15f7238f527f3f42e4a6ad327ee5cd8624eb6ad1
[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/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16
17 /*
18  * tunables
19  */
20 /* max queue in one round of service */
21 static const int cfq_quantum = 4;
22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23 /* maximum backwards seek, in KiB */
24 static const int cfq_back_max = 16 * 1024;
25 /* penalty of a backwards seek */
26 static const int cfq_back_penalty = 2;
27 static const int cfq_slice_sync = HZ / 10;
28 static int cfq_slice_async = HZ / 25;
29 static const int cfq_slice_async_rq = 2;
30 static int cfq_slice_idle = HZ / 125;
31 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
32 static const int cfq_hist_divisor = 4;
33
34 /*
35  * offset from end of service tree
36  */
37 #define CFQ_IDLE_DELAY          (HZ / 5)
38
39 /*
40  * below this threshold, we consider thinktime immediate
41  */
42 #define CFQ_MIN_TT              (2)
43
44 /*
45  * Allow merged cfqqs to perform this amount of seeky I/O before
46  * deciding to break the queues up again.
47  */
48 #define CFQQ_COOP_TOUT          (HZ)
49
50 #define CFQ_SLICE_SCALE         (5)
51 #define CFQ_HW_QUEUE_MIN        (5)
52
53 #define RQ_CIC(rq)              \
54         ((struct cfq_io_context *) (rq)->elevator_private)
55 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
56
57 static struct kmem_cache *cfq_pool;
58 static struct kmem_cache *cfq_ioc_pool;
59
60 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
61 static struct completion *ioc_gone;
62 static DEFINE_SPINLOCK(ioc_gone_lock);
63
64 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67
68 #define sample_valid(samples)   ((samples) > 80)
69
70 /*
71  * Most of our rbtree usage is for sorting with min extraction, so
72  * if we cache the leftmost node we don't have to walk down the tree
73  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74  * move this into the elevator for the rq sorting as well.
75  */
76 struct cfq_rb_root {
77         struct rb_root rb;
78         struct rb_node *left;
79         unsigned count;
80 };
81 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
82
83 /*
84  * Per process-grouping structure
85  */
86 struct cfq_queue {
87         /* reference count */
88         atomic_t ref;
89         /* various state flags, see below */
90         unsigned int flags;
91         /* parent cfq_data */
92         struct cfq_data *cfqd;
93         /* service_tree member */
94         struct rb_node rb_node;
95         /* service_tree key */
96         unsigned long rb_key;
97         /* prio tree member */
98         struct rb_node p_node;
99         /* prio tree root we belong to, if any */
100         struct rb_root *p_root;
101         /* sorted list of pending requests */
102         struct rb_root sort_list;
103         /* if fifo isn't expired, next request to serve */
104         struct request *next_rq;
105         /* requests queued in sort_list */
106         int queued[2];
107         /* currently allocated requests */
108         int allocated[2];
109         /* fifo list of requests in sort_list */
110         struct list_head fifo;
111
112         unsigned long slice_end;
113         long slice_resid;
114         unsigned int slice_dispatch;
115
116         /* pending metadata requests */
117         int meta_pending;
118         /* number of requests that are on the dispatch list or inside driver */
119         int dispatched;
120
121         /* io prio of this group */
122         unsigned short ioprio, org_ioprio;
123         unsigned short ioprio_class, org_ioprio_class;
124
125         unsigned int seek_samples;
126         u64 seek_total;
127         sector_t seek_mean;
128         sector_t last_request_pos;
129         unsigned long seeky_start;
130
131         pid_t pid;
132
133         struct cfq_rb_root *service_tree;
134         struct cfq_queue *new_cfqq;
135 };
136
137 /*
138  * First index in the service_trees.
139  * IDLE is handled separately, so it has negative index
140  */
141 enum wl_prio_t {
142         IDLE_WORKLOAD = -1,
143         BE_WORKLOAD = 0,
144         RT_WORKLOAD = 1
145 };
146
147 /*
148  * Second index in the service_trees.
149  */
150 enum wl_type_t {
151         ASYNC_WORKLOAD = 0,
152         SYNC_NOIDLE_WORKLOAD = 1,
153         SYNC_WORKLOAD = 2
154 };
155
156
157 /*
158  * Per block device queue structure
159  */
160 struct cfq_data {
161         struct request_queue *queue;
162
163         /*
164          * rr lists of queues with requests, onle rr for each priority class.
165          * Counts are embedded in the cfq_rb_root
166          */
167         struct cfq_rb_root service_trees[2][3];
168         struct cfq_rb_root service_tree_idle;
169         /*
170          * The priority currently being served
171          */
172         enum wl_prio_t serving_prio;
173         enum wl_type_t serving_type;
174         unsigned long workload_expires;
175
176         /*
177          * Each priority tree is sorted by next_request position.  These
178          * trees are used when determining if two or more queues are
179          * interleaving requests (see cfq_close_cooperator).
180          */
181         struct rb_root prio_trees[CFQ_PRIO_LISTS];
182
183         unsigned int busy_queues;
184         unsigned int busy_queues_avg[2];
185
186         int rq_in_driver[2];
187         int sync_flight;
188
189         /*
190          * queue-depth detection
191          */
192         int rq_queued;
193         int hw_tag;
194         /*
195          * hw_tag can be
196          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
197          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
198          *  0 => no NCQ
199          */
200         int hw_tag_est_depth;
201         unsigned int hw_tag_samples;
202
203         /*
204          * idle window management
205          */
206         struct timer_list idle_slice_timer;
207         struct work_struct unplug_work;
208
209         struct cfq_queue *active_queue;
210         struct cfq_io_context *active_cic;
211
212         /*
213          * async queue for each priority case
214          */
215         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
216         struct cfq_queue *async_idle_cfqq;
217
218         sector_t last_position;
219
220         /*
221          * tunables, see top of file
222          */
223         unsigned int cfq_quantum;
224         unsigned int cfq_fifo_expire[2];
225         unsigned int cfq_back_penalty;
226         unsigned int cfq_back_max;
227         unsigned int cfq_slice[2];
228         unsigned int cfq_slice_async_rq;
229         unsigned int cfq_slice_idle;
230         unsigned int cfq_latency;
231
232         struct list_head cic_list;
233
234         /*
235          * Fallback dummy cfqq for extreme OOM conditions
236          */
237         struct cfq_queue oom_cfqq;
238
239         unsigned long last_end_sync_rq;
240 };
241
242 static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio,
243                                             enum wl_type_t type,
244                                             struct cfq_data *cfqd)
245 {
246         if (prio == IDLE_WORKLOAD)
247                 return &cfqd->service_tree_idle;
248
249         return &cfqd->service_trees[prio][type];
250 }
251
252 enum cfqq_state_flags {
253         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
254         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
255         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
256         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
257         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
258         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
259         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
260         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
261         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
262         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
263         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
264 };
265
266 #define CFQ_CFQQ_FNS(name)                                              \
267 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
268 {                                                                       \
269         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
270 }                                                                       \
271 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
272 {                                                                       \
273         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
274 }                                                                       \
275 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
276 {                                                                       \
277         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
278 }
279
280 CFQ_CFQQ_FNS(on_rr);
281 CFQ_CFQQ_FNS(wait_request);
282 CFQ_CFQQ_FNS(must_dispatch);
283 CFQ_CFQQ_FNS(must_alloc_slice);
284 CFQ_CFQQ_FNS(fifo_expire);
285 CFQ_CFQQ_FNS(idle_window);
286 CFQ_CFQQ_FNS(prio_changed);
287 CFQ_CFQQ_FNS(slice_new);
288 CFQ_CFQQ_FNS(sync);
289 CFQ_CFQQ_FNS(coop);
290 CFQ_CFQQ_FNS(deep);
291 #undef CFQ_CFQQ_FNS
292
293 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
294         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
295 #define cfq_log(cfqd, fmt, args...)     \
296         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
297
298 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
299 {
300         if (cfq_class_idle(cfqq))
301                 return IDLE_WORKLOAD;
302         if (cfq_class_rt(cfqq))
303                 return RT_WORKLOAD;
304         return BE_WORKLOAD;
305 }
306
307
308 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
309 {
310         if (!cfq_cfqq_sync(cfqq))
311                 return ASYNC_WORKLOAD;
312         if (!cfq_cfqq_idle_window(cfqq))
313                 return SYNC_NOIDLE_WORKLOAD;
314         return SYNC_WORKLOAD;
315 }
316
317 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
318 {
319         if (wl == IDLE_WORKLOAD)
320                 return cfqd->service_tree_idle.count;
321
322         return cfqd->service_trees[wl][ASYNC_WORKLOAD].count
323                 + cfqd->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
324                 + cfqd->service_trees[wl][SYNC_WORKLOAD].count;
325 }
326
327 static void cfq_dispatch_insert(struct request_queue *, struct request *);
328 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
329                                        struct io_context *, gfp_t);
330 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
331                                                 struct io_context *);
332
333 static inline int rq_in_driver(struct cfq_data *cfqd)
334 {
335         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
336 }
337
338 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
339                                             bool is_sync)
340 {
341         return cic->cfqq[is_sync];
342 }
343
344 static inline void cic_set_cfqq(struct cfq_io_context *cic,
345                                 struct cfq_queue *cfqq, bool is_sync)
346 {
347         cic->cfqq[is_sync] = cfqq;
348 }
349
350 /*
351  * We regard a request as SYNC, if it's either a read or has the SYNC bit
352  * set (in which case it could also be direct WRITE).
353  */
354 static inline bool cfq_bio_sync(struct bio *bio)
355 {
356         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
357 }
358
359 /*
360  * scheduler run of queue, if there are requests pending and no one in the
361  * driver that will restart queueing
362  */
363 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
364 {
365         if (cfqd->busy_queues) {
366                 cfq_log(cfqd, "schedule dispatch");
367                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
368         }
369 }
370
371 static int cfq_queue_empty(struct request_queue *q)
372 {
373         struct cfq_data *cfqd = q->elevator->elevator_data;
374
375         return !cfqd->busy_queues;
376 }
377
378 /*
379  * Scale schedule slice based on io priority. Use the sync time slice only
380  * if a queue is marked sync and has sync io queued. A sync queue with async
381  * io only, should not get full sync slice length.
382  */
383 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
384                                  unsigned short prio)
385 {
386         const int base_slice = cfqd->cfq_slice[sync];
387
388         WARN_ON(prio >= IOPRIO_BE_NR);
389
390         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
391 }
392
393 static inline int
394 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
395 {
396         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
397 }
398
399 /*
400  * get averaged number of queues of RT/BE priority.
401  * average is updated, with a formula that gives more weight to higher numbers,
402  * to quickly follows sudden increases and decrease slowly
403  */
404
405 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
406 {
407         unsigned min_q, max_q;
408         unsigned mult  = cfq_hist_divisor - 1;
409         unsigned round = cfq_hist_divisor / 2;
410         unsigned busy = cfq_busy_queues_wl(rt, cfqd);
411
412         min_q = min(cfqd->busy_queues_avg[rt], busy);
413         max_q = max(cfqd->busy_queues_avg[rt], busy);
414         cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
415                 cfq_hist_divisor;
416         return cfqd->busy_queues_avg[rt];
417 }
418
419 static inline void
420 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
421 {
422         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
423         if (cfqd->cfq_latency) {
424                 /* interested queues (we consider only the ones with the same
425                  * priority class) */
426                 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
427                 unsigned sync_slice = cfqd->cfq_slice[1];
428                 unsigned expect_latency = sync_slice * iq;
429                 if (expect_latency > cfq_target_latency) {
430                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
431                         /* scale low_slice according to IO priority
432                          * and sync vs async */
433                         unsigned low_slice =
434                                 min(slice, base_low_slice * slice / sync_slice);
435                         /* the adapted slice value is scaled to fit all iqs
436                          * into the target latency */
437                         slice = max(slice * cfq_target_latency / expect_latency,
438                                     low_slice);
439                 }
440         }
441         cfqq->slice_end = jiffies + slice;
442         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
443 }
444
445 /*
446  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
447  * isn't valid until the first request from the dispatch is activated
448  * and the slice time set.
