8b5ba184cf9178e6497e13fbecfbe6529c463d7b
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
15
16 /*
17  * tunables
18  */
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
31 static const int cfq_hist_divisor = 4;
32
33 /*
34  * offset from end of service tree
35  */
36 #define CFQ_IDLE_DELAY          (HZ / 5)
37
38 /*
39  * below this threshold, we consider thinktime immediate
40  */
41 #define CFQ_MIN_TT              (2)
42
43 /*
44  * Allow merged cfqqs to perform this amount of seeky I/O before
45  * deciding to break the queues up again.
46  */
47 #define CFQQ_COOP_TOUT          (HZ)
48
49 #define CFQ_SLICE_SCALE         (5)
50 #define CFQ_HW_QUEUE_MIN        (5)
51
52 #define RQ_CIC(rq)              \
53         ((struct cfq_io_context *) (rq)->elevator_private)
54 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
55
56 static struct kmem_cache *cfq_pool;
57 static struct kmem_cache *cfq_ioc_pool;
58
59 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
60 static struct completion *ioc_gone;
61 static DEFINE_SPINLOCK(ioc_gone_lock);
62
63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
67 #define sample_valid(samples)   ((samples) > 80)
68
69 /*
70  * Most of our rbtree usage is for sorting with min extraction, so
71  * if we cache the leftmost node we don't have to walk down the tree
72  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
73  * move this into the elevator for the rq sorting as well.
74  */
75 struct cfq_rb_root {
76         struct rb_root rb;
77         struct rb_node *left;
78         unsigned count;
79 };
80 #define CFQ_RB_ROOT     (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
81
82 /*
83  * Per process-grouping structure
84  */
85 struct cfq_queue {
86         /* reference count */
87         atomic_t ref;
88         /* various state flags, see below */
89         unsigned int flags;
90         /* parent cfq_data */
91         struct cfq_data *cfqd;
92         /* service_tree member */
93         struct rb_node rb_node;
94         /* service_tree key */
95         unsigned long rb_key;
96         /* prio tree member */
97         struct rb_node p_node;
98         /* prio tree root we belong to, if any */
99         struct rb_root *p_root;
100         /* sorted list of pending requests */
101         struct rb_root sort_list;
102         /* if fifo isn't expired, next request to serve */
103         struct request *next_rq;
104         /* requests queued in sort_list */
105         int queued[2];
106         /* currently allocated requests */
107         int allocated[2];
108         /* fifo list of requests in sort_list */
109         struct list_head fifo;
110
111         unsigned long slice_end;
112         long slice_resid;
113         unsigned int slice_dispatch;
114
115         /* pending metadata requests */
116         int meta_pending;
117         /* number of requests that are on the dispatch list or inside driver */
118         int dispatched;
119
120         /* io prio of this group */
121         unsigned short ioprio, org_ioprio;
122         unsigned short ioprio_class, org_ioprio_class;
123
124         unsigned int seek_samples;
125         u64 seek_total;
126         sector_t seek_mean;
127         sector_t last_request_pos;
128         unsigned long seeky_start;
129
130         pid_t pid;
131
132         struct cfq_rb_root *service_tree;
133         struct cfq_queue *new_cfqq;
134 };
135
136 /*
137  * First index in the service_trees.
138  * IDLE is handled separately, so it has negative index
139  */
140 enum wl_prio_t {
141         IDLE_WORKLOAD = -1,
142         BE_WORKLOAD = 0,
143         RT_WORKLOAD = 1
144 };
145
146 /*
147  * Second index in the service_trees.
148  */
149 enum wl_type_t {
150         ASYNC_WORKLOAD = 0,
151         SYNC_NOIDLE_WORKLOAD = 1,
152         SYNC_WORKLOAD = 2
153 };
154
155
156 /*
157  * Per block device queue structure
158  */
159 struct cfq_data {
160         struct request_queue *queue;
161
162         /*
163          * rr lists of queues with requests, onle rr for each priority class.
164          * Counts are embedded in the cfq_rb_root
165          */
166         struct cfq_rb_root service_trees[2][3];
167         struct cfq_rb_root service_tree_idle;
168         /*
169          * The priority currently being served
170          */
171         enum wl_prio_t serving_prio;
172         enum wl_type_t serving_type;
173         unsigned long workload_expires;
174
175         /*
176          * Each priority tree is sorted by next_request position.  These
177          * trees are used when determining if two or more queues are
178          * interleaving requests (see cfq_close_cooperator).
