cfq-iosched: add a knob for desktop interactiveness
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY          (HZ / 5)
35
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT              (2)
40
41 #define CFQ_SLICE_SCALE         (5)
42 #define CFQ_HW_QUEUE_MIN        (5)
43
44 #define RQ_CIC(rq)              \
45         ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
47
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50
51 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
59 #define sample_valid(samples)   ((samples) > 80)
60
61 /*
62  * Most of our rbtree usage is for sorting with min extraction, so
63  * if we cache the leftmost node we don't have to walk down the tree
64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65  * move this into the elevator for the rq sorting as well.
66  */
67 struct cfq_rb_root {
68         struct rb_root rb;
69         struct rb_node *left;
70 };
71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73 /*
74  * Per process-grouping structure
75  */
76 struct cfq_queue {
77         /* reference count */
78         atomic_t ref;
79         /* various state flags, see below */
80         unsigned int flags;
81         /* parent cfq_data */
82         struct cfq_data *cfqd;
83         /* service_tree member */
84         struct rb_node rb_node;
85         /* service_tree key */
86         unsigned long rb_key;
87         /* prio tree member */
88         struct rb_node p_node;
89         /* prio tree root we belong to, if any */
90         struct rb_root *p_root;
91         /* sorted list of pending requests */
92         struct rb_root sort_list;
93         /* if fifo isn't expired, next request to serve */
94         struct request *next_rq;
95         /* requests queued in sort_list */
96         int queued[2];
97         /* currently allocated requests */
98         int allocated[2];
99         /* fifo list of requests in sort_list */
100         struct list_head fifo;
101
102         unsigned long slice_end;
103         long slice_resid;
104         unsigned int slice_dispatch;
105
106         /* pending metadata requests */
107         int meta_pending;
108         /* number of requests that are on the dispatch list or inside driver */
109         int dispatched;
110
111         /* io prio of this group */
112         unsigned short ioprio, org_ioprio;
113         unsigned short ioprio_class, org_ioprio_class;
114
115         pid_t pid;
116 };
117
118 /*
119  * Per block device queue structure
120  */
121 struct cfq_data {
122         struct request_queue *queue;
123
124         /*
125          * rr list of queues with requests and the count of them
126          */
127         struct cfq_rb_root service_tree;
128
129         /*
130          * Each priority tree is sorted by next_request position.  These
131          * trees are used when determining if two or more queues are
132          * interleaving requests (see cfq_close_cooperator).
133          */
134         struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
136         unsigned int busy_queues;
137
138         int rq_in_driver[2];
139         int sync_flight;
140
141         /*
142          * queue-depth detection
143          */
144         int rq_queued;
145         int hw_tag;
146         int hw_tag_samples;
147         int rq_in_driver_peak;
148
149         /*
150          * idle window management
151          */
152         struct timer_list idle_slice_timer;
153         struct 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         unsigned int cfq_desktop;
177
178         struct list_head cic_list;
179
180         /*
181          * Fallback dummy cfqq for extreme OOM conditions
182          */
183         struct cfq_queue oom_cfqq;
184 };
185
186 enum cfqq_state_flags {
187         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
188         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
189         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
190         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
191         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
192         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
193         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
194         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
195         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
196         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
197 };
198
199 #define CFQ_CFQQ_FNS(name)                                              \
200 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
201 {                                                                       \
202         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
203 }                                                                       \
204 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
205 {                                                                       \
206         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
207 }                                                                       \
208 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
209 {                                                                       \
210         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
211 }
212
213 CFQ_CFQQ_FNS(on_rr);
214 CFQ_CFQQ_FNS(wait_request);
215 CFQ_CFQQ_FNS(must_dispatch);
216 CFQ_CFQQ_FNS(must_alloc_slice);
217 CFQ_CFQQ_FNS(fifo_expire);
218 CFQ_CFQQ_FNS(idle_window);
219 CFQ_CFQQ_FNS(prio_changed);
220 CFQ_CFQQ_FNS(slice_new);
221 CFQ_CFQQ_FNS(sync);
222 CFQ_CFQQ_FNS(coop);
223 #undef CFQ_CFQQ_FNS
224
225 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
226         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
227 #define cfq_log(cfqd, fmt, args...)     \
228         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
229
230 static void cfq_dispatch_insert(struct request_queue *, struct request *);
231 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
232                                        struct io_context *, gfp_t);
233 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
234                                                 struct io_context *);
235
236 static inline int rq_in_driver(struct cfq_data *cfqd)
237 {
238         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
239 }
240
241 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
242                                             int is_sync)
243 {
244         return cic->cfqq[!!is_sync];
245 }
246
247 static inline void cic_set_cfqq(struct cfq_io_context *cic,
248                                 struct cfq_queue *cfqq, int is_sync)
249 {
250         cic->cfqq[!!is_sync] = cfqq;
251 }
252
253 /*
254  * We regard a request as SYNC, if it's either a read or has the SYNC bit
255  * set (in which case it could also be direct WRITE).
256  */
257 static inline int cfq_bio_sync(struct bio *bio)
258 {
259         if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
260                 return 1;
261
262         return 0;
263 }
264
265 /*
266  * scheduler run of queue, if there are requests pending and no one in the
267  * driver that will restart queueing
268  */
269 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
270 {
271         if (cfqd->busy_queues) {
272                 cfq_log(cfqd, "schedule dispatch");
273                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
274         }
275 }
276
277 static int cfq_queue_empty(struct request_queue *q)
278 {
279         struct cfq_data *cfqd = q->elevator->elevator_data;
280
281         return !cfqd->busy_queues;
282 }
283
284 /*
285  * Scale schedule slice based on io priority. Use the sync time slice only
286  * if a queue is marked sync and has sync io queued. A sync queue with async
287  * io only, should not get full sync slice length.
288  */
289 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
290                                  unsigned short prio)
291 {
292         const int base_slice = cfqd->cfq_slice[sync];
293
294         WARN_ON(prio >= IOPRIO_BE_NR);
295
296         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
297 }
298
299 static inline int
300 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
301 {
302         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
303 }
304
305 static inline void
306 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
307 {
308         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
309         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
310 }
311
312 /*
313  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
314  * isn't valid until the first request from the dispatch is activated
315  * and the slice time set.
316  */
317 static inline int cfq_slice_used(struct cfq_queue *cfqq)
318 {
319         if (cfq_cfqq_slice_new(cfqq))
320                 return 0;
321         if (time_before(jiffies, cfqq->slice_end))
322                 return 0;
323
324         return 1;
325 }
326
327 /*
328  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
329  * We choose the request that is closest to the head right now. Distance
330  * behind the head is penalized and only allowed to a certain extent.
331  */
332 static struct request *
333 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
334 {
335         sector_t last, s1, s2, d1 = 0, d2 = 0;
336         unsigned long back_max;
337 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
338 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
339         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
340
341         if (rq1 == NULL || rq1 == rq2)
342                 return rq2;
343         if (rq2 == NULL)
344                 return rq1;
345
346         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
347                 return rq1;
348         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
349                 return rq2;
350         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
351                 return rq1;
352         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
353                 return rq2;
354
355         s1 = blk_rq_pos(rq1);
356         s2 = blk_rq_pos(rq2);
357
358         last = cfqd->last_position;
359
360         /*
361          * by definition, 1KiB is 2 sectors
362          */
363         back_max = cfqd->cfq_back_max * 2;
364
365         /*
366          * Strict one way elevator _except_ in the case where we allow
367          * short backward seeks which are biased as twice the cost of a
368          * similar forward seek.
369          */
370         if (s1 >= last)
371                 d1 = s1 - last;
372         else if (s1 + back_max >= last)
373                 d1 = (last - s1) * cfqd->cfq_back_penalty;
374         else
375                 wrap |= CFQ_RQ1_WRAP;
376
377         if (s2 >= last)
378                 d2 = s2 - last;
379         else if (s2 + back_max >= last)
380                 d2 = (last - s2) * cfqd->cfq_back_penalty;
381         else
382                 wrap |= CFQ_RQ2_WRAP;
383
384         /* Found required data */
385
386         /*
387          * By doing switch() on the bit mask "wrap" we avoid having to
388          * check two variables for all permutations: --> faster!
389          */
390         switch (wrap) {
391         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
392                 if (d1 < d2)
393                         return rq1;
394                 else if (d2 < d1)
395                         return rq2;
396                 else {
397                         if (s1 >= s2)
398                                 return rq1;
399                         else
400                                 return rq2;
401                 }
402
403         case CFQ_RQ2_WRAP:
404                 return rq1;
405         case CFQ_RQ1_WRAP:
406                 return rq2;
407         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
408         default:
409                 /*
410                  * Since both rqs are wrapped,
411                  * start with the one that's further behind head
412                  * (--> only *one* back seek required),
413                  * since back seek takes more time than forward.