449  */
450 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
451 {
452         if (cfq_cfqq_slice_new(cfqq))
453                 return 0;
454         if (time_before(jiffies, cfqq->slice_end))
455                 return 0;
456
457         return 1;
458 }
459
460 /*
461  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
462  * We choose the request that is closest to the head right now. Distance
463  * behind the head is penalized and only allowed to a certain extent.
464  */
465 static struct request *
466 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
467 {
468         sector_t s1, s2, d1 = 0, d2 = 0;
469         unsigned long back_max;
470 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
471 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
472         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
473
474         if (rq1 == NULL || rq1 == rq2)
475                 return rq2;
476         if (rq2 == NULL)
477                 return rq1;
478
479         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
480                 return rq1;
481         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
482                 return rq2;
483         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
484                 return rq1;
485         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
486                 return rq2;
487
488         s1 = blk_rq_pos(rq1);
489         s2 = blk_rq_pos(rq2);
490
491         /*
492          * by definition, 1KiB is 2 sectors
493          */
494         back_max = cfqd->cfq_back_max * 2;
495
496         /*
497          * Strict one way elevator _except_ in the case where we allow
498          * short backward seeks which are biased as twice the cost of a
499          * similar forward seek.
500          */
501         if (s1 >= last)
502                 d1 = s1 - last;
503         else if (s1 + back_max >= last)
504                 d1 = (last - s1) * cfqd->cfq_back_penalty;
505         else
506                 wrap |= CFQ_RQ1_WRAP;
507
508         if (s2 >= last)
509                 d2 = s2 - last;
510         else if (s2 + back_max >= last)
511                 d2 = (last - s2) * cfqd->cfq_back_penalty;
512         else
513                 wrap |= CFQ_RQ2_WRAP;
514
515         /* Found required data */
516
517         /*
518          * By doing switch() on the bit mask "wrap" we avoid having to
519          * check two variables for all permutations: --> faster!
520          */
521         switch (wrap) {
522         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
523                 if (d1 < d2)
524                         return rq1;
525                 else if (d2 < d1)
526                         return rq2;
527                 else {
528                         if (s1 >= s2)
529                                 return rq1;
530                         else
531                                 return rq2;
532                 }
533
534         case CFQ_RQ2_WRAP:
535                 return rq1;
536         case CFQ_RQ1_WRAP:
537                 return rq2;
538         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
539         default:
540                 /*
541                  * Since both rqs are wrapped,
542                  * start with the one that's further behind head
543                  * (--> only *one* back seek required),
544                  * since back seek takes more time than forward.
545                  */
546                 if (s1 <= s2)
547                         return rq1;
548                 else
549                         return rq2;
550         }
551 }
552
553 /*
554  * The below is leftmost cache rbtree addon
555  */
556 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
557 {
558         if (!root->left)
559                 root->left = rb_first(&root->rb);
560
561         if (root->left)
562                 return rb_entry(root->left, struct cfq_queue, rb_node);
563
564         return NULL;
565 }
566
567 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
568 {
569         rb_erase(n, root);
570         RB_CLEAR_NODE(n);
571 }
572
573 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
574 {
575         if (root->left == n)
576                 root->left = NULL;
577         rb_erase_init(n, &root->rb);
578         --root->count;
579 }
580
581 /*
582  * would be nice to take fifo expire time into account as well
583  */
584 static struct request *
585 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
586                   struct request *last)
587 {
588         struct rb_node *rbnext = rb_next(&last->rb_node);
589         struct rb_node *rbprev = rb_prev(&last->rb_node);
590         struct request *next = NULL, *prev = NULL;
591
592         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
593
594         if (rbprev)
595                 prev = rb_entry_rq(rbprev);
596
597         if (rbnext)
598                 next = rb_entry_rq(rbnext);
599         else {
600                 rbnext = rb_first(&cfqq->sort_list);
601                 if (rbnext && rbnext != &last->rb_node)
602                         next = rb_entry_rq(rbnext);
603         }
604
605         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
606 }
607
608 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
609                                       struct cfq_queue *cfqq)
610 {
611         struct cfq_rb_root *service_tree;
612
613         service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
614
615         /*
616          * just an approximation, should be ok.
617          */
618         return  service_tree->count * (cfq_prio_slice(cfqd, 1, 0) -
619                    cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
620 }
621
622 /*
623  * The cfqd->service_trees holds all pending cfq_queue's that have
624  * requests waiting to be processed. It is sorted in the order that
625  * we will service the queues.
626  */
627 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
628                                  bool add_front)
629 {
630         struct rb_node **p, *parent;
631         struct cfq_queue *__cfqq;
632         unsigned long rb_key;
633         struct cfq_rb_root *service_tree;
634         int left;
635
636         service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
637         if (cfq_class_idle(cfqq)) {
638                 rb_key = CFQ_IDLE_DELAY;
639                 parent = rb_last(&service_tree->rb);
640                 if (parent && parent != &cfqq->rb_node) {
641                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
642                         rb_key += __cfqq->rb_key;
643                 } else
644                         rb_key += jiffies;
645         } else if (!add_front) {
646                 /*
647                  * Get our rb key offset. Subtract any residual slice
648                  * value carried from last service. A negative resid
649                  * count indicates slice overrun, and this should position
650                  * the next service time further away in the tree.
651                  */
652                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
653                 rb_key -= cfqq->slice_resid;
654                 cfqq->slice_resid = 0;
655         } else {
656                 rb_key = -HZ;
657                 __cfqq = cfq_rb_first(service_tree);
658                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
659         }
660
661         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
662                 /*
663                  * same position, nothing more to do
664                  */
665                 if (rb_key == cfqq->rb_key &&
666                     cfqq->service_tree == service_tree)
667                         return;
668
669                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
670                 cfqq->service_tree = NULL;
671         }
672
673         left = 1;
674         parent = NULL;
675         cfqq->service_tree = service_tree;
676         p = &service_tree->rb.rb_node;
677         while (*p) {
678                 struct rb_node **n;
679
680                 parent = *p;
681                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
682
683                 /*
684                  * sort by key, that represents service time.
685                  */
686                 if (time_before(rb_key, __cfqq->rb_key))
687                         n = &(*p)->rb_left;
688                 else {
689                         n = &(*p)->rb_right;
690                         left = 0;
691                 }
692
693                 p = n;
694         }
695
696         if (left)
697                 service_tree->left = &cfqq->rb_node;
698
699         cfqq->rb_key = rb_key;
700         rb_link_node(&cfqq->rb_node, parent, p);
701         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
702         service_tree->count++;
703 }
704
705 static struct cfq_queue *
706 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
707                      sector_t sector, struct rb_node **ret_parent,
708                      struct rb_node ***rb_link)
709 {
710         struct rb_node **p, *parent;
711         struct cfq_queue *cfqq = NULL;
712
713         parent = NULL;
714         p = &root->rb_node;
715         while (*p) {
716                 struct rb_node **n;
717
718                 parent = *p;
719                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
720
721                 /*
722                  * Sort strictly based on sector.  Smallest to the left,
723                  * largest to the right.
724                  */
725                 if (sector > blk_rq_pos(cfqq->next_rq))
726                         n = &(*p)->rb_right;
727                 else if (sector < blk_rq_pos(cfqq->next_rq))
728                         n = &(*p)->rb_left;
729                 else
730                         break;
731                 p = n;
732                 cfqq = NULL;
733         }
734
735         *ret_parent = parent;
736         if (rb_link)
737                 *rb_link = p;
738         return cfqq;
739 }
740
741 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
742 {
743         struct rb_node **p, *parent;
744         struct cfq_queue *__cfqq;
745
746         if (cfqq->p_root) {
747                 rb_erase(&cfqq->p_node, cfqq->p_root);
748                 cfqq->p_root = NULL;
749         }
750
751         if (cfq_class_idle(cfqq))
752                 return;
753         if (!cfqq->next_rq)
754                 return;
755
756         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
757         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
758                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
759         if (!__cfqq) {
760                 rb_link_node(&cfqq->p_node, parent, p);
761                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
762         } else
763                 cfqq->p_root = NULL;
764 }
765
766 /*
767  * Update cfqq's position in the service tree.
768  */
769 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
770 {
771         /*
772          * Resorting requires the cfqq to be on the RR list already.
773          */
774         if (cfq_cfqq_on_rr(cfqq)) {
775                 cfq_service_tree_add(cfqd, cfqq, 0);
776                 cfq_prio_tree_add(cfqd, cfqq);
777         }
778 }
779
780 /*
781  * add to busy list of queues for service, trying to be fair in ordering
782  * the pending list according to last request service
783  */
784 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
785 {
786         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
787         BUG_ON(cfq_cfqq_on_rr(cfqq));
788         cfq_mark_cfqq_on_rr(cfqq);
789         cfqd->busy_queues++;
790
791         cfq_resort_rr_list(cfqd, cfqq);
792 }
793
794 /*
795  * Called when the cfqq no longer has requests pending, remove it from
796  * the service tree.
797  */
798 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
799 {
800         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
801         BUG_ON(!cfq_cfqq_on_rr(cfqq));
802         cfq_clear_cfqq_on_rr(cfqq);
803
804         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
805                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
806                 cfqq->service_tree = NULL;
807         }
808         if (cfqq->p_root) {
809                 rb_erase(&cfqq->p_node, cfqq->p_root);
810                 cfqq->p_root = NULL;
811         }
812
813         BUG_ON(!cfqd->busy_queues);
814         cfqd->busy_queues--;
815 }
816
817 /*
818  * rb tree support functions
819  */
820 static void cfq_del_rq_rb(struct request *rq)
821 {
822         struct cfq_queue *cfqq = RQ_CFQQ(rq);
823         struct cfq_data *cfqd = cfqq->cfqd;
824         const int sync = rq_is_sync(rq);
825
826         BUG_ON(!cfqq->queued[sync]);
827         cfqq->queued[sync]--;
828
829         elv_rb_del(&cfqq->sort_list, rq);
830
831         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
832                 cfq_del_cfqq_rr(cfqd, cfqq);
833 }
834
835 static void cfq_add_rq_rb(struct request *rq)
836 {
837         struct cfq_queue *cfqq = RQ_CFQQ(rq);
838         struct cfq_data *cfqd = cfqq->cfqd;
839         struct request *__alias, *prev;
840
841         cfqq->queued[rq_is_sync(rq)]++;
842
843         /*
844          * looks a little odd, but the first insert might return an alias.