179          */
180         struct rb_root prio_trees[CFQ_PRIO_LISTS];
181
182         unsigned int busy_queues;
183         unsigned int busy_queues_avg[2];
184
185         int rq_in_driver[2];
186         int sync_flight;
187
188         /*
189          * queue-depth detection
190          */
191         int rq_queued;
192         int hw_tag;
193         int hw_tag_samples;
194         int rq_in_driver_peak;
195
196         /*
197          * idle window management
198          */
199         struct timer_list idle_slice_timer;
200         struct work_struct unplug_work;
201
202         struct cfq_queue *active_queue;
203         struct cfq_io_context *active_cic;
204
205         /*
206          * async queue for each priority case
207          */
208         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
209         struct cfq_queue *async_idle_cfqq;
210
211         sector_t last_position;
212
213         /*
214          * tunables, see top of file
215          */
216         unsigned int cfq_quantum;
217         unsigned int cfq_fifo_expire[2];
218         unsigned int cfq_back_penalty;
219         unsigned int cfq_back_max;
220         unsigned int cfq_slice[2];
221         unsigned int cfq_slice_async_rq;
222         unsigned int cfq_slice_idle;
223         unsigned int cfq_latency;
224
225         struct list_head cic_list;
226
227         /*
228          * Fallback dummy cfqq for extreme OOM conditions
229          */
230         struct cfq_queue oom_cfqq;
231
232         unsigned long last_end_sync_rq;
233 };
234
235 static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio,
236                                             enum wl_type_t type,
237                                             struct cfq_data *cfqd)
238 {
239         if (prio == IDLE_WORKLOAD)
240                 return &cfqd->service_tree_idle;
241
242         return &cfqd->service_trees[prio][type];
243 }
244
245 enum cfqq_state_flags {
246         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
247         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
248         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
249         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
250         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
251         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
252         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
253         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
254         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
255         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
256 };
257
258 #define CFQ_CFQQ_FNS(name)                                              \
259 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
260 {                                                                       \
261         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
262 }                                                                       \
263 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
264 {                                                                       \
265         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
266 }                                                                       \
267 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
268 {                                                                       \
269         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
270 }
271
272 CFQ_CFQQ_FNS(on_rr);
273 CFQ_CFQQ_FNS(wait_request);
274 CFQ_CFQQ_FNS(must_dispatch);
275 CFQ_CFQQ_FNS(must_alloc_slice);
276 CFQ_CFQQ_FNS(fifo_expire);
277 CFQ_CFQQ_FNS(idle_window);
278 CFQ_CFQQ_FNS(prio_changed);
279 CFQ_CFQQ_FNS(slice_new);
280 CFQ_CFQQ_FNS(sync);
281 CFQ_CFQQ_FNS(coop);
282 #undef CFQ_CFQQ_FNS
283
284 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
285         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
286 #define cfq_log(cfqd, fmt, args...)     \
287         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
288
289 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
290 {
291         if (cfq_class_idle(cfqq))
292                 return IDLE_WORKLOAD;
293         if (cfq_class_rt(cfqq))
294                 return RT_WORKLOAD;
295         return BE_WORKLOAD;
296 }
297
298
299 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
300 {
301         if (!cfq_cfqq_sync(cfqq))
302                 return ASYNC_WORKLOAD;
303         if (!cfq_cfqq_idle_window(cfqq))
304                 return SYNC_NOIDLE_WORKLOAD;
305         return SYNC_WORKLOAD;
306 }
307
308 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
309 {
310         if (wl == IDLE_WORKLOAD)
311                 return cfqd->service_tree_idle.count;
312
313         return cfqd->service_trees[wl][ASYNC_WORKLOAD].count
314                 + cfqd->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
315                 + cfqd->service_trees[wl][SYNC_WORKLOAD].count;
316 }
317
318 static void cfq_dispatch_insert(struct request_queue *, struct request *);
319 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
320                                        struct io_context *, gfp_t);
321 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
322                                                 struct io_context *);
323
324 static inline int rq_in_driver(struct cfq_data *cfqd)
325 {
326         return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
327 }
328
329 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
330                                             bool is_sync)
331 {
332         return cic->cfqq[is_sync];
333 }
334
335 static inline void cic_set_cfqq(struct cfq_io_context *cic,
336                                 struct cfq_queue *cfqq, bool is_sync)
337 {
338         cic->cfqq[is_sync] = cfqq;
339 }
340
341 /*
342  * We regard a request as SYNC, if it's either a read or has the SYNC bit
343  * set (in which case it could also be direct WRITE).