414                  */
415                 if (s1 <= s2)
416                         return rq1;
417                 else
418                         return rq2;
419         }
420 }
421
422 /*
423  * The below is leftmost cache rbtree addon
424  */
425 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
426 {
427         if (!root->left)
428                 root->left = rb_first(&root->rb);
429
430         if (root->left)
431                 return rb_entry(root->left, struct cfq_queue, rb_node);
432
433         return NULL;
434 }
435
436 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
437 {
438         rb_erase(n, root);
439         RB_CLEAR_NODE(n);
440 }
441
442 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
443 {
444         if (root->left == n)
445                 root->left = NULL;
446         rb_erase_init(n, &root->rb);
447 }
448
449 /*
450  * would be nice to take fifo expire time into account as well
451  */
452 static struct request *
453 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
454                   struct request *last)
455 {
456         struct rb_node *rbnext = rb_next(&last->rb_node);
457         struct rb_node *rbprev = rb_prev(&last->rb_node);
458         struct request *next = NULL, *prev = NULL;
459
460         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
461
462         if (rbprev)
463                 prev = rb_entry_rq(rbprev);
464
465         if (rbnext)
466                 next = rb_entry_rq(rbnext);
467         else {
468                 rbnext = rb_first(&cfqq->sort_list);
469                 if (rbnext && rbnext != &last->rb_node)
470                         next = rb_entry_rq(rbnext);
471         }
472
473         return cfq_choose_req(cfqd, next, prev);
474 }
475
476 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
477                                       struct cfq_queue *cfqq)
478 {
479         /*
480          * just an approximation, should be ok.
481          */
482         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
483                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
484 }
485
486 /*
487  * The cfqd->service_tree holds all pending cfq_queue's that have
488  * requests waiting to be processed. It is sorted in the order that
489  * we will service the queues.
490  */
491 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
492                                  int add_front)
493 {
494         struct rb_node **p, *parent;
495         struct cfq_queue *__cfqq;
496         unsigned long rb_key;
497         int left;
498
499         if (cfq_class_idle(cfqq)) {
500                 rb_key = CFQ_IDLE_DELAY;
501                 parent = rb_last(&cfqd->service_tree.rb);
502                 if (parent && parent != &cfqq->rb_node) {
503                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
504                         rb_key += __cfqq->rb_key;
505                 } else
506                         rb_key += jiffies;
507         } else if (!add_front) {
508                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
509                 rb_key += cfqq->slice_resid;
510                 cfqq->slice_resid = 0;
511         } else
512                 rb_key = 0;
513
514         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
515                 /*
516                  * same position, nothing more to do
517                  */
518                 if (rb_key == cfqq->rb_key)
519                         return;
520
521                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
522         }
523
524         left = 1;
525         parent = NULL;
526         p = &cfqd->service_tree.rb.rb_node;
527         while (*p) {
528                 struct rb_node **n;
529
530                 parent = *p;
531                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
532
533                 /*
534                  * sort RT queues first, we always want to give
535                  * preference to them. IDLE queues goes to the back.
536                  * after that, sort on the next service time.
537                  */
538                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
539                         n = &(*p)->rb_left;
540                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
541                         n = &(*p)->rb_right;
542                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
543                         n = &(*p)->rb_left;
544                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
545                         n = &(*p)->rb_right;
546                 else if (rb_key < __cfqq->rb_key)
547                         n = &(*p)->rb_left;
548                 else
549                         n = &(*p)->rb_right;
550
551                 if (n == &(*p)->rb_right)
552                         left = 0;
553
554                 p = n;
555         }
556
557         if (left)
558                 cfqd->service_tree.left = &cfqq->rb_node;
559
560         cfqq->rb_key = rb_key;
561         rb_link_node(&cfqq->rb_node, parent, p);
562         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
563 }
564
565 static struct cfq_queue *
566 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
567                      sector_t sector, struct rb_node **ret_parent,
568                      struct rb_node ***rb_link)
569 {
570         struct rb_node **p, *parent;
571         struct cfq_queue *cfqq = NULL;
572
573         parent = NULL;
574         p = &root->rb_node;
575         while (*p) {
576                 struct rb_node **n;
577
578                 parent = *p;
579                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
580
581                 /*
582                  * Sort strictly based on sector.  Smallest to the left,
583                  * largest to the right.
584                  */
585                 if (sector > blk_rq_pos(cfqq->next_rq))
586                         n = &(*p)->rb_right;
587                 else if (sector < blk_rq_pos(cfqq->next_rq))
588                         n = &(*p)->rb_left;
589                 else
590                         break;
591                 p = n;
592                 cfqq = NULL;
593         }
594
595         *ret_parent = parent;
596         if (rb_link)
597                 *rb_link = p;
598         return cfqq;
599 }
600
601 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
602 {
603         struct rb_node **p, *parent;
604         struct cfq_queue *__cfqq;
605
606         if (cfqq->p_root) {
607                 rb_erase(&cfqq->p_node, cfqq->p_root);
608                 cfqq->p_root = NULL;
609         }
610
611         if (cfq_class_idle(cfqq))
612                 return;
613         if (!cfqq->next_rq)
614                 return;
615
616         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
617         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
618                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
619         if (!__cfqq) {
620                 rb_link_node(&cfqq->p_node, parent, p);
621                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
622         } else
623                 cfqq->p_root = NULL;
624 }
625
626 /*
627  * Update cfqq's position in the service tree.
628  */
629 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
630 {
631         /*
632          * Resorting requires the cfqq to be on the RR list already.
633          */
634         if (cfq_cfqq_on_rr(cfqq)) {
635                 cfq_service_tree_add(cfqd, cfqq, 0);
636                 cfq_prio_tree_add(cfqd, cfqq);
637         }
638 }
639
640 /*
641  * add to busy list of queues for service, trying to be fair in ordering
642  * the pending list according to last request service
643  */
644 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
645 {
646         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
647         BUG_ON(cfq_cfqq_on_rr(cfqq));
648         cfq_mark_cfqq_on_rr(cfqq);
649         cfqd->busy_queues++;
650
651         cfq_resort_rr_list(cfqd, cfqq);
652 }
653
654 /*
655  * Called when the cfqq no longer has requests pending, remove it from
656  * the service tree.
657  */
658 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
659 {
660         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
661         BUG_ON(!cfq_cfqq_on_rr(cfqq));
662         cfq_clear_cfqq_on_rr(cfqq);
663
664         if (!RB_EMPTY_NODE(&cfqq->rb_node))
665                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
666         if (cfqq->p_root) {
667                 rb_erase(&cfqq->p_node, cfqq->p_root);
668                 cfqq->p_root = NULL;
669         }
670
671         BUG_ON(!cfqd->busy_queues);
672         cfqd->busy_queues--;
673 }
674
675 /*
676  * rb tree support functions
677  */
678 static void cfq_del_rq_rb(struct request *rq)
679 {
680         struct cfq_queue *cfqq = RQ_CFQQ(rq);
681         struct cfq_data *cfqd = cfqq->cfqd;
682         const int sync = rq_is_sync(rq);
683
684         BUG_ON(!cfqq->queued[sync]);
685         cfqq->queued[sync]--;
686
687         elv_rb_del(&cfqq->sort_list, rq);
688
689         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
690                 cfq_del_cfqq_rr(cfqd, cfqq);
691 }
692
693 static void cfq_add_rq_rb(struct request *rq)
694 {
695         struct cfq_queue *cfqq = RQ_CFQQ(rq);
696         struct cfq_data *cfqd = cfqq->cfqd;
697         struct request *__alias, *prev;
698
699         cfqq->queued[rq_is_sync(rq)]++;
700
701         /*
702          * looks a little odd, but the first insert might return an alias.
703          * if that happens, put the alias on the dispatch list
704          */
705         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
706                 cfq_dispatch_insert(cfqd->queue, __alias);
707
708         if (!cfq_cfqq_on_rr(cfqq))
709                 cfq_add_cfqq_rr(cfqd, cfqq);
710
711         /*
712          * check if this request is a better next-serve candidate
713          */
714         prev = cfqq->next_rq;
715         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
716
717         /*
718          * adjust priority tree position, if ->next_rq changes
719          */
720         if (prev != cfqq->next_rq)
721                 cfq_prio_tree_add(cfqd, cfqq);
722
723         BUG_ON(!cfqq->next_rq);
724 }
725
726 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
727 {
728         elv_rb_del(&cfqq->sort_list, rq);
729         cfqq->queued[rq_is_sync(rq)]--;
730         cfq_add_rq_rb(rq);
731 }
732
733 static struct request *
734 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
735 {
736         struct task_struct *tsk = current;
737         struct cfq_io_context *cic;
738         struct cfq_queue *cfqq;
739
740         cic = cfq_cic_lookup(cfqd, tsk->io_context);
741         if (!cic)
742                 return NULL;
743
744         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
745         if (cfqq) {
746                 sector_t sector = bio->bi_sector + bio_sectors(bio);
747
748                 return elv_rb_find(&cfqq->sort_list, sector);
749         }
750
751         return NULL;
752 }
753
754 static void cfq_activate_request(struct request_queue *q, struct request *rq)
755 {
756         struct cfq_data *cfqd = q->elevator->elevator_data;
757
758         cfqd->rq_in_driver[rq_is_sync(rq)]++;
759         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
760                                                 rq_in_driver(cfqd));
761
762         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
763 }
764
765 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
766 {
767         struct cfq_data *cfqd = q->elevator->elevator_data;
768         const int sync = rq_is_sync(rq);
769
770         WARN_ON(!cfqd->rq_in_driver[sync]);
771         cfqd->rq_in_driver[sync]--;
772         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
773                                                 rq_in_driver(cfqd));
774 }
775
776 static void cfq_remove_request(struct request *rq)
777 {
778         struct cfq_queue *cfqq = RQ_CFQQ(rq);
779
780         if (cfqq->next_rq == rq)
781                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
782
783         list_del_init(&rq->queuelist);
784         cfq_del_rq_rb(rq);
785
786         cfqq->cfqd->rq_queued--;
787         if (rq_is_meta(rq)) {
788                 WARN_ON(!cfqq->meta_pending);
789                 cfqq->meta_pending--;
790         }
791 }
792
793 static int cfq_merge(struct request_queue *q, struct request **req,
794                      struct bio *bio)
795 {
796         struct cfq_data *cfqd = q->elevator->elevator_data;
797         struct request *__rq;
798
799         __rq = cfq_find_rq_fmerge(cfqd, bio);
800         if (__rq && elv_rq_merge_ok(__rq, bio)) {
801                 *req = __rq;
802                 return ELEVATOR_FRONT_MERGE;
803         }
804
805         return ELEVATOR_NO_MERGE;
806 }
807
808 static void cfq_merged_request(struct request_queue *q, struct request *req,
809                                int type)
810 {
811         if (type == ELEVATOR_FRONT_MERGE) {
812                 struct cfq_queue *cfqq = RQ_CFQQ(req);
813
814                 cfq_reposition_rq_rb(cfqq, req);
815         }
816 }
817
818 static void
819 cfq_merged_requests(struct request_queue *q, struct request *rq,
820                     struct request *next)
821 {
822         /*
823          * reposition in fifo if next is older than rq
824          */
825         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
826             time_before(next->start_time, rq->start_time))
827                 list_move(&rq->queuelist, &next->queuelist);
828
829         cfq_remove_request(next);
830 }
831
832 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
833                            struct bio *bio)
834 {
835         struct cfq_data *cfqd = q->elevator->elevator_data;
836         struct cfq_io_context *cic;
837         struct cfq_queue *cfqq;
838
839         /*
840          * Disallow merge of a sync bio into an async request.
