cfq-iosched: apply bool value where we return 0/1
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30
31 /*
32  * offset from end of service tree
33  */
34 #define CFQ_IDLE_DELAY          (HZ / 5)
35
36 /*
37  * below this threshold, we consider thinktime immediate
38  */
39 #define CFQ_MIN_TT              (2)
40
41 #define CFQ_SLICE_SCALE         (5)
42 #define CFQ_HW_QUEUE_MIN        (5)
43
44 #define RQ_CIC(rq)              \
45         ((struct cfq_io_context *) (rq)->elevator_private)
46 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
47
48 static struct kmem_cache *cfq_pool;
49 static struct kmem_cache *cfq_ioc_pool;
50
51 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
52 static struct completion *ioc_gone;
53 static DEFINE_SPINLOCK(ioc_gone_lock);
54
55 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
56 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
57 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
58
59 #define sample_valid(samples)   ((samples) > 80)
60
61 /*
62  * Most of our rbtree usage is for sorting with min extraction, so
63  * if we cache the leftmost node we don't have to walk down the tree
64  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
65  * move this into the elevator for the rq sorting as well.
66  */
67 struct cfq_rb_root {
68         struct rb_root rb;
69         struct rb_node *left;
70 };
71 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, }
72
73 /*
74  * Per process-grouping structure
75  */
76 struct cfq_queue {
77         /* reference count */
78         atomic_t ref;
79         /* various state flags, see below */
80         unsigned int flags;
81         /* parent cfq_data */
82         struct cfq_data *cfqd;
83         /* service_tree member */
84         struct rb_node rb_node;
85         /* service_tree key */
86         unsigned long rb_key;
87         /* prio tree member */
88         struct rb_node p_node;
89         /* prio tree root we belong to, if any */
90         struct rb_root *p_root;
91         /* sorted list of pending requests */
92         struct rb_root sort_list;
93         /* if fifo isn't expired, next request to serve */
94         struct request *next_rq;
95         /* requests queued in sort_list */
96         int queued[2];
97         /* currently allocated requests */
98         int allocated[2];
99         /* fifo list of requests in sort_list */
100         struct list_head fifo;
101
102         unsigned long slice_end;
103         long slice_resid;
104         unsigned int slice_dispatch;
105
106         /* pending metadata requests */
107         int meta_pending;
108         /* number of requests that are on the dispatch list or inside driver */
109         int dispatched;
110
111         /* io prio of this group */
112         unsigned short ioprio, org_ioprio;
113         unsigned short ioprio_class, org_ioprio_class;
114
115         pid_t pid;
116 };
117
118 /*
119  * Per block device queue structure
120  */
121 struct cfq_data {
122         struct request_queue *queue;
123
124         /*
125          * rr list of queues with requests and the count of them
126          */
127         struct cfq_rb_root service_tree;
128
129         /*
130          * Each priority tree is sorted by next_request position.  These
131          * trees are used when determining if two or more queues are
132          * interleaving requests (see cfq_close_cooperator).
133          */
134         struct rb_root prio_trees[CFQ_PRIO_LISTS];
135
136         unsigned int busy_queues;
137
138         int rq_in_driver[2];
139         int sync_flight;
140
141         /*
142          * queue-depth detection
143          */
144         int rq_queued;
145         int hw_tag;
146         int hw_tag_samples;
147         int rq_in_driver_peak;
148
149         /*
150          * idle window management
151          */
152         struct timer_list idle_slice_timer;
153         struct work_struct unplug_work;
154
155         struct cfq_queue *active_queue;
156         struct cfq_io_context *active_cic;
157
158         /*
159          * async queue for each priority case
160          */
161         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
162         struct cfq_queue *async_idle_cfqq;
163
164         sector_t last_position;
165
166         /*
167          * tunables, see top of file
168          */
169         unsigned int cfq_quantum;
170         unsigned int cfq_fifo_expire[2];
171         unsigned int cfq_back_penalty;
172         unsigned int cfq_back_max;
173         unsigned int cfq_slice[2];
174         unsigned int cfq_slice_async_rq;
175         unsigned int cfq_slice_idle;
176         unsigned int cfq_latency;
177
178         struct list_head cic_list;
179
180         /*
181          * Fallback dummy cfqq for extreme OOM conditions
182          */
183         struct cfq_queue oom_cfqq;
184
185         unsigned long last_end_sync_rq;
186 };
187
188 enum cfqq_state_flags {
189         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
190         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
191         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
192         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
193         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
194         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
195         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
196         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
197         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
198         CFQ_CFQQ_FLAG_coop,             /* has done a coop jump of the queue */
199 };
200
201 #define CFQ_CFQQ_FNS(name)                                              \
202 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
203 {                                                                       \
204         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
205 }                                                                       \
206 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
207 {                                                                       \
208         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
209 }                                                                       \
210 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
211 {                                                                       \
212         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
213 }
214
215 CFQ_CFQQ_FNS(on_rr);
216 CFQ_CFQQ_FNS(wait_request);
217 CFQ_CFQQ_FNS(must_dispatch);
218 CFQ_CFQQ_FNS(must_alloc_slice);
219 CFQ_CFQQ_FNS(fifo_expire);
220 CFQ_CFQQ_FNS(idle_window);
221 CFQ_CFQQ_FNS(prio_changed);
222 CFQ_CFQQ_FNS(slice_new);
223 CFQ_CFQQ_FNS(sync);
224 CFQ_CFQQ_FNS(coop);
225 #undef CFQ_CFQQ_FNS
226
227 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
228         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
229 #define cfq_log(cfqd, fmt, args...)     \
230         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
231
232 static void cfq_dispatch_insert(struct request_queue *, struct request *);
233 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
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                                             bool 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, bool 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 bool cfq_bio_sync(struct bio *bio)
260 {
261         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
262 }
263
264 /*
265  * scheduler run of queue, if there are requests pending and no one in the
266  * driver that will restart queueing
267  */
268 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
269 {
270         if (cfqd->busy_queues) {
271                 cfq_log(cfqd, "schedule dispatch");
272                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
273         }
274 }
275
276 static int cfq_queue_empty(struct request_queue *q)
277 {
278         struct cfq_data *cfqd = q->elevator->elevator_data;
279
280         return !cfqd->busy_queues;
281 }
282
283 /*
284  * Scale schedule slice based on io priority. Use the sync time slice only
285  * if a queue is marked sync and has sync io queued. A sync queue with async
286  * io only, should not get full sync slice length.
287  */
288 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
289                                  unsigned short prio)
290 {
291         const int base_slice = cfqd->cfq_slice[sync];
292
293         WARN_ON(prio >= IOPRIO_BE_NR);
294
295         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
296 }
297
298 static inline int
299 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
300 {
301         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
302 }
303
304 static inline void
305 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
306 {
307         cfqq->slice_end = cfq_prio_to_slice(cfqd, cfqq) + jiffies;
308         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
309 }
310
311 /*
312  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
313  * isn't valid until the first request from the dispatch is activated
314  * and the slice time set.
315  */
316 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
317 {
318         if (cfq_cfqq_slice_new(cfqq))
319                 return 0;
320         if (time_before(jiffies, cfqq->slice_end))
321                 return 0;
322
323         return 1;
324 }
325
326 /*
327  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
328  * We choose the request that is closest to the head right now. Distance
329  * behind the head is penalized and only allowed to a certain extent.
330  */
331 static struct request *
332 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
333 {
334         sector_t last, s1, s2, d1 = 0, d2 = 0;
335         unsigned long back_max;
336 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
337 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
338         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
339
340         if (rq1 == NULL || rq1 == rq2)
341                 return rq2;
342         if (rq2 == NULL)
343                 return rq1;
344
345         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
346                 return rq1;
347         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
348                 return rq2;
349         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
350                 return rq1;
351         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
352                 return rq2;
353
354         s1 = blk_rq_pos(rq1);
355         s2 = blk_rq_pos(rq2);
356
357         last = cfqd->last_position;
358
359         /*
360          * by definition, 1KiB is 2 sectors
361          */
362         back_max = cfqd->cfq_back_max * 2;
363
364         /*
365          * Strict one way elevator _except_ in the case where we allow
366          * short backward seeks which are biased as twice the cost of a
367          * similar forward seek.
368          */
369         if (s1 >= last)
370                 d1 = s1 - last;
371         else if (s1 + back_max >= last)
372                 d1 = (last - s1) * cfqd->cfq_back_penalty;
373         else
374                 wrap |= CFQ_RQ1_WRAP;
375
376         if (s2 >= last)
377                 d2 = s2 - last;
378         else if (s2 + back_max >= last)
379                 d2 = (last - s2) * cfqd->cfq_back_penalty;
380         else
381                 wrap |= CFQ_RQ2_WRAP;
382
383         /* Found required data */
384
385         /*
386          * By doing switch() on the bit mask "wrap" we avoid having to
387          * check two variables for all permutations: --> faster!
388          */
389         switch (wrap) {
390         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
391                 if (d1 < d2)
392                         return rq1;
393                 else if (d2 < d1)
394                         return rq2;
395                 else {
396                         if (s1 >= s2)
397                                 return rq1;
398                         else
399                                 return rq2;
400                 }
401
402         case CFQ_RQ2_WRAP:
403                 return rq1;
404         case CFQ_RQ1_WRAP:
405                 return rq2;
406         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
407         default:
408                 /*
409                  * Since both rqs are wrapped,
410                  * start with the one that's further behind head
411                  * (--> only *one* back seek required),
412                  * since back seek takes more time than forward.
