cfq-iosched: fix possible problem with jiffies wraparound
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY          (HZ / 5)
35
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT              (2)
40
41 #define CFQ_SLICE_SCALE         (5)
42 #define CFQ_HW_QUEUE_MIN        (5)
43
44 #define RQ_CIC(rq)              \
45         ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
47
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50
51 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
59 #define sample_valid(samples)   ((samples) > 80)
60
61 /*
62  * Most of our rbtree usage is for sorting with min extraction, so
63  * if we cache the leftmost node we don't have to walk down the tree
64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65  * move this into the elevator for the rq sorting as well.
66  */
67 struct cfq_rb_root {
68         struct rb_root rb;
69         struct rb_node *left;
70 };
71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73 /*
74  * Per process-grouping structure
75  */
76 struct cfq_queue {
77         /* reference count */
78         atomic_t ref;
79         /* various state flags, see below */
80         unsigned int flags;
81         /* parent cfq_data */
82         struct cfq_data *cfqd;
83         /* service_tree member */
84         struct rb_node rb_node;
85         /* service_tree key */
86         unsigned long rb_key;
87         /* prio tree member */
88         struct rb_node p_node;
89         /* prio tree root we belong to, if any */
90         struct rb_root *p_root;
91         /* sorted list of pending requests */
92         struct rb_root sort_list;
93         /* if fifo isn't expired, next request to serve */
94         struct request *next_rq;
95         /* requests queued in sort_list */
96         int queued[2];
97         /* currently allocated requests */
98         int allocated[2];
99         /* fifo list of requests in sort_list */
100         struct list_head fifo;
101
102         unsigned long slice_end;
103         long slice_resid;
104         unsigned int slice_dispatch;
105
106         /* pending metadata requests */
107         int meta_pending;
108         /* number of requests that are on the dispatch list or inside driver */
109         int dispatched;
110
111         /* io prio of this group */
112         unsigned short ioprio, org_ioprio;
113         unsigned short ioprio_class, org_ioprio_class;
114
115         pid_t pid;
116 };
117
118 /*
119  * Per block device queue structure
120  */
121 struct cfq_data {
122         struct request_queue *queue;
123
124         /*
125          * rr list of queues with requests and the count of them
126          */
127         struct cfq_rb_root service_tree;
128
129         /*
130          * Each priority tree is sorted by next_request position.  These
131          * trees are used when determining if two or more queues are
132          * interleaving requests (see cfq_close_cooperator).
133          */
134         struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
136         unsigned int busy_queues;
137
138         int rq_in_driver[2];
139         int sync_flight;
140
141         /*
142          * queue-depth detection
143          */
144         int rq_queued;
145         int hw_tag;
146         int hw_tag_samples;
147         int rq_in_driver_peak;
148
149         /*
150          * idle window management
151          */
152         struct timer_list idle_slice_timer;
153         struct delayed_work unplug_work;
154
155         struct cfq_queue *active_queue;
156         struct cfq_io_context *active_cic;
157
158         /*
159          * async queue for each priority case
160          */
161         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162         struct cfq_queue *async_idle_cfqq;
163
164         sector_t last_position;
165
166         /*
167          * tunables, see top of file
168          */
169         unsigned int cfq_quantum;
170         unsigned int cfq_fifo_expire[2];
171         unsigned int cfq_back_penalty;
172         unsigned int cfq_back_max;
173         unsigned int cfq_slice[2];
174         unsigned int cfq_slice_async_rq;
175         unsigned int cfq_slice_idle;
176         unsigned int cfq_latency;
177
178         struct list_head cic_list;
179
180         /*
181          * Fallback dummy cfqq for extreme OOM conditions
182          */
183         struct cfq_queue oom_cfqq;
184
185         unsigned long last_end_sync_rq;
186 };
187
188 enum cfqq_state_flags {
189         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
190         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
191         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
192         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
193         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
194         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
195         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
196         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
197         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
198         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
199 };
200
201 #define CFQ_CFQQ_FNS(name)                                              \
202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
203 {                                                                       \
204         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
205 }                                                                       \
206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
207 {                                                                       \
208         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
209 }                                                                       \
210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
211 {                                                                       \
212         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
213 }
214
215 CFQ_CFQQ_FNS(on_rr);
216 CFQ_CFQQ_FNS(wait_request);
217 CFQ_CFQQ_FNS(must_dispatch);
218 CFQ_CFQQ_FNS(must_alloc_slice);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
222 CFQ_CFQQ_FNS(slice_new);
223 CFQ_CFQQ_FNS(sync);
224 CFQ_CFQQ_FNS(coop);
225 #undef CFQ_CFQQ_FNS
226
227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
228         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229 #define cfq_log(cfqd, fmt, args...)     \
230         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231
232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, int,
234                                        struct io_context *, gfp_t);
235 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
236                                                 struct io_context *);
237
238 static inline int rq_in_driver(struct cfq_data *cfqd)
239 {
240         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
241 }
242
243 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
244                                             int is_sync)
245 {
246         return cic->cfqq[!!is_sync];
247 }
248
249 static inline void cic_set_cfqq(struct cfq_io_context *cic,
250                                 struct cfq_queue *cfqq, int is_sync)
251 {
252         cic->cfqq[!!is_sync] = cfqq;
253 }
254
255 /*
256  * We regard a request as SYNC, if it's either a read or has the SYNC bit
257  * set (in which case it could also be direct WRITE).
258  */
259 static inline int cfq_bio_sync(struct bio *bio)
260 {
261         if (bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO))
262                 return 1;
263
264         return 0;
265 }
266
267 /*
268  * scheduler run of queue, if there are requests pending and no one in the
269  * driver that will restart queueing
270  */
271 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd,
272                                          unsigned long delay)
273 {
274         if (cfqd->busy_queues) {
275                 cfq_log(cfqd, "schedule dispatch");
276                 kblockd_schedule_delayed_work(cfqd->queue, &cfqd->unplug_work,
277                                                 delay);
278         }
279 }
280
281 static int cfq_queue_empty(struct request_queue *q)
282 {
283         struct cfq_data *cfqd = q->elevator->elevator_data;
284
285         return !cfqd->busy_queues;
286 }
287
288 /*
289  * Scale schedule slice based on io priority. Use the sync time slice only
290  * if a queue is marked sync and has sync io queued. A sync queue with async
291  * io only, should not get full sync slice length.
292  */
293 static inline int cfq_prio_slice(struct cfq_data *cfqd, int sync,
294                                  unsigned short prio)
295 {
296         const int base_slice = cfqd->cfq_slice[sync];
297
298         WARN_ON(prio >= IOPRIO_BE_NR);
299
300         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
301 }
302
303 static inline int
304 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
305 {
306         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
307 }
308
309 static inline void
310 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
311 {
312         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
313         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
314 }
315
316 /*
317  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
318  * isn't valid until the first request from the dispatch is activated
319  * and the slice time set.
320  */
321 static inline int cfq_slice_used(struct cfq_queue *cfqq)
322 {
323         if (cfq_cfqq_slice_new(cfqq))
324                 return 0;
325         if (time_before(jiffies, cfqq->slice_end))
326                 return 0;
327
328         return 1;
329 }
330
331 /*
332  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
333  * We choose the request that is closest to the head right now. Distance
334  * behind the head is penalized and only allowed to a certain extent.
335  */
336 static struct request *
337 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
338 {
339         sector_t last, s1, s2, d1 = 0, d2 = 0;
340         unsigned long back_max;
341 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
342 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
343         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
344
345         if (rq1 == NULL || rq1 == rq2)
346                 return rq2;
347         if (rq2 == NULL)
348                 return rq1;
349
350         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
351                 return rq1;
352         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
353                 return rq2;
354         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
355                 return rq1;
356         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
357                 return rq2;
358
359         s1 = blk_rq_pos(rq1);
360         s2 = blk_rq_pos(rq2);
361
362         last = cfqd->last_position;
363
364         /*
365          * by definition, 1KiB is 2 sectors
366          */
367         back_max = cfqd->cfq_back_max * 2;
368
369         /*
370          * Strict one way elevator _except_ in the case where we allow
371          * short backward seeks which are biased as twice the cost of a
372          * similar forward seek.
373          */
374         if (s1 >= last)
375                 d1 = s1 - last;
376         else if (s1 + back_max >= last)
377                 d1 = (last - s1) * cfqd->cfq_back_penalty;
378         else
379                 wrap |= CFQ_RQ1_WRAP;
380
381         if (s2 >= last)
382                 d2 = s2 - last;
383         else if (s2 + back_max >= last)
384                 d2 = (last - s2) * cfqd->cfq_back_penalty;
385         else
386                 wrap |= CFQ_RQ2_WRAP;
387
388         /* Found required data */
389
390         /*
391          * By doing switch() on the bit mask "wrap" we avoid having to
392          * check two variables for all permutations: --> faster!
393          */
394         switch (wrap) {
395         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
396                 if (d1 < d2)
397                         return rq1;
398                 else if (d2 < d1)
399                         return rq2;
400                 else {
401                         if (s1 >= s2)
402                                 return rq1;
403                         else
404                                 return rq2;
405                 }
406
407         case CFQ_RQ2_WRAP:
408                 return rq1;
409         case CFQ_RQ1_WRAP:
410                 return rq2;
411         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
412         default:
413                 /*
414                  * Since both rqs are wrapped,
415                  * start with the one that's further behind head
416                  * (--> only *one* back seek required),
417                  * since back seek takes more time than forward.
