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