413                  */
414                 if (s1 <= s2)
415                         return rq1;
416                 else
417                         return rq2;
418         }
419 }
420
421 /*
422  * The below is leftmost cache rbtree addon
423  */
424 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
425 {
426         if (!root->left)
427                 root->left = rb_first(&root->rb);
428
429         if (root->left)
430                 return rb_entry(root->left, struct cfq_queue, rb_node);
431
432         return NULL;
433 }
434
435 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
436 {
437         rb_erase(n, root);
438         RB_CLEAR_NODE(n);
439 }
440
441 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
442 {
443         if (root->left == n)
444                 root->left = NULL;
445         rb_erase_init(n, &root->rb);
446 }
447
448 /*
449  * would be nice to take fifo expire time into account as well
450  */
451 static struct request *
452 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
453                   struct request *last)
454 {
455         struct rb_node *rbnext = rb_next(&last->rb_node);
456         struct rb_node *rbprev = rb_prev(&last->rb_node);
457         struct request *next = NULL, *prev = NULL;
458
459         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
460
461         if (rbprev)
462                 prev = rb_entry_rq(rbprev);
463
464         if (rbnext)
465                 next = rb_entry_rq(rbnext);
466         else {
467                 rbnext = rb_first(&cfqq->sort_list);
468                 if (rbnext && rbnext != &last->rb_node)
469                         next = rb_entry_rq(rbnext);
470         }
471
472         return cfq_choose_req(cfqd, next, prev);
473 }
474
475 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
476                                       struct cfq_queue *cfqq)
477 {
478         /*
479          * just an approximation, should be ok.
480          */
481         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
482                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
483 }
484
485 /*
486  * The cfqd->service_tree holds all pending cfq_queue's that have
487  * requests waiting to be processed. It is sorted in the order that
488  * we will service the queues.
489  */
490 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
491                                  bool add_front)
492 {
493         struct rb_node **p, *parent;
494         struct cfq_queue *__cfqq;
495         unsigned long rb_key;
496         int left;
497
498         if (cfq_class_idle(cfqq)) {
499                 rb_key = CFQ_IDLE_DELAY;
500                 parent = rb_last(&cfqd->service_tree.rb);
501                 if (parent && parent != &cfqq->rb_node) {
502                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
503                         rb_key += __cfqq->rb_key;
504                 } else
505                         rb_key += jiffies;
506         } else if (!add_front) {
507                 /*
508                  * Get our rb key offset. Subtract any residual slice
509                  * value carried from last service. A negative resid
510                  * count indicates slice overrun, and this should position
511                  * the next service time further away in the tree.
512                  */
513                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
514                 rb_key -= cfqq->slice_resid;
515                 cfqq->slice_resid = 0;
516         } else {
517                 rb_key = -HZ;
518                 __cfqq = cfq_rb_first(&cfqd->service_tree);
519                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
520         }
521
522         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
523                 /*
524                  * same position, nothing more to do
525                  */
526                 if (rb_key == cfqq->rb_key)
527                         return;
528
529                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
530         }
531
532         left = 1;
533         parent = NULL;
534         p = &cfqd->service_tree.rb.rb_node;
535         while (*p) {
536                 struct rb_node **n;
537
538                 parent = *p;
539                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
540
541                 /*
542                  * sort RT queues first, we always want to give
543                  * preference to them. IDLE queues goes to the back.
544                  * after that, sort on the next service time.
545                  */
546                 if (cfq_class_rt(cfqq) > cfq_class_rt(__cfqq))
547                         n = &(*p)->rb_left;
548                 else if (cfq_class_rt(cfqq) < cfq_class_rt(__cfqq))
549                         n = &(*p)->rb_right;
550                 else if (cfq_class_idle(cfqq) < cfq_class_idle(__cfqq))
551                         n = &(*p)->rb_left;
552                 else if (cfq_class_idle(cfqq) > cfq_class_idle(__cfqq))
553                         n = &(*p)->rb_right;
554                 else if (time_before(rb_key, __cfqq->rb_key))
555                         n = &(*p)->rb_left;
556                 else
557                         n = &(*p)->rb_right;
558
559                 if (n == &(*p)->rb_right)
560                         left = 0;
561
562                 p = n;
563         }
564
565         if (left)
566                 cfqd->service_tree.left = &cfqq->rb_node;
567
568         cfqq->rb_key = rb_key;
569         rb_link_node(&cfqq->rb_node, parent, p);
570         rb_insert_color(&cfqq->rb_node, &cfqd->service_tree.rb);
571 }
572
573 static struct cfq_queue *
574 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
575                      sector_t sector, struct rb_node **ret_parent,
576                      struct rb_node ***rb_link)
577 {
578         struct rb_node **p, *parent;
579         struct cfq_queue *cfqq = NULL;
580
581         parent = NULL;
582         p = &root->rb_node;
583         while (*p) {
584                 struct rb_node **n;
585
586                 parent = *p;
587                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
588
589                 /*
590                  * Sort strictly based on sector.  Smallest to the left,
591                  * largest to the right.
592                  */
593                 if (sector > blk_rq_pos(cfqq->next_rq))
594                         n = &(*p)->rb_right;
595                 else if (sector < blk_rq_pos(cfqq->next_rq))
596                         n = &(*p)->rb_left;
597                 else
598                         break;
599                 p = n;
600                 cfqq = NULL;
601         }
602
603         *ret_parent = parent;
604         if (rb_link)
605                 *rb_link = p;
606         return cfqq;
607 }
608
609 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
610 {
611         struct rb_node **p, *parent;
612         struct cfq_queue *__cfqq;
613
614         if (cfqq->p_root) {
615                 rb_erase(&cfqq->p_node, cfqq->p_root);
616                 cfqq->p_root = NULL;
617         }
618
619         if (cfq_class_idle(cfqq))
620                 return;
621         if (!cfqq->next_rq)
622                 return;
623
624         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
625         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
626                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
627         if (!__cfqq) {
628                 rb_link_node(&cfqq->p_node, parent, p);
629                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
630         } else
631                 cfqq->p_root = NULL;
632 }
633
634 /*
635  * Update cfqq's position in the service tree.
636  */
637 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
638 {
639         /*
640          * Resorting requires the cfqq to be on the RR list already.
641          */
642         if (cfq_cfqq_on_rr(cfqq)) {
643                 cfq_service_tree_add(cfqd, cfqq, 0);
644                 cfq_prio_tree_add(cfqd, cfqq);
645         }
646 }
647
648 /*
649  * add to busy list of queues for service, trying to be fair in ordering
650  * the pending list according to last request service
651  */
652 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
653 {
654         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
655         BUG_ON(cfq_cfqq_on_rr(cfqq));
656         cfq_mark_cfqq_on_rr(cfqq);
657         cfqd->busy_queues++;
658
659         cfq_resort_rr_list(cfqd, cfqq);
660 }
661
662 /*
663  * Called when the cfqq no longer has requests pending, remove it from
664  * the service tree.
665  */
666 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
667 {
668         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
669         BUG_ON(!cfq_cfqq_on_rr(cfqq));
670         cfq_clear_cfqq_on_rr(cfqq);
671
672         if (!RB_EMPTY_NODE(&cfqq->rb_node))
673                 cfq_rb_erase(&cfqq->rb_node, &cfqd->service_tree);
674         if (cfqq->p_root) {
675                 rb_erase(&cfqq->p_node, cfqq->p_root);
676                 cfqq->p_root = NULL;
677         }
678
679         BUG_ON(!cfqd->busy_queues);
680         cfqd->busy_queues--;
681 }
682
683 /*
684  * rb tree support functions
685  */
686 static void cfq_del_rq_rb(struct request *rq)
687 {
688         struct cfq_queue *cfqq = RQ_CFQQ(rq);
689         struct cfq_data *cfqd = cfqq->cfqd;
690         const int sync = rq_is_sync(rq);
691
692         BUG_ON(!cfqq->queued[sync]);
693         cfqq->queued[sync]--;
694
695         elv_rb_del(&cfqq->sort_list, rq);
696
697         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
698                 cfq_del_cfqq_rr(cfqd, cfqq);
699 }
700
701 static void cfq_add_rq_rb(struct request *rq)
702 {
703         struct cfq_queue *cfqq = RQ_CFQQ(rq);
704         struct cfq_data *cfqd = cfqq->cfqd;
705         struct request *__alias, *prev;
706
707         cfqq->queued[rq_is_sync(rq)]++;
708
709         /*
710          * looks a little odd, but the first insert might return an alias.