418                  */
419                 if (s1 <= s2)
420                         return rq1;
421                 else
422                         return rq2;
423         }
424 }
425
426 /*
427  * The below is leftmost cache rbtree addon
428  */
429 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
430 {
431         if (!root->left)
432                 root->left = rb_first(&root->rb);
433
434         if (root->left)
435                 return rb_entry(root->left, struct cfq_queue, rb_node);
436
437         return NULL;
438 }
439
440 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
441 {
442         rb_erase(n, root);
443         RB_CLEAR_NODE(n);
444 }
445
446 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
447 {
448         if (root->left == n)
449                 root->left = NULL;
450         rb_erase_init(n, &root->rb);
451 }
452
453 /*
454  * would be nice to take fifo expire time into account as well
455  */
456 static struct request *
457 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
458                   struct request *last)
459 {
460         struct rb_node *rbnext = rb_next(&last->rb_node);
461         struct rb_node *rbprev = rb_prev(&last->rb_node);
462         struct request *next = NULL, *prev = NULL;
463
464         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
465
466         if (rbprev)
467                 prev = rb_entry_rq(rbprev);
468
469         if (rbnext)
470                 next = rb_entry_rq(rbnext);
471         else {
472                 rbnext = rb_first(&cfqq->sort_list);
473                 if (rbnext && rbnext != &last->rb_node)
474                         next = rb_entry_rq(rbnext);
475         }
476
477         return cfq_choose_req(cfqd, next, prev);
478 }
479
480 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
481                                       struct cfq_queue *cfqq)
482 {
483         /*
484          * just an approximation, should be ok.
485          */
486         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
487                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
488 }
489
490 /*
491  * The cfqd->service_tree holds all pending cfq_queue's that have
492  * requests waiting to be processed. It is sorted in the order that
493  * we will service the queues.
494  */
495 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
496                                  int add_front)
497 {
498         struct rb_node **p, *parent;
499         struct cfq_queue *__cfqq;
500         unsigned long rb_key;
501         int left;
502
503         if (cfq_class_idle(cfqq)) {
504                 rb_key = CFQ_IDLE_DELAY;
505                 parent = rb_last(&cfqd->service_tree.rb);
506                 if (parent && parent != &cfqq->rb_node) {
507                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
508                         rb_key += __cfqq->rb_key;
509                 } else
510                         rb_key += jiffies;
511         } else if (!add_front) {
512                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
513                 rb_key += cfqq->slice_resid;
514                 cfqq->slice_resid = 0;
515         } else {
516                 rb_key = -HZ;
517                 __cfqq = cfq_rb_first(&cfqd->service_tree);
518                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
519         }
520
521         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
522                 /*
523                  * same position, nothing more to do
524                  */
525                 if (rb_key == cfqq->rb_key)
526                         return;
527
528                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
529         }
530
531         left = 1;
532         parent = NULL;
533         p = &cfqd->service_tree.rb.rb_node;
534         while (*p) {
535                 struct rb_node **n;
536
537                 parent = *p;
538                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
539
540                 /*
541                  * sort RT queues first, we always want to give
542                  * preference to them. IDLE queues goes to the back.
543                  * after that, sort on the next service time.
544                  */
545                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
546                         n = &(*p)->rb_left;
547                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
548                         n = &(*p)->rb_right;
549                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
550                         n = &(*p)->rb_left;
551                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
552                         n = &(*p)->rb_right;
553                 else if (time_before(rb_key, __cfqq->rb_key))
554                         n = &(*p)->rb_left;
555                 else
556                         n = &(*p)->rb_right;
557
558                 if (n == &(*p)->rb_right)
559                         left = 0;
560
561                 p = n;
562         }
563
564         if (left)
565                 cfqd->service_tree.left = &cfqq->rb_node;
566
567         cfqq->rb_key = rb_key;
568         rb_link_node(&cfqq->rb_node, parent, p);
569         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
570 }
571
572 static struct cfq_queue *
573 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
574                      sector_t sector, struct rb_node **ret_parent,
575                      struct rb_node ***rb_link)
576 {
577         struct rb_node **p, *parent;
578         struct cfq_queue *cfqq = NULL;
579
580         parent = NULL;
581         p = &root->rb_node;
582         while (*p) {
583                 struct rb_node **n;
584
585                 parent = *p;
586                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
587
588                 /*
589                  * Sort strictly based on sector.  Smallest to the left,
590                  * largest to the right.
591                  */
592                 if (sector > blk_rq_pos(cfqq->next_rq))
593                         n = &(*p)->rb_right;
594                 else if (sector < blk_rq_pos(cfqq->next_rq))
595                         n = &(*p)->rb_left;
596                 else
597                         break;
598                 p = n;
599                 cfqq = NULL;
600         }
601
602         *ret_parent = parent;
603         if (rb_link)
604                 *rb_link = p;
605         return cfqq;
606 }
607
608 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
609 {
610         struct rb_node **p, *parent;
611         struct cfq_queue *__cfqq;
612
613         if (cfqq->p_root) {
614                 rb_erase(&cfqq->p_node, cfqq->p_root);
615                 cfqq->p_root = NULL;
616         }
617
618         if (cfq_class_idle(cfqq))
619                 return;
620         if (!cfqq->next_rq)
621                 return;
622
623         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
624         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
625                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
626         if (!__cfqq) {
627                 rb_link_node(&cfqq->p_node, parent, p);
628                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
629         } else
630                 cfqq->p_root = NULL;
631 }
632
633 /*
634  * Update cfqq's position in the service tree.
635  */
636 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
637 {
638         /*
639          * Resorting requires the cfqq to be on the RR list already.
640          */
641         if (cfq_cfqq_on_rr(cfqq)) {
642                 cfq_service_tree_add(cfqd, cfqq, 0);
643                 cfq_prio_tree_add(cfqd, cfqq);
644         }
645 }
646
647 /*
648  * add to busy list of queues for service, trying to be fair in ordering
649  * the pending list according to last request service
650  */
651 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
652 {
653         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
654         BUG_ON(cfq_cfqq_on_rr(cfqq));
655         cfq_mark_cfqq_on_rr(cfqq);
656         cfqd->busy_queues++;
657
658         cfq_resort_rr_list(cfqd, cfqq);
659 }
660
661 /*
662  * Called when the cfqq no longer has requests pending, remove it from
663  * the service tree.
664  */
665 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
666 {
667         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
668         BUG_ON(!cfq_cfqq_on_rr(cfqq));
669         cfq_clear_cfqq_on_rr(cfqq);
670
671         if (!RB_EMPTY_NODE(&cfqq->rb_node))
672                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
673         if (cfqq->p_root) {
674                 rb_erase(&cfqq->p_node, cfqq->p_root);
675                 cfqq->p_root = NULL;
676         }
677
678         BUG_ON(!cfqd->busy_queues);
679         cfqd->busy_queues--;
680 }
681
682 /*
683  * rb tree support functions
684  */
685 static void cfq_del_rq_rb(struct request *rq)
686 {
687         struct cfq_queue *cfqq = RQ_CFQQ(rq);
688         struct cfq_data *cfqd = cfqq->cfqd;
689         const int sync = rq_is_sync(rq);
690
691         BUG_ON(!cfqq->queued[sync]);
692         cfqq->queued[sync]--;
693
694         elv_rb_del(&cfqq->sort_list, rq);
695
696         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
697                 cfq_del_cfqq_rr(cfqd, cfqq);
698 }
699
700 static void cfq_add_rq_rb(struct request *rq)
701 {
702         struct cfq_queue *cfqq = RQ_CFQQ(rq);
703         struct cfq_data *cfqd = cfqq->cfqd;
704         struct request *__alias, *prev;
705
706         cfqq->queued[rq_is_sync(rq)]++;
707
708         /*
709          * looks a little odd, but the first insert might return an alias.