841          */
842         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
843                 return 0;
844
845         /*
846          * Lookup the cfqq that this bio will be queued with. Allow
847          * merge only if rq is queued there.
848          */
849         cic = cfq_cic_lookup(cfqd, current->io_context);
850         if (!cic)
851                 return 0;
852
853         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
854         if (cfqq == RQ_CFQQ(rq))
855                 return 1;
856
857         return 0;
858 }
859
860 static void __cfq_set_active_queue(struct cfq_data *cfqd,
861                                    struct cfq_queue *cfqq)
862 {
863         if (cfqq) {
864                 cfq_log_cfqq(cfqd, cfqq, "set_active");
865                 cfqq->slice_end = 0;
866                 cfqq->slice_dispatch = 0;
867
868                 cfq_clear_cfqq_wait_request(cfqq);
869                 cfq_clear_cfqq_must_dispatch(cfqq);
870                 cfq_clear_cfqq_must_alloc_slice(cfqq);
871                 cfq_clear_cfqq_fifo_expire(cfqq);
872                 cfq_mark_cfqq_slice_new(cfqq);
873
874                 del_timer(&cfqd->idle_slice_timer);
875         }
876
877         cfqd->active_queue = cfqq;
878 }
879
880 /*
881  * current cfqq expired its slice (or was too idle), select new one
882  */
883 static void
884 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
885                     int timed_out)
886 {
887         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
888
889         if (cfq_cfqq_wait_request(cfqq))
890                 del_timer(&cfqd->idle_slice_timer);
891
892         cfq_clear_cfqq_wait_request(cfqq);
893
894         /*
895          * store what was left of this slice, if the queue idled/timed out
896          */
897         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
898                 cfqq->slice_resid = cfqq->slice_end - jiffies;
899                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
900         }
901
902         cfq_resort_rr_list(cfqd, cfqq);
903
904         if (cfqq == cfqd->active_queue)
905                 cfqd->active_queue = NULL;
906
907         if (cfqd->active_cic) {
908                 put_io_context(cfqd->active_cic->ioc);
909                 cfqd->active_cic = NULL;
910         }
911 }
912
913 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
914 {
915         struct cfq_queue *cfqq = cfqd->active_queue;
916
917         if (cfqq)
918                 __cfq_slice_expired(cfqd, cfqq, timed_out);
919 }
920
921 /*
922  * Get next queue for service. Unless we have a queue preemption,
923  * we'll simply select the first cfqq in the service tree.
924  */
925 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
926 {
927         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
928                 return NULL;
929
930         return cfq_rb_first(&cfqd->service_tree);
931 }
932
933 /*
934  * Get and set a new active queue for service.
935  */
936 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
937                                               struct cfq_queue *cfqq)
938 {
939         if (!cfqq) {
940                 cfqq = cfq_get_next_queue(cfqd);
941                 if (cfqq)
942                         cfq_clear_cfqq_coop(cfqq);
943         }
944
945         __cfq_set_active_queue(cfqd, cfqq);
946         return cfqq;
947 }
948
949 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
950                                           struct request *rq)
951 {
952         if (blk_rq_pos(rq) >= cfqd->last_position)
953                 return blk_rq_pos(rq) - cfqd->last_position;
954         else
955                 return cfqd->last_position - blk_rq_pos(rq);
956 }
957
958 #define CIC_SEEK_THR    8 * 1024
959 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
960
961 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
962 {
963         struct cfq_io_context *cic = cfqd->active_cic;
964         sector_t sdist = cic->seek_mean;
965
966         if (!sample_valid(cic->seek_samples))
967                 sdist = CIC_SEEK_THR;
968
969         return cfq_dist_from_last(cfqd, rq) <= sdist;
970 }
971
972 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
973                                     struct cfq_queue *cur_cfqq)
974 {
975         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
976         struct rb_node *parent, *node;
977         struct cfq_queue *__cfqq;
978         sector_t sector = cfqd->last_position;
979
980         if (RB_EMPTY_ROOT(root))
981                 return NULL;
982
983         /*
984          * First, if we find a request starting at the end of the last
985          * request, choose it.
986          */
987         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
988         if (__cfqq)
989                 return __cfqq;
990
991         /*
992          * If the exact sector wasn't found, the parent of the NULL leaf
993          * will contain the closest sector.
994          */
995         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
996         if (cfq_rq_close(cfqd, __cfqq->next_rq))
997                 return __cfqq;
998
999         if (blk_rq_pos(__cfqq->next_rq) < sector)
1000                 node = rb_next(&__cfqq->p_node);
1001         else
1002                 node = rb_prev(&__cfqq->p_node);
1003         if (!node)
1004                 return NULL;
1005
1006         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1007         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1008                 return __cfqq;
1009
1010         return NULL;
1011 }
1012
1013 /*
1014  * cfqd - obvious
1015  * cur_cfqq - passed in so that we don't decide that the current queue is
1016  *            closely cooperating with itself.
1017  *
1018  * So, basically we're assuming that that cur_cfqq has dispatched at least
1019  * one request, and that cfqd->last_position reflects a position on the disk
1020  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1021  * assumption.
1022  */
1023 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1024                                               struct cfq_queue *cur_cfqq,
1025                                               int probe)
1026 {
1027         struct cfq_queue *cfqq;
1028
1029         /*
1030          * A valid cfq_io_context is necessary to compare requests against
1031          * the seek_mean of the current cfqq.
1032          */
1033         if (!cfqd->active_cic)
1034                 return NULL;
1035
1036         /*
1037          * We should notice if some of the queues are cooperating, eg
1038          * working closely on the same area of the disk. In that case,
1039          * we can group them together and don't waste time idling.
1040          */
1041         cfqq = cfqq_close(cfqd, cur_cfqq);
1042         if (!cfqq)
1043                 return NULL;
1044
1045         if (cfq_cfqq_coop(cfqq))
1046                 return NULL;
1047
1048         if (!probe)
1049                 cfq_mark_cfqq_coop(cfqq);
1050         return cfqq;
1051 }
1052
1053 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1054 {
1055         struct cfq_queue *cfqq = cfqd->active_queue;
1056         struct cfq_io_context *cic;
1057         unsigned long sl;
1058
1059         /*
1060          * SSD device without seek penalty, disable idling. But only do so
1061          * for devices that support queuing, otherwise we still have a problem
1062          * with sync vs async workloads.
1063          */
1064         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1065                 return;
1066
1067         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1068         WARN_ON(cfq_cfqq_slice_new(cfqq));
1069
1070         /*
1071          * idle is disabled, either manually or by past process history
1072          */
1073         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1074                 return;
1075
1076         /*
1077          * still requests with the driver, don't idle
1078          */
1079         if (rq_in_driver(cfqd))
1080                 return;
1081
1082         /*
1083          * task has exited, don't wait
1084          */
1085         cic = cfqd->active_cic;
1086         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1087                 return;
1088
1089         cfq_mark_cfqq_wait_request(cfqq);
1090
1091         /*
1092          * we don't want to idle for seeks, but we do want to allow
1093          * fair distribution of slice time for a process doing back-to-back
1094          * seeks. so allow a little bit of time for him to submit a new rq
1095          */
1096         sl = cfqd->cfq_slice_idle;
1097         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1098                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1099
1100         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1101         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1102 }
1103
1104 /*
1105  * Move request from internal lists to the request queue dispatch list.
