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