710          * if that happens, put the alias on the dispatch list
711          */
712         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
713                 cfq_dispatch_insert(cfqd->queue, __alias);
714
715         if (!cfq_cfqq_on_rr(cfqq))
716                 cfq_add_cfqq_rr(cfqd, cfqq);
717
718         /*
719          * check if this request is a better next-serve candidate
720          */
721         prev = cfqq->next_rq;
722         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
723
724         /*
725          * adjust priority tree position, if ->next_rq changes
726          */
727         if (prev != cfqq->next_rq)
728                 cfq_prio_tree_add(cfqd, cfqq);
729
730         BUG_ON(!cfqq->next_rq);
731 }
732
733 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
734 {
735         elv_rb_del(&cfqq->sort_list, rq);
736         cfqq->queued[rq_is_sync(rq)]--;
737         cfq_add_rq_rb(rq);
738 }
739
740 static struct request *
741 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
742 {
743         struct task_struct *tsk = current;
744         struct cfq_io_context *cic;
745         struct cfq_queue *cfqq;
746
747         cic = cfq_cic_lookup(cfqd, tsk->io_context);
748         if (!cic)
749                 return NULL;
750
751         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
752         if (cfqq) {
753                 sector_t sector = bio->bi_sector + bio_sectors(bio);
754
755                 return elv_rb_find(&cfqq->sort_list, sector);
756         }
757
758         return NULL;
759 }
760
761 static void cfq_activate_request(struct request_queue *q, struct request *rq)
762 {
763         struct cfq_data *cfqd = q->elevator->elevator_data;
764
765         cfqd->rq_in_driver[rq_is_sync(rq)]++;
766         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
767                                                 rq_in_driver(cfqd));
768
769         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
770 }
771
772 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
773 {
774         struct cfq_data *cfqd = q->elevator->elevator_data;
775         const int sync = rq_is_sync(rq);
776
777         WARN_ON(!cfqd->rq_in_driver[sync]);
778         cfqd->rq_in_driver[sync]--;
779         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
780                                                 rq_in_driver(cfqd));
781 }
782
783 static void cfq_remove_request(struct request *rq)
784 {
785         struct cfq_queue *cfqq = RQ_CFQQ(rq);
786
787         if (cfqq->next_rq == rq)
788                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
789
790         list_del_init(&rq->queuelist);
791         cfq_del_rq_rb(rq);
792
793         cfqq->cfqd->rq_queued--;
794         if (rq_is_meta(rq)) {
795                 WARN_ON(!cfqq->meta_pending);
796                 cfqq->meta_pending--;
797         }
798 }
799
800 static int cfq_merge(struct request_queue *q, struct request **req,
801                      struct bio *bio)
802 {
803         struct cfq_data *cfqd = q->elevator->elevator_data;
804         struct request *__rq;
805
806         __rq = cfq_find_rq_fmerge(cfqd, bio);
807         if (__rq && elv_rq_merge_ok(__rq, bio)) {
808                 *req = __rq;
809                 return ELEVATOR_FRONT_MERGE;
810         }
811
812         return ELEVATOR_NO_MERGE;
813 }
814
815 static void cfq_merged_request(struct request_queue *q, struct request *req,
816                                int type)
817 {
818         if (type == ELEVATOR_FRONT_MERGE) {
819                 struct cfq_queue *cfqq = RQ_CFQQ(req);
820
821                 cfq_reposition_rq_rb(cfqq, req);
822         }
823 }
824
825 static void
826 cfq_merged_requests(struct request_queue *q, struct request *rq,
827                     struct request *next)
828 {
829         /*
830          * reposition in fifo if next is older than rq
831          */
832         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
833             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
834                 list_move(&rq->queuelist, &next->queuelist);
835                 rq_set_fifo_time(rq, rq_fifo_time(next));
836         }
837
838         cfq_remove_request(next);
839 }
840
841 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
842                            struct bio *bio)
843 {
844         struct cfq_data *cfqd = q->elevator->elevator_data;
845         struct cfq_io_context *cic;
846         struct cfq_queue *cfqq;
847
848         /*
849          * Disallow merge of a sync bio into an async request.
850          */
851         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
852                 return 0;
853
854         /*
855          * Lookup the cfqq that this bio will be queued with. Allow
856          * merge only if rq is queued there.
857          */
858         cic = cfq_cic_lookup(cfqd, current->io_context);
859         if (!cic)
860                 return 0;
861
862         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
863         if (cfqq == RQ_CFQQ(rq))
864                 return 1;
865
866         return 0;
867 }
868
869 static void __cfq_set_active_queue(struct cfq_data *cfqd,
870                                    struct cfq_queue *cfqq)
871 {
872         if (cfqq) {
873                 cfq_log_cfqq(cfqd, cfqq, "set_active");
874                 cfqq->slice_end = 0;
875                 cfqq->slice_dispatch = 0;
876
877                 cfq_clear_cfqq_wait_request(cfqq);
878                 cfq_clear_cfqq_must_dispatch(cfqq);
879                 cfq_clear_cfqq_must_alloc_slice(cfqq);
880                 cfq_clear_cfqq_fifo_expire(cfqq);
881                 cfq_mark_cfqq_slice_new(cfqq);
882
883                 del_timer(&cfqd->idle_slice_timer);
884         }
885
886         cfqd->active_queue = cfqq;
887 }
888
889 /*
890  * current cfqq expired its slice (or was too idle), select new one
891  */
892 static void
893 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
894                     int timed_out)
895 {
896         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
897
898         if (cfq_cfqq_wait_request(cfqq))
899                 del_timer(&cfqd->idle_slice_timer);
900
901         cfq_clear_cfqq_wait_request(cfqq);
902
903         /*
904          * store what was left of this slice, if the queue idled/timed out
905          */
906         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
907                 cfqq->slice_resid = cfqq->slice_end - jiffies;
908                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
909         }
910
911         cfq_resort_rr_list(cfqd, cfqq);
912
913         if (cfqq == cfqd->active_queue)
914                 cfqd->active_queue = NULL;
915
916         if (cfqd->active_cic) {
917                 put_io_context(cfqd->active_cic->ioc);
918                 cfqd->active_cic = NULL;
919         }
920 }
921
922 static inline void cfq_slice_expired(struct cfq_data *cfqd, int timed_out)
923 {
924         struct cfq_queue *cfqq = cfqd->active_queue;
925
926         if (cfqq)
927                 __cfq_slice_expired(cfqd, cfqq, timed_out);
928 }
929
930 /*
931  * Get next queue for service. Unless we have a queue preemption,
932  * we'll simply select the first cfqq in the service tree.
933  */
934 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
935 {
936         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
937                 return NULL;
938
939         return cfq_rb_first(&cfqd->service_tree);
940 }
941
942 /*
943  * Get and set a new active queue for service.
944  */
945 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
946                                               struct cfq_queue *cfqq)
947 {
948         if (!cfqq) {
949                 cfqq = cfq_get_next_queue(cfqd);
950                 if (cfqq)
951                         cfq_clear_cfqq_coop(cfqq);
952         }
953
954         __cfq_set_active_queue(cfqd, cfqq);
955         return cfqq;
956 }
957
958 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
959                                           struct request *rq)
960 {
961         if (blk_rq_pos(rq) >= cfqd->last_position)
962                 return blk_rq_pos(rq) - cfqd->last_position;
963         else
964                 return cfqd->last_position - blk_rq_pos(rq);
965 }
966
967 #define CIC_SEEK_THR    8 * 1024
968 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
969
970 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
971 {
972         struct cfq_io_context *cic = cfqd->active_cic;
973         sector_t sdist = cic->seek_mean;
974
975         if (!sample_valid(cic->seek_samples))
976                 sdist = CIC_SEEK_THR;
977
978         return cfq_dist_from_last(cfqd, rq) <= sdist;
979 }
980
981 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
982                                     struct cfq_queue *cur_cfqq)
983 {
984         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
985         struct rb_node *parent, *node;
986         struct cfq_queue *__cfqq;
987         sector_t sector = cfqd->last_position;
988
989         if (RB_EMPTY_ROOT(root))
990                 return NULL;
991
992         /*
993          * First, if we find a request starting at the end of the last
994          * request, choose it.
995          */
996         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
997         if (__cfqq)
998                 return __cfqq;
999
1000         /*
1001          * If the exact sector wasn't found, the parent of the NULL leaf
1002          * will contain the closest sector.
1003          */
1004         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1005         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1006                 return __cfqq;
1007
1008         if (blk_rq_pos(__cfqq->next_rq) < sector)
1009                 node = rb_next(&__cfqq->p_node);
1010         else
1011                 node = rb_prev(&__cfqq->p_node);
1012         if (!node)
1013                 return NULL;
1014
1015         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1016         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1017                 return __cfqq;
1018
1019         return NULL;
1020 }
1021
1022 /*
1023  * cfqd - obvious
1024  * cur_cfqq - passed in so that we don't decide that the current queue is
1025  *            closely cooperating with itself.
1026  *
1027  * So, basically we're assuming that that cur_cfqq has dispatched at least
1028  * one request, and that cfqd->last_position reflects a position on the disk
1029  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1030  * assumption.
1031  */
1032 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1033                                               struct cfq_queue *cur_cfqq,
1034                                               int probe)
1035 {
1036         struct cfq_queue *cfqq;
1037
1038         /*
1039          * A valid cfq_io_context is necessary to compare requests against
1040          * the seek_mean of the current cfqq.
1041          */
1042         if (!cfqd->active_cic)
1043                 return NULL;
1044
1045         /*
1046          * We should notice if some of the queues are cooperating, eg
1047          * working closely on the same area of the disk. In that case,
1048          * we can group them together and don't waste time idling.
1049          */
1050         cfqq = cfqq_close(cfqd, cur_cfqq);
1051         if (!cfqq)
1052                 return NULL;
1053
1054         if (cfq_cfqq_coop(cfqq))
1055                 return NULL;
1056
1057         if (!probe)
1058                 cfq_mark_cfqq_coop(cfqq);
1059         return cfqq;
1060 }
1061
1062 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1063 {
1064         struct cfq_queue *cfqq = cfqd->active_queue;
1065         struct cfq_io_context *cic;
1066         unsigned long sl;
1067
1068         /*
1069          * SSD device without seek penalty, disable idling. But only do so
1070          * for devices that support queuing, otherwise we still have a problem
1071          * with sync vs async workloads.
1072          */
1073         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1074                 return;
1075
1076         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1077         WARN_ON(cfq_cfqq_slice_new(cfqq));
1078
1079         /*
1080          * idle is disabled, either manually or by past process history
1081          */
1082         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1083                 return;
1084
1085         /*
1086          * still requests with the driver, don't idle
1087          */
1088         if (rq_in_driver(cfqd))
1089                 return;
1090
1091         /*
1092          * task has exited, don't wait
1093          */
1094         cic = cfqd->active_cic;
1095         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1096                 return;
1097
1098         cfq_mark_cfqq_wait_request(cfqq);
1099
1100         /*
1101          * we don't want to idle for seeks, but we do want to allow
1102          * fair distribution of slice time for a process doing back-to-back
1103          * seeks. so allow a little bit of time for him to submit a new rq
1104          */
1105         sl = cfqd->cfq_slice_idle;
1106         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1107                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1108
1109         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1110         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1111 }
1112
1113 /*
1114  * Move request from internal lists to the request queue dispatch list.