1106  */
1107 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1108 {
1109         struct cfq_data *cfqd = q->elevator->elevator_data;
1110         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1111
1112         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1113
1114         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1115         cfq_remove_request(rq);
1116         cfqq->dispatched++;
1117         elv_dispatch_sort(q, rq);
1118
1119         if (cfq_cfqq_sync(cfqq))
1120                 cfqd->sync_flight++;
1121 }
1122
1123 /*
1124  * return expired entry, or NULL to just start from scratch in rbtree
1125  */
1126 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1127 {
1128         struct cfq_data *cfqd = cfqq->cfqd;
1129         struct request *rq;
1130         int fifo;
1131
1132         if (cfq_cfqq_fifo_expire(cfqq))
1133                 return NULL;
1134
1135         cfq_mark_cfqq_fifo_expire(cfqq);
1136
1137         if (list_empty(&cfqq->fifo))
1138                 return NULL;
1139
1140         fifo = cfq_cfqq_sync(cfqq);
1141         rq = rq_entry_fifo(cfqq->fifo.next);
1142
1143         if (time_before(jiffies, rq->start_time + cfqd->cfq_fifo_expire[fifo]))
1144                 rq = NULL;
1145
1146         cfq_log_cfqq(cfqd, cfqq, "fifo=%p", rq);
1147         return rq;
1148 }
1149
1150 static inline int
1151 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1152 {
1153         const int base_rq = cfqd->cfq_slice_async_rq;
1154
1155         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1156
1157         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1158 }
1159
1160 /*
1161  * Select a queue for service. If we have a current active queue,
1162  * check whether to continue servicing it, or retrieve and set a new one.
1163  */
1164 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1165 {
1166         struct cfq_queue *cfqq, *new_cfqq = NULL;
1167
1168         cfqq = cfqd->active_queue;
1169         if (!cfqq)
1170                 goto new_queue;
1171
1172         /*
1173          * The active queue has run out of time, expire it and select new.
1174          */
1175         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1176                 goto expire;
1177
1178         /*
1179          * The active queue has requests and isn't expired, allow it to
1180          * dispatch.
1181          */
1182         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1183                 goto keep_queue;
1184
1185         /*
1186          * If another queue has a request waiting within our mean seek
1187          * distance, let it run.  The expire code will check for close
1188          * cooperators and put the close queue at the front of the service
1189          * tree.
1190          */
1191         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1192         if (new_cfqq)
1193                 goto expire;
1194
1195         /*
1196          * No requests pending. If the active queue still has requests in
1197          * flight or is idling for a new request, allow either of these
1198          * conditions to happen (or time out) before selecting a new queue.
1199          */
1200         if (timer_pending(&cfqd->idle_slice_timer) ||
1201             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1202                 cfqq = NULL;
1203                 goto keep_queue;
1204         }
1205
1206 expire:
1207         cfq_slice_expired(cfqd, 0);
1208 new_queue:
1209         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1210 keep_queue:
1211         return cfqq;
1212 }
1213
1214 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1215 {
1216         int dispatched = 0;
1217
1218         while (cfqq->next_rq) {
1219                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1220                 dispatched++;
1221         }
1222
1223         BUG_ON(!list_empty(&cfqq->fifo));
1224         return dispatched;
1225 }
1226
1227 /*
1228  * Drain our current requests. Used for barriers and when switching
1229  * io schedulers on-the-fly.
1230  */
1231 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1232 {
1233         struct cfq_queue *cfqq;
1234         int dispatched = 0;
1235
1236         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1237                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1238
1239         cfq_slice_expired(cfqd, 0);
1240
1241         BUG_ON(cfqd->busy_queues);
1242
1243         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1244         return dispatched;
1245 }
1246
1247 /*
1248  * Dispatch a request from cfqq, moving them to the request queue
1249  * dispatch list.
1250  */
1251 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1252 {
1253         struct request *rq;
1254
1255         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1256
1257         /*
1258          * follow expired path, else get first next available
1259          */
1260         rq = cfq_check_fifo(cfqq);
1261         if (!rq)
1262                 rq = cfqq->next_rq;
1263
1264         /*
1265          * insert request into driver dispatch list
1266          */
1267         cfq_dispatch_insert(cfqd->queue, rq);
1268
1269         if (!cfqd->active_cic) {
1270                 struct cfq_io_context *cic = RQ_CIC(rq);
1271
1272                 atomic_long_inc(&cic->ioc->refcount);
1273                 cfqd->active_cic = cic;
1274         }
1275 }
1276
1277 /*
1278  * Find the cfqq that we need to service and move a request from that to the
1279  * dispatch list
1280  */
1281 static int cfq_dispatch_requests(struct request_queue *q, int force)
1282 {
1283         struct cfq_data *cfqd = q->elevator->elevator_data;
1284         struct cfq_queue *cfqq;
1285         unsigned int max_dispatch;
1286
1287         if (!cfqd->busy_queues)
1288                 return 0;
1289
1290         if (unlikely(force))
1291                 return cfq_forced_dispatch(cfqd);
1292
1293         cfqq = cfq_select_queue(cfqd);
1294         if (!cfqq)
1295                 return 0;
1296
1297         /*
1298          * Drain async requests before we start sync IO
1299          */
1300         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1301                 return 0;
1302
1303         /*
1304          * If this is an async queue and we have sync IO in flight, let it wait
1305          */
1306         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1307                 return 0;
1308
1309         max_dispatch = cfqd->cfq_quantum;
1310         if (cfq_class_idle(cfqq))
1311                 max_dispatch = 1;
1312
1313         /*
1314          * Does this cfqq already have too much IO in flight?
1315          */
1316         if (cfqq->dispatched >= max_dispatch) {
1317                 /*
1318                  * idle queue must always only have a single IO in flight
1319                  */
1320                 if (cfq_class_idle(cfqq))
1321                         return 0;
1322
1323                 /*
1324                  * We have other queues, don't allow more IO from this one
1325                  */
1326                 if (cfqd->busy_queues > 1)
1327                         return 0;
1328
1329                 /*
1330                  * we are the only queue, allow up to 4 times of 'quantum'
1331                  */
1332                 if (cfqq->dispatched >= 4 * max_dispatch)
1333                         return 0;
1334         }
1335
1336         /*
1337          * Dispatch a request from this cfqq
1338          */
1339         cfq_dispatch_request(cfqd, cfqq);
1340         cfqq->slice_dispatch++;
1341         cfq_clear_cfqq_must_dispatch(cfqq);
1342
1343         /*
1344          * expire an async queue immediately if it has used up its slice. idle
1345          * queue always expire after 1 dispatch round.
1346          */
1347         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1348             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1349             cfq_class_idle(cfqq))) {
1350                 cfqq->slice_end = jiffies + 1;
1351                 cfq_slice_expired(cfqd, 0);
1352         }
1353
1354         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1355         return 1;
1356 }
1357
1358 /*
1359  * task holds one reference to the queue, dropped when task exits. each rq
1360  * in-flight on this queue also holds a reference, dropped when rq is freed.
1361  *
1362  * queue lock must be held here.
1363  */
1364 static void cfq_put_queue(struct cfq_queue *cfqq)
1365 {
1366         struct cfq_data *cfqd = cfqq->cfqd;
1367
1368         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1369
1370         if (!atomic_dec_and_test(&cfqq->ref))
1371                 return;
1372
1373         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1374         BUG_ON(rb_first(&cfqq->sort_list));
1375         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1376         BUG_ON(cfq_cfqq_on_rr(cfqq));
1377
1378         if (unlikely(cfqd->active_queue == cfqq)) {
1379                 __cfq_slice_expired(cfqd, cfqq, 0);
1380                 cfq_schedule_dispatch(cfqd);
1381         }
1382
1383         kmem_cache_free(cfq_pool, cfqq);
1384 }
1385
1386 /*
1387  * Must always be called with the rcu_read_lock() held
1388  */
1389 static void
1390 __call_for_each_cic(struct io_context *ioc,
1391                     void (*func)(struct io_context *, struct cfq_io_context *))
1392 {
1393         struct cfq_io_context *cic;
1394         struct hlist_node *n;
1395
1396         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1397                 func(ioc, cic);
1398 }
1399
1400 /*
1401  * Call func for each cic attached to this ioc.
1402  */
1403 static void
1404 call_for_each_cic(struct io_context *ioc,
1405                   void (*func)(struct io_context *, struct cfq_io_context *))
1406 {
1407         rcu_read_lock();
1408         __call_for_each_cic(ioc, func);
1409         rcu_read_unlock();
1410 }
1411
1412 static void cfq_cic_free_rcu(struct rcu_head *head)
1413 {
1414         struct cfq_io_context *cic;
1415
1416         cic = container_of(head, struct cfq_io_context, rcu_head);
1417
1418         kmem_cache_free(cfq_ioc_pool, cic);
1419         elv_ioc_count_dec(cfq_ioc_count);
1420
1421         if (ioc_gone) {
1422                 /*
1423                  * CFQ scheduler is exiting, grab exit lock and check
1424                  * the pending io context count. If it hits zero,
1425                  * complete ioc_gone and set it back to NULL
1426                  */
1427                 spin_lock(&ioc_gone_lock);
1428                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1429                         complete(ioc_gone);
1430                         ioc_gone = NULL;
1431                 }
1432                 spin_unlock(&ioc_gone_lock);
1433         }
1434 }
1435
1436 static void cfq_cic_free(struct cfq_io_context *cic)
1437 {
1438         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1439 }
1440
1441 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1442 {
1443         unsigned long flags;
1444
1445         BUG_ON(!cic->dead_key);
1446
1447         spin_lock_irqsave(&ioc->lock, flags);
1448         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1449         hlist_del_rcu(&cic->cic_list);
1450         spin_unlock_irqrestore(&ioc->lock, flags);
1451
1452         cfq_cic_free(cic);
1453 }
1454
1455 /*
1456  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1457  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1458  * and ->trim() which is called with the task lock held
1459  */
1460 static void cfq_free_io_context(struct io_context *ioc)
1461 {
1462         /*
1463          * ioc->refcount is zero here, or we are called from elv_unregister(),
1464          * so no more cic's are allowed to be linked into this ioc.  So it
1465          * should be ok to iterate over the known list, we will see all cic's
1466          * since no new ones are added.