344  */
345 static inline bool cfq_bio_sync(struct bio *bio)
346 {
347         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
348 }
349
350 /*
351  * scheduler run of queue, if there are requests pending and no one in the
352  * driver that will restart queueing
353  */
354 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
355 {
356         if (cfqd->busy_queues) {
357                 cfq_log(cfqd, "schedule dispatch");
358                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
359         }
360 }
361
362 static int cfq_queue_empty(struct request_queue *q)
363 {
364         struct cfq_data *cfqd = q->elevator->elevator_data;
365
366         return !cfqd->busy_queues;
367 }
368
369 /*
370  * Scale schedule slice based on io priority. Use the sync time slice only
371  * if a queue is marked sync and has sync io queued. A sync queue with async
372  * io only, should not get full sync slice length.
373  */
374 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
375                                  unsigned short prio)
376 {
377         const int base_slice = cfqd->cfq_slice[sync];
378
379         WARN_ON(prio >= IOPRIO_BE_NR);
380
381         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
382 }
383
384 static inline int
385 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
386 {
387         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
388 }
389
390 /*
391  * get averaged number of queues of RT/BE priority.
392  * average is updated, with a formula that gives more weight to higher numbers,
393  * to quickly follows sudden increases and decrease slowly
394  */
395
396 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
397 {
398         unsigned min_q, max_q;
399         unsigned mult  = cfq_hist_divisor - 1;
400         unsigned round = cfq_hist_divisor / 2;
401         unsigned busy = cfq_busy_queues_wl(rt, cfqd);
402
403         min_q = min(cfqd->busy_queues_avg[rt], busy);
404         max_q = max(cfqd->busy_queues_avg[rt], busy);
405         cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
406                 cfq_hist_divisor;
407         return cfqd->busy_queues_avg[rt];
408 }
409
410 static inline void
411 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
412 {
413         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
414         if (cfqd->cfq_latency) {
415                 /* interested queues (we consider only the ones with the same
416                  * priority class) */
417                 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
418                 unsigned sync_slice = cfqd->cfq_slice[1];
419                 unsigned expect_latency = sync_slice * iq;
420                 if (expect_latency > cfq_target_latency) {
421                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
422                         /* scale low_slice according to IO priority
423                          * and sync vs async */
424                         unsigned low_slice =
425                                 min(slice, base_low_slice * slice / sync_slice);
426                         /* the adapted slice value is scaled to fit all iqs
427                          * into the target latency */
428                         slice = max(slice * cfq_target_latency / expect_latency,
429                                     low_slice);
430                 }
431         }
432         cfqq->slice_end = jiffies + slice;
433         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
434 }
435
436 /*
437  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
438  * isn't valid until the first request from the dispatch is activated
439  * and the slice time set.
440  */
441 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
442 {
443         if (cfq_cfqq_slice_new(cfqq))
444                 return 0;
445         if (time_before(jiffies, cfqq->slice_end))
446                 return 0;
447
448         return 1;
449 }
450
451 /*
452  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
453  * We choose the request that is closest to the head right now. Distance
454  * behind the head is penalized and only allowed to a certain extent.
455  */
456 static struct request *
457 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
458 {
459         sector_t last, s1, s2, d1 = 0, d2 = 0;
460         unsigned long back_max;
461 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
462 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
463         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
464
465         if (rq1 == NULL || rq1 == rq2)
466                 return rq2;
467         if (rq2 == NULL)
468                 return rq1;
469
470         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
471                 return rq1;
472         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
473                 return rq2;
474         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
475                 return rq1;
476         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
477                 return rq2;
478
479         s1 = blk_rq_pos(rq1);
480         s2 = blk_rq_pos(rq2);
481
482         last = cfqd->last_position;
483
484         /*
485          * by definition, 1KiB is 2 sectors
486          */
487         back_max = cfqd->cfq_back_max * 2;
488
489         /*
490          * Strict one way elevator _except_ in the case where we allow
491          * short backward seeks which are biased as twice the cost of a
492          * similar forward seek.
493          */
494         if (s1 >= last)
495                 d1 = s1 - last;
496         else if (s1 + back_max >= last)
497                 d1 = (last - s1) * cfqd->cfq_back_penalty;
498         else
499                 wrap |= CFQ_RQ1_WRAP;
500
501         if (s2 >= last)
502                 d2 = s2 - last;
503         else if (s2 + back_max >= last)
504                 d2 = (last - s2) * cfqd->cfq_back_penalty;
505         else
506                 wrap |= CFQ_RQ2_WRAP;
507
508         /* Found required data */
509
510         /*
511          * By doing switch() on the bit mask "wrap" we avoid having to
512          * check two variables for all permutations: --> faster!