1115  */
1116 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1117 {
1118         struct cfq_data *cfqd = q->elevator->elevator_data;
1119         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1120
1121         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1122
1123         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1124         cfq_remove_request(rq);
1125         cfqq->dispatched++;
1126         elv_dispatch_sort(q, rq);
1127
1128         if (cfq_cfqq_sync(cfqq))
1129                 cfqd->sync_flight++;
1130 }
1131
1132 /*
1133  * return expired entry, or NULL to just start from scratch in rbtree
1134  */
1135 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1136 {
1137         struct request *rq = NULL;
1138
1139         if (cfq_cfqq_fifo_expire(cfqq))
1140                 return NULL;
1141
1142         cfq_mark_cfqq_fifo_expire(cfqq);
1143
1144         if (list_empty(&cfqq->fifo))
1145                 return NULL;
1146
1147         rq = rq_entry_fifo(cfqq->fifo.next);
1148         if (time_before(jiffies, rq_fifo_time(rq)))
1149                 rq = NULL;
1150
1151         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1152         return rq;
1153 }
1154
1155 static inline int
1156 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1157 {
1158         const int base_rq = cfqd->cfq_slice_async_rq;
1159
1160         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1161
1162         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1163 }
1164
1165 /*
1166  * Select a queue for service. If we have a current active queue,
1167  * check whether to continue servicing it, or retrieve and set a new one.
1168  */
1169 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1170 {
1171         struct cfq_queue *cfqq, *new_cfqq = NULL;
1172
1173         cfqq = cfqd->active_queue;
1174         if (!cfqq)
1175                 goto new_queue;
1176
1177         /*
1178          * The active queue has run out of time, expire it and select new.
1179          */
1180         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1181                 goto expire;
1182
1183         /*
1184          * The active queue has requests and isn't expired, allow it to
1185          * dispatch.
1186          */
1187         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1188                 goto keep_queue;
1189
1190         /*
1191          * If another queue has a request waiting within our mean seek
1192          * distance, let it run.  The expire code will check for close
1193          * cooperators and put the close queue at the front of the service
1194          * tree.
1195          */
1196         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1197         if (new_cfqq)
1198                 goto expire;
1199
1200         /*
1201          * No requests pending. If the active queue still has requests in
1202          * flight or is idling for a new request, allow either of these
1203          * conditions to happen (or time out) before selecting a new queue.
1204          */
1205         if (timer_pending(&cfqd->idle_slice_timer) ||
1206             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1207                 cfqq = NULL;
1208                 goto keep_queue;
1209         }
1210
1211 expire:
1212         cfq_slice_expired(cfqd, 0);
1213 new_queue:
1214         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1215 keep_queue:
1216         return cfqq;
1217 }
1218
1219 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1220 {
1221         int dispatched = 0;
1222
1223         while (cfqq->next_rq) {
1224                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1225                 dispatched++;
1226         }
1227
1228         BUG_ON(!list_empty(&cfqq->fifo));
1229         return dispatched;
1230 }
1231
1232 /*
1233  * Drain our current requests. Used for barriers and when switching
1234  * io schedulers on-the-fly.
1235  */
1236 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1237 {
1238         struct cfq_queue *cfqq;
1239         int dispatched = 0;
1240
1241         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1242                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1243
1244         cfq_slice_expired(cfqd, 0);
1245
1246         BUG_ON(cfqd->busy_queues);
1247
1248         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1249         return dispatched;
1250 }
1251
1252 /*
1253  * Dispatch a request from cfqq, moving them to the request queue
1254  * dispatch list.
1255  */
1256 static void cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1257 {
1258         struct request *rq;
1259
1260         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1261
1262         /*
1263          * follow expired path, else get first next available
1264          */
1265         rq = cfq_check_fifo(cfqq);
1266         if (!rq)
1267                 rq = cfqq->next_rq;
1268
1269         /*
1270          * insert request into driver dispatch list
1271          */
1272         cfq_dispatch_insert(cfqd->queue, rq);
1273
1274         if (!cfqd->active_cic) {
1275                 struct cfq_io_context *cic = RQ_CIC(rq);
1276
1277                 atomic_long_inc(&cic->ioc->refcount);
1278                 cfqd->active_cic = cic;
1279         }
1280 }
1281
1282 /*
1283  * Find the cfqq that we need to service and move a request from that to the
1284  * dispatch list
1285  */
1286 static int cfq_dispatch_requests(struct request_queue *q, int force)
1287 {
1288         struct cfq_data *cfqd = q->elevator->elevator_data;
1289         struct cfq_queue *cfqq;
1290         unsigned int max_dispatch;
1291
1292         if (!cfqd->busy_queues)
1293                 return 0;
1294
1295         if (unlikely(force))
1296                 return cfq_forced_dispatch(cfqd);
1297
1298         cfqq = cfq_select_queue(cfqd);
1299         if (!cfqq)
1300                 return 0;
1301
1302         /*
1303          * Drain async requests before we start sync IO
1304          */
1305         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1306                 return 0;
1307
1308         /*
1309          * If this is an async queue and we have sync IO in flight, let it wait
1310          */
1311         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1312                 return 0;
1313
1314         max_dispatch = cfqd->cfq_quantum;
1315         if (cfq_class_idle(cfqq))
1316                 max_dispatch = 1;
1317
1318         /*
1319          * Does this cfqq already have too much IO in flight?
1320          */
1321         if (cfqq->dispatched >= max_dispatch) {
1322                 /*
1323                  * idle queue must always only have a single IO in flight
1324                  */
1325                 if (cfq_class_idle(cfqq))
1326                         return 0;
1327
1328                 /*
1329                  * We have other queues, don't allow more IO from this one
1330                  */
1331                 if (cfqd->busy_queues > 1)
1332                         return 0;
1333
1334                 /*
1335                  * Sole queue user, allow bigger slice
1336                  */
1337                 max_dispatch *= 4;
1338         }
1339
1340         /*
1341          * Async queues must wait a bit before being allowed dispatch.
1342          * We also ramp up the dispatch depth gradually for async IO,
1343          * based on the last sync IO we serviced
1344          */
1345         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1346                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1347                 unsigned int depth;
1348
1349                 depth = last_sync / cfqd->cfq_slice[1];
1350                 if (!depth && !cfqq->dispatched)
1351                         depth = 1;
1352                 if (depth < max_dispatch)
1353                         max_dispatch = depth;
1354         }
1355
1356         if (cfqq->dispatched >= max_dispatch)
1357                 return 0;
1358
1359         /*
1360          * Dispatch a request from this cfqq
1361          */
1362         cfq_dispatch_request(cfqd, cfqq);
1363         cfqq->slice_dispatch++;
1364         cfq_clear_cfqq_must_dispatch(cfqq);
1365
1366         /*
1367          * expire an async queue immediately if it has used up its slice. idle
1368          * queue always expire after 1 dispatch round.
1369          */
1370         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1371             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1372             cfq_class_idle(cfqq))) {
1373                 cfqq->slice_end = jiffies + 1;
1374                 cfq_slice_expired(cfqd, 0);
1375         }
1376
1377         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1378         return 1;
1379 }
1380
1381 /*
1382  * task holds one reference to the queue, dropped when task exits. each rq
1383  * in-flight on this queue also holds a reference, dropped when rq is freed.
1384  *
1385  * queue lock must be held here.
1386  */
1387 static void cfq_put_queue(struct cfq_queue *cfqq)
1388 {
1389         struct cfq_data *cfqd = cfqq->cfqd;
1390
1391         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1392
1393         if (!atomic_dec_and_test(&cfqq->ref))
1394                 return;
1395
1396         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1397         BUG_ON(rb_first(&cfqq->sort_list));
1398         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1399         BUG_ON(cfq_cfqq_on_rr(cfqq));
1400
1401         if (unlikely(cfqd->active_queue == cfqq)) {
1402                 __cfq_slice_expired(cfqd, cfqq, 0);
1403                 cfq_schedule_dispatch(cfqd, 0);
1404         }
1405
1406         kmem_cache_free(cfq_pool, cfqq);
1407 }
1408
1409 /*
1410  * Must always be called with the rcu_read_lock() held
1411  */
1412 static void
1413 __call_for_each_cic(struct io_context *ioc,
1414                     void (*func)(struct io_context *, struct cfq_io_context *))
1415 {
1416         struct cfq_io_context *cic;
1417         struct hlist_node *n;
1418
1419         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1420                 func(ioc, cic);
1421 }
1422
1423 /*
1424  * Call func for each cic attached to this ioc.