1467          */
1468         __call_for_each_cic(ioc, cic_free_func);
1469 }
1470
1471 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1472 {
1473         if (unlikely(cfqq == cfqd->active_queue)) {
1474                 __cfq_slice_expired(cfqd, cfqq, 0);
1475                 cfq_schedule_dispatch(cfqd);
1476         }
1477
1478         cfq_put_queue(cfqq);
1479 }
1480
1481 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1482                                          struct cfq_io_context *cic)
1483 {
1484         struct io_context *ioc = cic->ioc;
1485
1486         list_del_init(&cic->queue_list);
1487
1488         /*
1489          * Make sure key == NULL is seen for dead queues
1490          */
1491         smp_wmb();
1492         cic->dead_key = (unsigned long) cic->key;
1493         cic->key = NULL;
1494
1495         if (ioc->ioc_data == cic)
1496                 rcu_assign_pointer(ioc->ioc_data, NULL);
1497
1498         if (cic->cfqq[BLK_RW_ASYNC]) {
1499                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1500                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1501         }
1502
1503         if (cic->cfqq[BLK_RW_SYNC]) {
1504                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1505                 cic->cfqq[BLK_RW_SYNC] = NULL;
1506         }
1507 }
1508
1509 static void cfq_exit_single_io_context(struct io_context *ioc,
1510                                        struct cfq_io_context *cic)
1511 {
1512         struct cfq_data *cfqd = cic->key;
1513
1514         if (cfqd) {
1515                 struct request_queue *q = cfqd->queue;
1516                 unsigned long flags;
1517
1518                 spin_lock_irqsave(q->queue_lock, flags);
1519
1520                 /*
1521                  * Ensure we get a fresh copy of the ->key to prevent
1522                  * race between exiting task and queue
1523                  */
1524                 smp_read_barrier_depends();
1525                 if (cic->key)
1526                         __cfq_exit_single_io_context(cfqd, cic);
1527
1528                 spin_unlock_irqrestore(q->queue_lock, flags);
1529         }
1530 }
1531
1532 /*
1533  * The process that ioc belongs to has exited, we need to clean up
1534  * and put the internal structures we have that belongs to that process.
1535  */
1536 static void cfq_exit_io_context(struct io_context *ioc)
1537 {
1538         call_for_each_cic(ioc, cfq_exit_single_io_context);
1539 }
1540
1541 static struct cfq_io_context *
1542 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1543 {
1544         struct cfq_io_context *cic;
1545
1546         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1547                                                         cfqd->queue->node);
1548         if (cic) {
1549                 cic->last_end_request = jiffies;
1550                 INIT_LIST_HEAD(&cic->queue_list);
1551                 INIT_HLIST_NODE(&cic->cic_list);
1552                 cic->dtor = cfq_free_io_context;
1553                 cic->exit = cfq_exit_io_context;
1554                 elv_ioc_count_inc(cfq_ioc_count);
1555         }
1556
1557         return cic;
1558 }
1559
1560 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1561 {
1562         struct task_struct *tsk = current;
1563         int ioprio_class;
1564
1565         if (!cfq_cfqq_prio_changed(cfqq))
1566                 return;
1567
1568         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1569         switch (ioprio_class) {
1570         default:
1571                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1572         case IOPRIO_CLASS_NONE:
1573                 /*
1574                  * no prio set, inherit CPU scheduling settings
1575                  */
1576                 cfqq->ioprio = task_nice_ioprio(tsk);
1577                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1578                 break;
1579         case IOPRIO_CLASS_RT:
1580                 cfqq->ioprio = task_ioprio(ioc);
1581                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1582                 break;
1583         case IOPRIO_CLASS_BE:
1584                 cfqq->ioprio = task_ioprio(ioc);
1585                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1586                 break;
1587         case IOPRIO_CLASS_IDLE:
1588                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1589                 cfqq->ioprio = 7;
1590                 cfq_clear_cfqq_idle_window(cfqq);
1591                 break;
1592         }
1593
1594         /*
1595          * keep track of original prio settings in case we have to temporarily
1596          * elevate the priority of this queue
1597          */
1598         cfqq->org_ioprio = cfqq->ioprio;
1599         cfqq->org_ioprio_class = cfqq->ioprio_class;
1600         cfq_clear_cfqq_prio_changed(cfqq);
1601 }
1602
1603 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1604 {
1605         struct cfq_data *cfqd = cic->key;
1606         struct cfq_queue *cfqq;
1607         unsigned long flags;
1608
1609         if (unlikely(!cfqd))
1610                 return;
1611
1612         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1613
1614         cfqq = cic->cfqq[BLK_RW_ASYNC];
1615         if (cfqq) {
1616                 struct cfq_queue *new_cfqq;
1617                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1618                                                 GFP_ATOMIC);
1619                 if (new_cfqq) {
1620                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1621                         cfq_put_queue(cfqq);
1622                 }
1623         }
1624
1625         cfqq = cic->cfqq[BLK_RW_SYNC];
1626         if (cfqq)
1627                 cfq_mark_cfqq_prio_changed(cfqq);
1628
1629         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1630 }
1631
1632 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1633 {
1634         call_for_each_cic(ioc, changed_ioprio);
1635         ioc->ioprio_changed = 0;
1636 }
1637
1638 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1639                           pid_t pid, int is_sync)
1640 {
1641         RB_CLEAR_NODE(&cfqq->rb_node);
1642         RB_CLEAR_NODE(&cfqq->p_node);
1643         INIT_LIST_HEAD(&cfqq->fifo);
1644
1645         atomic_set(&cfqq->ref, 0);
1646         cfqq->cfqd = cfqd;
1647
1648         cfq_mark_cfqq_prio_changed(cfqq);
1649
1650         if (is_sync) {
1651                 if (!cfq_class_idle(cfqq))
1652                         cfq_mark_cfqq_idle_window(cfqq);
1653                 cfq_mark_cfqq_sync(cfqq);
1654         }
1655         cfqq->pid = pid;
1656 }
1657
1658 static struct cfq_queue *
1659 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1660                      struct io_context *ioc, gfp_t gfp_mask)
1661 {
1662         struct cfq_queue *cfqq, *new_cfqq = NULL;
1663         struct cfq_io_context *cic;
1664
1665 retry:
1666         cic = cfq_cic_lookup(cfqd, ioc);
1667         /* cic always exists here */
1668         cfqq = cic_to_cfqq(cic, is_sync);
1669
1670         /*
1671          * Always try a new alloc if we fell back to the OOM cfqq
1672          * originally, since it should just be a temporary situation.
1673          */
1674         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1675                 cfqq = NULL;
1676                 if (new_cfqq) {
1677                         cfqq = new_cfqq;
1678                         new_cfqq = NULL;
1679                 } else if (gfp_mask & __GFP_WAIT) {
1680                         spin_unlock_irq(cfqd->queue->queue_lock);
1681                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1682                                         gfp_mask | __GFP_ZERO,
1683                                         cfqd->queue->node);
1684                         spin_lock_irq(cfqd->queue->queue_lock);
1685                         if (new_cfqq)
1686                                 goto retry;
1687                 } else {
1688                         cfqq = kmem_cache_alloc_node(cfq_pool,
1689                                         gfp_mask | __GFP_ZERO,
1690                                         cfqd->queue->node);
1691                 }
1692
1693                 if (cfqq) {
1694                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1695                         cfq_init_prio_data(cfqq, ioc);
1696                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1697                 } else
1698                         cfqq = &cfqd->oom_cfqq;
1699         }
1700
1701         if (new_cfqq)
1702                 kmem_cache_free(cfq_pool, new_cfqq);
1703
1704         return cfqq;
1705 }
1706
1707 static struct cfq_queue **
1708 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1709 {
1710         switch (ioprio_class) {
1711         case IOPRIO_CLASS_RT:
1712                 return &cfqd->async_cfqq[0][ioprio];
1713         case IOPRIO_CLASS_BE:
1714                 return &cfqd->async_cfqq[1][ioprio];
1715         case IOPRIO_CLASS_IDLE:
1716                 return &cfqd->async_idle_cfqq;
1717         default:
1718                 BUG();
1719         }
1720 }
1721
1722 static struct cfq_queue *
1723 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1724               gfp_t gfp_mask)
1725 {
1726         const int ioprio = task_ioprio(ioc);
1727         const int ioprio_class = task_ioprio_class(ioc);
1728         struct cfq_queue **async_cfqq = NULL;
1729         struct cfq_queue *cfqq = NULL;
1730
1731         if (!is_sync) {
1732                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1733                 cfqq = *async_cfqq;
1734         }
1735
1736         if (!cfqq)
1737                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1738
1739         /*
1740          * pin the queue now that it's allocated, scheduler exit will prune it
1741          */
1742         if (!is_sync && !(*async_cfqq)) {
1743                 atomic_inc(&cfqq->ref);
1744                 *async_cfqq = cfqq;
1745         }
1746
1747         atomic_inc(&cfqq->ref);
1748         return cfqq;
1749 }
1750
1751 /*
1752  * We drop cfq io contexts lazily, so we may find a dead one.