513          */
514         switch (wrap) {
515         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
516                 if (d1 < d2)
517                         return rq1;
518                 else if (d2 < d1)
519                         return rq2;
520                 else {
521                         if (s1 >= s2)
522                                 return rq1;
523                         else
524                                 return rq2;
525                 }
526
527         case CFQ_RQ2_WRAP:
528                 return rq1;
529         case CFQ_RQ1_WRAP:
530                 return rq2;
531         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
532         default:
533                 /*
534                  * Since both rqs are wrapped,
535                  * start with the one that's further behind head
536                  * (--> only *one* back seek required),
537                  * since back seek takes more time than forward.
538                  */
539                 if (s1 <= s2)
540                         return rq1;
541                 else
542                         return rq2;
543         }
544 }
545
546 /*
547  * The below is leftmost cache rbtree addon
548  */
549 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
550 {
551         if (!root->left)
552                 root->left = rb_first(&root->rb);
553
554         if (root->left)
555                 return rb_entry(root->left, struct cfq_queue, rb_node);
556
557         return NULL;
558 }
559
560 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
561 {
562         rb_erase(n, root);
563         RB_CLEAR_NODE(n);
564 }
565
566 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
567 {
568         if (root->left == n)
569                 root->left = NULL;
570         rb_erase_init(n, &root->rb);
571         --root->count;
572 }
573
574 /*
575  * would be nice to take fifo expire time into account as well
576  */
577 static struct request *
578 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
579                   struct request *last)
580 {
581         struct rb_node *rbnext = rb_next(&last->rb_node);
582         struct rb_node *rbprev = rb_prev(&last->rb_node);
583         struct request *next = NULL, *prev = NULL;
584
585         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
586
587         if (rbprev)
588                 prev = rb_entry_rq(rbprev);
589
590         if (rbnext)
591                 next = rb_entry_rq(rbnext);
592         else {
593                 rbnext = rb_first(&cfqq->sort_list);
594                 if (rbnext && rbnext != &last->rb_node)
595                         next = rb_entry_rq(rbnext);
596         }
597
598         return cfq_choose_req(cfqd, next, prev);
599 }
600
601 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
602                                       struct cfq_queue *cfqq)
603 {
604         /*
605          * just an approximation, should be ok.
606          */
607         return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
608                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
609 }
610
611 /*
612  * The cfqd->service_trees holds all pending cfq_queue's that have
613  * requests waiting to be processed. It is sorted in the order that
614  * we will service the queues.
615  */
616 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
617                                  bool add_front)
618 {
619         struct rb_node **p, *parent;
620         struct cfq_queue *__cfqq;
621         unsigned long rb_key;
622         struct cfq_rb_root *service_tree;
623         int left;
624
625         service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
626         if (cfq_class_idle(cfqq)) {
627                 rb_key = CFQ_IDLE_DELAY;
628                 parent = rb_last(&service_tree->rb);
629                 if (parent && parent != &cfqq->rb_node) {
630                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
631                         rb_key += __cfqq->rb_key;
632                 } else
633                         rb_key += jiffies;
634         } else if (!add_front) {
635                 /*
636                  * Get our rb key offset. Subtract any residual slice
637                  * value carried from last service. A negative resid
638                  * count indicates slice overrun, and this should position
639                  * the next service time further away in the tree.
640                  */
641                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
642                 rb_key -= cfqq->slice_resid;
643                 cfqq->slice_resid = 0;
644         } else {
645                 rb_key = -HZ;
646                 __cfqq = cfq_rb_first(service_tree);
647                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
648         }
649
650         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
651                 /*
652                  * same position, nothing more to do
653                  */
654                 if (rb_key == cfqq->rb_key &&
655                     cfqq->service_tree == service_tree)
656                         return;
657
658                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
659                 cfqq->service_tree = NULL;
660         }
661
662         left = 1;
663         parent = NULL;
664         cfqq->service_tree = service_tree;
665         p = &service_tree->rb.rb_node;
666         while (*p) {
667                 struct rb_node **n;
668
669                 parent = *p;
670                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
671
672                 /*
673                  * sort by key, that represents service time.
674                  */
675                 if (time_before(rb_key, __cfqq->rb_key))
676                         n = &(*p)->rb_left;
677                 else {
678                         n = &(*p)->rb_right;
679                         left = 0;
680                 }
681
682                 p = n;
683         }
684
685         if (left)
686                 service_tree->left = &cfqq->rb_node;
687
688         cfqq->rb_key = rb_key;
689         rb_link_node(&cfqq->rb_node, parent, p);
690         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
691         service_tree->count++;
692 }
693
694 static struct cfq_queue *
695 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
696                      sector_t sector, struct rb_node **ret_parent,
697                      struct rb_node ***rb_link)
698 {
699         struct rb_node **p, *parent;
700         struct cfq_queue *cfqq = NULL;
701
702         parent = NULL;
703         p = &root->rb_node;
704         while (*p) {
705                 struct rb_node **n;
706
707                 parent = *p;
708                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
709
710                 /*
711                  * Sort strictly based on sector.  Smallest to the left,
712                  * largest to the right.