711          * if that happens, put the alias on the dispatch list
712          */
713         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
714                 cfq_dispatch_insert(cfqd->queue, __alias);
715
716         if (!cfq_cfqq_on_rr(cfqq))
717                 cfq_add_cfqq_rr(cfqd, cfqq);
718
719         /*
720          * check if this request is a better next-serve candidate
721          */
722         prev = cfqq->next_rq;
723         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
724
725         /*
726          * adjust priority tree position, if ->next_rq changes
727          */
728         if (prev != cfqq->next_rq)
729                 cfq_prio_tree_add(cfqd, cfqq);
730
731         BUG_ON(!cfqq->next_rq);
732 }
733
734 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
735 {
736         elv_rb_del(&cfqq->sort_list, rq);
737         cfqq->queued[rq_is_sync(rq)]--;
738         cfq_add_rq_rb(rq);
739 }
740
741 static struct request *
742 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
743 {
744         struct task_struct *tsk = current;
745         struct cfq_io_context *cic;
746         struct cfq_queue *cfqq;
747
748         cic = cfq_cic_lookup(cfqd, tsk->io_context);
749         if (!cic)
750                 return NULL;
751
752         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
753         if (cfqq) {
754                 sector_t sector = bio->bi_sector + bio_sectors(bio);
755
756                 return elv_rb_find(&cfqq->sort_list, sector);
757         }
758
759         return NULL;
760 }
761
762 static void cfq_activate_request(struct request_queue *q, struct request *rq)
763 {
764         struct cfq_data *cfqd = q->elevator->elevator_data;
765
766         cfqd->rq_in_driver[rq_is_sync(rq)]++;
767         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
768                                                 rq_in_driver(cfqd));
769
770         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
771 }
772
773 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
774 {
775         struct cfq_data *cfqd = q->elevator->elevator_data;
776         const int sync = rq_is_sync(rq);
777
778         WARN_ON(!cfqd->rq_in_driver[sync]);
779         cfqd->rq_in_driver[sync]--;
780         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
781                                                 rq_in_driver(cfqd));
782 }
783
784 static void cfq_remove_request(struct request *rq)
785 {
786         struct cfq_queue *cfqq = RQ_CFQQ(rq);
787
788         if (cfqq->next_rq == rq)
789                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
790
791         list_del_init(&rq->queuelist);
792         cfq_del_rq_rb(rq);
793
794         cfqq->cfqd->rq_queued--;
795         if (rq_is_meta(rq)) {
796                 WARN_ON(!cfqq->meta_pending);
797                 cfqq->meta_pending--;
798         }
799 }
800
801 static int cfq_merge(struct request_queue *q, struct request **req,
802                      struct bio *bio)
803 {
804         struct cfq_data *cfqd = q->elevator->elevator_data;
805         struct request *__rq;
806
807         __rq = cfq_find_rq_fmerge(cfqd, bio);
808         if (__rq && elv_rq_merge_ok(__rq, bio)) {
809                 *req = __rq;
810                 return ELEVATOR_FRONT_MERGE;
811         }
812
813         return ELEVATOR_NO_MERGE;
814 }
815
816 static void cfq_merged_request(struct request_queue *q, struct request *req,
817                                int type)
818 {
819         if (type == ELEVATOR_FRONT_MERGE) {
820                 struct cfq_queue *cfqq = RQ_CFQQ(req);
821
822                 cfq_reposition_rq_rb(cfqq, req);
823         }
824 }
825
826 static void
827 cfq_merged_requests(struct request_queue *q, struct request *rq,
828                     struct request *next)
829 {
830         /*
831          * reposition in fifo if next is older than rq
832          */
833         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
834             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
835                 list_move(&rq->queuelist, &next->queuelist);
836                 rq_set_fifo_time(rq, rq_fifo_time(next));
837         }
838
839         cfq_remove_request(next);
840 }
841
842 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
843                            struct bio *bio)
844 {
845         struct cfq_data *cfqd = q->elevator->elevator_data;
846         struct cfq_io_context *cic;
847         struct cfq_queue *cfqq;
848
849         /*
850          * Disallow merge of a sync bio into an async request.
851          */
852         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
853                 return false;
854
855         /*
856          * Lookup the cfqq that this bio will be queued with. Allow
857          * merge only if rq is queued there.
858          */
859         cic = cfq_cic_lookup(cfqd, current->io_context);
860         if (!cic)
861                 return false;
862
863         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
864         return cfqq == RQ_CFQQ(rq);
865 }
866
867 static void __cfq_set_active_queue(struct cfq_data *cfqd,
868                                    struct cfq_queue *cfqq)
869 {
870         if (cfqq) {
871                 cfq_log_cfqq(cfqd, cfqq, "set_active");
872                 cfqq->slice_end = 0;
873                 cfqq->slice_dispatch = 0;
874
875                 cfq_clear_cfqq_wait_request(cfqq);
876                 cfq_clear_cfqq_must_dispatch(cfqq);
877                 cfq_clear_cfqq_must_alloc_slice(cfqq);
878                 cfq_clear_cfqq_fifo_expire(cfqq);
879                 cfq_mark_cfqq_slice_new(cfqq);
880
881                 del_timer(&cfqd->idle_slice_timer);
882         }
883
884         cfqd->active_queue = cfqq;
885 }
886
887 /*
888  * current cfqq expired its slice (or was too idle), select new one
889  */
890 static void
891 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
892                     bool timed_out)
893 {
894         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
895
896         if (cfq_cfqq_wait_request(cfqq))
897                 del_timer(&cfqd->idle_slice_timer);
898
899         cfq_clear_cfqq_wait_request(cfqq);
900
901         /*
902          * store what was left of this slice, if the queue idled/timed out
903          */
904         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
905                 cfqq->slice_resid = cfqq->slice_end - jiffies;
906                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
907         }
908
909         cfq_resort_rr_list(cfqd, cfqq);
910
911         if (cfqq == cfqd->active_queue)
912                 cfqd->active_queue = NULL;
913
914         if (cfqd->active_cic) {
915                 put_io_context(cfqd->active_cic->ioc);
916                 cfqd->active_cic = NULL;
917         }
918 }
919
920 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
921 {
922         struct cfq_queue *cfqq = cfqd->active_queue;
923
924         if (cfqq)
925                 __cfq_slice_expired(cfqd, cfqq, timed_out);
926 }
927
928 /*
929  * Get next queue for service. Unless we have a queue preemption,
930  * we'll simply select the first cfqq in the service tree.
931  */
932 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
933 {
934         if (RB_EMPTY_ROOT(&cfqd->service_tree.rb))
935                 return NULL;
936
937         return cfq_rb_first(&cfqd->service_tree);
938 }
939
940 /*
941  * Get and set a new active queue for service.
942  */
943 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
944                                               struct cfq_queue *cfqq)
945 {
946         if (!cfqq) {
947                 cfqq = cfq_get_next_queue(cfqd);
948                 if (cfqq)
949                         cfq_clear_cfqq_coop(cfqq);
950         }
951
952         __cfq_set_active_queue(cfqd, cfqq);
953         return cfqq;
954 }
955
956 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
957                                           struct request *rq)
958 {
959         if (blk_rq_pos(rq) >= cfqd->last_position)
960                 return blk_rq_pos(rq) - cfqd->last_position;
961         else
962                 return cfqd->last_position - blk_rq_pos(rq);
963 }
964
965 #define CIC_SEEK_THR    8 * 1024
966 #define CIC_SEEKY(cic)  ((cic)->seek_mean > CIC_SEEK_THR)
967
968 static inline int cfq_rq_close(struct cfq_data *cfqd, struct request *rq)
969 {
970         struct cfq_io_context *cic = cfqd->active_cic;
971         sector_t sdist = cic->seek_mean;
972
973         if (!sample_valid(cic->seek_samples))
974                 sdist = CIC_SEEK_THR;
975
976         return cfq_dist_from_last(cfqd, rq) <= sdist;
977 }
978
979 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
980                                     struct cfq_queue *cur_cfqq)
981 {
982         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
983         struct rb_node *parent, *node;
984         struct cfq_queue *__cfqq;
985         sector_t sector = cfqd->last_position;
986
987         if (RB_EMPTY_ROOT(root))
988                 return NULL;
989
990         /*
991          * First, if we find a request starting at the end of the last
992          * request, choose it.
993          */
994         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
995         if (__cfqq)
996                 return __cfqq;
997
998         /*
999          * If the exact sector wasn't found, the parent of the NULL leaf
1000          * will contain the closest sector.
1001          */
1002         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1003         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1004                 return __cfqq;
1005
1006         if (blk_rq_pos(__cfqq->next_rq) < sector)
1007                 node = rb_next(&__cfqq->p_node);
1008         else
1009                 node = rb_prev(&__cfqq->p_node);
1010         if (!node)
1011                 return NULL;
1012
1013         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1014         if (cfq_rq_close(cfqd, __cfqq->next_rq))
1015                 return __cfqq;
1016
1017         return NULL;
1018 }
1019
1020 /*
1021  * cfqd - obvious
1022  * cur_cfqq - passed in so that we don't decide that the current queue is
1023  *            closely cooperating with itself.
1024  *
1025  * So, basically we're assuming that that cur_cfqq has dispatched at least
1026  * one request, and that cfqd->last_position reflects a position on the disk
1027  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1028  * assumption.
1029  */
1030 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1031                                               struct cfq_queue *cur_cfqq,
1032                                               bool probe)
1033 {
1034         struct cfq_queue *cfqq;
1035
1036         /*
1037          * A valid cfq_io_context is necessary to compare requests against
1038          * the seek_mean of the current cfqq.
1039          */
1040         if (!cfqd->active_cic)
1041                 return NULL;
1042
1043         /*
1044          * We should notice if some of the queues are cooperating, eg
1045          * working closely on the same area of the disk. In that case,
1046          * we can group them together and don't waste time idling.
1047          */
1048         cfqq = cfqq_close(cfqd, cur_cfqq);
1049         if (!cfqq)
1050                 return NULL;
1051
1052         if (cfq_cfqq_coop(cfqq))
1053                 return NULL;
1054
1055         if (!probe)
1056                 cfq_mark_cfqq_coop(cfqq);
1057         return cfqq;
1058 }
1059
1060 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1061 {
1062         struct cfq_queue *cfqq = cfqd->active_queue;
1063         struct cfq_io_context *cic;
1064         unsigned long sl;
1065
1066         /*
1067          * SSD device without seek penalty, disable idling. But only do so
1068          * for devices that support queuing, otherwise we still have a problem
1069          * with sync vs async workloads.
1070          */
1071         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1072                 return;
1073
1074         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1075         WARN_ON(cfq_cfqq_slice_new(cfqq));
1076
1077         /*
1078          * idle is disabled, either manually or by past process history
1079          */
1080         if (!cfqd->cfq_slice_idle || !cfq_cfqq_idle_window(cfqq))
1081                 return;
1082
1083         /*
1084          * still requests with the driver, don't idle
1085          */
1086         if (rq_in_driver(cfqd))
1087                 return;
1088
1089         /*
1090          * task has exited, don't wait
1091          */
1092         cic = cfqd->active_cic;
1093         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1094                 return;
1095
1096         cfq_mark_cfqq_wait_request(cfqq);
1097
1098         /*
1099          * we don't want to idle for seeks, but we do want to allow
1100          * fair distribution of slice time for a process doing back-to-back
1101          * seeks. so allow a little bit of time for him to submit a new rq
1102          */
1103         sl = cfqd->cfq_slice_idle;
1104         if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1105                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1106
1107         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1108         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1109 }
1110
1111 /*
1112  * Move request from internal lists to the request queue dispatch list.