845          * if that happens, put the alias on the dispatch list
846          */
847         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
848                 cfq_dispatch_insert(cfqd->queue, __alias);
849
850         if (!cfq_cfqq_on_rr(cfqq))
851                 cfq_add_cfqq_rr(cfqd, cfqq);
852
853         /*
854          * check if this request is a better next-serve candidate
855          */
856         prev = cfqq->next_rq;
857         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
858
859         /*
860          * adjust priority tree position, if ->next_rq changes
861          */
862         if (prev != cfqq->next_rq)
863                 cfq_prio_tree_add(cfqd, cfqq);
864
865         BUG_ON(!cfqq->next_rq);
866 }
867
868 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
869 {
870         elv_rb_del(&cfqq->sort_list, rq);
871         cfqq->queued[rq_is_sync(rq)]--;
872         cfq_add_rq_rb(rq);
873 }
874
875 static struct request *
876 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
877 {
878         struct task_struct *tsk = current;
879         struct cfq_io_context *cic;
880         struct cfq_queue *cfqq;
881
882         cic = cfq_cic_lookup(cfqd, tsk->io_context);
883         if (!cic)
884                 return NULL;
885
886         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
887         if (cfqq) {
888                 sector_t sector = bio->bi_sector + bio_sectors(bio);
889
890                 return elv_rb_find(&cfqq->sort_list, sector);
891         }
892
893         return NULL;
894 }
895
896 static void cfq_activate_request(struct request_queue *q, struct request *rq)
897 {
898         struct cfq_data *cfqd = q->elevator->elevator_data;
899
900         cfqd->rq_in_driver[rq_is_sync(rq)]++;
901         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
902                                                 rq_in_driver(cfqd));
903
904         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
905 }
906
907 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
908 {
909         struct cfq_data *cfqd = q->elevator->elevator_data;
910         const int sync = rq_is_sync(rq);
911
912         WARN_ON(!cfqd->rq_in_driver[sync]);
913         cfqd->rq_in_driver[sync]--;
914         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
915                                                 rq_in_driver(cfqd));
916 }
917
918 static void cfq_remove_request(struct request *rq)
919 {
920         struct cfq_queue *cfqq = RQ_CFQQ(rq);
921
922         if (cfqq->next_rq == rq)
923                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
924
925         list_del_init(&rq->queuelist);
926         cfq_del_rq_rb(rq);
927
928         cfqq->cfqd->rq_queued--;
929         if (rq_is_meta(rq)) {
930                 WARN_ON(!cfqq->meta_pending);
931                 cfqq->meta_pending--;
932         }
933 }
934
935 static int cfq_merge(struct request_queue *q, struct request **req,
936                      struct bio *bio)
937 {
938         struct cfq_data *cfqd = q->elevator->elevator_data;
939         struct request *__rq;
940
941         __rq = cfq_find_rq_fmerge(cfqd, bio);
942         if (__rq && elv_rq_merge_ok(__rq, bio)) {
943                 *req = __rq;
944                 return ELEVATOR_FRONT_MERGE;
945         }
946
947         return ELEVATOR_NO_MERGE;
948 }
949
950 static void cfq_merged_request(struct request_queue *q, struct request *req,
951                                int type)
952 {
953         if (type == ELEVATOR_FRONT_MERGE) {
954                 struct cfq_queue *cfqq = RQ_CFQQ(req);
955
956                 cfq_reposition_rq_rb(cfqq, req);
957         }
958 }
959
960 static void
961 cfq_merged_requests(struct request_queue *q, struct request *rq,
962                     struct request *next)
963 {
964         struct cfq_queue *cfqq = RQ_CFQQ(rq);
965         /*
966          * reposition in fifo if next is older than rq
967          */
968         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
969             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
970                 list_move(&rq->queuelist, &next->queuelist);
971                 rq_set_fifo_time(rq, rq_fifo_time(next));
972         }
973
974         if (cfqq->next_rq == next)
975                 cfqq->next_rq = rq;
976         cfq_remove_request(next);
977 }
978
979 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
980                            struct bio *bio)
981 {
982         struct cfq_data *cfqd = q->elevator->elevator_data;
983         struct cfq_io_context *cic;
984         struct cfq_queue *cfqq;
985
986         /*
987          * Disallow merge of a sync bio into an async request.
988          */
989         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
990                 return false;
991
992         /*
993          * Lookup the cfqq that this bio will be queued with. Allow
994          * merge only if rq is queued there.
995          */
996         cic = cfq_cic_lookup(cfqd, current->io_context);
997         if (!cic)
998                 return false;
999
1000         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1001         return cfqq == RQ_CFQQ(rq);
1002 }
1003
1004 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1005                                    struct cfq_queue *cfqq)
1006 {
1007         if (cfqq) {
1008                 cfq_log_cfqq(cfqd, cfqq, "set_active");
1009                 cfqq->slice_end = 0;
1010                 cfqq->slice_dispatch = 0;
1011
1012                 cfq_clear_cfqq_wait_request(cfqq);
1013                 cfq_clear_cfqq_must_dispatch(cfqq);
1014                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1015                 cfq_clear_cfqq_fifo_expire(cfqq);
1016                 cfq_mark_cfqq_slice_new(cfqq);
1017
1018                 del_timer(&cfqd->idle_slice_timer);
1019         }
1020
1021         cfqd->active_queue = cfqq;
1022 }
1023
1024 /*
1025  * current cfqq expired its slice (or was too idle), select new one
1026  */
1027 static void
1028 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1029                     bool timed_out)
1030 {
1031         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1032
1033         if (cfq_cfqq_wait_request(cfqq))
1034                 del_timer(&cfqd->idle_slice_timer);
1035
1036         cfq_clear_cfqq_wait_request(cfqq);
1037
1038         /*
1039          * store what was left of this slice, if the queue idled/timed out
1040          */
1041         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1042                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1043                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1044         }
1045
1046         cfq_resort_rr_list(cfqd, cfqq);
1047
1048         if (cfqq == cfqd->active_queue)
1049                 cfqd->active_queue = NULL;
1050
1051         if (cfqd->active_cic) {
1052                 put_io_context(cfqd->active_cic->ioc);
1053                 cfqd->active_cic = NULL;
1054         }
1055 }
1056
1057 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1058 {
1059         struct cfq_queue *cfqq = cfqd->active_queue;
1060
1061         if (cfqq)
1062                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1063 }
1064
1065 /*
1066  * Get next queue for service. Unless we have a queue preemption,
1067  * we'll simply select the first cfqq in the service tree.
1068  */
1069 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1070 {
1071         struct cfq_rb_root *service_tree =
1072                 service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd);
1073
1074         if (RB_EMPTY_ROOT(&service_tree->rb))
1075                 return NULL;
1076         return cfq_rb_first(service_tree);
1077 }
1078
1079 /*
1080  * Get and set a new active queue for service.
1081  */
1082 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1083                                               struct cfq_queue *cfqq)
1084 {
1085         if (!cfqq)
1086                 cfqq = cfq_get_next_queue(cfqd);
1087
1088         __cfq_set_active_queue(cfqd, cfqq);
1089         return cfqq;
1090 }
1091
1092 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1093                                           struct request *rq)
1094 {
1095         if (blk_rq_pos(rq) >= cfqd->last_position)
1096                 return blk_rq_pos(rq) - cfqd->last_position;
1097         else
1098                 return cfqd->last_position - blk_rq_pos(rq);
1099 }
1100
1101 #define CFQQ_SEEK_THR           8 * 1024
1102 #define CFQQ_SEEKY(cfqq)        ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1103
1104 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1105                                struct request *rq)
1106 {
1107         sector_t sdist = cfqq->seek_mean;
1108
1109         if (!sample_valid(cfqq->seek_samples))
1110                 sdist = CFQQ_SEEK_THR;
1111
1112         return cfq_dist_from_last(cfqd, rq) <= sdist;
1113 }
1114
1115 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1116                                     struct cfq_queue *cur_cfqq)
1117 {
1118         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1119         struct rb_node *parent, *node;
1120         struct cfq_queue *__cfqq;
1121         sector_t sector = cfqd->last_position;
1122
1123         if (RB_EMPTY_ROOT(root))
1124                 return NULL;
1125
1126         /*
1127          * First, if we find a request starting at the end of the last
1128          * request, choose it.
1129          */
1130         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1131         if (__cfqq)
1132                 return __cfqq;
1133
1134         /*
1135          * If the exact sector wasn't found, the parent of the NULL leaf
1136          * will contain the closest sector.
1137          */
1138         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1139         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1140                 return __cfqq;
1141
1142         if (blk_rq_pos(__cfqq->next_rq) < sector)
1143                 node = rb_next(&__cfqq->p_node);
1144         else
1145                 node = rb_prev(&__cfqq->p_node);
1146         if (!node)
1147                 return NULL;
1148
1149         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1150         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1151                 return __cfqq;
1152
1153         return NULL;
1154 }
1155
1156 /*
1157  * cfqd - obvious
1158  * cur_cfqq - passed in so that we don't decide that the current queue is
1159  *            closely cooperating with itself.
1160  *
1161  * So, basically we're assuming that that cur_cfqq has dispatched at least
1162  * one request, and that cfqd->last_position reflects a position on the disk
1163  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1164  * assumption.
1165  */
1166 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1167                                               struct cfq_queue *cur_cfqq)
1168 {
1169         struct cfq_queue *cfqq;
1170
1171         if (!cfq_cfqq_sync(cur_cfqq))
1172                 return NULL;
1173         if (CFQQ_SEEKY(cur_cfqq))
1174                 return NULL;
1175
1176         /*
1177          * We should notice if some of the queues are cooperating, eg
1178          * working closely on the same area of the disk. In that case,
1179          * we can group them together and don't waste time idling.
1180          */
1181         cfqq = cfqq_close(cfqd, cur_cfqq);
1182         if (!cfqq)
1183                 return NULL;
1184
1185         /*
1186          * It only makes sense to merge sync queues.
1187          */
1188         if (!cfq_cfqq_sync(cfqq))
1189                 return NULL;
1190         if (CFQQ_SEEKY(cfqq))
1191                 return NULL;
1192
1193         /*
1194          * Do not merge queues of different priority classes
1195          */
1196         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1197                 return NULL;
1198
1199         return cfqq;
1200 }
1201
1202 /*
1203  * Determine whether we should enforce idle window for this queue.
1204  */
1205
1206 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1207 {
1208         enum wl_prio_t prio = cfqq_prio(cfqq);
1209         struct cfq_rb_root *service_tree = cfqq->service_tree;
1210
1211         /* We never do for idle class queues. */
1212         if (prio == IDLE_WORKLOAD)
1213                 return false;
1214
1215         /* We do for queues that were marked with idle window flag. */
1216         if (cfq_cfqq_idle_window(cfqq))
1217                 return true;
1218
1219         /*
1220          * Otherwise, we do only if they are the last ones
1221          * in their service tree.
1222          */
1223         if (!service_tree)
1224                 service_tree = service_tree_for(prio, cfqq_type(cfqq), cfqd);
1225
1226         if (service_tree->count == 0)
1227                 return true;
1228
1229         return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq);
1230 }
1231
1232 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1233 {
1234         struct cfq_queue *cfqq = cfqd->active_queue;
1235         struct cfq_io_context *cic;
1236         unsigned long sl;
1237
1238         /*
1239          * SSD device without seek penalty, disable idling. But only do so
1240          * for devices that support queuing, otherwise we still have a problem
1241          * with sync vs async workloads.