1425  */
1426 static void
1427 call_for_each_cic(struct io_context *ioc,
1428                   void (*func)(struct io_context *, struct cfq_io_context *))
1429 {
1430         rcu_read_lock();
1431         __call_for_each_cic(ioc, func);
1432         rcu_read_unlock();
1433 }
1434
1435 static void cfq_cic_free_rcu(struct rcu_head *head)
1436 {
1437         struct cfq_io_context *cic;
1438
1439         cic = container_of(head, struct cfq_io_context, rcu_head);
1440
1441         kmem_cache_free(cfq_ioc_pool, cic);
1442         elv_ioc_count_dec(cfq_ioc_count);
1443
1444         if (ioc_gone) {
1445                 /*
1446                  * CFQ scheduler is exiting, grab exit lock and check
1447                  * the pending io context count. If it hits zero,
1448                  * complete ioc_gone and set it back to NULL
1449                  */
1450                 spin_lock(&ioc_gone_lock);
1451                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1452                         complete(ioc_gone);
1453                         ioc_gone = NULL;
1454                 }
1455                 spin_unlock(&ioc_gone_lock);
1456         }
1457 }
1458
1459 static void cfq_cic_free(struct cfq_io_context *cic)
1460 {
1461         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1462 }
1463
1464 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1465 {
1466         unsigned long flags;
1467
1468         BUG_ON(!cic->dead_key);
1469
1470         spin_lock_irqsave(&ioc->lock, flags);
1471         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1472         hlist_del_rcu(&cic->cic_list);
1473         spin_unlock_irqrestore(&ioc->lock, flags);
1474
1475         cfq_cic_free(cic);
1476 }
1477
1478 /*
1479  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1480  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1481  * and ->trim() which is called with the task lock held
1482  */
1483 static void cfq_free_io_context(struct io_context *ioc)
1484 {
1485         /*
1486          * ioc->refcount is zero here, or we are called from elv_unregister(),
1487          * so no more cic's are allowed to be linked into this ioc.  So it
1488          * should be ok to iterate over the known list, we will see all cic's
1489          * since no new ones are added.
1490          */
1491         __call_for_each_cic(ioc, cic_free_func);
1492 }
1493
1494 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1495 {
1496         if (unlikely(cfqq == cfqd->active_queue)) {
1497                 __cfq_slice_expired(cfqd, cfqq, 0);
1498                 cfq_schedule_dispatch(cfqd, 0);
1499         }
1500
1501         cfq_put_queue(cfqq);
1502 }
1503
1504 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1505                                          struct cfq_io_context *cic)
1506 {
1507         struct io_context *ioc = cic->ioc;
1508
1509         list_del_init(&cic->queue_list);
1510
1511         /*
1512          * Make sure key == NULL is seen for dead queues
1513          */
1514         smp_wmb();
1515         cic->dead_key = (unsigned long) cic->key;
1516         cic->key = NULL;
1517
1518         if (ioc->ioc_data == cic)
1519                 rcu_assign_pointer(ioc->ioc_data, NULL);
1520
1521         if (cic->cfqq[BLK_RW_ASYNC]) {
1522                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1523                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1524         }
1525
1526         if (cic->cfqq[BLK_RW_SYNC]) {
1527                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1528                 cic->cfqq[BLK_RW_SYNC] = NULL;
1529         }
1530 }
1531
1532 static void cfq_exit_single_io_context(struct io_context *ioc,
1533                                        struct cfq_io_context *cic)
1534 {
1535         struct cfq_data *cfqd = cic->key;
1536
1537         if (cfqd) {
1538                 struct request_queue *q = cfqd->queue;
1539                 unsigned long flags;
1540
1541                 spin_lock_irqsave(q->queue_lock, flags);
1542
1543                 /*
1544                  * Ensure we get a fresh copy of the ->key to prevent
1545                  * race between exiting task and queue
1546                  */
1547                 smp_read_barrier_depends();
1548                 if (cic->key)
1549                         __cfq_exit_single_io_context(cfqd, cic);
1550
1551                 spin_unlock_irqrestore(q->queue_lock, flags);
1552         }
1553 }
1554
1555 /*
1556  * The process that ioc belongs to has exited, we need to clean up
1557  * and put the internal structures we have that belongs to that process.
1558  */
1559 static void cfq_exit_io_context(struct io_context *ioc)
1560 {
1561         call_for_each_cic(ioc, cfq_exit_single_io_context);
1562 }
1563
1564 static struct cfq_io_context *
1565 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1566 {
1567         struct cfq_io_context *cic;
1568
1569         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1570                                                         cfqd->queue->node);
1571         if (cic) {
1572                 cic->last_end_request = jiffies;
1573                 INIT_LIST_HEAD(&cic->queue_list);
1574                 INIT_HLIST_NODE(&cic->cic_list);
1575                 cic->dtor = cfq_free_io_context;
1576                 cic->exit = cfq_exit_io_context;
1577                 elv_ioc_count_inc(cfq_ioc_count);
1578         }
1579
1580         return cic;
1581 }
1582
1583 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1584 {
1585         struct task_struct *tsk = current;
1586         int ioprio_class;
1587
1588         if (!cfq_cfqq_prio_changed(cfqq))
1589                 return;
1590
1591         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1592         switch (ioprio_class) {
1593         default:
1594                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1595         case IOPRIO_CLASS_NONE:
1596                 /*
1597                  * no prio set, inherit CPU scheduling settings
1598                  */
1599                 cfqq->ioprio = task_nice_ioprio(tsk);
1600                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1601                 break;
1602         case IOPRIO_CLASS_RT:
1603                 cfqq->ioprio = task_ioprio(ioc);
1604                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1605                 break;
1606         case IOPRIO_CLASS_BE:
1607                 cfqq->ioprio = task_ioprio(ioc);
1608                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1609                 break;
1610         case IOPRIO_CLASS_IDLE:
1611                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1612                 cfqq->ioprio = 7;
1613                 cfq_clear_cfqq_idle_window(cfqq);
1614                 break;
1615         }
1616
1617         /*
1618          * keep track of original prio settings in case we have to temporarily
1619          * elevate the priority of this queue
1620          */
1621         cfqq->org_ioprio = cfqq->ioprio;
1622         cfqq->org_ioprio_class = cfqq->ioprio_class;
1623         cfq_clear_cfqq_prio_changed(cfqq);
1624 }
1625
1626 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1627 {
1628         struct cfq_data *cfqd = cic->key;
1629         struct cfq_queue *cfqq;
1630         unsigned long flags;
1631
1632         if (unlikely(!cfqd))
1633                 return;
1634
1635         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1636
1637         cfqq = cic->cfqq[BLK_RW_ASYNC];
1638         if (cfqq) {
1639                 struct cfq_queue *new_cfqq;
1640                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1641                                                 GFP_ATOMIC);
1642                 if (new_cfqq) {
1643                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1644                         cfq_put_queue(cfqq);
1645                 }
1646         }
1647
1648         cfqq = cic->cfqq[BLK_RW_SYNC];
1649         if (cfqq)
1650                 cfq_mark_cfqq_prio_changed(cfqq);
1651
1652         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1653 }
1654
1655 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1656 {
1657         call_for_each_cic(ioc, changed_ioprio);
1658         ioc->ioprio_changed = 0;
1659 }
1660
1661 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1662                           pid_t pid, int is_sync)
1663 {
1664         RB_CLEAR_NODE(&cfqq->rb_node);
1665         RB_CLEAR_NODE(&cfqq->p_node);
1666         INIT_LIST_HEAD(&cfqq->fifo);
1667
1668         atomic_set(&cfqq->ref, 0);
1669         cfqq->cfqd = cfqd;
1670
1671         cfq_mark_cfqq_prio_changed(cfqq);
1672
1673         if (is_sync) {
1674                 if (!cfq_class_idle(cfqq))
1675                         cfq_mark_cfqq_idle_window(cfqq);
1676                 cfq_mark_cfqq_sync(cfqq);
1677         }
1678         cfqq->pid = pid;
1679 }
1680
1681 static struct cfq_queue *
1682 cfq_find_alloc_queue(struct cfq_data *cfqd, int is_sync,
1683                      struct io_context *ioc, gfp_t gfp_mask)
1684 {
1685         struct cfq_queue *cfqq, *new_cfqq = NULL;
1686         struct cfq_io_context *cic;
1687
1688 retry:
1689         cic = cfq_cic_lookup(cfqd, ioc);
1690         /* cic always exists here */
1691         cfqq = cic_to_cfqq(cic, is_sync);
1692
1693         /*
1694          * Always try a new alloc if we fell back to the OOM cfqq
1695          * originally, since it should just be a temporary situation.
1696          */
1697         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1698                 cfqq = NULL;
1699                 if (new_cfqq) {
1700                         cfqq = new_cfqq;
1701                         new_cfqq = NULL;
1702                 } else if (gfp_mask & __GFP_WAIT) {
1703                         spin_unlock_irq(cfqd->queue->queue_lock);
1704                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1705                                         gfp_mask | __GFP_ZERO,
1706                                         cfqd->queue->node);
1707                         spin_lock_irq(cfqd->queue->queue_lock);
1708                         if (new_cfqq)
1709                                 goto retry;
1710                 } else {
1711                         cfqq = kmem_cache_alloc_node(cfq_pool,
1712                                         gfp_mask | __GFP_ZERO,
1713                                         cfqd->queue->node);
1714                 }
1715
1716                 if (cfqq) {
1717                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1718                         cfq_init_prio_data(cfqq, ioc);
1719                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1720                 } else
1721                         cfqq = &cfqd->oom_cfqq;
1722         }
1723
1724         if (new_cfqq)
1725                 kmem_cache_free(cfq_pool, new_cfqq);
1726
1727         return cfqq;
1728 }
1729
1730 static struct cfq_queue **
1731 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1732 {
1733         switch (ioprio_class) {
1734         case IOPRIO_CLASS_RT:
1735                 return &cfqd->async_cfqq[0][ioprio];
1736         case IOPRIO_CLASS_BE:
1737                 return &cfqd->async_cfqq[1][ioprio];
1738         case IOPRIO_CLASS_IDLE:
1739                 return &cfqd->async_idle_cfqq;
1740         default:
1741                 BUG();
1742         }
1743 }
1744
1745 static struct cfq_queue *
1746 cfq_get_queue(struct cfq_data *cfqd, int is_sync, struct io_context *ioc,
1747               gfp_t gfp_mask)
1748 {
1749         const int ioprio = task_ioprio(ioc);
1750         const int ioprio_class = task_ioprio_class(ioc);
1751         struct cfq_queue **async_cfqq = NULL;
1752         struct cfq_queue *cfqq = NULL;
1753
1754         if (!is_sync) {
1755                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1756                 cfqq = *async_cfqq;
1757         }
1758
1759         if (!cfqq)
1760                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1761
1762         /*
1763          * pin the queue now that it's allocated, scheduler exit will prune it
1764          */
1765         if (!is_sync && !(*async_cfqq)) {
1766                 atomic_inc(&cfqq->ref);
1767                 *async_cfqq = cfqq;
1768         }
1769
1770         atomic_inc(&cfqq->ref);
1771         return cfqq;
1772 }
1773
1774 /*
1775  * We drop cfq io contexts lazily, so we may find a dead one.