713                  */
714                 if (sector > blk_rq_pos(cfqq->next_rq))
715                         n = &(*p)->rb_right;
716                 else if (sector < blk_rq_pos(cfqq->next_rq))
717                         n = &(*p)->rb_left;
718                 else
719                         break;
720                 p = n;
721                 cfqq = NULL;
722         }
723
724         *ret_parent = parent;
725         if (rb_link)
726                 *rb_link = p;
727         return cfqq;
728 }
729
730 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
731 {
732         struct rb_node **p, *parent;
733         struct cfq_queue *__cfqq;
734
735         if (cfqq->p_root) {
736                 rb_erase(&cfqq->p_node, cfqq->p_root);
737                 cfqq->p_root = NULL;
738         }
739
740         if (cfq_class_idle(cfqq))
741                 return;
742         if (!cfqq->next_rq)
743                 return;
744
745         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
746         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
747                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
748         if (!__cfqq) {
749                 rb_link_node(&cfqq->p_node, parent, p);
750                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
751         } else
752                 cfqq->p_root = NULL;
753 }
754
755 /*
756  * Update cfqq's position in the service tree.
757  */
758 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
759 {
760         /*
761          * Resorting requires the cfqq to be on the RR list already.
762          */
763         if (cfq_cfqq_on_rr(cfqq)) {
764                 cfq_service_tree_add(cfqd, cfqq, 0);
765                 cfq_prio_tree_add(cfqd, cfqq);
766         }
767 }
768
769 /*
770  * add to busy list of queues for service, trying to be fair in ordering
771  * the pending list according to last request service
772  */
773 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
774 {
775         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
776         BUG_ON(cfq_cfqq_on_rr(cfqq));
777         cfq_mark_cfqq_on_rr(cfqq);
778         cfqd->busy_queues++;
779
780         cfq_resort_rr_list(cfqd, cfqq);
781 }
782
783 /*
784  * Called when the cfqq no longer has requests pending, remove it from
785  * the service tree.
786  */
787 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
788 {
789         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
790         BUG_ON(!cfq_cfqq_on_rr(cfqq));
791         cfq_clear_cfqq_on_rr(cfqq);
792
793         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
794                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
795                 cfqq->service_tree = NULL;
796         }
797         if (cfqq->p_root) {
798                 rb_erase(&cfqq->p_node, cfqq->p_root);
799                 cfqq->p_root = NULL;
800         }
801
802         BUG_ON(!cfqd->busy_queues);
803         cfqd->busy_queues--;
804 }
805
806 /*
807  * rb tree support functions
808  */
809 static void cfq_del_rq_rb(struct request *rq)
810 {
811         struct cfq_queue *cfqq = RQ_CFQQ(rq);
812         struct cfq_data *cfqd = cfqq->cfqd;
813         const int sync = rq_is_sync(rq);
814
815         BUG_ON(!cfqq->queued[sync]);
816         cfqq->queued[sync]--;
817
818         elv_rb_del(&cfqq->sort_list, rq);
819
820         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
821                 cfq_del_cfqq_rr(cfqd, cfqq);
822 }
823
824 static void cfq_add_rq_rb(struct request *rq)
825 {
826         struct cfq_queue *cfqq = RQ_CFQQ(rq);
827         struct cfq_data *cfqd = cfqq->cfqd;
828         struct request *__alias, *prev;
829
830         cfqq->queued[rq_is_sync(rq)]++;
831
832         /*
833          * looks a little odd, but the first insert might return an alias.