1242          */
1243         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1244                 return;
1245
1246         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1247         WARN_ON(cfq_cfqq_slice_new(cfqq));
1248
1249         /*
1250          * idle is disabled, either manually or by past process history
1251          */
1252         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1253                 return;
1254
1255         /*
1256          * still requests with the driver, don't idle
1257          */
1258         if (rq_in_driver(cfqd))
1259                 return;
1260
1261         /*
1262          * task has exited, don't wait
1263          */
1264         cic = cfqd->active_cic;
1265         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1266                 return;
1267
1268         /*
1269          * If our average think time is larger than the remaining time
1270          * slice, then don't idle. This avoids overrunning the allotted
1271          * time slice.
1272          */
1273         if (sample_valid(cic->ttime_samples) &&
1274             (cfqq->slice_end - jiffies < cic->ttime_mean))
1275                 return;
1276
1277         cfq_mark_cfqq_wait_request(cfqq);
1278
1279         sl = cfqd->cfq_slice_idle;
1280
1281         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1282         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1283 }
1284
1285 /*
1286  * Move request from internal lists to the request queue dispatch list.
1287  */
1288 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1289 {
1290         struct cfq_data *cfqd = q->elevator->elevator_data;
1291         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1292
1293         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1294
1295         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1296         cfq_remove_request(rq);
1297         cfqq->dispatched++;
1298         elv_dispatch_sort(q, rq);
1299
1300         if (cfq_cfqq_sync(cfqq))
1301                 cfqd->sync_flight++;
1302 }
1303
1304 /*
1305  * return expired entry, or NULL to just start from scratch in rbtree
1306  */
1307 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1308 {
1309         struct request *rq = NULL;
1310
1311         if (cfq_cfqq_fifo_expire(cfqq))
1312                 return NULL;
1313
1314         cfq_mark_cfqq_fifo_expire(cfqq);
1315
1316         if (list_empty(&cfqq->fifo))
1317                 return NULL;
1318
1319         rq = rq_entry_fifo(cfqq->fifo.next);
1320         if (time_before(jiffies, rq_fifo_time(rq)))
1321                 rq = NULL;
1322
1323         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1324         return rq;
1325 }
1326
1327 static inline int
1328 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1329 {
1330         const int base_rq = cfqd->cfq_slice_async_rq;
1331
1332         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1333
1334         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1335 }
1336
1337 /*
1338  * Must be called with the queue_lock held.
1339  */
1340 static int cfqq_process_refs(struct cfq_queue *cfqq)
1341 {
1342         int process_refs, io_refs;
1343
1344         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1345         process_refs = atomic_read(&cfqq->ref) - io_refs;
1346         BUG_ON(process_refs < 0);
1347         return process_refs;
1348 }
1349
1350 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1351 {
1352         int process_refs, new_process_refs;
1353         struct cfq_queue *__cfqq;
1354
1355         /* Avoid a circular list and skip interim queue merges */
1356         while ((__cfqq = new_cfqq->new_cfqq)) {
1357                 if (__cfqq == cfqq)
1358                         return;
1359                 new_cfqq = __cfqq;
1360         }
1361
1362         process_refs = cfqq_process_refs(cfqq);
1363         /*
1364          * If the process for the cfqq has gone away, there is no
1365          * sense in merging the queues.
1366          */
1367         if (process_refs == 0)
1368                 return;
1369
1370         /*
1371          * Merge in the direction of the lesser amount of work.
1372          */
1373         new_process_refs = cfqq_process_refs(new_cfqq);
1374         if (new_process_refs >= process_refs) {
1375                 cfqq->new_cfqq = new_cfqq;
1376                 atomic_add(process_refs, &new_cfqq->ref);
1377         } else {
1378                 new_cfqq->new_cfqq = cfqq;
1379                 atomic_add(new_process_refs, &cfqq->ref);
1380         }
1381 }
1382
1383 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, enum wl_prio_t prio,
1384                                     bool prio_changed)
1385 {
1386         struct cfq_queue *queue;
1387         int i;
1388         bool key_valid = false;
1389         unsigned long lowest_key = 0;
1390         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1391
1392         if (prio_changed) {
1393                 /*
1394                  * When priorities switched, we prefer starting
1395                  * from SYNC_NOIDLE (first choice), or just SYNC
1396                  * over ASYNC
1397                  */
1398                 if (service_tree_for(prio, cur_best, cfqd)->count)
1399                         return cur_best;
1400                 cur_best = SYNC_WORKLOAD;
1401                 if (service_tree_for(prio, cur_best, cfqd)->count)
1402                         return cur_best;
1403
1404                 return ASYNC_WORKLOAD;
1405         }
1406
1407         for (i = 0; i < 3; ++i) {
1408                 /* otherwise, select the one with lowest rb_key */
1409                 queue = cfq_rb_first(service_tree_for(prio, i, cfqd));
1410                 if (queue &&
1411                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1412                         lowest_key = queue->rb_key;
1413                         cur_best = i;
1414                         key_valid = true;
1415                 }
1416         }
1417
1418         return cur_best;
1419 }
1420
1421 static void choose_service_tree(struct cfq_data *cfqd)
1422 {
1423         enum wl_prio_t previous_prio = cfqd->serving_prio;
1424         bool prio_changed;
1425         unsigned slice;
1426         unsigned count;
1427
1428         /* Choose next priority. RT > BE > IDLE */
1429         if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1430                 cfqd->serving_prio = RT_WORKLOAD;
1431         else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1432                 cfqd->serving_prio = BE_WORKLOAD;
1433         else {
1434                 cfqd->serving_prio = IDLE_WORKLOAD;
1435                 cfqd->workload_expires = jiffies + 1;
1436                 return;
1437         }
1438
1439         /*
1440          * For RT and BE, we have to choose also the type
1441          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1442          * expiration time
1443          */
1444         prio_changed = (cfqd->serving_prio != previous_prio);
1445         count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1446                 ->count;
1447
1448         /*
1449          * If priority didn't change, check workload expiration,
1450          * and that we still have other queues ready
1451          */
1452         if (!prio_changed && count &&
1453             !time_after(jiffies, cfqd->workload_expires))
1454                 return;
1455
1456         /* otherwise select new workload type */
1457         cfqd->serving_type =
1458                 cfq_choose_wl(cfqd, cfqd->serving_prio, prio_changed);
1459         count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1460                 ->count;
1461
1462         /*
1463          * the workload slice is computed as a fraction of target latency
1464          * proportional to the number of queues in that workload, over
1465          * all the queues in the same priority class
1466          */
1467         slice = cfq_target_latency * count /
1468                 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1469                       cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1470
1471         if (cfqd->serving_type == ASYNC_WORKLOAD)
1472                 /* async workload slice is scaled down according to
1473                  * the sync/async slice ratio. */
1474                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1475         else
1476                 /* sync workload slice is at least 2 * cfq_slice_idle */
1477                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1478
1479         slice = max_t(unsigned, slice, CFQ_MIN_TT);
1480         cfqd->workload_expires = jiffies + slice;
1481 }
1482
1483 /*
1484  * Select a queue for service. If we have a current active queue,
1485  * check whether to continue servicing it, or retrieve and set a new one.
1486  */
1487 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1488 {
1489         struct cfq_queue *cfqq, *new_cfqq = NULL;
1490
1491         cfqq = cfqd->active_queue;
1492         if (!cfqq)
1493                 goto new_queue;
1494
1495         /*
1496          * The active queue has run out of time, expire it and select new.
1497          */
1498         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1499                 goto expire;
1500
1501         /*
1502          * The active queue has requests and isn't expired, allow it to
1503          * dispatch.
1504          */
1505         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1506                 goto keep_queue;
1507
1508         /*
1509          * If another queue has a request waiting within our mean seek
1510          * distance, let it run.  The expire code will check for close
1511          * cooperators and put the close queue at the front of the service
1512          * tree.  If possible, merge the expiring queue with the new cfqq.
1513          */
1514         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1515         if (new_cfqq) {
1516                 if (!cfqq->new_cfqq)
1517                         cfq_setup_merge(cfqq, new_cfqq);
1518                 goto expire;
1519         }
1520
1521         /*
1522          * No requests pending. If the active queue still has requests in
1523          * flight or is idling for a new request, allow either of these
1524          * conditions to happen (or time out) before selecting a new queue.
1525          */
1526         if (timer_pending(&cfqd->idle_slice_timer) ||
1527             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1528                 cfqq = NULL;
1529                 goto keep_queue;
1530         }
1531
1532 expire:
1533         cfq_slice_expired(cfqd, 0);
1534 new_queue:
1535         /*
1536          * Current queue expired. Check if we have to switch to a new
1537          * service tree
1538          */
1539         if (!new_cfqq)
1540                 choose_service_tree(cfqd);
1541
1542         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1543 keep_queue:
1544         return cfqq;
1545 }
1546
1547 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1548 {
1549         int dispatched = 0;
1550
1551         while (cfqq->next_rq) {
1552                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1553                 dispatched++;
1554         }
1555
1556         BUG_ON(!list_empty(&cfqq->fifo));
1557         return dispatched;
1558 }
1559
1560 /*
1561  * Drain our current requests. Used for barriers and when switching
1562  * io schedulers on-the-fly.
1563  */
1564 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1565 {
1566         struct cfq_queue *cfqq;
1567         int dispatched = 0;
1568         int i, j;
1569         for (i = 0; i < 2; ++i)
1570                 for (j = 0; j < 3; ++j)
1571                         while ((cfqq = cfq_rb_first(&cfqd->service_trees[i][j]))
1572                                 != NULL)
1573                                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1574
1575         while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
1576                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1577
1578         cfq_slice_expired(cfqd, 0);
1579
1580         BUG_ON(cfqd->busy_queues);
1581
1582         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1583         return dispatched;
1584 }
1585
1586 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1587 {
1588         unsigned int max_dispatch;
1589
1590         /*
1591          * Drain async requests before we start sync IO
1592          */
1593         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1594                 return false;
1595
1596         /*
1597          * If this is an async queue and we have sync IO in flight, let it wait
1598          */
1599         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1600                 return false;
1601
1602         max_dispatch = cfqd->cfq_quantum;
1603         if (cfq_class_idle(cfqq))
1604                 max_dispatch = 1;
1605
1606         /*
1607          * Does this cfqq already have too much IO in flight?
1608          */
1609         if (cfqq->dispatched >= max_dispatch) {
1610                 /*
1611                  * idle queue must always only have a single IO in flight
1612                  */
1613                 if (cfq_class_idle(cfqq))
1614                         return false;
1615
1616                 /*
1617                  * We have other queues, don't allow more IO from this one
1618                  */
1619                 if (cfqd->busy_queues > 1)
1620                         return false;
1621
1622                 /*
1623                  * Sole queue user, allow bigger slice
1624                  */
1625                 max_dispatch *= 4;
1626         }
1627
1628         /*
1629          * Async queues must wait a bit before being allowed dispatch.
1630          * We also ramp up the dispatch depth gradually for async IO,
1631          * based on the last sync IO we serviced
1632          */
1633         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1634                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1635                 unsigned int depth;
1636
1637                 depth = last_sync / cfqd->cfq_slice[1];
1638                 if (!depth && !cfqq->dispatched)
1639                         depth = 1;
1640                 if (depth < max_dispatch)
1641                         max_dispatch = depth;
1642         }
1643
1644         /*
1645          * If we're below the current max, allow a dispatch
1646          */
1647         return cfqq->dispatched < max_dispatch;
1648 }
1649
1650 /*
1651  * Dispatch a request from cfqq, moving them to the request queue
1652  * dispatch list.