1113  */
1114 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1115 {
1116         struct cfq_data *cfqd = q->elevator->elevator_data;
1117         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1118
1119         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1120
1121         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1122         cfq_remove_request(rq);
1123         cfqq->dispatched++;
1124         elv_dispatch_sort(q, rq);
1125
1126         if (cfq_cfqq_sync(cfqq))
1127                 cfqd->sync_flight++;
1128 }
1129
1130 /*
1131  * return expired entry, or NULL to just start from scratch in rbtree
1132  */
1133 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1134 {
1135         struct request *rq = NULL;
1136
1137         if (cfq_cfqq_fifo_expire(cfqq))
1138                 return NULL;
1139
1140         cfq_mark_cfqq_fifo_expire(cfqq);
1141
1142         if (list_empty(&cfqq->fifo))
1143                 return NULL;
1144
1145         rq = rq_entry_fifo(cfqq->fifo.next);
1146         if (time_before(jiffies, rq_fifo_time(rq)))
1147                 rq = NULL;
1148
1149         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1150         return rq;
1151 }
1152
1153 static inline int
1154 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1155 {
1156         const int base_rq = cfqd->cfq_slice_async_rq;
1157
1158         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1159
1160         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1161 }
1162
1163 /*
1164  * Select a queue for service. If we have a current active queue,
1165  * check whether to continue servicing it, or retrieve and set a new one.
1166  */
1167 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1168 {
1169         struct cfq_queue *cfqq, *new_cfqq = NULL;
1170
1171         cfqq = cfqd->active_queue;
1172         if (!cfqq)
1173                 goto new_queue;
1174
1175         /*
1176          * The active queue has run out of time, expire it and select new.
1177          */
1178         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1179                 goto expire;
1180
1181         /*
1182          * The active queue has requests and isn't expired, allow it to
1183          * dispatch.
1184          */
1185         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1186                 goto keep_queue;
1187
1188         /*
1189          * If another queue has a request waiting within our mean seek
1190          * distance, let it run.  The expire code will check for close
1191          * cooperators and put the close queue at the front of the service
1192          * tree.
1193          */
1194         new_cfqq = cfq_close_cooperator(cfqd, cfqq, 0);
1195         if (new_cfqq)
1196                 goto expire;
1197
1198         /*
1199          * No requests pending. If the active queue still has requests in
1200          * flight or is idling for a new request, allow either of these
1201          * conditions to happen (or time out) before selecting a new queue.
1202          */
1203         if (timer_pending(&cfqd->idle_slice_timer) ||
1204             (cfqq->dispatched && cfq_cfqq_idle_window(cfqq))) {
1205                 cfqq = NULL;
1206                 goto keep_queue;
1207         }
1208
1209 expire:
1210         cfq_slice_expired(cfqd, 0);
1211 new_queue:
1212         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1213 keep_queue:
1214         return cfqq;
1215 }
1216
1217 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1218 {
1219         int dispatched = 0;
1220
1221         while (cfqq->next_rq) {
1222                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1223                 dispatched++;
1224         }
1225
1226         BUG_ON(!list_empty(&cfqq->fifo));
1227         return dispatched;
1228 }
1229
1230 /*
1231  * Drain our current requests. Used for barriers and when switching
1232  * io schedulers on-the-fly.
1233  */
1234 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1235 {
1236         struct cfq_queue *cfqq;
1237         int dispatched = 0;
1238
1239         while ((cfqq = cfq_rb_first(&cfqd->service_tree)) != NULL)
1240                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1241
1242         cfq_slice_expired(cfqd, 0);
1243
1244         BUG_ON(cfqd->busy_queues);
1245
1246         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1247         return dispatched;
1248 }
1249
1250 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1251 {
1252         unsigned int max_dispatch;
1253
1254         /*
1255          * Drain async requests before we start sync IO
1256          */
1257         if (cfq_cfqq_idle_window(cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1258                 return false;
1259
1260         /*
1261          * If this is an async queue and we have sync IO in flight, let it wait
1262          */
1263         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1264                 return false;
1265
1266         max_dispatch = cfqd->cfq_quantum;
1267         if (cfq_class_idle(cfqq))
1268                 max_dispatch = 1;
1269
1270         /*
1271          * Does this cfqq already have too much IO in flight?
1272          */
1273         if (cfqq->dispatched >= max_dispatch) {
1274                 /*
1275                  * idle queue must always only have a single IO in flight
1276                  */
1277                 if (cfq_class_idle(cfqq))
1278                         return false;
1279
1280                 /*
1281                  * We have other queues, don't allow more IO from this one
1282                  */
1283                 if (cfqd->busy_queues > 1)
1284                         return false;
1285
1286                 /*
1287                  * Sole queue user, allow bigger slice
1288                  */
1289                 max_dispatch *= 4;
1290         }
1291
1292         /*
1293          * Async queues must wait a bit before being allowed dispatch.
1294          * We also ramp up the dispatch depth gradually for async IO,
1295          * based on the last sync IO we serviced
1296          */
1297         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1298                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1299                 unsigned int depth;
1300
1301                 depth = last_sync / cfqd->cfq_slice[1];
1302                 if (!depth && !cfqq->dispatched)
1303                         depth = 1;
1304                 if (depth < max_dispatch)
1305                         max_dispatch = depth;
1306         }
1307
1308         /*
1309          * If we're below the current max, allow a dispatch
1310          */
1311         return cfqq->dispatched < max_dispatch;
1312 }
1313
1314 /*
1315  * Dispatch a request from cfqq, moving them to the request queue
1316  * dispatch list.
1317  */
1318 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1319 {
1320         struct request *rq;
1321
1322         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1323
1324         if (!cfq_may_dispatch(cfqd, cfqq))
1325                 return false;
1326
1327         /*
1328          * follow expired path, else get first next available
1329          */
1330         rq = cfq_check_fifo(cfqq);
1331         if (!rq)
1332                 rq = cfqq->next_rq;
1333
1334         /*
1335          * insert request into driver dispatch list
1336          */
1337         cfq_dispatch_insert(cfqd->queue, rq);
1338
1339         if (!cfqd->active_cic) {
1340                 struct cfq_io_context *cic = RQ_CIC(rq);
1341
1342                 atomic_long_inc(&cic->ioc->refcount);
1343                 cfqd->active_cic = cic;
1344         }
1345
1346         return true;
1347 }
1348
1349 /*
1350  * Find the cfqq that we need to service and move a request from that to the
1351  * dispatch list
1352  */
1353 static int cfq_dispatch_requests(struct request_queue *q, int force)
1354 {
1355         struct cfq_data *cfqd = q->elevator->elevator_data;
1356         struct cfq_queue *cfqq;
1357
1358         if (!cfqd->busy_queues)
1359                 return 0;
1360
1361         if (unlikely(force))
1362                 return cfq_forced_dispatch(cfqd);
1363
1364         cfqq = cfq_select_queue(cfqd);
1365         if (!cfqq)
1366                 return 0;
1367
1368         /*
1369          * Dispatch a request from this cfqq, if it is allowed
1370          */
1371         if (!cfq_dispatch_request(cfqd, cfqq))
1372                 return 0;
1373
1374         cfqq->slice_dispatch++;
1375         cfq_clear_cfqq_must_dispatch(cfqq);
1376
1377         /*
1378          * expire an async queue immediately if it has used up its slice. idle
1379          * queue always expire after 1 dispatch round.
1380          */
1381         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1382             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1383             cfq_class_idle(cfqq))) {
1384                 cfqq->slice_end = jiffies + 1;
1385                 cfq_slice_expired(cfqd, 0);
1386         }
1387
1388         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1389         return 1;
1390 }
1391
1392 /*
1393  * task holds one reference to the queue, dropped when task exits. each rq
1394  * in-flight on this queue also holds a reference, dropped when rq is freed.
1395  *
1396  * queue lock must be held here.
1397  */
1398 static void cfq_put_queue(struct cfq_queue *cfqq)
1399 {
1400         struct cfq_data *cfqd = cfqq->cfqd;
1401
1402         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1403
1404         if (!atomic_dec_and_test(&cfqq->ref))
1405                 return;
1406
1407         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1408         BUG_ON(rb_first(&cfqq->sort_list));
1409         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1410         BUG_ON(cfq_cfqq_on_rr(cfqq));
1411
1412         if (unlikely(cfqd->active_queue == cfqq)) {
1413                 __cfq_slice_expired(cfqd, cfqq, 0);
1414                 cfq_schedule_dispatch(cfqd);
1415         }
1416
1417         kmem_cache_free(cfq_pool, cfqq);
1418 }
1419
1420 /*
1421  * Must always be called with the rcu_read_lock() held
1422  */
1423 static void
1424 __call_for_each_cic(struct io_context *ioc,
1425                     void (*func)(struct io_context *, struct cfq_io_context *))
1426 {
1427         struct cfq_io_context *cic;
1428         struct hlist_node *n;
1429
1430         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1431                 func(ioc, cic);
1432 }
1433
1434 /*
1435  * Call func for each cic attached to this ioc.