834          * if that happens, put the alias on the dispatch list
835          */
836         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
837                 cfq_dispatch_insert(cfqd->queue, __alias);
838
839         if (!cfq_cfqq_on_rr(cfqq))
840                 cfq_add_cfqq_rr(cfqd, cfqq);
841
842         /*
843          * check if this request is a better next-serve candidate
844          */
845         prev = cfqq->next_rq;
846         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
847
848         /*
849          * adjust priority tree position, if ->next_rq changes
850          */
851         if (prev != cfqq->next_rq)
852                 cfq_prio_tree_add(cfqd, cfqq);
853
854         BUG_ON(!cfqq->next_rq);
855 }
856
857 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
858 {
859         elv_rb_del(&cfqq->sort_list, rq);
860         cfqq->queued[rq_is_sync(rq)]--;
861         cfq_add_rq_rb(rq);
862 }
863
864 static struct request *
865 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
866 {
867         struct task_struct *tsk = current;
868         struct cfq_io_context *cic;
869         struct cfq_queue *cfqq;
870
871         cic = cfq_cic_lookup(cfqd, tsk->io_context);
872         if (!cic)
873                 return NULL;
874
875         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
876         if (cfqq) {
877                 sector_t sector = bio->bi_sector + bio_sectors(bio);
878
879                 return elv_rb_find(&cfqq->sort_list, sector);
880         }
881
882         return NULL;
883 }
884
885 static void cfq_activate_request(struct request_queue *q, struct request *rq)
886 {
887         struct cfq_data *cfqd = q->elevator->elevator_data;
888
889         cfqd->rq_in_driver[rq_is_sync(rq)]++;
890         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
891                                                 rq_in_driver(cfqd));
892
893         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
894 }
895
896 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
897 {
898         struct cfq_data *cfqd = q->elevator->elevator_data;
899         const int sync = rq_is_sync(rq);
900
901         WARN_ON(!cfqd->rq_in_driver[sync]);
902         cfqd->rq_in_driver[sync]--;
903         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
904                                                 rq_in_driver(cfqd));
905 }
906
907 static void cfq_remove_request(struct request *rq)
908 {
909         struct cfq_queue *cfqq = RQ_CFQQ(rq);
910
911         if (cfqq->next_rq == rq)
912                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
913
914         list_del_init(&rq->queuelist);
915         cfq_del_rq_rb(rq);
916
917         cfqq->cfqd->rq_queued--;
918         if (rq_is_meta(rq)) {
919                 WARN_ON(!cfqq->meta_pending);
920                 cfqq->meta_pending--;
921         }
922 }
923
924 static int cfq_merge(struct request_queue *q, struct request **req,
925                      struct bio *bio)
926 {
927         struct cfq_data *cfqd = q->elevator->elevator_data;
928         struct request *__rq;
929
930         __rq = cfq_find_rq_fmerge(cfqd, bio);
931         if (__rq && elv_rq_merge_ok(__rq, bio)) {
932                 *req = __rq;
933                 return ELEVATOR_FRONT_MERGE;
934         }
935
936         return ELEVATOR_NO_MERGE;
937 }
938
939 static void cfq_merged_request(struct request_queue *q, struct request *req,
940                                int type)
941 {
942         if (type == ELEVATOR_FRONT_MERGE) {
943                 struct cfq_queue *cfqq = RQ_CFQQ(req);
944
945                 cfq_reposition_rq_rb(cfqq, req);
946         }
947 }
948
949 static void
950 cfq_merged_requests(struct request_queue *q, struct request *rq,
951                     struct request *next)
952 {
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         cfq_remove_request(next);
963 }
964
965 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
966                            struct bio *bio)
967 {
968         struct cfq_data *cfqd = q->elevator->elevator_data;
969         struct cfq_io_context *cic;
970         struct cfq_queue *cfqq;
971
972         /*
973          * Disallow merge of a sync bio into an async request.
974          */
975         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
976                 return false;
977
978         /*
979          * Lookup the cfqq that this bio will be queued with. Allow
980          * merge only if rq is queued there.
981          */
982         cic = cfq_cic_lookup(cfqd, current->io_context);
983         if (!cic)
984                 return false;
985
986         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
987         return cfqq == RQ_CFQQ(rq);
988 }
989
990 static void __cfq_set_active_queue(struct cfq_data *cfqd,
991                                    struct cfq_queue *cfqq)
992 {
993         if (cfqq) {
994                 cfq_log_cfqq(cfqd, cfqq, "set_active");
995                 cfqq->slice_end = 0;
996                 cfqq->slice_dispatch = 0;
997
998                 cfq_clear_cfqq_wait_request(cfqq);
999                 cfq_clear_cfqq_must_dispatch(cfqq);
1000                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1001                 cfq_clear_cfqq_fifo_expire(cfqq);
1002                 cfq_mark_cfqq_slice_new(cfqq);
1003
1004                 del_timer(&cfqd->idle_slice_timer);
1005         }
1006
1007         cfqd->active_queue = cfqq;
1008 }
1009
1010 /*
1011  * current cfqq expired its slice (or was too idle), select new one
1012  */
1013 static void
1014 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1015                     bool timed_out)
1016 {
1017         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1018
1019         if (cfq_cfqq_wait_request(cfqq))
1020                 del_timer(&cfqd->idle_slice_timer);
1021
1022         cfq_clear_cfqq_wait_request(cfqq);
1023
1024         /*
1025          * store what was left of this slice, if the queue idled/timed out
1026          */
1027         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1028                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1029                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1030         }
1031
1032         cfq_resort_rr_list(cfqd, cfqq);
1033
1034         if (cfqq == cfqd->active_queue)
1035                 cfqd->active_queue = NULL;
1036
1037         if (cfqd->active_cic) {
1038                 put_io_context(cfqd->active_cic->ioc);
1039                 cfqd->active_cic = NULL;
1040         }
1041 }
1042
1043 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1044 {
1045         struct cfq_queue *cfqq = cfqd->active_queue;
1046
1047         if (cfqq)
1048                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1049 }
1050
1051 /*
1052  * Get next queue for service. Unless we have a queue preemption,
1053  * we'll simply select the first cfqq in the service tree.