1653  */
1654 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1655 {
1656         struct request *rq;
1657
1658         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1659
1660         if (!cfq_may_dispatch(cfqd, cfqq))
1661                 return false;
1662
1663         /*
1664          * follow expired path, else get first next available
1665          */
1666         rq = cfq_check_fifo(cfqq);
1667         if (!rq)
1668                 rq = cfqq->next_rq;
1669
1670         /*
1671          * insert request into driver dispatch list
1672          */
1673         cfq_dispatch_insert(cfqd->queue, rq);
1674
1675         if (!cfqd->active_cic) {
1676                 struct cfq_io_context *cic = RQ_CIC(rq);
1677
1678                 atomic_long_inc(&cic->ioc->refcount);
1679                 cfqd->active_cic = cic;
1680         }
1681
1682         return true;
1683 }
1684
1685 /*
1686  * Find the cfqq that we need to service and move a request from that to the
1687  * dispatch list
1688  */
1689 static int cfq_dispatch_requests(struct request_queue *q, int force)
1690 {
1691         struct cfq_data *cfqd = q->elevator->elevator_data;
1692         struct cfq_queue *cfqq;
1693
1694         if (!cfqd->busy_queues)
1695                 return 0;
1696
1697         if (unlikely(force))
1698                 return cfq_forced_dispatch(cfqd);
1699
1700         cfqq = cfq_select_queue(cfqd);
1701         if (!cfqq)
1702                 return 0;
1703
1704         /*
1705          * Dispatch a request from this cfqq, if it is allowed
1706          */
1707         if (!cfq_dispatch_request(cfqd, cfqq))
1708                 return 0;
1709
1710         cfqq->slice_dispatch++;
1711         cfq_clear_cfqq_must_dispatch(cfqq);
1712
1713         /*
1714          * expire an async queue immediately if it has used up its slice. idle
1715          * queue always expire after 1 dispatch round.
1716          */
1717         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1718             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1719             cfq_class_idle(cfqq))) {
1720                 cfqq->slice_end = jiffies + 1;
1721                 cfq_slice_expired(cfqd, 0);
1722         }
1723
1724         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1725         return 1;
1726 }
1727
1728 /*
1729  * task holds one reference to the queue, dropped when task exits. each rq
1730  * in-flight on this queue also holds a reference, dropped when rq is freed.
1731  *
1732  * queue lock must be held here.
1733  */
1734 static void cfq_put_queue(struct cfq_queue *cfqq)
1735 {
1736         struct cfq_data *cfqd = cfqq->cfqd;
1737
1738         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1739
1740         if (!atomic_dec_and_test(&cfqq->ref))
1741                 return;
1742
1743         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1744         BUG_ON(rb_first(&cfqq->sort_list));
1745         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1746         BUG_ON(cfq_cfqq_on_rr(cfqq));
1747
1748         if (unlikely(cfqd->active_queue == cfqq)) {
1749                 __cfq_slice_expired(cfqd, cfqq, 0);
1750                 cfq_schedule_dispatch(cfqd);
1751         }
1752
1753         kmem_cache_free(cfq_pool, cfqq);
1754 }
1755
1756 /*
1757  * Must always be called with the rcu_read_lock() held
1758  */
1759 static void
1760 __call_for_each_cic(struct io_context *ioc,
1761                     void (*func)(struct io_context *, struct cfq_io_context *))
1762 {
1763         struct cfq_io_context *cic;
1764         struct hlist_node *n;
1765
1766         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1767                 func(ioc, cic);
1768 }
1769
1770 /*
1771  * Call func for each cic attached to this ioc.
1772  */
1773 static void
1774 call_for_each_cic(struct io_context *ioc,
1775                   void (*func)(struct io_context *, struct cfq_io_context *))
1776 {
1777         rcu_read_lock();
1778         __call_for_each_cic(ioc, func);
1779         rcu_read_unlock();
1780 }
1781
1782 static void cfq_cic_free_rcu(struct rcu_head *head)
1783 {
1784         struct cfq_io_context *cic;
1785
1786         cic = container_of(head, struct cfq_io_context, rcu_head);
1787
1788         kmem_cache_free(cfq_ioc_pool, cic);
1789         elv_ioc_count_dec(cfq_ioc_count);
1790
1791         if (ioc_gone) {
1792                 /*
1793                  * CFQ scheduler is exiting, grab exit lock and check
1794                  * the pending io context count. If it hits zero,
1795                  * complete ioc_gone and set it back to NULL
1796                  */
1797                 spin_lock(&ioc_gone_lock);
1798                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1799                         complete(ioc_gone);
1800                         ioc_gone = NULL;
1801                 }
1802                 spin_unlock(&ioc_gone_lock);
1803         }
1804 }
1805
1806 static void cfq_cic_free(struct cfq_io_context *cic)
1807 {
1808         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1809 }
1810
1811 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1812 {
1813         unsigned long flags;
1814
1815         BUG_ON(!cic->dead_key);
1816
1817         spin_lock_irqsave(&ioc->lock, flags);
1818         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1819         hlist_del_rcu(&cic->cic_list);
1820         spin_unlock_irqrestore(&ioc->lock, flags);
1821
1822         cfq_cic_free(cic);
1823 }
1824
1825 /*
1826  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1827  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1828  * and ->trim() which is called with the task lock held
1829  */
1830 static void cfq_free_io_context(struct io_context *ioc)
1831 {
1832         /*
1833          * ioc->refcount is zero here, or we are called from elv_unregister(),
1834          * so no more cic's are allowed to be linked into this ioc.  So it
1835          * should be ok to iterate over the known list, we will see all cic's
1836          * since no new ones are added.
1837          */
1838         __call_for_each_cic(ioc, cic_free_func);
1839 }
1840
1841 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1842 {
1843         struct cfq_queue *__cfqq, *next;
1844
1845         if (unlikely(cfqq == cfqd->active_queue)) {
1846                 __cfq_slice_expired(cfqd, cfqq, 0);
1847                 cfq_schedule_dispatch(cfqd);
1848         }
1849
1850         /*
1851          * If this queue was scheduled to merge with another queue, be
1852          * sure to drop the reference taken on that queue (and others in
1853          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
1854          */
1855         __cfqq = cfqq->new_cfqq;
1856         while (__cfqq) {
1857                 if (__cfqq == cfqq) {
1858                         WARN(1, "cfqq->new_cfqq loop detected\n");
1859                         break;
1860                 }
1861                 next = __cfqq->new_cfqq;
1862                 cfq_put_queue(__cfqq);
1863                 __cfqq = next;
1864         }
1865
1866         cfq_put_queue(cfqq);
1867 }
1868
1869 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1870                                          struct cfq_io_context *cic)
1871 {
1872         struct io_context *ioc = cic->ioc;
1873
1874         list_del_init(&cic->queue_list);
1875
1876         /*
1877          * Make sure key == NULL is seen for dead queues
1878          */
1879         smp_wmb();
1880         cic->dead_key = (unsigned long) cic->key;
1881         cic->key = NULL;
1882
1883         if (ioc->ioc_data == cic)
1884                 rcu_assign_pointer(ioc->ioc_data, NULL);
1885
1886         if (cic->cfqq[BLK_RW_ASYNC]) {
1887                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1888                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1889         }
1890
1891         if (cic->cfqq[BLK_RW_SYNC]) {
1892                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1893                 cic->cfqq[BLK_RW_SYNC] = NULL;
1894         }
1895 }
1896
1897 static void cfq_exit_single_io_context(struct io_context *ioc,
1898                                        struct cfq_io_context *cic)
1899 {
1900         struct cfq_data *cfqd = cic->key;
1901
1902         if (cfqd) {
1903                 struct request_queue *q = cfqd->queue;
1904                 unsigned long flags;
1905
1906                 spin_lock_irqsave(q->queue_lock, flags);
1907
1908                 /*
1909                  * Ensure we get a fresh copy of the ->key to prevent
1910                  * race between exiting task and queue
1911                  */
1912                 smp_read_barrier_depends();
1913                 if (cic->key)
1914                         __cfq_exit_single_io_context(cfqd, cic);
1915
1916                 spin_unlock_irqrestore(q->queue_lock, flags);
1917         }
1918 }
1919
1920 /*
1921  * The process that ioc belongs to has exited, we need to clean up
1922  * and put the internal structures we have that belongs to that process.
1923  */
1924 static void cfq_exit_io_context(struct io_context *ioc)
1925 {
1926         call_for_each_cic(ioc, cfq_exit_single_io_context);
1927 }
1928
1929 static struct cfq_io_context *
1930 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1931 {
1932         struct cfq_io_context *cic;
1933
1934         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1935                                                         cfqd->queue->node);
1936         if (cic) {
1937                 cic->last_end_request = jiffies;
1938                 INIT_LIST_HEAD(&cic->queue_list);
1939                 INIT_HLIST_NODE(&cic->cic_list);
1940                 cic->dtor = cfq_free_io_context;
1941                 cic->exit = cfq_exit_io_context;
1942                 elv_ioc_count_inc(cfq_ioc_count);
1943         }
1944
1945         return cic;
1946 }
1947
1948 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1949 {
1950         struct task_struct *tsk = current;
1951         int ioprio_class;
1952
1953         if (!cfq_cfqq_prio_changed(cfqq))
1954                 return;
1955
1956         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1957         switch (ioprio_class) {
1958         default:
1959                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1960         case IOPRIO_CLASS_NONE:
1961                 /*
1962                  * no prio set, inherit CPU scheduling settings
1963                  */
1964                 cfqq->ioprio = task_nice_ioprio(tsk);
1965                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1966                 break;
1967         case IOPRIO_CLASS_RT:
1968                 cfqq->ioprio = task_ioprio(ioc);
1969                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1970                 break;
1971         case IOPRIO_CLASS_BE:
1972                 cfqq->ioprio = task_ioprio(ioc);
1973                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1974                 break;
1975         case IOPRIO_CLASS_IDLE:
1976                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1977                 cfqq->ioprio = 7;
1978                 cfq_clear_cfqq_idle_window(cfqq);
1979                 break;
1980         }
1981
1982         /*
1983          * keep track of original prio settings in case we have to temporarily
1984          * elevate the priority of this queue
1985          */
1986         cfqq->org_ioprio = cfqq->ioprio;
1987         cfqq->org_ioprio_class = cfqq->ioprio_class;
1988         cfq_clear_cfqq_prio_changed(cfqq);
1989 }
1990
1991 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1992 {
1993         struct cfq_data *cfqd = cic->key;
1994         struct cfq_queue *cfqq;
1995         unsigned long flags;
1996
1997         if (unlikely(!cfqd))
1998                 return;
1999
2000         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2001
2002         cfqq = cic->cfqq[BLK_RW_ASYNC];
2003         if (cfqq) {
2004                 struct cfq_queue *new_cfqq;
2005                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2006                                                 GFP_ATOMIC);
2007                 if (new_cfqq) {
2008                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2009                         cfq_put_queue(cfqq);
2010                 }
2011         }
2012
2013         cfqq = cic->cfqq[BLK_RW_SYNC];
2014         if (cfqq)
2015                 cfq_mark_cfqq_prio_changed(cfqq);
2016
2017         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2018 }
2019
2020 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2021 {
2022         call_for_each_cic(ioc, changed_ioprio);
2023         ioc->ioprio_changed = 0;
2024 }
2025
2026 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2027                           pid_t pid, bool is_sync)
2028 {
2029         RB_CLEAR_NODE(&cfqq->rb_node);
2030         RB_CLEAR_NODE(&cfqq->p_node);
2031         INIT_LIST_HEAD(&cfqq->fifo);
2032
2033         atomic_set(&cfqq->ref, 0);
2034         cfqq->cfqd = cfqd;
2035
2036         cfq_mark_cfqq_prio_changed(cfqq);
2037
2038         if (is_sync) {
2039                 if (!cfq_class_idle(cfqq))
2040                         cfq_mark_cfqq_idle_window(cfqq);
2041                 cfq_mark_cfqq_sync(cfqq);
2042         }
2043         cfqq->pid = pid;
2044 }
2045
2046 static struct cfq_queue *
2047 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2048                      struct io_context *ioc, gfp_t gfp_mask)
2049 {
2050         struct cfq_queue *cfqq, *new_cfqq = NULL;
2051         struct cfq_io_context *cic;
2052
2053 retry:
2054         cic = cfq_cic_lookup(cfqd, ioc);
2055         /* cic always exists here */
2056         cfqq = cic_to_cfqq(cic, is_sync);
2057
2058         /*
2059          * Always try a new alloc if we fell back to the OOM cfqq
2060          * originally, since it should just be a temporary situation.
