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