1436  */
1437 static void
1438 call_for_each_cic(struct io_context *ioc,
1439                   void (*func)(struct io_context *, struct cfq_io_context *))
1440 {
1441         rcu_read_lock();
1442         __call_for_each_cic(ioc, func);
1443         rcu_read_unlock();
1444 }
1445
1446 static void cfq_cic_free_rcu(struct rcu_head *head)
1447 {
1448         struct cfq_io_context *cic;
1449
1450         cic = container_of(head, struct cfq_io_context, rcu_head);
1451
1452         kmem_cache_free(cfq_ioc_pool, cic);
1453         elv_ioc_count_dec(cfq_ioc_count);
1454
1455         if (ioc_gone) {
1456                 /*
1457                  * CFQ scheduler is exiting, grab exit lock and check
1458                  * the pending io context count. If it hits zero,
1459                  * complete ioc_gone and set it back to NULL
1460                  */
1461                 spin_lock(&ioc_gone_lock);
1462                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1463                         complete(ioc_gone);
1464                         ioc_gone = NULL;
1465                 }
1466                 spin_unlock(&ioc_gone_lock);
1467         }
1468 }
1469
1470 static void cfq_cic_free(struct cfq_io_context *cic)
1471 {
1472         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1473 }
1474
1475 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1476 {
1477         unsigned long flags;
1478
1479         BUG_ON(!cic->dead_key);
1480
1481         spin_lock_irqsave(&ioc->lock, flags);
1482         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1483         hlist_del_rcu(&cic->cic_list);
1484         spin_unlock_irqrestore(&ioc->lock, flags);
1485
1486         cfq_cic_free(cic);
1487 }
1488
1489 /*
1490  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1491  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1492  * and ->trim() which is called with the task lock held
1493  */
1494 static void cfq_free_io_context(struct io_context *ioc)
1495 {
1496         /*
1497          * ioc->refcount is zero here, or we are called from elv_unregister(),
1498          * so no more cic's are allowed to be linked into this ioc.  So it
1499          * should be ok to iterate over the known list, we will see all cic's
1500          * since no new ones are added.
1501          */
1502         __call_for_each_cic(ioc, cic_free_func);
1503 }
1504
1505 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1506 {
1507         if (unlikely(cfqq == cfqd->active_queue)) {
1508                 __cfq_slice_expired(cfqd, cfqq, 0);
1509                 cfq_schedule_dispatch(cfqd);
1510         }
1511
1512         cfq_put_queue(cfqq);
1513 }
1514
1515 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1516                                          struct cfq_io_context *cic)
1517 {
1518         struct io_context *ioc = cic->ioc;
1519
1520         list_del_init(&cic->queue_list);
1521
1522         /*
1523          * Make sure key == NULL is seen for dead queues
1524          */
1525         smp_wmb();
1526         cic->dead_key = (unsigned long) cic->key;
1527         cic->key = NULL;
1528
1529         if (ioc->ioc_data == cic)
1530                 rcu_assign_pointer(ioc->ioc_data, NULL);
1531
1532         if (cic->cfqq[BLK_RW_ASYNC]) {
1533                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1534                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1535         }
1536
1537         if (cic->cfqq[BLK_RW_SYNC]) {
1538                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1539                 cic->cfqq[BLK_RW_SYNC] = NULL;
1540         }
1541 }
1542
1543 static void cfq_exit_single_io_context(struct io_context *ioc,
1544                                        struct cfq_io_context *cic)
1545 {
1546         struct cfq_data *cfqd = cic->key;
1547
1548         if (cfqd) {
1549                 struct request_queue *q = cfqd->queue;
1550                 unsigned long flags;
1551
1552                 spin_lock_irqsave(q->queue_lock, flags);
1553
1554                 /*
1555                  * Ensure we get a fresh copy of the ->key to prevent
1556                  * race between exiting task and queue
1557                  */
1558                 smp_read_barrier_depends();
1559                 if (cic->key)
1560                         __cfq_exit_single_io_context(cfqd, cic);
1561
1562                 spin_unlock_irqrestore(q->queue_lock, flags);
1563         }
1564 }
1565
1566 /*
1567  * The process that ioc belongs to has exited, we need to clean up
1568  * and put the internal structures we have that belongs to that process.
1569  */
1570 static void cfq_exit_io_context(struct io_context *ioc)
1571 {
1572         call_for_each_cic(ioc, cfq_exit_single_io_context);
1573 }
1574
1575 static struct cfq_io_context *
1576 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1577 {
1578         struct cfq_io_context *cic;
1579
1580         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1581                                                         cfqd->queue->node);
1582         if (cic) {
1583                 cic->last_end_request = jiffies;
1584                 INIT_LIST_HEAD(&cic->queue_list);
1585                 INIT_HLIST_NODE(&cic->cic_list);
1586                 cic->dtor = cfq_free_io_context;
1587                 cic->exit = cfq_exit_io_context;
1588                 elv_ioc_count_inc(cfq_ioc_count);
1589         }
1590
1591         return cic;
1592 }
1593
1594 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1595 {
1596         struct task_struct *tsk = current;
1597         int ioprio_class;
1598
1599         if (!cfq_cfqq_prio_changed(cfqq))
1600                 return;
1601
1602         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1603         switch (ioprio_class) {
1604         default:
1605                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1606         case IOPRIO_CLASS_NONE:
1607                 /*
1608                  * no prio set, inherit CPU scheduling settings
1609                  */
1610                 cfqq->ioprio = task_nice_ioprio(tsk);
1611                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1612                 break;
1613         case IOPRIO_CLASS_RT:
1614                 cfqq->ioprio = task_ioprio(ioc);
1615                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1616                 break;
1617         case IOPRIO_CLASS_BE:
1618                 cfqq->ioprio = task_ioprio(ioc);
1619                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1620                 break;
1621         case IOPRIO_CLASS_IDLE:
1622                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1623                 cfqq->ioprio = 7;
1624                 cfq_clear_cfqq_idle_window(cfqq);
1625                 break;
1626         }
1627
1628         /*
1629          * keep track of original prio settings in case we have to temporarily
1630          * elevate the priority of this queue
1631          */
1632         cfqq->org_ioprio = cfqq->ioprio;
1633         cfqq->org_ioprio_class = cfqq->ioprio_class;
1634         cfq_clear_cfqq_prio_changed(cfqq);
1635 }
1636
1637 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1638 {
1639         struct cfq_data *cfqd = cic->key;
1640         struct cfq_queue *cfqq;
1641         unsigned long flags;
1642
1643         if (unlikely(!cfqd))
1644                 return;
1645
1646         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1647
1648         cfqq = cic->cfqq[BLK_RW_ASYNC];
1649         if (cfqq) {
1650                 struct cfq_queue *new_cfqq;
1651                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
1652                                                 GFP_ATOMIC);
1653                 if (new_cfqq) {
1654                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
1655                         cfq_put_queue(cfqq);
1656                 }
1657         }
1658
1659         cfqq = cic->cfqq[BLK_RW_SYNC];
1660         if (cfqq)
1661                 cfq_mark_cfqq_prio_changed(cfqq);
1662
1663         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1664 }
1665
1666 static void cfq_ioc_set_ioprio(struct io_context *ioc)
1667 {
1668         call_for_each_cic(ioc, changed_ioprio);
1669         ioc->ioprio_changed = 0;
1670 }
1671
1672 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1673                           pid_t pid, bool is_sync)
1674 {
1675         RB_CLEAR_NODE(&cfqq->rb_node);
1676         RB_CLEAR_NODE(&cfqq->p_node);
1677         INIT_LIST_HEAD(&cfqq->fifo);
1678
1679         atomic_set(&cfqq->ref, 0);
1680         cfqq->cfqd = cfqd;
1681
1682         cfq_mark_cfqq_prio_changed(cfqq);
1683
1684         if (is_sync) {
1685                 if (!cfq_class_idle(cfqq))
1686                         cfq_mark_cfqq_idle_window(cfqq);
1687                 cfq_mark_cfqq_sync(cfqq);
1688         }
1689         cfqq->pid = pid;
1690 }
1691
1692 static struct cfq_queue *
1693 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
1694                      struct io_context *ioc, gfp_t gfp_mask)
1695 {
1696         struct cfq_queue *cfqq, *new_cfqq = NULL;
1697         struct cfq_io_context *cic;
1698
1699 retry:
1700         cic = cfq_cic_lookup(cfqd, ioc);
1701         /* cic always exists here */
1702         cfqq = cic_to_cfqq(cic, is_sync);
1703
1704         /*
1705          * Always try a new alloc if we fell back to the OOM cfqq
1706          * originally, since it should just be a temporary situation.
1707          */
1708         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
1709                 cfqq = NULL;
1710                 if (new_cfqq) {
1711                         cfqq = new_cfqq;
1712                         new_cfqq = NULL;
1713                 } else if (gfp_mask & __GFP_WAIT) {
1714                         spin_unlock_irq(cfqd->queue->queue_lock);
1715                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
1716                                         gfp_mask | __GFP_ZERO,
1717                                         cfqd->queue->node);
1718                         spin_lock_irq(cfqd->queue->queue_lock);
1719                         if (new_cfqq)
1720                                 goto retry;
1721                 } else {
1722                         cfqq = kmem_cache_alloc_node(cfq_pool,
1723                                         gfp_mask | __GFP_ZERO,
1724                                         cfqd->queue->node);
1725                 }
1726
1727                 if (cfqq) {
1728                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
1729                         cfq_init_prio_data(cfqq, ioc);
1730                         cfq_log_cfqq(cfqd, cfqq, "alloced");
1731                 } else
1732                         cfqq = &cfqd->oom_cfqq;
1733         }
1734
1735         if (new_cfqq)
1736                 kmem_cache_free(cfq_pool, new_cfqq);
1737
1738         return cfqq;
1739 }
1740
1741 static struct cfq_queue **
1742 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
1743 {
1744         switch (ioprio_class) {
1745         case IOPRIO_CLASS_RT:
1746                 return &cfqd->async_cfqq[0][ioprio];
1747         case IOPRIO_CLASS_BE:
1748                 return &cfqd->async_cfqq[1][ioprio];
1749         case IOPRIO_CLASS_IDLE:
1750                 return &cfqd->async_idle_cfqq;
1751         default:
1752                 BUG();
1753         }
1754 }
1755
1756 static struct cfq_queue *
1757 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
1758               gfp_t gfp_mask)
1759 {
1760         const int ioprio = task_ioprio(ioc);
1761         const int ioprio_class = task_ioprio_class(ioc);
1762         struct cfq_queue **async_cfqq = NULL;
1763         struct cfq_queue *cfqq = NULL;
1764
1765         if (!is_sync) {
1766                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
1767                 cfqq = *async_cfqq;
1768         }
1769
1770         if (!cfqq)
1771                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
1772
1773         /*
1774          * pin the queue now that it's allocated, scheduler exit will prune it
1775          */
1776         if (!is_sync && !(*async_cfqq)) {
1777                 atomic_inc(&cfqq->ref);
1778                 *async_cfqq = cfqq;
1779         }
1780
1781         atomic_inc(&cfqq->ref);
1782         return cfqq;
1783 }
1784
1785 /*
1786  * We drop cfq io contexts lazily, so we may find a dead one.