1776  */
1777 static void
1778 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1779                   struct cfq_io_context *cic)
1780 {
1781         unsigned long flags;
1782
1783         WARN_ON(!list_empty(&cic->queue_list));
1784
1785         spin_lock_irqsave(&ioc->lock, flags);
1786
1787         BUG_ON(ioc->ioc_data == cic);
1788
1789         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1790         hlist_del_rcu(&cic->cic_list);
1791         spin_unlock_irqrestore(&ioc->lock, flags);
1792
1793         cfq_cic_free(cic);
1794 }
1795
1796 static struct cfq_io_context *
1797 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1798 {
1799         struct cfq_io_context *cic;
1800         unsigned long flags;
1801         void *k;
1802
1803         if (unlikely(!ioc))
1804                 return NULL;
1805
1806         rcu_read_lock();
1807
1808         /*
1809          * we maintain a last-hit cache, to avoid browsing over the tree
1810          */
1811         cic = rcu_dereference(ioc->ioc_data);
1812         if (cic && cic->key == cfqd) {
1813                 rcu_read_unlock();
1814                 return cic;
1815         }
1816
1817         do {
1818                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1819                 rcu_read_unlock();
1820                 if (!cic)
1821                         break;
1822                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1823                 k = cic->key;
1824                 if (unlikely(!k)) {
1825                         cfq_drop_dead_cic(cfqd, ioc, cic);
1826                         rcu_read_lock();
1827                         continue;
1828                 }
1829
1830                 spin_lock_irqsave(&ioc->lock, flags);
1831                 rcu_assign_pointer(ioc->ioc_data, cic);
1832                 spin_unlock_irqrestore(&ioc->lock, flags);
1833                 break;
1834         } while (1);
1835
1836         return cic;
1837 }
1838
1839 /*
1840  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1841  * the process specific cfq io context when entered from the block layer.
1842  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1843  */
1844 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1845                         struct cfq_io_context *cic, gfp_t gfp_mask)
1846 {
1847         unsigned long flags;
1848         int ret;
1849
1850         ret = radix_tree_preload(gfp_mask);
1851         if (!ret) {
1852                 cic->ioc = ioc;
1853                 cic->key = cfqd;
1854
1855                 spin_lock_irqsave(&ioc->lock, flags);
1856                 ret = radix_tree_insert(&ioc->radix_root,
1857                                                 (unsigned long) cfqd, cic);
1858                 if (!ret)
1859                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1860                 spin_unlock_irqrestore(&ioc->lock, flags);
1861
1862                 radix_tree_preload_end();
1863
1864                 if (!ret) {
1865                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1866                         list_add(&cic->queue_list, &cfqd->cic_list);
1867                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1868                 }
1869         }
1870
1871         if (ret)
1872                 printk(KERN_ERR "cfq: cic link failed!\n");
1873
1874         return ret;
1875 }
1876
1877 /*
1878  * Setup general io context and cfq io context. There can be several cfq
1879  * io contexts per general io context, if this process is doing io to more
1880  * than one device managed by cfq.
1881  */
1882 static struct cfq_io_context *
1883 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1884 {
1885         struct io_context *ioc = NULL;
1886         struct cfq_io_context *cic;
1887
1888         might_sleep_if(gfp_mask & __GFP_WAIT);
1889
1890         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1891         if (!ioc)
1892                 return NULL;
1893
1894         cic = cfq_cic_lookup(cfqd, ioc);
1895         if (cic)
1896                 goto out;
1897
1898         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1899         if (cic == NULL)
1900                 goto err;
1901
1902         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1903                 goto err_free;
1904
1905 out:
1906         smp_read_barrier_depends();
1907         if (unlikely(ioc->ioprio_changed))
1908                 cfq_ioc_set_ioprio(ioc);
1909
1910         return cic;
1911 err_free:
1912         cfq_cic_free(cic);
1913 err:
1914         put_io_context(ioc);
1915         return NULL;
1916 }
1917
1918 static void
1919 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1920 {
1921         unsigned long elapsed = jiffies - cic->last_end_request;
1922         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1923
1924         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1925         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1926         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1927 }
1928
1929 static void
1930 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1931                        struct request *rq)
1932 {
1933         sector_t sdist;
1934         u64 total;
1935
1936         if (!cic->last_request_pos)
1937                 sdist = 0;
1938         else if (cic->last_request_pos < blk_rq_pos(rq))
1939                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1940         else
1941                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1942
1943         /*
1944          * Don't allow the seek distance to get too large from the
1945          * odd fragment, pagein, etc
1946          */
1947         if (cic->seek_samples <= 60) /* second&third seek */
1948                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1949         else
1950                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1951
1952         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1953         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1954         total = cic->seek_total + (cic->seek_samples/2);
1955         do_div(total, cic->seek_samples);
1956         cic->seek_mean = (sector_t)total;
1957 }
1958
1959 /*
1960  * Disable idle window if the process thinks too long or seeks so much that
1961  * it doesn't matter
1962  */
1963 static void
1964 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1965                        struct cfq_io_context *cic)
1966 {
1967         int old_idle, enable_idle;
1968
1969         /*
1970          * Don't idle for async or idle io prio class
1971          */
1972         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1973                 return;
1974
1975         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1976
1977         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1978             (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1979                 enable_idle = 0;
1980         else if (sample_valid(cic->ttime_samples)) {
1981                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
1982                         enable_idle = 0;
1983                 else
1984                         enable_idle = 1;
1985         }
1986
1987         if (old_idle != enable_idle) {
1988                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
1989                 if (enable_idle)
1990                         cfq_mark_cfqq_idle_window(cfqq);
1991                 else
1992                         cfq_clear_cfqq_idle_window(cfqq);
1993         }
1994 }
1995
1996 /*
1997  * Check if new_cfqq should preempt the currently active queue. Return 0 for
1998  * no or if we aren't sure, a 1 will cause a preempt.
1999  */
2000 static int
2001 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2002                    struct request *rq)
2003 {
2004         struct cfq_queue *cfqq;
2005
2006         cfqq = cfqd->active_queue;
2007         if (!cfqq)
2008                 return 0;
2009
2010         if (cfq_slice_used(cfqq))
2011                 return 1;
2012
2013         if (cfq_class_idle(new_cfqq))
2014                 return 0;
2015
2016         if (cfq_class_idle(cfqq))
2017                 return 1;
2018
2019         /*
2020          * if the new request is sync, but the currently running queue is
2021          * not, let the sync request have priority.
2022          */
2023         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2024                 return 1;
2025
2026         /*
2027          * So both queues are sync. Let the new request get disk time if
2028          * it's a metadata request and the current queue is doing regular IO.
2029          */
2030         if (rq_is_meta(rq) && !cfqq->meta_pending)
2031                 return 1;
2032
2033         /*
2034          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2035          */
2036         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2037                 return 1;
2038
2039         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2040                 return 0;
2041
2042         /*
2043          * if this request is as-good as one we would expect from the
2044          * current cfqq, let it preempt
2045          */
2046         if (cfq_rq_close(cfqd, rq))
2047                 return 1;
2048
2049         return 0;
2050 }
2051
2052 /*
2053  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2054  * let it have half of its nominal slice.
2055  */
2056 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2057 {
2058         cfq_log_cfqq(cfqd, cfqq, "preempt");
2059         cfq_slice_expired(cfqd, 1);
2060
2061         /*
2062          * Put the new queue at the front of the of the current list,
2063          * so we know that it will be selected next.
2064          */
2065         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2066
2067         cfq_service_tree_add(cfqd, cfqq, 1);
2068
2069         cfqq->slice_end = 0;
2070         cfq_mark_cfqq_slice_new(cfqq);
2071 }
2072
2073 /*
2074  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2075  * something we should do about it
2076  */
2077 static void
2078 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2079                 struct request *rq)
2080 {
2081         struct cfq_io_context *cic = RQ_CIC(rq);
2082
2083         cfqd->rq_queued++;
2084         if (rq_is_meta(rq))
2085                 cfqq->meta_pending++;
2086
2087         cfq_update_io_thinktime(cfqd, cic);
2088         cfq_update_io_seektime(cfqd, cic, rq);
2089         cfq_update_idle_window(cfqd, cfqq, cic);
2090
2091         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2092
2093         if (cfqq == cfqd->active_queue) {
2094                 /*
2095                  * Remember that we saw a request from this process, but
2096                  * don't start queuing just yet. Otherwise we risk seeing lots
2097                  * of tiny requests, because we disrupt the normal plugging
2098                  * and merging. If the request is already larger than a single
2099                  * page, let it rip immediately. For that case we assume that
2100                  * merging is already done. Ditto for a busy system that
2101                  * has other work pending, don't risk delaying until the
2102                  * idle timer unplug to continue working.
