            \
2559         int ret = cfq_var_store(&__data, (page), count);                \
2560         if (__data < (MIN))                                             \
2561                 __data = (MIN);                                         \
2562         else if (__data > (MAX))                                        \
2563                 __data = (MAX);                                         \
2564         if (__CONV)                                                     \
2565                 *(__PTR) = msecs_to_jiffies(__data);                    \
2566         else                                                            \
2567                 *(__PTR) = __data;                                      \
2568         return ret;                                                     \
2569 }
2570 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2571 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2572                 UINT_MAX, 1);
2573 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2574                 UINT_MAX, 1);
2575 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2576 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2577                 UINT_MAX, 0);
2578 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2579 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2580 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2581 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2582                 UINT_MAX, 0);
2583 #undef STORE_FUNCTION
2584
2585 #define CFQ_ATTR(name) \
2586         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2587
2588 static struct elv_fs_entry cfq_attrs[] = {
2589         CFQ_ATTR(quantum),
2590         CFQ_ATTR(fifo_expire_sync),
2591         CFQ_ATTR(fifo_expire_async),
2592         CFQ_ATTR(back_seek_max),
2593         CFQ_ATTR(back_seek_penalty),
2594         CFQ_ATTR(slice_sync),
2595         CFQ_ATTR(slice_async),
2596         CFQ_ATTR(slice_async_rq),
2597         CFQ_ATTR(slice_idle),
2598         __ATTR_NULL
2599 };
2600
2601 static struct elevator_type iosched_cfq = {
2602         .ops = {
2603                 .elevator_merge_fn =            cfq_merge,
2604                 .elevator_merged_fn =           cfq_merged_request,
2605                 .elevator_merge_req_fn =        cfq_merged_requests,
2606                 .elevator_allow_merge_fn =      cfq_allow_merge,
2607                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2608                 .elevator_add_req_fn =          cfq_insert_request,
2609                 .elevator_activate_req_fn =     cfq_activate_request,
2610                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2611                 .elevator_queue_empty_fn =      cfq_queue_empty,
2612                 .elevator_completed_req_fn =    cfq_completed_request,
2613                 .elevator_former_req_fn =       elv_rb_former_request,
2614                 .elevator_latter_req_fn =       elv_rb_latter_request,
2615                 .elevator_set_req_fn =          cfq_set_request,
2616                 .elevator_put_req_fn =          cfq_put_request,
2617                 .elevator_may_queue_fn =        cfq_may_queue,
2618                 .elevator_init_fn =             cfq_init_queue,
2619                 .elevator_exit_fn =             cfq_exit_queue,
2620                 .trim =                         cfq_free_io_context,
2621         },
2622         .elevator_attrs =       cfq_attrs,
2623         .elevator_name =        "cfq",
2624         .elevator_owner =       THIS_MODULE,
2625 };
2626
2627 static int __init cfq_init(void)
2628 {
2629         /*
2630          * could be 0 on HZ < 1000 setups
2631          */
2632         if (!cfq_slice_async)
2633                 cfq_slice_async = 1;
2634         if (!cfq_slice_idle)
2635                 cfq_slice_idle = 1;
2636
2637         if (cfq_slab_setup())
2638                 return -ENOMEM;
2639
2640         elv_register(&iosched_cfq);
2641
2642         return 0;
2643 }
2644
2645 static void __exit cfq_exit(void)
2646 {
2647         DECLARE_COMPLETION_ONSTACK(all_gone);
2648         elv_unregister(&iosched_cfq);
2649         ioc_gone = &all_gone;
2650         /* ioc_gone's update must be visible before reading ioc_count */
2651         smp_wmb();
2652
2653         /*
2654          * this also protects us from entering cfq_slab_kill() with
2655          * pending RCU callbacks
2656          */
2657         if (elv_ioc_count_read(cfq_ioc_count))
2658                 wait_for_completion(&all_gone);
2659         cfq_slab_kill();
2660 }
2661
2662 module_init(cfq_init);
2663 module_exit(cfq_exit);
2664
2665 MODULE_AUTHOR("Jens Axboe");
2666 MODULE_LICENSE("GPL");
2667 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");