1787  */
1788 static void
1789 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
1790                   struct cfq_io_context *cic)
1791 {
1792         unsigned long flags;
1793
1794         WARN_ON(!list_empty(&cic->queue_list));
1795
1796         spin_lock_irqsave(&ioc->lock, flags);
1797
1798         BUG_ON(ioc->ioc_data == cic);
1799
1800         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
1801         hlist_del_rcu(&cic->cic_list);
1802         spin_unlock_irqrestore(&ioc->lock, flags);
1803
1804         cfq_cic_free(cic);
1805 }
1806
1807 static struct cfq_io_context *
1808 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
1809 {
1810         struct cfq_io_context *cic;
1811         unsigned long flags;
1812         void *k;
1813
1814         if (unlikely(!ioc))
1815                 return NULL;
1816
1817         rcu_read_lock();
1818
1819         /*
1820          * we maintain a last-hit cache, to avoid browsing over the tree
1821          */
1822         cic = rcu_dereference(ioc->ioc_data);
1823         if (cic && cic->key == cfqd) {
1824                 rcu_read_unlock();
1825                 return cic;
1826         }
1827
1828         do {
1829                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
1830                 rcu_read_unlock();
1831                 if (!cic)
1832                         break;
1833                 /* ->key must be copied to avoid race with cfq_exit_queue() */
1834                 k = cic->key;
1835                 if (unlikely(!k)) {
1836                         cfq_drop_dead_cic(cfqd, ioc, cic);
1837                         rcu_read_lock();
1838                         continue;
1839                 }
1840
1841                 spin_lock_irqsave(&ioc->lock, flags);
1842                 rcu_assign_pointer(ioc->ioc_data, cic);
1843                 spin_unlock_irqrestore(&ioc->lock, flags);
1844                 break;
1845         } while (1);
1846
1847         return cic;
1848 }
1849
1850 /*
1851  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
1852  * the process specific cfq io context when entered from the block layer.
1853  * Also adds the cic to a per-cfqd list, used when this queue is removed.
1854  */
1855 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
1856                         struct cfq_io_context *cic, gfp_t gfp_mask)
1857 {
1858         unsigned long flags;
1859         int ret;
1860
1861         ret = radix_tree_preload(gfp_mask);
1862         if (!ret) {
1863                 cic->ioc = ioc;
1864                 cic->key = cfqd;
1865
1866                 spin_lock_irqsave(&ioc->lock, flags);
1867                 ret = radix_tree_insert(&ioc->radix_root,
1868                                                 (unsigned long) cfqd, cic);
1869                 if (!ret)
1870                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
1871                 spin_unlock_irqrestore(&ioc->lock, flags);
1872
1873                 radix_tree_preload_end();
1874
1875                 if (!ret) {
1876                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1877                         list_add(&cic->queue_list, &cfqd->cic_list);
1878                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1879                 }
1880         }
1881
1882         if (ret)
1883                 printk(KERN_ERR "cfq: cic link failed!\n");
1884
1885         return ret;
1886 }
1887
1888 /*
1889  * Setup general io context and cfq io context. There can be several cfq
1890  * io contexts per general io context, if this process is doing io to more
1891  * than one device managed by cfq.
1892  */
1893 static struct cfq_io_context *
1894 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1895 {
1896         struct io_context *ioc = NULL;
1897         struct cfq_io_context *cic;
1898
1899         might_sleep_if(gfp_mask & __GFP_WAIT);
1900
1901         ioc = get_io_context(gfp_mask, cfqd->queue->node);
1902         if (!ioc)
1903                 return NULL;
1904
1905         cic = cfq_cic_lookup(cfqd, ioc);
1906         if (cic)
1907                 goto out;
1908
1909         cic = cfq_alloc_io_context(cfqd, gfp_mask);
1910         if (cic == NULL)
1911                 goto err;
1912
1913         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
1914                 goto err_free;
1915
1916 out:
1917         smp_read_barrier_depends();
1918         if (unlikely(ioc->ioprio_changed))
1919                 cfq_ioc_set_ioprio(ioc);
1920
1921         return cic;
1922 err_free:
1923         cfq_cic_free(cic);
1924 err:
1925         put_io_context(ioc);
1926         return NULL;
1927 }
1928
1929 static void
1930 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
1931 {
1932         unsigned long elapsed = jiffies - cic->last_end_request;
1933         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
1934
1935         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
1936         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
1937         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
1938 }
1939
1940 static void
1941 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_io_context *cic,
1942                        struct request *rq)
1943 {
1944         sector_t sdist;
1945         u64 total;
1946
1947         if (!cic->last_request_pos)
1948                 sdist = 0;
1949         else if (cic->last_request_pos < blk_rq_pos(rq))
1950                 sdist = blk_rq_pos(rq) - cic->last_request_pos;
1951         else
1952                 sdist = cic->last_request_pos - blk_rq_pos(rq);
1953
1954         /*
1955          * Don't allow the seek distance to get too large from the
1956          * odd fragment, pagein, etc
1957          */
1958         if (cic->seek_samples <= 60) /* second&third seek */
1959                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*1024);
1960         else
1961                 sdist = min(sdist, (cic->seek_mean * 4) + 2*1024*64);
1962
1963         cic->seek_samples = (7*cic->seek_samples + 256) / 8;
1964         cic->seek_total = (7*cic->seek_total + (u64)256*sdist) / 8;
1965         total = cic->seek_total + (cic->seek_samples/2);
1966         do_div(total, cic->seek_samples);
1967         cic->seek_mean = (sector_t)total;
1968 }
1969
1970 /*
1971  * Disable idle window if the process thinks too long or seeks so much that
1972  * it doesn't matter
1973  */
1974 static void
1975 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1976                        struct cfq_io_context *cic)
1977 {
1978         int old_idle, enable_idle;
1979
1980         /*
1981          * Don't idle for async or idle io prio class
1982          */
1983         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
1984                 return;
1985
1986         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
1987
1988         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
1989             (!cfqd->cfq_latency && cfqd->hw_tag && CIC_SEEKY(cic)))
1990                 enable_idle = 0;
1991         else if (sample_valid(cic->ttime_samples)) {
1992                 unsigned int slice_idle = cfqd->cfq_slice_idle;
1993                 if (sample_valid(cic->seek_samples) && CIC_SEEKY(cic))
1994                         slice_idle = msecs_to_jiffies(CFQ_MIN_TT);
1995                 if (cic->ttime_mean > slice_idle)
1996                         enable_idle = 0;
1997                 else
1998                         enable_idle = 1;
1999         }
2000
2001         if (old_idle != enable_idle) {
2002                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2003                 if (enable_idle)
2004                         cfq_mark_cfqq_idle_window(cfqq);
2005                 else
2006                         cfq_clear_cfqq_idle_window(cfqq);
2007         }
2008 }
2009
2010 /*
2011  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2012  * no or if we aren't sure, a 1 will cause a preempt.
2013  */
2014 static bool
2015 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2016                    struct request *rq)
2017 {
2018         struct cfq_queue *cfqq;
2019
2020         cfqq = cfqd->active_queue;
2021         if (!cfqq)
2022                 return false;
2023
2024         if (cfq_slice_used(cfqq))
2025                 return true;
2026
2027         if (cfq_class_idle(new_cfqq))
2028                 return false;
2029
2030         if (cfq_class_idle(cfqq))
2031                 return true;
2032
2033         /*
2034          * if the new request is sync, but the currently running queue is
2035          * not, let the sync request have priority.
2036          */
2037         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2038                 return true;
2039
2040         /*
2041          * So both queues are sync. Let the new request get disk time if
2042          * it's a metadata request and the current queue is doing regular IO.
2043          */
2044         if (rq_is_meta(rq) && !cfqq->meta_pending)
2045                 return false;
2046
2047         /*
2048          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2049          */
2050         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2051                 return true;
2052
2053         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2054                 return false;
2055
2056         /*
2057          * if this request is as-good as one we would expect from the
2058          * current cfqq, let it preempt
2059          */
2060         if (cfq_rq_close(cfqd, rq))
2061                 return true;
2062
2063         return false;
2064 }
2065
2066 /*
2067  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2068  * let it have half of its nominal slice.
2069  */
2070 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2071 {
2072         cfq_log_cfqq(cfqd, cfqq, "preempt");
2073         cfq_slice_expired(cfqd, 1);
2074
2075         /*
2076          * Put the new queue at the front of the of the current list,
2077          * so we know that it will be selected next.