1753  */
1754 static void
1755 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1756                   struct cfq_io_context *cic)
1757 {
1758         unsigned long flags;
1759
1760         WARN_ON(!list_empty(&cic->queue_list));
1761
1762         spin_lock_irqsave(&ioc->lock, flags);
1763
1764         BUG_ON(ioc->ioc_data == cic);
1765
1766         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1767         hlist_del_rcu(&cic->cic_list);
1768         spin_unlock_irqrestore(&ioc->lock, flags);
1769
1770         cfq_cic_free(cic);
1771 }
1772
1773 static struct cfq_io_context *
1774 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1775 {
1776         struct cfq_io_context *cic;
1777         unsigned long flags;
1778         void *k;
1779
1780         if (unlikely(!ioc))
1781                 return NULL;
1782
1783         rcu_read_lock();
1784
1785         /*
1786          * we maintain a last-hit cache, to avoid browsing over the tree
1787          */
1788         cic = rcu_dereference(ioc->ioc_data);
1789         if (cic && cic->key == cfqd) {
1790                 rcu_read_unlock();
1791                 return cic;
1792         }
1793
1794         do {
1795                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1796                 rcu_read_unlock();
1797                 if (!cic)
1798                         break;
1799                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1800                 k = cic->key;
1801                 if (unlikely(!k)) {
1802                         cfq_drop_dead_cic(cfqd, ioc, cic);
1803                         rcu_read_lock();
1804                         continue;
1805                 }
1806
1807                 spin_lock_irqsave(&ioc->lock, flags);
1808                 rcu_assign_pointer(ioc->ioc_data, cic);
1809                 spin_unlock_irqrestore(&ioc->lock, flags);
1810                 break;
1811         } while (1);
1812
1813         return cic;
1814 }
1815
1816 /*
1817  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1818  * the process specific cfq io context when entered from the block layer.
1819  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1820  */
1821 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1822                         struct cfq_io_context *cic, gfp_t gfp_mask)
1823 {
1824         unsigned long flags;
1825         int ret;
1826
1827         ret = radix_tree_preload(gfp_mask);
1828         if (!ret) {
1829                 cic->ioc = ioc;
1830                 cic->key = cfqd;
1831
1832                 spin_lock_irqsave(&ioc->lock, flags);
1833                 ret = radix_tree_insert(&ioc->radix_root,
1834                                                 (unsigned long) cfqd, cic);
1835                 if (!ret)
1836                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1837                 spin_unlock_irqrestore(&ioc->lock, flags);
1838
1839                 radix_tree_preload_end();
1840
1841                 if (!ret) {
1842                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1843                         list_add(&cic->queue_list, &cfqd->cic_list);
1844                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1845                 }
1846         }
1847
1848         if (ret)
1849                 printk(KERN_ERR "cfq: cic link failed!\n");
1850
1851         return ret;
1852 }
1853
1854 /*
1855  * Setup general io context and cfq io context. There can be several cfq
1856  * io contexts per general io context, if this process is doing io to more
1857  * than one device managed by cfq.
1858  */
1859 static struct cfq_io_context *
1860 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1861 {
1862         struct io_context *ioc = NULL;
1863         struct cfq_io_context *cic;
1864
1865         might_sleep_if(gfp_mask & __GFP_WAIT);
1866
1867         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1868         if (!ioc)
1869                 return NULL;
1870
1871         cic = cfq_cic_lookup(cfqd, ioc);
1872         if (cic)
1873                 goto out;
1874
1875         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1876         if (cic == NULL)
1877                 goto err;
1878
1879         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1880                 goto err_free;
1881
1882 out:
1883         smp_read_barrier_depends();
1884         if (unlikely(ioc->ioprio_changed))
1885                 cfq_ioc_set_ioprio(ioc);
1886
1887         return cic;
1888 err_free:
1889         cfq_cic_free(cic);
1890 err:
1891         put_io_context(ioc);
1892         return NULL;
1893 }
1894
1895 static void
1896 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1897 {
1898         unsigned long elapsed = jiffies - cic->last_end_request;
1899         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1900
1901         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1902         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1903         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1904 }
1905
1906 static void
1907 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1908                        struct request *rq)
1909 {
1910         sector_t sdist;
1911         u64 total;
1912
1913         if (!cic->last_request_pos)
1914                 sdist = 0;
1915         else if (cic->last_request_pos < blk_rq_pos(rq))
1916                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1917         else
1918                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1919
1920         /*
1921          * Don't allow the seek distance to get too large from the
1922          * odd fragment, pagein, etc
1923          */
1924         if (cic->seek_samples <= 60) /* second&third seek */
1925                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1926         else
1927                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1928
1929         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1930         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1931         total = cic->seek_total + (cic->seek_samples/2);
1932         do_div(total, cic->seek_samples);
1933         cic->seek_mean = (sector_t)total;
1934 }
1935
1936 /*
1937  * Disable idle window if the process thinks too long or seeks so much that
1938  * it doesn't matter
1939  */
1940 static void
1941 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1942                        struct cfq_io_context *cic)
1943 {
1944         int old_idle, enable_idle;
1945
1946         /*
1947          * Don't idle for async or idle io prio class
1948          */
1949         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1950                 return;
1951
1952         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1953
1954         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1955             (!cfqd->cfq_desktop && cfqd->hw_tag && CIC_SEEKY(cic)))
1956                 enable_idle = 0;
1957         else if (sample_valid(cic->ttime_samples)) {
1958                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1959                         enable_idle = 0;
1960                 else
1961                         enable_idle = 1;
1962         }
1963
1964         if (old_idle != enable_idle) {
1965                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1966                 if (enable_idle)
1967                         cfq_mark_cfqq_idle_window(cfqq);
1968                 else
1969                         cfq_clear_cfqq_idle_window(cfqq);
1970         }
1971 }
1972
1973 /*
1974  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1975  * no or if we aren't sure, a 1 will cause a preempt.
1976  */
1977 static int
1978 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
1979                    struct request *rq)
1980 {
1981         struct cfq_queue *cfqq;
1982
1983         cfqq = cfqd->active_queue;
1984         if (!cfqq)
1985                 return 0;
1986
1987         if (cfq_slice_used(cfqq))
1988                 return 1;
1989
1990         if (cfq_class_idle(new_cfqq))
1991                 return 0;
1992
1993         if (cfq_class_idle(cfqq))
1994                 return 1;
1995
1996         /*
1997          * if the new request is sync, but the currently running queue is
1998          * not, let the sync request have priority.
1999          */
2000         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2001                 return 1;
2002
2003         /*
2004          * So both queues are sync. Let the new request get disk time if
2005          * it's a metadata request and the current queue is doing regular IO.
2006          */
2007         if (rq_is_meta(rq) && !cfqq->meta_pending)
2008                 return 1;
2009
2010         /*
2011          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2012          */
2013         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2014                 return 1;
2015
2016         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2017                 return 0;
2018
2019         /*
2020          * if this request is as-good as one we would expect from the
2021          * current cfqq, let it preempt
2022          */
2023         if (cfq_rq_close(cfqd, rq))
2024                 return 1;
2025
2026         return 0;
2027 }
2028
2029 /*
2030  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2031  * let it have half of its nominal slice.
2032  */
2033 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2034 {
2035         cfq_log_cfqq(cfqd, cfqq, "preempt");
2036         cfq_slice_expired(cfqd, 1);
2037
2038         /*
2039          * Put the new queue at the front of the of the current list,
2040          * so we know that it will be selected next.
2041          */
2042         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2043
2044         cfq_service_tree_add(cfqd, cfqq, 1);
2045
2046         cfqq->slice_end = 0;
2047         cfq_mark_cfqq_slice_new(cfqq);
2048 }
2049
2050 /*
2051  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2052  * something we should do about it
2053  */
2054 static void
2055 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2056                 struct request *rq)
2057 {
2058         struct cfq_io_context *cic = RQ_CIC(rq);
2059
2060         cfqd->rq_queued++;
2061         if (rq_is_meta(rq))
2062                 cfqq->meta_pending++;
2063
2064         cfq_update_io_thinktime(cfqd, cic);
2065         cfq_update_io_seektime(cfqd, cic, rq);
2066         cfq_update_idle_window(cfqd, cfqq, cic);
2067
2068         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2069
2070         if (cfqq == cfqd->active_queue) {
2071                 /*
2072                  * Remember that we saw a request from this process, but
2073                  * don't start queuing just yet. Otherwise we risk seeing lots
2074                  * of tiny requests, because we disrupt the normal plugging
2075                  * and merging. If the request is already larger than a single
2076                  * page, let it rip immediately. For that case we assume that
2077                  * merging is already done. Ditto for a busy system that
2078                  * has other work pending, don't risk delaying until the
2079                  * idle timer unplug to continue working.