1054  */
1055 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1056 {
1057         struct cfq_rb_root *service_tree =
1058                 service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd);
1059
1060         if (RB_EMPTY_ROOT(&service_tree->rb))
1061                 return NULL;
1062         return cfq_rb_first(service_tree);
1063 }
1064
1065 /*
1066  * Get and set a new active queue for service.
1067  */
1068 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1069                                               struct cfq_queue *cfqq)
1070 {
1071         if (!cfqq)
1072                 cfqq = cfq_get_next_queue(cfqd);
1073
1074         __cfq_set_active_queue(cfqd, cfqq);
1075         return cfqq;
1076 }
1077
1078 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1079                                           struct request *rq)
1080 {
1081         if (blk_rq_pos(rq) >= cfqd->last_position)
1082                 return blk_rq_pos(rq) - cfqd->last_position;
1083         else
1084                 return cfqd->last_position - blk_rq_pos(rq);
1085 }
1086
1087 #define CFQQ_SEEK_THR           8 * 1024
1088 #define CFQQ_SEEKY(cfqq)        ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1089
1090 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1091                                struct request *rq)
1092 {
1093         sector_t sdist = cfqq->seek_mean;
1094
1095         if (!sample_valid(cfqq->seek_samples))
1096                 sdist = CFQQ_SEEK_THR;
1097
1098         return cfq_dist_from_last(cfqd, rq) <= sdist;
1099 }
1100
1101 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1102                                     struct cfq_queue *cur_cfqq)
1103 {
1104         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1105         struct rb_node *parent, *node;
1106         struct cfq_queue *__cfqq;
1107         sector_t sector = cfqd->last_position;
1108
1109         if (RB_EMPTY_ROOT(root))
1110                 return NULL;
1111
1112         /*
1113          * First, if we find a request starting at the end of the last
1114          * request, choose it.
1115          */
1116         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1117         if (__cfqq)
1118                 return __cfqq;
1119
1120         /*
1121          * If the exact sector wasn't found, the parent of the NULL leaf
1122          * will contain the closest sector.
1123          */
1124         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1125         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1126                 return __cfqq;
1127
1128         if (blk_rq_pos(__cfqq->next_rq) < sector)
1129                 node = rb_next(&__cfqq->p_node);
1130         else
1131                 node = rb_prev(&__cfqq->p_node);
1132         if (!node)
1133                 return NULL;
1134
1135         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1136         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1137                 return __cfqq;
1138
1139         return NULL;
1140 }
1141
1142 /*
1143  * cfqd - obvious
1144  * cur_cfqq - passed in so that we don't decide that the current queue is
1145  *            closely cooperating with itself.
1146  *
1147  * So, basically we're assuming that that cur_cfqq has dispatched at least
1148  * one request, and that cfqd->last_position reflects a position on the disk
1149  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1150  * assumption.
1151  */
1152 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1153                                               struct cfq_queue *cur_cfqq)
1154 {
1155         struct cfq_queue *cfqq;
1156
1157         if (!cfq_cfqq_sync(cur_cfqq))
1158                 return NULL;
1159         if (CFQQ_SEEKY(cur_cfqq))
1160                 return NULL;
1161
1162         /*
1163          * We should notice if some of the queues are cooperating, eg
1164          * working closely on the same area of the disk. In that case,
1165          * we can group them together and don't waste time idling.
1166          */
1167         cfqq = cfqq_close(cfqd, cur_cfqq);
1168         if (!cfqq)
1169                 return NULL;
1170
1171         /*
1172          * It only makes sense to merge sync queues.
1173          */
1174         if (!cfq_cfqq_sync(cfqq))
1175                 return NULL;
1176         if (CFQQ_SEEKY(cfqq))
1177                 return NULL;
1178
1179         /*
1180          * Do not merge queues of different priority classes
1181          */
1182         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1183                 return NULL;
1184
1185         return cfqq;
1186 }
1187
1188 /*
1189  * Determine whether we should enforce idle window for this queue.