2078          */
2079         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2080
2081         cfq_service_tree_add(cfqd, cfqq, 1);
2082
2083         cfqq->slice_end = 0;
2084         cfq_mark_cfqq_slice_new(cfqq);
2085 }
2086
2087 /*
2088  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2089  * something we should do about it
2090  */
2091 static void
2092 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2093                 struct request *rq)
2094 {
2095         struct cfq_io_context *cic = RQ_CIC(rq);
2096
2097         cfqd->rq_queued++;
2098         if (rq_is_meta(rq))
2099                 cfqq->meta_pending++;
2100
2101         cfq_update_io_thinktime(cfqd, cic);
2102         cfq_update_io_seektime(cfqd, cic, rq);
2103         cfq_update_idle_window(cfqd, cfqq, cic);
2104
2105         cic->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2106
2107         if (cfqq == cfqd->active_queue) {
2108                 /*
2109                  * Remember that we saw a request from this process, but
2110                  * don't start queuing just yet. Otherwise we risk seeing lots
2111                  * of tiny requests, because we disrupt the normal plugging
2112                  * and merging. If the request is already larger than a single
2113                  * page, let it rip immediately. For that case we assume that
2114                  * merging is already done. Ditto for a busy system that
2115                  * has other work pending, don't risk delaying until the
2116                  * idle timer unplug to continue working.
2117                  */
2118                 if (cfq_cfqq_wait_request(cfqq)) {
2119                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2120                             cfqd->busy_queues > 1) {
2121                                 del_timer(&cfqd->idle_slice_timer);
2122                         __blk_run_queue(cfqd->queue);
2123                         }
2124                         cfq_mark_cfqq_must_dispatch(cfqq);
2125                 }
2126         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2127                 /*
2128                  * not the active queue - expire current slice if it is
2129                  * idle and has expired it's mean thinktime or this new queue
2130                  * has some old slice time left and is of higher priority or
2131                  * this new queue is RT and the current one is BE
2132                  */
2133                 cfq_preempt_queue(cfqd, cfqq);
2134                 __blk_run_queue(cfqd->queue);
2135         }
2136 }
2137
2138 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2139 {
2140         struct cfq_data *cfqd = q->elevator->elevator_data;
2141         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2142
2143         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2144         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2145
2146         cfq_add_rq_rb(rq);
2147
2148         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2149         list_add_tail(&rq->queuelist, &cfqq->fifo);
2150
2151         cfq_rq_enqueued(cfqd, cfqq, rq);
2152 }
2153
2154 /*
2155  * Update hw_tag based on peak queue depth over 50 samples under
2156  * sufficient load.
2157  */
2158 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2159 {
2160         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2161                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2162
2163         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2164             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2165                 return;
2166
2167         if (cfqd->hw_tag_samples++ < 50)
2168                 return;
2169
2170         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2171                 cfqd->hw_tag = 1;
2172         else
2173                 cfqd->hw_tag = 0;
2174
2175         cfqd->hw_tag_samples = 0;
2176         cfqd->rq_in_driver_peak = 0;
2177 }
2178
2179 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2180 {
2181         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2182         struct cfq_data *cfqd = cfqq->cfqd;
2183         const int sync = rq_is_sync(rq);
2184         unsigned long now;
2185
2186         now = jiffies;
2187         cfq_log_cfqq(cfqd, cfqq, "complete");
2188
2189         cfq_update_hw_tag(cfqd);
2190
2191         WARN_ON(!cfqd->rq_in_driver[sync]);
2192         WARN_ON(!cfqq->dispatched);
2193         cfqd->rq_in_driver[sync]--;
2194         cfqq->dispatched--;
2195
2196         if (cfq_cfqq_sync(cfqq))
2197                 cfqd->sync_flight--;
2198
2199         if (sync) {
2200                 RQ_CIC(rq)->last_end_request = now;
2201                 cfqd->last_end_sync_rq = now;
2202         }
2203
2204         /*
2205          * If this is the active queue, check if it needs to be expired,
2206          * or if we want to idle in case it has no pending requests.
2207          */
2208         if (cfqd->active_queue == cfqq) {
2209                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2210
2211                 if (cfq_cfqq_slice_new(cfqq)) {
2212                         cfq_set_prio_slice(cfqd, cfqq);
2213                         cfq_clear_cfqq_slice_new(cfqq);
2214                 }
2215                 /*
2216                  * If there are no requests waiting in this queue, and
2217                  * there are other queues ready to issue requests, AND
2218                  * those other queues are issuing requests within our
2219                  * mean seek distance, give them a chance to run instead
2220                  * of idling.
2221                  */
2222                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2223                         cfq_slice_expired(cfqd, 1);
2224                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq, 1) &&
2225                          sync && !rq_noidle(rq))
2226                         cfq_arm_slice_timer(cfqd);
2227         }
2228
2229         if (!rq_in_driver(cfqd))
2230                 cfq_schedule_dispatch(cfqd);
2231 }
2232
2233 /*
2234  * we temporarily boost lower priority queues if they are holding fs exclusive
2235  * resources. they are boosted to normal prio (CLASS_BE/4)
2236  */
2237 static void cfq_prio_boost(struct cfq_queue *cfqq)
2238 {
2239         if (has_fs_excl()) {
2240                 /*
2241                  * boost idle prio on transactions that would lock out other
2242                  * users of the filesystem
2243                  */
2244                 if (cfq_class_idle(cfqq))
2245                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2246                 if (cfqq->ioprio > IOPRIO_NORM)
2247                         cfqq->ioprio = IOPRIO_NORM;
2248         } else {
2249                 /*
2250                  * check if we need to unboost the queue
2251                  */
2252                 if (cfqq->ioprio_class != cfqq->org_ioprio_class)
2253                         cfqq->ioprio_class = cfqq->org_ioprio_class;
2254                 if (cfqq->ioprio != cfqq->org_ioprio)
2255                         cfqq->ioprio = cfqq->org_ioprio;
2256         }
2257 }
2258
2259 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2260 {
2261         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2262                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2263                 return ELV_MQUEUE_MUST;
2264         }
2265
2266         return ELV_MQUEUE_MAY;
2267 }
2268
2269 static int cfq_may_queue(struct request_queue *q, int rw)
2270 {
2271         struct cfq_data *cfqd = q->elevator->elevator_data;
2272         struct task_struct *tsk = current;
2273         struct cfq_io_context *cic;
2274         struct cfq_queue *cfqq;
2275
2276         /*
2277          * don't force setup of a queue from here, as a call to may_queue
2278          * does not necessarily imply that a request actually will be queued.
2279          * so just lookup a possibly existing queue, or return 'may queue'
2280          * if that fails
2281          */
2282         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2283         if (!cic)
2284                 return ELV_MQUEUE_MAY;
2285
2286         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2287         if (cfqq) {
2288                 cfq_init_prio_data(cfqq, cic->ioc);
2289                 cfq_prio_boost(cfqq);
2290
2291                 return __cfq_may_queue(cfqq);
2292         }
2293
2294         return ELV_MQUEUE_MAY;
2295 }
2296
2297 /*
2298  * queue lock held here
2299  */
2300 static void cfq_put_request(struct request *rq)
2301 {
2302         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2303
2304         if (cfqq) {
2305                 const int rw = rq_data_dir(rq);
2306
2307                 BUG_ON(!cfqq->allocated[rw]);
2308                 cfqq->allocated[rw]--;
2309
2310                 put_io_context(RQ_CIC(rq)->ioc);
2311
2312                 rq->elevator_private = NULL;
2313                 rq->elevator_private2 = NULL;
2314
2315                 cfq_put_queue(cfqq);
2316         }
2317 }
2318
2319 /*
2320  * Allocate cfq data structures associated with this request.