2080                  */
2081                 if (cfq_cfqq_wait_request(cfqq)) {
2082                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2083                             cfqd->busy_queues > 1) {
2084                                 del_timer(&cfqd->idle_slice_timer);
2085                         __blk_run_queue(cfqd->queue);
2086                         }
2087                         cfq_mark_cfqq_must_dispatch(cfqq);
2088                 }
2089         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2090                 /*
2091                  * not the active queue - expire current slice if it is
2092                  * idle and has expired it's mean thinktime or this new queue
2093                  * has some old slice time left and is of higher priority or
2094                  * this new queue is RT and the current one is BE
2095                  */
2096                 cfq_preempt_queue(cfqd, cfqq);
2097                 __blk_run_queue(cfqd->queue);
2098         }
2099 }
2100
2101 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2102 {
2103         struct cfq_data *cfqd = q->elevator->elevator_data;
2104         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2105
2106         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2107         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2108
2109         cfq_add_rq_rb(rq);
2110
2111         list_add_tail(&rq->queuelist, &cfqq->fifo);
2112
2113         cfq_rq_enqueued(cfqd, cfqq, rq);
2114 }
2115
2116 /*
2117  * Update hw_tag based on peak queue depth over 50 samples under
2118  * sufficient load.
2119  */
2120 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2121 {
2122         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2123                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2124
2125         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2126             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2127                 return;
2128
2129         if (cfqd->hw_tag_samples++ < 50)
2130                 return;
2131
2132         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2133                 cfqd->hw_tag = 1;
2134         else
2135                 cfqd->hw_tag = 0;
2136
2137         cfqd->hw_tag_samples = 0;
2138         cfqd->rq_in_driver_peak = 0;
2139 }
2140
2141 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2142 {
2143         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2144         struct cfq_data *cfqd = cfqq->cfqd;
2145         const int sync = rq_is_sync(rq);
2146         unsigned long now;
2147
2148         now = jiffies;
2149         cfq_log_cfqq(cfqd, cfqq, "complete");
2150
2151         cfq_update_hw_tag(cfqd);
2152
2153         WARN_ON(!cfqd->rq_in_driver[sync]);
2154         WARN_ON(!cfqq->dispatched);
2155         cfqd->rq_in_driver[sync]--;
2156         cfqq->dispatched--;
2157
2158         if (cfq_cfqq_sync(cfqq))
2159                 cfqd->sync_flight--;
2160
2161         if (sync)
2162                 RQ_CIC(rq)->last_end_request = now;
2163
2164         /*
2165          * If this is the active queue, check if it needs to be expired,
2166          * or if we want to idle in case it has no pending requests.
2167          */
2168         if (cfqd->active_queue == cfqq) {
2169                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2170
2171                 if (cfq_cfqq_slice_new(cfqq)) {
2172                         cfq_set_prio_slice(cfqd, cfqq);
2173                         cfq_clear_cfqq_slice_new(cfqq);
2174                 }
2175                 /*
2176                  * If there are no requests waiting in this queue, and
2177                  * there are other queues ready to issue requests, AND
2178                  * those other queues are issuing requests within our
2179                  * mean seek distance, give them a chance to run instead
2180                  * of idling.
2181                  */
2182                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2183                         cfq_slice_expired(cfqd, 1);
2184                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2185                          sync && !rq_noidle(rq))
2186                         cfq_arm_slice_timer(cfqd);
2187         }
2188
2189         if (!rq_in_driver(cfqd))
2190                 cfq_schedule_dispatch(cfqd);
2191 }
2192
2193 /*
2194  * we temporarily boost lower priority queues if they are holding fs exclusive
2195  * resources. they are boosted to normal prio (CLASS_BE/4)
2196  */
2197 static void cfq_prio_boost(struct cfq_queue *cfqq)
2198 {
2199         if (has_fs_excl()) {
2200                 /*
2201                  * boost idle prio on transactions that would lock out other
2202                  * users of the filesystem
2203                  */
2204                 if (cfq_class_idle(cfqq))
2205                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2206                 if (cfqq->ioprio > IOPRIO_NORM)
2207                         cfqq->ioprio = IOPRIO_NORM;
2208         } else {
2209                 /*
2210                  * check if we need to unboost the queue
2211                  */
2212                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2213                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2214                 if (cfqq->ioprio != cfqq->org_ioprio)
2215                         cfqq->ioprio = cfqq->org_ioprio;
2216         }
2217 }
2218
2219 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2220 {
2221         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2222                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2223                 return ELV_MQUEUE_MUST;
2224         }
2225
2226         return ELV_MQUEUE_MAY;
2227 }
2228
2229 static int cfq_may_queue(struct request_queue *q, int rw)
2230 {
2231         struct cfq_data *cfqd = q->elevator->elevator_data;
2232         struct task_struct *tsk = current;
2233         struct cfq_io_context *cic;
2234         struct cfq_queue *cfqq;
2235
2236         /*
2237          * don't force setup of a queue from here, as a call to may_queue
2238          * does not necessarily imply that a request actually will be queued.
2239          * so just lookup a possibly existing queue, or return 'may queue'
2240          * if that fails
2241          */
2242         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2243         if (!cic)
2244                 return ELV_MQUEUE_MAY;
2245
2246         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2247         if (cfqq) {
2248                 cfq_init_prio_data(cfqq, cic->ioc);
2249                 cfq_prio_boost(cfqq);
2250
2251                 return __cfq_may_queue(cfqq);
2252         }
2253
2254         return ELV_MQUEUE_MAY;
2255 }
2256
2257 /*
2258  * queue lock held here
2259  */
2260 static void cfq_put_request(struct request *rq)
2261 {
2262         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2263
2264         if (cfqq) {
2265                 const int rw = rq_data_dir(rq);
2266
2267                 BUG_ON(!cfqq->allocated[rw]);
2268                 cfqq->allocated[rw]--;
2269
2270                 put_io_context(RQ_CIC(rq)->ioc);
2271
2272                 rq->elevator_private = NULL;
2273                 rq->elevator_private2 = NULL;
2274
2275                 cfq_put_queue(cfqq);
2276         }
2277 }
2278
2279 /*
2280  * Allocate cfq data structures associated with this request.
2281  */
2282 static int
2283 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2284 {
2285         struct cfq_data *cfqd = q->elevator->elevator_data;
2286         struct cfq_io_context *cic;
2287         const int rw = rq_data_dir(rq);
2288         const int is_sync = rq_is_sync(rq);
2289         struct cfq_queue *cfqq;
2290         unsigned long flags;
2291
2292         might_sleep_if(gfp_mask & __GFP_WAIT);
2293
2294         cic = cfq_get_io_context(cfqd, gfp_mask);
2295
2296         spin_lock_irqsave(q->queue_lock, flags);
2297
2298         if (!cic)
2299                 goto queue_fail;
2300
2301         cfqq = cic_to_cfqq(cic, is_sync);
2302         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2303                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2304                 cic_set_cfqq(cic, cfqq, is_sync);
2305         }
2306
2307         cfqq->allocated[rw]++;
2308         atomic_inc(&cfqq->ref);
2309
2310         spin_unlock_irqrestore(q->queue_lock, flags);
2311
2312         rq->elevator_private = cic;
2313         rq->elevator_private2 = cfqq;
2314         return 0;
2315
2316 queue_fail:
2317         if (cic)
2318                 put_io_context(cic->ioc);
2319
2320         cfq_schedule_dispatch(cfqd);
2321         spin_unlock_irqrestore(q->queue_lock, flags);
2322         cfq_log(cfqd, "set_request fail");
2323         return 1;
2324 }
2325
2326 static void cfq_kick_queue(struct work_struct *work)
2327 {
2328         struct cfq_data *cfqd =
2329                 container_of(work, struct cfq_data, unplug_work);
2330         struct request_queue *q = cfqd->queue;
2331
2332         spin_lock_irq(q->queue_lock);
2333         __blk_run_queue(cfqd->queue);
2334         spin_unlock_irq(q->queue_lock);
2335 }
2336
2337 /*
2338  * Timer running if the active_queue is currently idling inside its time slice
2339  */
2340 static void cfq_idle_slice_timer(unsigned long data)
2341 {
2342         struct cfq_data *cfqd = (struct cfq_data *) data;
2343         struct cfq_queue *cfqq;
2344         unsigned long flags;
2345         int timed_out = 1;
2346
2347         cfq_log(cfqd, "idle timer fired");
2348
2349         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2350
2351         cfqq = cfqd->active_queue;
2352         if (cfqq) {
2353                 timed_out = 0;
2354
2355                 /*
2356                  * We saw a request before the queue expired, let it through
2357                  */
2358                 if (cfq_cfqq_must_dispatch(cfqq))
2359                         goto out_kick;
2360
2361                 /*
2362                  * expired
2363                  */
2364                 if (cfq_slice_used(cfqq))
2365                         goto expire;
2366
2367                 /*
2368                  * only expire and reinvoke request handler, if there are
2369                  * other queues with pending requests
2370                  */
2371                 if (!cfqd->busy_queues)
2372                         goto out_cont;
2373
2374                 /*
2375                  * not expired and it has a request pending, let it dispatch
2376                  */
2377                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2378                         goto out_kick;
2379         }
2380 expire:
2381         cfq_slice_expired(cfqd, timed_out);
2382 out_kick:
2383         cfq_schedule_dispatch(cfqd);
2384 out_cont:
2385         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2386 }
2387
2388 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2389 {
2390         del_timer_sync(&cfqd->idle_slice_timer);
2391         cancel_work_sync(&cfqd->unplug_work);
2392 }
2393
2394 static void cfq_put_async_queues(struct cfq_data *cfqd)
2395 {
2396         int i;
2397
2398         for (i = 0; i < IOPRIO_BE_NR; i++) {
2399                 if (cfqd->async_cfqq[0][i])
2400                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2401                 if (cfqd->async_cfqq[1][i])
2402                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2403         }
2404
2405         if (cfqd->async_idle_cfqq)
2406                 cfq_put_queue(cfqd->async_idle_cfqq);
2407 }
2408
2409 static void cfq_exit_queue(struct elevator_queue *e)
2410 {
2411         struct cfq_data *cfqd = e->elevator_data;
2412         struct request_queue *q = cfqd->queue;
2413
2414         cfq_shutdown_timer_wq(cfqd);
2415
2416         spin_lock_irq(q->queue_lock);
2417
2418         if (cfqd->active_queue)
2419                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2420
2421         while (!list_empty(&cfqd->cic_list)) {
2422                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2423                                                         struct cfq_io_context,
2424                                                         queue_list);
2425
2426                 __cfq_exit_single_io_context(cfqd, cic);
2427         }
2428
2429         cfq_put_async_queues(cfqd);
2430
2431         spin_unlock_irq(q->queue_lock);
2432
2433         cfq_shutdown_timer_wq(cfqd);
2434
2435         kfree(cfqd);
2436 }
2437
2438 static void *cfq_init_queue(struct request_queue *q)
2439 {
2440         struct cfq_data *cfqd;
2441         int i;
2442
2443         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2444         if (!cfqd)
2445                 return NULL;
2446
2447         cfqd->service_tree = CFQ_RB_ROOT;
2448
2449         /*
2450          * Not strictly needed (since RB_ROOT just clears the node and we
2451          * zeroed cfqd on alloc), but better be safe in case someone decides
2452          * to add magic to the rb code
2453          */
2454         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2455                 cfqd->prio_trees[i] = RB_ROOT;
2456
2457         /*
2458          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2459          * Grab a permanent reference to it, so that the normal code flow
2460          * will not attempt to free it.