2103                  */
2104                 if (cfq_cfqq_wait_request(cfqq)) {
2105                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2106                             cfqd->busy_queues > 1) {
2107                                 del_timer(&cfqd->idle_slice_timer);
2108                         __blk_run_queue(cfqd->queue);
2109                         }
2110                         cfq_mark_cfqq_must_dispatch(cfqq);
2111                 }
2112         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2113                 /*
2114                  * not the active queue - expire current slice if it is
2115                  * idle and has expired it's mean thinktime or this new queue
2116                  * has some old slice time left and is of higher priority or
2117                  * this new queue is RT and the current one is BE
2118                  */
2119                 cfq_preempt_queue(cfqd, cfqq);
2120                 __blk_run_queue(cfqd->queue);
2121         }
2122 }
2123
2124 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2125 {
2126         struct cfq_data *cfqd = q->elevator->elevator_data;
2127         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2128
2129         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2130         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2131
2132         cfq_add_rq_rb(rq);
2133
2134         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2135         list_add_tail(&rq->queuelist, &cfqq->fifo);
2136
2137         cfq_rq_enqueued(cfqd, cfqq, rq);
2138 }
2139
2140 /*
2141  * Update hw_tag based on peak queue depth over 50 samples under
2142  * sufficient load.
2143  */
2144 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2145 {
2146         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2147                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2148
2149         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2150             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2151                 return;
2152
2153         if (cfqd->hw_tag_samples++ < 50)
2154                 return;
2155
2156         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2157                 cfqd->hw_tag = 1;
2158         else
2159                 cfqd->hw_tag = 0;
2160
2161         cfqd->hw_tag_samples = 0;
2162         cfqd->rq_in_driver_peak = 0;
2163 }
2164
2165 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2166 {
2167         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2168         struct cfq_data *cfqd = cfqq->cfqd;
2169         const int sync = rq_is_sync(rq);
2170         unsigned long now;
2171
2172         now = jiffies;
2173         cfq_log_cfqq(cfqd, cfqq, "complete");
2174
2175         cfq_update_hw_tag(cfqd);
2176
2177         WARN_ON(!cfqd->rq_in_driver[sync]);
2178         WARN_ON(!cfqq->dispatched);
2179         cfqd->rq_in_driver[sync]--;
2180         cfqq->dispatched--;
2181
2182         if (cfq_cfqq_sync(cfqq))
2183                 cfqd->sync_flight--;
2184
2185         if (sync) {
2186                 RQ_CIC(rq)->last_end_request = now;
2187                 cfqd->last_end_sync_rq = now;
2188         }
2189
2190         /*
2191          * If this is the active queue, check if it needs to be expired,
2192          * or if we want to idle in case it has no pending requests.
2193          */
2194         if (cfqd->active_queue == cfqq) {
2195                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2196
2197                 if (cfq_cfqq_slice_new(cfqq)) {
2198                         cfq_set_prio_slice(cfqd, cfqq);
2199                         cfq_clear_cfqq_slice_new(cfqq);
2200                 }
2201                 /*
2202                  * If there are no requests waiting in this queue, and
2203                  * there are other queues ready to issue requests, AND
2204                  * those other queues are issuing requests within our
2205                  * mean seek distance, give them a chance to run instead
2206                  * of idling.
2207                  */
2208                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2209                         cfq_slice_expired(cfqd, 1);
2210                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2211                          sync && !rq_noidle(rq))
2212                         cfq_arm_slice_timer(cfqd);
2213         }
2214
2215         if (!rq_in_driver(cfqd))
2216                 cfq_schedule_dispatch(cfqd, 0);
2217 }
2218
2219 /*
2220  * we temporarily boost lower priority queues if they are holding fs exclusive
2221  * resources. they are boosted to normal prio (CLASS_BE/4)
2222  */
2223 static void cfq_prio_boost(struct cfq_queue *cfqq)
2224 {
2225         if (has_fs_excl()) {
2226                 /*
2227                  * boost idle prio on transactions that would lock out other
2228                  * users of the filesystem
2229                  */
2230                 if (cfq_class_idle(cfqq))
2231                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2232                 if (cfqq->ioprio > IOPRIO_NORM)
2233                         cfqq->ioprio = IOPRIO_NORM;
2234         } else {
2235                 /*
2236                  * check if we need to unboost the queue
2237                  */
2238                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2239                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2240                 if (cfqq->ioprio != cfqq->org_ioprio)
2241                         cfqq->ioprio = cfqq->org_ioprio;
2242         }
2243 }
2244
2245 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2246 {
2247         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2248                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2249                 return ELV_MQUEUE_MUST;
2250         }
2251
2252         return ELV_MQUEUE_MAY;
2253 }
2254
2255 static int cfq_may_queue(struct request_queue *q, int rw)
2256 {
2257         struct cfq_data *cfqd = q->elevator->elevator_data;
2258         struct task_struct *tsk = current;
2259         struct cfq_io_context *cic;
2260         struct cfq_queue *cfqq;
2261
2262         /*
2263          * don't force setup of a queue from here, as a call to may_queue
2264          * does not necessarily imply that a request actually will be queued.
2265          * so just lookup a possibly existing queue, or return 'may queue'
2266          * if that fails
2267          */
2268         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2269         if (!cic)
2270                 return ELV_MQUEUE_MAY;
2271
2272         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2273         if (cfqq) {
2274                 cfq_init_prio_data(cfqq, cic->ioc);
2275                 cfq_prio_boost(cfqq);
2276
2277                 return __cfq_may_queue(cfqq);
2278         }
2279
2280         return ELV_MQUEUE_MAY;
2281 }
2282
2283 /*
2284  * queue lock held here
2285  */
2286 static void cfq_put_request(struct request *rq)
2287 {
2288         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2289
2290         if (cfqq) {
2291                 const int rw = rq_data_dir(rq);
2292
2293                 BUG_ON(!cfqq->allocated[rw]);
2294                 cfqq->allocated[rw]--;
2295
2296                 put_io_context(RQ_CIC(rq)->ioc);
2297
2298                 rq->elevator_private = NULL;
2299                 rq->elevator_private2 = NULL;
2300
2301                 cfq_put_queue(cfqq);
2302         }
2303 }
2304
2305 /*
2306  * Allocate cfq data structures associated with this request.
2307  */
2308 static int
2309 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2310 {
2311         struct cfq_data *cfqd = q->elevator->elevator_data;
2312         struct cfq_io_context *cic;
2313         const int rw = rq_data_dir(rq);
2314         const int is_sync = rq_is_sync(rq);
2315         struct cfq_queue *cfqq;
2316         unsigned long flags;
2317
2318         might_sleep_if(gfp_mask & __GFP_WAIT);
2319
2320         cic = cfq_get_io_context(cfqd, gfp_mask);
2321
2322         spin_lock_irqsave(q->queue_lock, flags);
2323
2324         if (!cic)
2325                 goto queue_fail;
2326
2327         cfqq = cic_to_cfqq(cic, is_sync);
2328         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2329                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2330                 cic_set_cfqq(cic, cfqq, is_sync);
2331         }
2332
2333         cfqq->allocated[rw]++;
2334         atomic_inc(&cfqq->ref);
2335
2336         spin_unlock_irqrestore(q->queue_lock, flags);
2337
2338         rq->elevator_private = cic;
2339         rq->elevator_private2 = cfqq;
2340         return 0;
2341
2342 queue_fail:
2343         if (cic)
2344                 put_io_context(cic->ioc);
2345
2346         cfq_schedule_dispatch(cfqd, 0);
2347         spin_unlock_irqrestore(q->queue_lock, flags);
2348         cfq_log(cfqd, "set_request fail");
2349         return 1;
2350 }
2351
2352 static void cfq_kick_queue(struct work_struct *work)
2353 {
2354         struct cfq_data *cfqd =
2355                 container_of(work, struct cfq_data, unplug_work.work);
2356         struct request_queue *q = cfqd->queue;
2357
2358         spin_lock_irq(q->queue_lock);
2359         __blk_run_queue(cfqd->queue);
2360         spin_unlock_irq(q->queue_lock);
2361 }
2362
2363 /*
2364  * Timer running if the active_queue is currently idling inside its time slice
2365  */
2366 static void cfq_idle_slice_timer(unsigned long data)
2367 {
2368         struct cfq_data *cfqd = (struct cfq_data *) data;
2369         struct cfq_queue *cfqq;
2370         unsigned long flags;
2371         int timed_out = 1;
2372
2373         cfq_log(cfqd, "idle timer fired");
2374
2375         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2376
2377         cfqq = cfqd->active_queue;
2378         if (cfqq) {
2379                 timed_out = 0;
2380
2381                 /*
2382                  * We saw a request before the queue expired, let it through
2383                  */
2384                 if (cfq_cfqq_must_dispatch(cfqq))
2385                         goto out_kick;
2386
2387                 /*
2388                  * expired
2389                  */
2390                 if (cfq_slice_used(cfqq))
2391                         goto expire;
2392
2393                 /*
2394                  * only expire and reinvoke request handler, if there are
2395                  * other queues with pending requests
2396                  */
2397                 if (!cfqd->busy_queues)
2398                         goto out_cont;
2399
2400                 /*
2401                  * not expired and it has a request pending, let it dispatch
2402                  */
2403                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2404                         goto out_kick;
2405         }
2406 expire:
2407         cfq_slice_expired(cfqd, timed_out);
2408 out_kick:
2409         cfq_schedule_dispatch(cfqd, 0);
2410 out_cont:
2411         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2412 }
2413
2414 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2415 {
2416         del_timer_sync(&cfqd->idle_slice_timer);
2417         cancel_delayed_work_sync(&cfqd->unplug_work);
2418 }
2419
2420 static void cfq_put_async_queues(struct cfq_data *cfqd)
2421 {
2422         int i;
2423
2424         for (i = 0; i < IOPRIO_BE_NR; i++) {
2425                 if (cfqd->async_cfqq[0][i])
2426                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2427                 if (cfqd->async_cfqq[1][i])
2428                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2429         }
2430
2431         if (cfqd->async_idle_cfqq)
2432                 cfq_put_queue(cfqd->async_idle_cfqq);
2433 }
2434
2435 static void cfq_exit_queue(struct elevator_queue *e)
2436 {
2437         struct cfq_data *cfqd = e->elevator_data;
2438         struct request_queue *q = cfqd->queue;
2439
2440         cfq_shutdown_timer_wq(cfqd);
2441
2442         spin_lock_irq(q->queue_lock);
2443
2444         if (cfqd->active_queue)
2445                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2446
2447         while (!list_empty(&cfqd->cic_list)) {
2448                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2449                                                         struct cfq_io_context,
2450                                                         queue_list);
2451
2452                 __cfq_exit_single_io_context(cfqd, cic);
2453         }
2454
2455         cfq_put_async_queues(cfqd);
2456
2457         spin_unlock_irq(q->queue_lock);
2458
2459         cfq_shutdown_timer_wq(cfqd);
2460
2461         kfree(cfqd);
2462 }
2463
2464 static void *cfq_init_queue(struct request_queue *q)
2465 {
2466         struct cfq_data *cfqd;
2467         int i;
2468
2469         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2470         if (!cfqd)
2471                 return NULL;
2472
2473         cfqd->service_tree = CFQ_RB_ROOT;
2474
2475         /*
2476          * Not strictly needed (since RB_ROOT just clears the node and we
2477          * zeroed cfqd on alloc), but better be safe in case someone decides
2478          * to add magic to the rb code
2479          */
2480         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2481                 cfqd->prio_trees[i] = RB_ROOT;
2482
2483         /*
2484          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2485          * Grab a permanent reference to it, so that the normal code flow
2486          * will not attempt to free it.