2321  */
2322 static int
2323 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2324 {
2325         struct cfq_data *cfqd = q->elevator->elevator_data;
2326         struct cfq_io_context *cic;
2327         const int rw = rq_data_dir(rq);
2328         const bool is_sync = rq_is_sync(rq);
2329         struct cfq_queue *cfqq;
2330         unsigned long flags;
2331
2332         might_sleep_if(gfp_mask & __GFP_WAIT);
2333
2334         cic = cfq_get_io_context(cfqd, gfp_mask);
2335
2336         spin_lock_irqsave(q->queue_lock, flags);
2337
2338         if (!cic)
2339                 goto queue_fail;
2340
2341         cfqq = cic_to_cfqq(cic, is_sync);
2342         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2343                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2344                 cic_set_cfqq(cic, cfqq, is_sync);
2345         }
2346
2347         cfqq->allocated[rw]++;
2348         atomic_inc(&cfqq->ref);
2349
2350         spin_unlock_irqrestore(q->queue_lock, flags);
2351
2352         rq->elevator_private = cic;
2353         rq->elevator_private2 = cfqq;
2354         return 0;
2355
2356 queue_fail:
2357         if (cic)
2358                 put_io_context(cic->ioc);
2359
2360         cfq_schedule_dispatch(cfqd);
2361         spin_unlock_irqrestore(q->queue_lock, flags);
2362         cfq_log(cfqd, "set_request fail");
2363         return 1;
2364 }
2365
2366 static void cfq_kick_queue(struct work_struct *work)
2367 {
2368         struct cfq_data *cfqd =
2369                 container_of(work, struct cfq_data, unplug_work);
2370         struct request_queue *q = cfqd->queue;
2371
2372         spin_lock_irq(q->queue_lock);
2373         __blk_run_queue(cfqd->queue);
2374         spin_unlock_irq(q->queue_lock);
2375 }
2376
2377 /*
2378  * Timer running if the active_queue is currently idling inside its time slice
2379  */
2380 static void cfq_idle_slice_timer(unsigned long data)
2381 {
2382         struct cfq_data *cfqd = (struct cfq_data *) data;
2383         struct cfq_queue *cfqq;
2384         unsigned long flags;
2385         int timed_out = 1;
2386
2387         cfq_log(cfqd, "idle timer fired");
2388
2389         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2390
2391         cfqq = cfqd->active_queue;
2392         if (cfqq) {
2393                 timed_out = 0;
2394
2395                 /*
2396                  * We saw a request before the queue expired, let it through
2397                  */
2398                 if (cfq_cfqq_must_dispatch(cfqq))
2399                         goto out_kick;
2400
2401                 /*
2402                  * expired
2403                  */
2404                 if (cfq_slice_used(cfqq))
2405                         goto expire;
2406
2407                 /*
2408                  * only expire and reinvoke request handler, if there are
2409                  * other queues with pending requests
2410                  */
2411                 if (!cfqd->busy_queues)
2412                         goto out_cont;
2413
2414                 /*
2415                  * not expired and it has a request pending, let it dispatch
2416                  */
2417                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2418                         goto out_kick;
2419         }
2420 expire:
2421         cfq_slice_expired(cfqd, timed_out);
2422 out_kick:
2423         cfq_schedule_dispatch(cfqd);
2424 out_cont:
2425         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2426 }
2427
2428 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2429 {
2430         del_timer_sync(&cfqd->idle_slice_timer);
2431         cancel_work_sync(&cfqd->unplug_work);
2432 }
2433
2434 static void cfq_put_async_queues(struct cfq_data *cfqd)
2435 {
2436         int i;
2437
2438         for (i = 0; i < IOPRIO_BE_NR; i++) {
2439                 if (cfqd->async_cfqq[0][i])
2440                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2441                 if (cfqd->async_cfqq[1][i])
2442                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2443         }
2444
2445         if (cfqd->async_idle_cfqq)
2446                 cfq_put_queue(cfqd->async_idle_cfqq);
2447 }
2448
2449 static void cfq_exit_queue(struct elevator_queue *e)
2450 {
2451         struct cfq_data *cfqd = e->elevator_data;
2452         struct request_queue *q = cfqd->queue;
2453
2454         cfq_shutdown_timer_wq(cfqd);
2455
2456         spin_lock_irq(q->queue_lock);
2457
2458         if (cfqd->active_queue)
2459                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2460
2461         while (!list_empty(&cfqd->cic_list)) {
2462                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2463                                                         struct cfq_io_context,
2464                                                         queue_list);
2465
2466                 __cfq_exit_single_io_context(cfqd, cic);
2467         }
2468
2469         cfq_put_async_queues(cfqd);
2470
2471         spin_unlock_irq(q->queue_lock);
2472
2473         cfq_shutdown_timer_wq(cfqd);
2474
2475         kfree(cfqd);
2476 }
2477
2478 static void *cfq_init_queue(struct request_queue *q)
2479 {
2480         struct cfq_data *cfqd;
2481         int i;
2482
2483         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2484         if (!cfqd)
2485                 return NULL;
2486
2487         cfqd->service_tree = CFQ_RB_ROOT;
2488
2489         /*
2490          * Not strictly needed (since RB_ROOT just clears the node and we
2491          * zeroed cfqd on alloc), but better be safe in case someone decides
2492          * to add magic to the rb code
2493          */
2494         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2495                 cfqd->prio_trees[i] = RB_ROOT;
2496
2497         /*
2498          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2499          * Grab a permanent reference to it, so that the normal code flow
2500          * will not attempt to free it.
2501          */
2502         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2503         atomic_inc(&cfqd->oom_cfqq.ref);
2504
2505         INIT_LIST_HEAD(&cfqd->cic_list);
2506
2507         cfqd->queue = q;
2508
2509         init_timer(&cfqd->idle_slice_timer);
2510         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2511         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2512
2513         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2514
2515         cfqd->cfq_quantum = cfq_quantum;
2516         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2517         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2518         cfqd->cfq_back_max = cfq_back_max;
2519         cfqd->cfq_back_penalty = cfq_back_penalty;
2520         cfqd->cfq_slice[0] = cfq_slice_async;
2521         cfqd->cfq_slice[1] = cfq_slice_sync;
2522         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2523         cfqd->cfq_slice_idle = cfq_slice_idle;
2524         cfqd->cfq_latency = 1;
2525         cfqd->hw_tag = 1;
2526         cfqd->last_end_sync_rq = jiffies;
2527         return cfqd;
2528 }
2529
2530 static void cfq_slab_kill(void)
2531 {
2532         /*
2533          * Caller already ensured that pending RCU callbacks are completed,
2534          * so we should have no busy allocations at this point.
2535          */
2536         if (cfq_pool)
2537                 kmem_cache_destroy(cfq_pool);
2538         if (cfq_ioc_pool)
2539                 kmem_cache_destroy(cfq_ioc_pool);
2540 }
2541
2542 static int __init cfq_slab_setup(void)
2543 {
2544         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2545         if (!cfq_pool)
2546                 goto fail;
2547
2548         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2549         if (!cfq_ioc_pool)
2550                 goto fail;
2551
2552         return 0;
2553 fail:
2554         cfq_slab_kill();
2555         return -ENOMEM;
2556 }
2557
2558 /*
2559  * sysfs parts below -->
2560  */
2561 static ssize_t
2562 cfq_var_show(unsigned int var, char *page)
2563 {
2564         return sprintf(page, "%d\n", var);
2565 }
2566
2567 static ssize_t
2568 cfq_var_store(unsigned int *var, const char *page, size_t count)
2569 {
2570         char *p = (char *) page;
2571
2572         *var = simple_strtoul(p, &p, 10);
2573         return count;
2574 }
2575
2576 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
2577 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
2578 {                                                                       \
2579         struct cfq_data *cfqd = e->elevator_data;                       \
2580         unsigned int __data = __VAR;                                    \
2581         if (__CONV)                                                     \
2582                 __data = jiffies_to_msecs(__data);                      \
2583         return cfq_var_show(__data, (page));                            \
2584 }
2585 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
2586 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
2587 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
2588 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
2589 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
2590 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
2591 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
2592 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
2593 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
2594 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
2595 #undef SHOW_FUNCTION
2596
2597 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
2598 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
2599 {                                                                       \
2600         struct cfq_data *cfqd = e->elevator_data;                       \
2601         unsigned int __data;                                            \
2602         int ret = cfq_var_store(&__data, (page), count);                \
2603         if (__data < (MIN))                                             \
2604                 __data = (MIN);                                         \
2605         else if (__data > (MAX))                                        \
2606                 __data = (MAX);                                         \
2607         if (__CONV)                                                     \
2608                 *(__PTR) = msecs_to_jiffies(__data);                    \
2609         else                                                            \
2610                 *(__PTR) = __data;                                      \
2611         return ret;                                                     \
2612 }
2613 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
2614 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
2615                 UINT_MAX, 1);
2616 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
2617                 UINT_MAX, 1);
2618 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
2619 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
2620                 UINT_MAX, 0);
2621 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
2622 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
2623 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
2624 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
2625                 UINT_MAX, 0);
2626 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
2627 #undef STORE_FUNCTION
2628
2629 #define CFQ_ATTR(name) \
2630         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
2631
2632 static struct elv_fs_entry cfq_attrs[] = {
2633         CFQ_ATTR(quantum),
2634         CFQ_ATTR(fifo_expire_sync),
2635         CFQ_ATTR(fifo_expire_async),
2636         CFQ_ATTR(back_seek_max),
2637         CFQ_ATTR(back_seek_penalty),
2638         CFQ_ATTR(slice_sync),
2639         CFQ_ATTR(slice_async),
2640         CFQ_ATTR(slice_async_rq),
2641         CFQ_ATTR(slice_idle),
2642         CFQ_ATTR(low_latency),
2643         __ATTR_NULL
2644 };
2645
2646 static struct elevator_type iosched_cfq = {
2647         .ops = {
2648                 .elevator_merge_fn =            cfq_merge,
2649                 .elevator_merged_fn =           cfq_merged_request,
2650                 .elevator_merge_req_fn =        cfq_merged_requests,
2651                 .elevator_allow_merge_fn =      cfq_allow_merge,
2652                 .elevator_dispatch_fn =         cfq_dispatch_requests,
2653                 .elevator_add_req_fn =          cfq_insert_request,
2654                 .elevator_activate_req_fn =     cfq_activate_request,
2655                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
2656                 .elevator_queue_empty_fn =      cfq_queue_empty,
2657                 .elevator_completed_req_fn =    cfq_completed_request,
2658                 .elevator_former_req_fn =       elv_rb_former_request,
2659                 .elevator_latter_req_fn =       elv_rb_latter_request,
2660                 .elevator_set_req_fn =          cfq_set_request,
2661                 .elevator_put_req_fn =          cfq_put_request,
2662                 .elevator_may_queue_fn =        cfq_may_queue,
2663                 .elevator_init_fn =             cfq_init_queue,
2664                 .elevator_exit_fn =             cfq_exit_queue,
2665                 .trim =                         cfq_free_io_context,
2666         },
2667         .elevator_attrs =       cfq_attrs,
2668         .elevator_name =        "cfq",
2669         .elevator_owner =       THIS_MODULE,
2670 };
2671
2672 static int __init cfq_init(void)
2673 {
2674         /*
2675          * could be 0 on HZ < 1000 setups
2676          */
2677         if (!cfq_slice_async)
2678                 cfq_slice_async = 1;
2679         if (!cfq_slice_idle)
2680                 cfq_slice_idle = 1;
2681
2682         if (cfq_slab_setup())
2683                 return -ENOMEM;
2684
2685         elv_register(&iosched_cfq);
2686
2687         return 0;
2688 }
2689
2690 static void __exit cfq_exit(void)
2691 {
2692         DECLARE_COMPLETION_ONSTACK(all_gone);
2693         elv_unregister(&iosched_cfq);
2694         ioc_gone = &all_gone;
2695         /* ioc_gone's update must be visible before reading ioc_count */
2696         smp_wmb();
2697
2698         /*
2699          * this also protects us from entering cfq_slab_kill() with
2700          * pending RCU callbacks
2701          */
2702         if (elv_ioc_count_read(cfq_ioc_count))
2703                 wait_for_completion(&all_gone);
2704         cfq_slab_kill();
2705 }
2706
2707 module_init(cfq_init);
2708 module_exit(cfq_exit);
2709
2710 MODULE_AUTHOR("Jens Axboe");
2711 MODULE_LICENSE("GPL");
2712 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");