2061          */
2062         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2063                 cfqq = NULL;
2064                 if (new_cfqq) {
2065                         cfqq = new_cfqq;
2066                         new_cfqq = NULL;
2067                 } else if (gfp_mask & __GFP_WAIT) {
2068                         spin_unlock_irq(cfqd->queue->queue_lock);
2069                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2070                                         gfp_mask | __GFP_ZERO,
2071                                         cfqd->queue->node);
2072                         spin_lock_irq(cfqd->queue->queue_lock);
2073                         if (new_cfqq)
2074                                 goto retry;
2075                 } else {
2076                         cfqq = kmem_cache_alloc_node(cfq_pool,
2077                                         gfp_mask | __GFP_ZERO,
2078                                         cfqd->queue->node);
2079                 }
2080
2081                 if (cfqq) {
2082                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2083                         cfq_init_prio_data(cfqq, ioc);
2084                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2085                 } else
2086                         cfqq = &cfqd->oom_cfqq;
2087         }
2088
2089         if (new_cfqq)
2090                 kmem_cache_free(cfq_pool, new_cfqq);
2091
2092         return cfqq;
2093 }
2094
2095 static struct cfq_queue **
2096 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2097 {
2098         switch (ioprio_class) {
2099         case IOPRIO_CLASS_RT:
2100                 return &cfqd->async_cfqq[0][ioprio];
2101         case IOPRIO_CLASS_BE:
2102                 return &cfqd->async_cfqq[1][ioprio];
2103         case IOPRIO_CLASS_IDLE:
2104                 return &cfqd->async_idle_cfqq;
2105         default:
2106                 BUG();
2107         }
2108 }
2109
2110 static struct cfq_queue *
2111 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2112               gfp_t gfp_mask)
2113 {
2114         const int ioprio = task_ioprio(ioc);
2115         const int ioprio_class = task_ioprio_class(ioc);
2116         struct cfq_queue **async_cfqq = NULL;
2117         struct cfq_queue *cfqq = NULL;
2118
2119         if (!is_sync) {
2120                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2121                 cfqq = *async_cfqq;
2122         }
2123
2124         if (!cfqq)
2125                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2126
2127         /*
2128          * pin the queue now that it's allocated, scheduler exit will prune it
2129          */
2130         if (!is_sync && !(*async_cfqq)) {
2131                 atomic_inc(&cfqq->ref);
2132                 *async_cfqq = cfqq;
2133         }
2134
2135         atomic_inc(&cfqq->ref);
2136         return cfqq;
2137 }
2138
2139 /*
2140  * We drop cfq io contexts lazily, so we may find a dead one.
2141  */
2142 static void
2143 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2144                   struct cfq_io_context *cic)
2145 {
2146         unsigned long flags;
2147
2148         WARN_ON(!list_empty(&cic->queue_list));
2149
2150         spin_lock_irqsave(&ioc->lock, flags);
2151
2152         BUG_ON(ioc->ioc_data == cic);
2153
2154         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2155         hlist_del_rcu(&cic->cic_list);
2156         spin_unlock_irqrestore(&ioc->lock, flags);
2157
2158         cfq_cic_free(cic);
2159 }
2160
2161 static struct cfq_io_context *
2162 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2163 {
2164         struct cfq_io_context *cic;
2165         unsigned long flags;
2166         void *k;
2167
2168         if (unlikely(!ioc))
2169                 return NULL;
2170
2171         rcu_read_lock();
2172
2173         /*
2174          * we maintain a last-hit cache, to avoid browsing over the tree
2175          */
2176         cic = rcu_dereference(ioc->ioc_data);
2177         if (cic && cic->key == cfqd) {
2178                 rcu_read_unlock();
2179                 return cic;
2180         }
2181
2182         do {
2183                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2184                 rcu_read_unlock();
2185                 if (!cic)
2186                         break;
2187                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2188                 k = cic->key;
2189                 if (unlikely(!k)) {
2190                         cfq_drop_dead_cic(cfqd, ioc, cic);
2191                         rcu_read_lock();
2192                         continue;
2193                 }
2194
2195                 spin_lock_irqsave(&ioc->lock, flags);
2196                 rcu_assign_pointer(ioc->ioc_data, cic);
2197                 spin_unlock_irqrestore(&ioc->lock, flags);
2198                 break;
2199         } while (1);
2200
2201         return cic;
2202 }
2203
2204 /*
2205  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2206  * the process specific cfq io context when entered from the block layer.
2207  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2208  */
2209 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2210                         struct cfq_io_context *cic, gfp_t gfp_mask)
2211 {
2212         unsigned long flags;
2213         int ret;
2214
2215         ret = radix_tree_preload(gfp_mask);
2216         if (!ret) {
2217                 cic->ioc = ioc;
2218                 cic->key = cfqd;
2219
2220                 spin_lock_irqsave(&ioc->lock, flags);
2221                 ret = radix_tree_insert(&ioc->radix_root,
2222                                                 (unsigned long) cfqd, cic);
2223                 if (!ret)
2224                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2225                 spin_unlock_irqrestore(&ioc->lock, flags);
2226
2227                 radix_tree_preload_end();
2228
2229                 if (!ret) {
2230                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2231                         list_add(&cic->queue_list, &cfqd->cic_list);
2232                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2233                 }
2234         }
2235
2236         if (ret)
2237                 printk(KERN_ERR "cfq: cic link failed!\n");
2238
2239         return ret;
2240 }
2241
2242 /*
2243  * Setup general io context and cfq io context. There can be several cfq
2244  * io contexts per general io context, if this process is doing io to more
2245  * than one device managed by cfq.
2246  */
2247 static struct cfq_io_context *
2248 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2249 {
2250         struct io_context *ioc = NULL;
2251         struct cfq_io_context *cic;
2252
2253         might_sleep_if(gfp_mask & __GFP_WAIT);
2254
2255         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2256         if (!ioc)
2257                 return NULL;
2258
2259         cic = cfq_cic_lookup(cfqd, ioc);
2260         if (cic)
2261                 goto out;
2262
2263         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2264         if (cic == NULL)
2265                 goto err;
2266
2267         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2268                 goto err_free;
2269
2270 out:
2271         smp_read_barrier_depends();
2272         if (unlikely(ioc->ioprio_changed))
2273                 cfq_ioc_set_ioprio(ioc);
2274
2275         return cic;
2276 err_free:
2277         cfq_cic_free(cic);
2278 err:
2279         put_io_context(ioc);
2280         return NULL;
2281 }
2282
2283 static void
2284 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2285 {
2286         unsigned long elapsed = jiffies - cic->last_end_request;
2287         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2288
2289         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2290         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2291         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2292 }
2293
2294 static void
2295 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2296                        struct request *rq)
2297 {
2298         sector_t sdist;
2299         u64 total;
2300
2301         if (!cfqq->last_request_pos)
2302                 sdist = 0;
2303         else if (cfqq->last_request_pos < blk_rq_pos(rq))
2304                 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2305         else
2306                 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2307
2308         /*
2309          * Don't allow the seek distance to get too large from the
2310          * odd fragment, pagein, etc
2311          */
2312         if (cfqq->seek_samples <= 60) /* second&third seek */
2313                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2314         else
2315                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2316
2317         cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2318         cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2319         total = cfqq->seek_total + (cfqq->seek_samples/2);
2320         do_div(total, cfqq->seek_samples);
2321         cfqq->seek_mean = (sector_t)total;
2322
2323         /*
2324          * If this cfqq is shared between multiple processes, check to
2325          * make sure that those processes are still issuing I/Os within
2326          * the mean seek distance.  If not, it may be time to break the
2327          * queues apart again.
2328          */
2329         if (cfq_cfqq_coop(cfqq)) {
2330                 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2331                         cfqq->seeky_start = jiffies;
2332                 else if (!CFQQ_SEEKY(cfqq))
2333                         cfqq->seeky_start = 0;
2334         }
2335 }
2336
2337 /*
2338  * Disable idle window if the process thinks too long or seeks so much that
2339  * it doesn't matter
2340  */
2341 static void
2342 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2343                        struct cfq_io_context *cic)
2344 {
2345         int old_idle, enable_idle;
2346
2347         /*
2348          * Don't idle for async or idle io prio class
2349          */
2350         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2351                 return;
2352
2353         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2354
2355         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2356                 cfq_mark_cfqq_deep(cfqq);
2357
2358         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2359             (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2360              && CFQQ_SEEKY(cfqq)))
2361                 enable_idle = 0;
2362         else if (sample_valid(cic->ttime_samples)) {
2363                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2364                         enable_idle = 0;
2365                 else
2366                         enable_idle = 1;
2367         }
2368
2369         if (old_idle != enable_idle) {
2370                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2371                 if (enable_idle)
2372                         cfq_mark_cfqq_idle_window(cfqq);
2373                 else
2374                         cfq_clear_cfqq_idle_window(cfqq);
2375         }
2376 }
2377
2378 /*
2379  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2380  * no or if we aren't sure, a 1 will cause a preempt.
2381  */
2382 static bool
2383 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2384                    struct request *rq)
2385 {
2386         struct cfq_queue *cfqq;
2387
2388         cfqq = cfqd->active_queue;
2389         if (!cfqq)
2390                 return false;
2391
2392         if (cfq_slice_used(cfqq))
2393                 return true;
2394
2395         if (cfq_class_idle(new_cfqq))
2396                 return false;
2397
2398         if (cfq_class_idle(cfqq))
2399                 return true;
2400
2401         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2402             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2403             new_cfqq->service_tree->count == 1)
2404                 return true;
2405
2406         /*
2407          * if the new request is sync, but the currently running queue is
2408          * not, let the sync request have priority.
2409          */
2410         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2411                 return true;
2412
2413         /*
2414          * So both queues are sync. Let the new request get disk time if
2415          * it's a metadata request and the current queue is doing regular IO.