2487          */
2488         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2489         atomic_inc(&cfqd->oom_cfqq.ref);
2490
2491         INIT_LIST_HEAD(&cfqd->cic_list);
2492
2493         cfqd->queue = q;
2494
2495         init_timer(&cfqd->idle_slice_timer);
2496         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2497         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2498
2499         INIT_DELAYED_WORK(&cfqd->unplug_work, cfq_kick_queue);
2500
2501         cfqd->cfq_quantum = cfq_quantum;
2502         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2503         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2504         cfqd->cfq_back_max = cfq_back_max;
2505         cfqd->cfq_back_penalty = cfq_back_penalty;
2506         cfqd->cfq_slice[0] = cfq_slice_async;
2507         cfqd->cfq_slice[1] = cfq_slice_sync;
2508         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2509         cfqd->cfq_slice_idle = cfq_slice_idle;
2510         cfqd->cfq_latency = 1;
2511         cfqd->hw_tag = 1;
2512         cfqd->last_end_sync_rq = jiffies;
2513         return cfqd;
2514 }
2515
2516 static void cfq_slab_kill(void)
2517 {
2518         /*
2519          * Caller already ensured that pending RCU callbacks are completed,
2520          * so we should have no busy allocations at this point.
2521          */
2522         if (cfq_pool)
2523                 kmem_cache_destroy(cfq_pool);
2524         if (cfq_ioc_pool)
2525                 kmem_cache_destroy(cfq_ioc_pool);
2526 }
2527
2528 static int __init cfq_slab_setup(void)
2529 {
2530         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2531         if (!cfq_pool)
2532                 goto fail;
2533
2534         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2535         if (!cfq_ioc_pool)
2536                 goto fail;
2537
2538         return 0;
2539 fail:
2540         cfq_slab_kill();
2541         return -ENOMEM;
2542 }
2543
2544 /*
2545  * sysfs parts below -->
2546  */
2547 static ssize_t
2548 cfq_var_show(unsigned int var, char *page)
2549 {
2550         return sprintf(page, "%d\n", var);
2551 }
2552
2553 static ssize_t
2554 cfq_var_store(unsigned int *var, const char *page, size_t count)
2555 {
2556         char *p = (char *) page;
2557
2558         *var = simple_strtoul(p, &p, 10);
2559         return count;
2560 }
2561
2562 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2563 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2564 {                                                                       \
2565         struct cfq_data *cfqd = e->elevator_data;                       \
2566         unsigned int __data = __VAR;                                    \
2567         if (__CONV)                                                     \
2568                 __data = jiffies_to_msecs(__data);                      \
2569         return cfq_var_show(__data, (page));                            \
2570 }
2571 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2572 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2573 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2574 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2575 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2576 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2577 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2578 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2579 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2580 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2581 #undef SHOW_FUNCTION
2582
2583 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2584 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2585 {                                                                       \
2586         struct cfq_data *cfqd = e->elevator_data;                       \
2587         unsigned int __data;                                            \
2588         int ret = cfq_var_store(&__data, (page), count);                \
2589         if (__data < (MIN))                                             \
2590                 __data = (MIN);                                         \
2591         else if (__data > (MAX))                                        \
2592                 __data = (MAX);                                         \
2593         if (__CONV)                                                     \
2594                 *(__PTR) = msecs_to_jiffies(__data);                    \
2595         else                                                            \
2596                 *(__PTR) = __data;                                      \
2597         return ret;                                                     \
2598 }
2599 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2600 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2601                 UINT_MAX, 1);
2602 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2603                 UINT_MAX, 1);
2604 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2605 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2606                 UINT_MAX, 0);
2607 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2608 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2609 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2610 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2611                 UINT_MAX, 0);
2612 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2613 #undef STORE_FUNCTION
2614
2615 #define CFQ_ATTR(name) \
2616         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2617
2618 static struct elv_fs_entry cfq_attrs[] = {
2619         CFQ_ATTR(quantum),
2620         CFQ_ATTR(fifo_expire_sync),
2621         CFQ_ATTR(fifo_expire_async),
2622         CFQ_ATTR(back_seek_max),
2623         CFQ_ATTR(back_seek_penalty),
2624         CFQ_ATTR(slice_sync),
2625         CFQ_ATTR(slice_async),
2626         CFQ_ATTR(slice_async_rq),
2627         CFQ_ATTR(slice_idle),
2628         CFQ_ATTR(low_latency),
2629         __ATTR_NULL
2630 };
2631
2632 static struct elevator_type iosched_cfq = {
2633         .ops = {
2634                 .elevator_merge_fn =            cfq_merge,
2635                 .elevator_merged_fn =           cfq_merged_request,
2636                 .elevator_merge_req_fn =        cfq_merged_requests,
2637                 .elevator_allow_merge_fn =      cfq_allow_merge,
2638                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2639                 .elevator_add_req_fn =          cfq_insert_request,
2640                 .elevator_activate_req_fn =     cfq_activate_request,
2641                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2642                 .elevator_queue_empty_fn =      cfq_queue_empty,
2643                 .elevator_completed_req_fn =    cfq_completed_request,
2644                 .elevator_former_req_fn =       elv_rb_former_request,
2645                 .elevator_latter_req_fn =       elv_rb_latter_request,
2646                 .elevator_set_req_fn =          cfq_set_request,
2647                 .elevator_put_req_fn =          cfq_put_request,
2648                 .elevator_may_queue_fn =        cfq_may_queue,
2649                 .elevator_init_fn =             cfq_init_queue,
2650                 .elevator_exit_fn =             cfq_exit_queue,
2651                 .trim =                         cfq_free_io_context,
2652         },
2653         .elevator_attrs =       cfq_attrs,
2654         .elevator_name =        "cfq",
2655         .elevator_owner =       THIS_MODULE,
2656 };
2657
2658 static int __init cfq_init(void)
2659 {
2660         /*
2661          * could be 0 on HZ < 1000 setups
2662          */
2663         if (!cfq_slice_async)
2664                 cfq_slice_async = 1;
2665         if (!cfq_slice_idle)
2666                 cfq_slice_idle = 1;
2667
2668         if (cfq_slab_setup())
2669                 return -ENOMEM;
2670
2671         elv_register(&iosched_cfq);
2672
2673         return 0;
2674 }
2675
2676 static void __exit cfq_exit(void)
2677 {
2678         DECLARE_COMPLETION_ONSTACK(all_gone);
2679         elv_unregister(&iosched_cfq);
2680         ioc_gone = &all_gone;
2681         /* ioc_gone's update must be visible before reading ioc_count */
2682         smp_wmb();
2683
2684         /*
2685          * this also protects us from entering cfq_slab_kill() with
2686          * pending RCU callbacks
2687          */
2688         if (elv_ioc_count_read(cfq_ioc_count))
2689                 wait_for_completion(&all_gone);
2690         cfq_slab_kill();
2691 }
2692
2693 module_init(cfq_init);
2694 module_exit(cfq_exit);
2695
2696 MODULE_AUTHOR("Jens Axboe");
2697 MODULE_LICENSE("GPL");
2698 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");