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