2416          */
2417         if (rq_is_meta(rq) && !cfqq->meta_pending)
2418                 return true;
2419
2420         /*
2421          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2422          */
2423         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2424                 return true;
2425
2426         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2427                 return false;
2428
2429         /*
2430          * if this request is as-good as one we would expect from the
2431          * current cfqq, let it preempt
2432          */
2433         if (cfq_rq_close(cfqd, cfqq, rq))
2434                 return true;
2435
2436         return false;
2437 }
2438
2439 /*
2440  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2441  * let it have half of its nominal slice.
2442  */
2443 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2444 {
2445         cfq_log_cfqq(cfqd, cfqq, "preempt");
2446         cfq_slice_expired(cfqd, 1);
2447
2448         /*
2449          * Put the new queue at the front of the of the current list,
2450          * so we know that it will be selected next.
2451          */
2452         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2453
2454         cfq_service_tree_add(cfqd, cfqq, 1);
2455
2456         cfqq->slice_end = 0;
2457         cfq_mark_cfqq_slice_new(cfqq);
2458 }
2459
2460 /*
2461  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2462  * something we should do about it
2463  */
2464 static void
2465 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2466                 struct request *rq)
2467 {
2468         struct cfq_io_context *cic = RQ_CIC(rq);
2469
2470         cfqd->rq_queued++;
2471         if (rq_is_meta(rq))
2472                 cfqq->meta_pending++;
2473
2474         cfq_update_io_thinktime(cfqd, cic);
2475         cfq_update_io_seektime(cfqd, cfqq, rq);
2476         cfq_update_idle_window(cfqd, cfqq, cic);
2477
2478         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2479
2480         if (cfqq == cfqd->active_queue) {
2481                 /*
2482                  * Remember that we saw a request from this process, but
2483                  * don't start queuing just yet. Otherwise we risk seeing lots
2484                  * of tiny requests, because we disrupt the normal plugging
2485                  * and merging. If the request is already larger than a single
2486                  * page, let it rip immediately. For that case we assume that
2487                  * merging is already done. Ditto for a busy system that
2488                  * has other work pending, don't risk delaying until the
2489                  * idle timer unplug to continue working.
2490                  */
2491                 if (cfq_cfqq_wait_request(cfqq)) {
2492                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2493                             cfqd->busy_queues > 1) {
2494                                 del_timer(&cfqd->idle_slice_timer);
2495                         __blk_run_queue(cfqd->queue);
2496                         }
2497                         cfq_mark_cfqq_must_dispatch(cfqq);
2498                 }
2499         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2500                 /*
2501                  * not the active queue - expire current slice if it is
2502                  * idle and has expired it's mean thinktime or this new queue
2503                  * has some old slice time left and is of higher priority or
2504                  * this new queue is RT and the current one is BE
2505                  */
2506                 cfq_preempt_queue(cfqd, cfqq);
2507                 __blk_run_queue(cfqd->queue);
2508         }
2509 }
2510
2511 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2512 {
2513         struct cfq_data *cfqd = q->elevator->elevator_data;
2514         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2515
2516         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2517         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2518
2519         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2520         list_add_tail(&rq->queuelist, &cfqq->fifo);
2521         cfq_add_rq_rb(rq);
2522
2523         cfq_rq_enqueued(cfqd, cfqq, rq);
2524 }
2525
2526 /*
2527  * Update hw_tag based on peak queue depth over 50 samples under
2528  * sufficient load.
2529  */
2530 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2531 {
2532         struct cfq_queue *cfqq = cfqd->active_queue;
2533
2534         if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2535                 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2536
2537         if (cfqd->hw_tag == 1)
2538                 return;
2539
2540         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2541             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2542                 return;
2543
2544         /*
2545          * If active queue hasn't enough requests and can idle, cfq might not
2546          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2547          * case
2548          */
2549         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2550             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2551             CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2552                 return;
2553
2554         if (cfqd->hw_tag_samples++ < 50)
2555                 return;
2556
2557         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2558                 cfqd->hw_tag = 1;
2559         else
2560                 cfqd->hw_tag = 0;
2561 }
2562
2563 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2564 {
2565         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2566         struct cfq_data *cfqd = cfqq->cfqd;
2567         const int sync = rq_is_sync(rq);
2568         unsigned long now;
2569
2570         now = jiffies;
2571         cfq_log_cfqq(cfqd, cfqq, "complete");
2572
2573         cfq_update_hw_tag(cfqd);
2574
2575         WARN_ON(!cfqd->rq_in_driver[sync]);
2576         WARN_ON(!cfqq->dispatched);
2577         cfqd->rq_in_driver[sync]--;
2578         cfqq->dispatched--;
2579
2580         if (cfq_cfqq_sync(cfqq))
2581                 cfqd->sync_flight--;
2582
2583         if (sync) {
2584                 RQ_CIC(rq)->last_end_request = now;
2585                 cfqd->last_end_sync_rq = now;
2586         }
2587
2588         /*
2589          * If this is the active queue, check if it needs to be expired,
2590          * or if we want to idle in case it has no pending requests.
2591          */
2592         if (cfqd->active_queue == cfqq) {
2593                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2594
2595                 if (cfq_cfqq_slice_new(cfqq)) {
2596                         cfq_set_prio_slice(cfqd, cfqq);
2597                         cfq_clear_cfqq_slice_new(cfqq);
2598                 }
2599                 /*
2600                  * If there are no requests waiting in this queue, and
2601                  * there are other queues ready to issue requests, AND
2602                  * those other queues are issuing requests within our
2603                  * mean seek distance, give them a chance to run instead
2604                  * of idling.
2605                  */
2606                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2607                         cfq_slice_expired(cfqd, 1);
2608                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2609                          sync && !rq_noidle(rq))
2610                         cfq_arm_slice_timer(cfqd);
2611         }
2612
2613         if (!rq_in_driver(cfqd))
2614                 cfq_schedule_dispatch(cfqd);
2615 }
2616
2617 /*
2618  * we temporarily boost lower priority queues if they are holding fs exclusive
2619  * resources. they are boosted to normal prio (CLASS_BE/4)
2620  */
2621 static void cfq_prio_boost(struct cfq_queue *cfqq)
2622 {
2623         if (has_fs_excl()) {
2624                 /*
2625                  * boost idle prio on transactions that would lock out other
2626                  * users of the filesystem
2627                  */
2628                 if (cfq_class_idle(cfqq))
2629                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2630                 if (cfqq->ioprio > IOPRIO_NORM)
2631                         cfqq->ioprio = IOPRIO_NORM;
2632         } else {
2633                 /*
2634                  * unboost the queue (if needed)
2635                  */
2636                 cfqq->ioprio_class = cfqq->org_ioprio_class;
2637                 cfqq->ioprio = cfqq->org_ioprio;
2638         }
2639 }
2640
2641 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2642 {
2643         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2644                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2645                 return ELV_MQUEUE_MUST;
2646         }
2647
2648         return ELV_MQUEUE_MAY;
2649 }
2650
2651 static int cfq_may_queue(struct request_queue *q, int rw)
2652 {
2653         struct cfq_data *cfqd = q->elevator->elevator_data;
2654         struct task_struct *tsk = current;
2655         struct cfq_io_context *cic;
2656         struct cfq_queue *cfqq;
2657
2658         /*
2659          * don't force setup of a queue from here, as a call to may_queue
2660          * does not necessarily imply that a request actually will be queued.
2661          * so just lookup a possibly existing queue, or return 'may queue'
2662          * if that fails
2663          */
2664         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2665         if (!cic)
2666                 return ELV_MQUEUE_MAY;
2667
2668         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2669         if (cfqq) {
2670                 cfq_init_prio_data(cfqq, cic->ioc);
2671                 cfq_prio_boost(cfqq);
2672
2673                 return __cfq_may_queue(cfqq);
2674         }
2675
2676         return ELV_MQUEUE_MAY;
2677 }
2678
2679 /*
2680  * queue lock held here
2681  */
2682 static void cfq_put_request(struct request *rq)
2683 {
2684         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2685
2686         if (cfqq) {
2687                 const int rw = rq_data_dir(rq);
2688
2689                 BUG_ON(!cfqq->allocated[rw]);
2690                 cfqq->allocated[rw]--;
2691
2692                 put_io_context(RQ_CIC(rq)->ioc);
2693
2694                 rq->elevator_private = NULL;
2695                 rq->elevator_private2 = NULL;
2696
2697                 cfq_put_queue(cfqq);
2698         }
2699 }
2700
2701 static struct cfq_queue *
2702 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2703                 struct cfq_queue *cfqq)
2704 {
2705         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2706         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2707         cfq_mark_cfqq_coop(cfqq->new_cfqq);
2708         cfq_put_queue(cfqq);
2709         return cic_to_cfqq(cic, 1);
2710 }
2711
2712 static int should_split_cfqq(struct cfq_queue *cfqq)
2713 {
2714         if (cfqq->seeky_start &&
2715             time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2716                 return 1;
2717         return 0;
2718 }
2719
2720 /*
2721  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2722  * was the last process referring to said cfqq.
2723  */
2724 static struct cfq_queue *
2725 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2726 {
2727         if (cfqq_process_refs(cfqq) == 1) {
2728                 cfqq->seeky_start = 0;
2729                 cfqq->pid = current->pid;
2730                 cfq_clear_cfqq_coop(cfqq);
2731                 return cfqq;
2732         }
2733
2734         cic_set_cfqq(cic, NULL, 1);
2735         cfq_put_queue(cfqq);
2736         return NULL;
2737 }
2738 /*
2739  * Allocate cfq data structures associated with this request.
2740  */
2741 static int
2742 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2743 {
2744         struct cfq_data *cfqd = q->elevator->elevator_data;
2745         struct cfq_io_context *cic;
2746         const int rw = rq_data_dir(rq);
2747         const bool is_sync = rq_is_sync(rq);
2748         struct cfq_queue *cfqq;
2749         unsigned long flags;
2750
2751         might_sleep_if(gfp_mask & __GFP_WAIT);
2752
2753         cic = cfq_get_io_context(cfqd, gfp_mask);
2754
2755         spin_lock_irqsave(q->queue_lock, flags);
2756
2757         if (!cic)
2758                 goto queue_fail;
2759
2760 new_queue:
2761         cfqq = cic_to_cfqq(cic, is_sync);
2762         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2763                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2764                 cic_set_cfqq(cic, cfqq, is_sync);
2765         } else {
2766                 /*
2767                  * If the queue was seeky for too long, break it apart.
2768                  */
2769                 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2770                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2771                         cfqq = split_cfqq(cic, cfqq);
2772                         if (!cfqq)
2773                                 goto new_queue;
2774                 }
2775
2776                 /*
2777                  * Check to see if this queue is scheduled to merge with
2778                  * another, closely cooperating queue.  The merging of
2779                  * queues happens here as it must be done in process context.
2780                  * The reference on new_cfqq was taken in merge_cfqqs.