2461          */
2462         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2463         atomic_inc(&cfqd->oom_cfqq.ref);
2464
2465         INIT_LIST_HEAD(&cfqd->cic_list);
2466
2467         cfqd->queue = q;
2468
2469         init_timer(&cfqd->idle_slice_timer);
2470         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2471         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2472
2473         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2474
2475         cfqd->cfq_quantum = cfq_quantum;
2476         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2477         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2478         cfqd->cfq_back_max = cfq_back_max;
2479         cfqd->cfq_back_penalty = cfq_back_penalty;
2480         cfqd->cfq_slice[0] = cfq_slice_async;
2481         cfqd->cfq_slice[1] = cfq_slice_sync;
2482         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2483         cfqd->cfq_slice_idle = cfq_slice_idle;
2484         cfqd->cfq_desktop = 1;
2485         cfqd->hw_tag = 1;
2486
2487         return cfqd;
2488 }
2489
2490 static void cfq_slab_kill(void)
2491 {
2492         /*
2493          * Caller already ensured that pending RCU callbacks are completed,
2494          * so we should have no busy allocations at this point.
2495          */
2496         if (cfq_pool)
2497                 kmem_cache_destroy(cfq_pool);
2498         if (cfq_ioc_pool)
2499                 kmem_cache_destroy(cfq_ioc_pool);
2500 }
2501
2502 static int __init cfq_slab_setup(void)
2503 {
2504         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2505         if (!cfq_pool)
2506                 goto fail;
2507
2508         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2509         if (!cfq_ioc_pool)
2510                 goto fail;
2511
2512         return 0;
2513 fail:
2514         cfq_slab_kill();
2515         return -ENOMEM;
2516 }
2517
2518 /*
2519  * sysfs parts below -->
2520  */
2521 static ssize_t
2522 cfq_var_show(unsigned int var, char *page)
2523 {
2524         return sprintf(page, "%d\n", var);
2525 }
2526
2527 static ssize_t
2528 cfq_var_store(unsigned int *var, const char *page, size_t count)
2529 {
2530         char *p = (char *) page;
2531
2532         *var = simple_strtoul(p, &p, 10);
2533         return count;
2534 }
2535
2536 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2537 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2538 {                                                                       \
2539         struct cfq_data *cfqd = e->elevator_data;                       \
2540         unsigned int __data = __VAR;                                    \
2541         if (__CONV)                                                     \
2542                 __data = jiffies_to_msecs(__data);                      \
2543         return cfq_var_show(__data, (page));                            \
2544 }
2545 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2546 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2547 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2548 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2549 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2550 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2551 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2552 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2553 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2554 SHOW_FUNCTION(cfq_desktop_show, cfqd->cfq_desktop, 0);
2555 #undef SHOW_FUNCTION
2556
2557 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2558 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2559 {                                                                       \
2560         struct cfq_data *cfqd = e->elevator_data;                       \
2561         unsigned int __data;                                            \
2562         int ret = cfq_var_store(&__data, (page), count);                \
2563         if (__data < (MIN))                                             \
2564                 __data = (MIN);                                         \
2565         else if (__data > (MAX))                                        \
2566                 __data = (MAX);                                         \
2567         if (__CONV)                                                     \
2568                 *(__PTR) = msecs_to_jiffies(__data);                    \
2569         else                                                            \
2570                 *(__PTR) = __data;                                      \
2571         return ret;                                                     \
2572 }
2573 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2574 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2575                 UINT_MAX, 1);
2576 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2577                 UINT_MAX, 1);
2578 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2579 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2580                 UINT_MAX, 0);
2581 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2582 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2583 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2584 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2585                 UINT_MAX, 0);
2586 STORE_FUNCTION(cfq_desktop_store, &cfqd->cfq_desktop, 0, 1, 0);
2587 #undef STORE_FUNCTION
2588
2589 #define CFQ_ATTR(name) \
2590         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2591
2592 static struct elv_fs_entry cfq_attrs[] = {
2593         CFQ_ATTR(quantum),
2594         CFQ_ATTR(fifo_expire_sync),
2595         CFQ_ATTR(fifo_expire_async),
2596         CFQ_ATTR(back_seek_max),
2597         CFQ_ATTR(back_seek_penalty),
2598         CFQ_ATTR(slice_sync),
2599         CFQ_ATTR(slice_async),
2600         CFQ_ATTR(slice_async_rq),
2601         CFQ_ATTR(slice_idle),
2602         CFQ_ATTR(desktop),
2603         __ATTR_NULL
2604 };
2605
2606 static struct elevator_type iosched_cfq = {
2607         .ops = {
2608                 .elevator_merge_fn =            cfq_merge,
2609                 .elevator_merged_fn =           cfq_merged_request,
2610                 .elevator_merge_req_fn =        cfq_merged_requests,
2611                 .elevator_allow_merge_fn =      cfq_allow_merge,
2612                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2613                 .elevator_add_req_fn =          cfq_insert_request,
2614                 .elevator_activate_req_fn =     cfq_activate_request,
2615                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2616                 .elevator_queue_empty_fn =      cfq_queue_empty,
2617                 .elevator_completed_req_fn =    cfq_completed_request,
2618                 .elevator_former_req_fn =       elv_rb_former_request,
2619                 .elevator_latter_req_fn =       elv_rb_latter_request,
2620                 .elevator_set_req_fn =          cfq_set_request,
2621                 .elevator_put_req_fn =          cfq_put_request,
2622                 .elevator_may_queue_fn =        cfq_may_queue,
2623                 .elevator_init_fn =             cfq_init_queue,
2624                 .elevator_exit_fn =             cfq_exit_queue,
2625                 .trim =                         cfq_free_io_context,
2626         },
2627         .elevator_attrs =       cfq_attrs,
2628         .elevator_name =        "cfq",
2629         .elevator_owner =       THIS_MODULE,
2630 };
2631
2632 static int __init cfq_init(void)
2633 {
2634         /*
2635          * could be 0 on HZ < 1000 setups
2636          */
2637         if (!cfq_slice_async)
2638                 cfq_slice_async = 1;
2639         if (!cfq_slice_idle)
2640                 cfq_slice_idle = 1;
2641
2642         if (cfq_slab_setup())
2643                 return -ENOMEM;
2644
2645         elv_register(&iosched_cfq);
2646
2647         return 0;
2648 }
2649
2650 static void __exit cfq_exit(void)
2651 {
2652         DECLARE_COMPLETION_ONSTACK(all_gone);
2653         elv_unregister(&iosched_cfq);
2654         ioc_gone = &all_gone;
2655         /* ioc_gone's update must be visible before reading ioc_count */
2656         smp_wmb();
2657
2658         /*
2659          * this also protects us from entering cfq_slab_kill() with
2660          * pending RCU callbacks
2661          */
2662         if (elv_ioc_count_read(cfq_ioc_count))
2663                 wait_for_completion(&all_gone);
2664         cfq_slab_kill();
2665 }
2666
2667 module_init(cfq_init);
2668 module_exit(cfq_exit);
2669
2670 MODULE_AUTHOR("Jens Axboe");
2671 MODULE_LICENSE("GPL");
2672 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");