2781                  */
2782                 if (cfqq->new_cfqq)
2783                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2784         }
2785
2786         cfqq->allocated[rw]++;
2787         atomic_inc(&cfqq->ref);
2788
2789         spin_unlock_irqrestore(q->queue_lock, flags);
2790
2791         rq->elevator_private = cic;
2792         rq->elevator_private2 = cfqq;
2793         return 0;
2794
2795 queue_fail:
2796         if (cic)
2797                 put_io_context(cic->ioc);
2798
2799         cfq_schedule_dispatch(cfqd);
2800         spin_unlock_irqrestore(q->queue_lock, flags);
2801         cfq_log(cfqd, "set_request fail");
2802         return 1;
2803 }
2804
2805 static void cfq_kick_queue(struct work_struct *work)
2806 {
2807         struct cfq_data *cfqd =
2808                 container_of(work, struct cfq_data, unplug_work);
2809         struct request_queue *q = cfqd->queue;
2810
2811         spin_lock_irq(q->queue_lock);
2812         __blk_run_queue(cfqd->queue);
2813         spin_unlock_irq(q->queue_lock);
2814 }
2815
2816 /*
2817  * Timer running if the active_queue is currently idling inside its time slice
2818  */
2819 static void cfq_idle_slice_timer(unsigned long data)
2820 {
2821         struct cfq_data *cfqd = (struct cfq_data *) data;
2822         struct cfq_queue *cfqq;
2823         unsigned long flags;
2824         int timed_out = 1;
2825
2826         cfq_log(cfqd, "idle timer fired");
2827
2828         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2829
2830         cfqq = cfqd->active_queue;
2831         if (cfqq) {
2832                 timed_out = 0;
2833
2834                 /*
2835                  * We saw a request before the queue expired, let it through
2836                  */
2837                 if (cfq_cfqq_must_dispatch(cfqq))
2838                         goto out_kick;
2839
2840                 /*
2841                  * expired
2842                  */
2843                 if (cfq_slice_used(cfqq))
2844                         goto expire;
2845
2846                 /*
2847                  * only expire and reinvoke request handler, if there are
2848                  * other queues with pending requests
2849                  */
2850                 if (!cfqd->busy_queues)
2851                         goto out_cont;
2852
2853                 /*
2854                  * not expired and it has a request pending, let it dispatch
2855                  */
2856                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2857                         goto out_kick;
2858
2859                 /*
2860                  * Queue depth flag is reset only when the idle didn't succeed
2861                  */
2862                 cfq_clear_cfqq_deep(cfqq);
2863         }
2864 expire:
2865         cfq_slice_expired(cfqd, timed_out);
2866 out_kick:
2867         cfq_schedule_dispatch(cfqd);
2868 out_cont:
2869         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2870 }
2871
2872 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2873 {
2874         del_timer_sync(&cfqd->idle_slice_timer);
2875         cancel_work_sync(&cfqd->unplug_work);
2876 }
2877
2878 static void cfq_put_async_queues(struct cfq_data *cfqd)
2879 {
2880         int i;
2881
2882         for (i = 0; i < IOPRIO_BE_NR; i++) {
2883                 if (cfqd->async_cfqq[0][i])
2884                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2885                 if (cfqd->async_cfqq[1][i])
2886                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2887         }
2888
2889         if (cfqd->async_idle_cfqq)
2890                 cfq_put_queue(cfqd->async_idle_cfqq);
2891 }
2892
2893 static void cfq_exit_queue(struct elevator_queue *e)
2894 {
2895         struct cfq_data *cfqd = e->elevator_data;
2896         struct request_queue *q = cfqd->queue;
2897
2898         cfq_shutdown_timer_wq(cfqd);
2899
2900         spin_lock_irq(q->queue_lock);
2901
2902         if (cfqd->active_queue)
2903                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2904
2905         while (!list_empty(&cfqd->cic_list)) {
2906                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2907                                                         struct cfq_io_context,
2908                                                         queue_list);
2909
2910                 __cfq_exit_single_io_context(cfqd, cic);
2911         }
2912
2913         cfq_put_async_queues(cfqd);
2914
2915         spin_unlock_irq(q->queue_lock);
2916
2917         cfq_shutdown_timer_wq(cfqd);
2918
2919         kfree(cfqd);
2920 }
2921
2922 static void *cfq_init_queue(struct request_queue *q)
2923 {
2924         struct cfq_data *cfqd;
2925         int i, j;
2926
2927         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2928         if (!cfqd)
2929                 return NULL;
2930
2931         for (i = 0; i < 2; ++i)
2932                 for (j = 0; j < 3; ++j)
2933                         cfqd->service_trees[i][j] = CFQ_RB_ROOT;
2934         cfqd->service_tree_idle = CFQ_RB_ROOT;
2935
2936         /*
2937          * Not strictly needed (since RB_ROOT just clears the node and we
2938          * zeroed cfqd on alloc), but better be safe in case someone decides
2939          * to add magic to the rb code
2940          */
2941         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2942                 cfqd->prio_trees[i] = RB_ROOT;
2943
2944         /*
2945          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2946          * Grab a permanent reference to it, so that the normal code flow
2947          * will not attempt to free it.
2948          */
2949         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2950         atomic_inc(&cfqd->oom_cfqq.ref);
2951
2952         INIT_LIST_HEAD(&cfqd->cic_list);
2953
2954         cfqd->queue = q;
2955
2956         init_timer(&cfqd->idle_slice_timer);
2957         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2958         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2959
2960         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2961
2962         cfqd->cfq_quantum = cfq_quantum;
2963         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2964         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2965         cfqd->cfq_back_max = cfq_back_max;
2966         cfqd->cfq_back_penalty = cfq_back_penalty;
2967         cfqd->cfq_slice[0] = cfq_slice_async;
2968         cfqd->cfq_slice[1] = cfq_slice_sync;
2969         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2970         cfqd->cfq_slice_idle = cfq_slice_idle;
2971         cfqd->cfq_latency = 1;
2972         cfqd->hw_tag = -1;
2973         cfqd->last_end_sync_rq = jiffies;
2974         return cfqd;
2975 }
2976
2977 static void cfq_slab_kill(void)
2978 {
2979         /*
2980          * Caller already ensured that pending RCU callbacks are completed,
2981          * so we should have no busy allocations at this point.
2982          */
2983         if (cfq_pool)
2984                 kmem_cache_destroy(cfq_pool);
2985         if (cfq_ioc_pool)
2986                 kmem_cache_destroy(cfq_ioc_pool);
2987 }
2988
2989 static int __init cfq_slab_setup(void)
2990 {
2991         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2992         if (!cfq_pool)
2993                 goto fail;
2994
2995         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2996         if (!cfq_ioc_pool)
2997                 goto fail;
2998
2999         return 0;
3000 fail:
3001         cfq_slab_kill();
3002         return -ENOMEM;
3003 }
3004
3005 /*
3006  * sysfs parts below -->
3007  */
3008 static ssize_t
3009 cfq_var_show(unsigned int var, char *page)
3010 {
3011         return sprintf(page, "%d\n", var);
3012 }
3013
3014 static ssize_t
3015 cfq_var_store(unsigned int *var, const char *page, size_t count)
3016 {
3017         char *p = (char *) page;
3018
3019         *var = simple_strtoul(p, &p, 10);
3020         return count;
3021 }
3022
3023 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3024 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3025 {                                                                       \
3026         struct cfq_data *cfqd = e->elevator_data;                       \
3027         unsigned int __data = __VAR;                                    \
3028         if (__CONV)                                                     \
3029                 __data = jiffies_to_msecs(__data);                      \
3030         return cfq_var_show(__data, (page));                            \
3031 }
3032 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3033 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3034 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3035 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3036 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3037 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3038 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3039 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3040 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3041 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3042 #undef SHOW_FUNCTION
3043
3044 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3045 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3046 {                                                                       \
3047         struct cfq_data *cfqd = e->elevator_data;                       \
3048         unsigned int __data;                                            \
3049         int ret = cfq_var_store(&__data, (page), count);                \
3050         if (__data < (MIN))                                             \
3051                 __data = (MIN);                                         \
3052         else if (__data > (MAX))                                        \
3053                 __data = (MAX);                                         \
3054         if (__CONV)                                                     \
3055                 *(__PTR) = msecs_to_jiffies(__data);                    \
3056         else                                                            \
3057                 *(__PTR) = __data;                                      \
3058         return ret;                                                     \
3059 }
3060 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3061 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3062                 UINT_MAX, 1);
3063 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3064                 UINT_MAX, 1);
3065 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3066 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3067                 UINT_MAX, 0);
3068 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3069 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3070 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3071 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3072                 UINT_MAX, 0);
3073 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3074 #undef STORE_FUNCTION
3075
3076 #define CFQ_ATTR(name) \
3077         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3078
3079 static struct elv_fs_entry cfq_attrs[] = {
3080         CFQ_ATTR(quantum),
3081         CFQ_ATTR(fifo_expire_sync),
3082         CFQ_ATTR(fifo_expire_async),
3083         CFQ_ATTR(back_seek_max),
3084         CFQ_ATTR(back_seek_penalty),
3085         CFQ_ATTR(slice_sync),
3086         CFQ_ATTR(slice_async),
3087         CFQ_ATTR(slice_async_rq),
3088         CFQ_ATTR(slice_idle),
3089         CFQ_ATTR(low_latency),
3090         __ATTR_NULL
3091 };
3092
3093 static struct elevator_type iosched_cfq = {
3094         .ops = {
3095                 .elevator_merge_fn =            cfq_merge,
3096                 .elevator_merged_fn =           cfq_merged_request,
3097                 .elevator_merge_req_fn =        cfq_merged_requests,
3098                 .elevator_allow_merge_fn =      cfq_allow_merge,
3099                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3100                 .elevator_add_req_fn =          cfq_insert_request,
3101                 .elevator_activate_req_fn =     cfq_activate_request,
3102                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3103                 .elevator_queue_empty_fn =      cfq_queue_empty,
3104                 .elevator_completed_req_fn =    cfq_completed_request,
3105                 .elevator_former_req_fn =       elv_rb_former_request,
3106                 .elevator_latter_req_fn =       elv_rb_latter_request,
3107                 .elevator_set_req_fn =          cfq_set_request,
3108                 .elevator_put_req_fn =          cfq_put_request,
3109                 .elevator_may_queue_fn =        cfq_may_queue,
3110                 .elevator_init_fn =             cfq_init_queue,
3111                 .elevator_exit_fn =             cfq_exit_queue,
3112                 .trim =                         cfq_free_io_context,
3113         },
3114         .elevator_attrs =       cfq_attrs,
3115         .elevator_name =        "cfq",
3116         .elevator_owner =       THIS_MODULE,
3117 };
3118
3119 static int __init cfq_init(void)
3120 {
3121         /*
3122          * could be 0 on HZ < 1000 setups
3123          */
3124         if (!cfq_slice_async)
3125                 cfq_slice_async = 1;
3126         if (!cfq_slice_idle)
3127                 cfq_slice_idle = 1;
3128
3129         if (cfq_slab_setup())
3130                 return -ENOMEM;
3131
3132         elv_register(&iosched_cfq);
3133
3134         return 0;
3135 }
3136
3137 static void __exit cfq_exit(void)
3138 {
3139         DECLARE_COMPLETION_ONSTACK(all_gone);
3140         elv_unregister(&iosched_cfq);
3141         ioc_gone = &all_gone;
3142         /* ioc_gone's update must be visible before reading ioc_count */
3143         smp_wmb();
3144
3145         /*
3146          * this also protects us from entering cfq_slab_kill() with
3147          * pending RCU callbacks
3148          */
3149         if (elv_ioc_count_read(cfq_ioc_count))
3150                 wait_for_completion(&all_gone);
3151         cfq_slab_kill();
3152 }
3153
3154 module_init(cfq_init);
3155 module_exit(cfq_exit);
3156
3157 MODULE_AUTHOR("Jens Axboe");
3158 MODULE_LICENSE("GPL");
3159 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");