1190  */
1191
1192 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1193 {
1194         enum wl_prio_t prio = cfqq_prio(cfqq);
1195         struct cfq_rb_root *service_tree = cfqq->service_tree;
1196
1197         /* We never do for idle class queues. */
1198         if (prio == IDLE_WORKLOAD)
1199                 return false;
1200
1201         /* We do for queues that were marked with idle window flag. */
1202         if (cfq_cfqq_idle_window(cfqq))
1203                 return true;
1204
1205         /*
1206          * Otherwise, we do only if they are the last ones
1207          * in their service tree.
1208          */
1209         if (!service_tree)
1210                 service_tree = service_tree_for(prio, cfqq_type(cfqq), cfqd);
1211
1212         if (service_tree->count == 0)
1213                 return true;
1214
1215         return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq);
1216 }
1217
1218 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1219 {
1220         struct cfq_queue *cfqq = cfqd->active_queue;
1221         struct cfq_io_context *cic;
1222         unsigned long sl;
1223
1224         /*
1225          * SSD device without seek penalty, disable idling. But only do so
1226          * for devices that support queuing, otherwise we still have a problem
1227          * with sync vs async workloads.
1228          */
1229         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1230                 return;
1231
1232         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1233         WARN_ON(cfq_cfqq_slice_new(cfqq));
1234
1235         /*
1236          * idle is disabled, either manually or by past process history
1237          */
1238         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1239                 return;
1240
1241         /*
1242          * still requests with the driver, don't idle
1243          */
1244         if (rq_in_driver(cfqd))
1245                 return;
1246
1247         /*
1248          * task has exited, don't wait
1249          */
1250         cic = cfqd->active_cic;
1251         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1252                 return;
1253
1254         /*
1255          * If our average think time is larger than the remaining time
1256          * slice, then don't idle. This avoids overrunning the allotted
1257          * time slice.
1258          */
1259         if (sample_valid(cic->ttime_samples) &&
1260             (cfqq->slice_end - jiffies < cic->ttime_mean))
1261                 return;
1262
1263         cfq_mark_cfqq_wait_request(cfqq);
1264
1265         sl = cfqd->cfq_slice_idle;
1266         /* are we servicing noidle tree, and there are more queues?
1267          * non-rotational or NCQ: no idle
1268          * non-NCQ rotational : very small idle, to allow
1269          *     fair distribution of slice time for a process doing back-to-back
1270          *     seeks.
1271          */
1272         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
1273             service_tree_for(cfqd->serving_prio, SYNC_NOIDLE_WORKLOAD, cfqd)
1274                 ->count > 0) {
1275                 if (blk_queue_nonrot(cfqd->queue) || cfqd->hw_tag)
1276                         return;
1277                 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1278         }
1279
1280         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1281         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1282 }
1283
1284 /*
1285  * Move request from internal lists to the request queue dispatch list.
1286  */
1287 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1288 {
1289         struct cfq_data *cfqd = q->elevator->elevator_data;
1290         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1291
1292         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1293
1294         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1295         cfq_remove_request(rq);
1296         cfqq->dispatched++;
1297         elv_dispatch_sort(q, rq);
1298
1299         if (cfq_cfqq_sync(cfqq))
1300                 cfqd->sync_flight++;
1301 }
1302
1303 /*
1304  * return expired entry, or NULL to just start from scratch in rbtree
1305  */
1306 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1307 {
1308         struct request *rq = NULL;
1309
1310         if (cfq_cfqq_fifo_expire(cfqq))
1311                 return NULL;
1312
1313         cfq_mark_cfqq_fifo_expire(cfqq);
1314
1315         if (list_empty(&cfqq->fifo))
1316                 return NULL;
1317
1318         rq = rq_entry_fifo(cfqq->fifo.next);
1319         if (time_before(jiffies, rq_fifo_time(rq)))
1320                 rq = NULL;
1321
1322         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1323         return rq;
1324 }
1325
1326 static inline int
1327 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328 {
1329         const int base_rq = cfqd->cfq_slice_async_rq;
1330
1331         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1332
1333         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1334 }
1335
1336 /*
1337  * Must be called with the queue_lock held.
1338  */
1339 static int cfqq_process_refs(struct cfq_queue *cfqq)
1340 {
1341         int process_refs, io_refs;
1342
1343         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1344         process_refs = atomic_read(&cfqq->ref) - io_refs;
1345         BUG_ON(process_refs < 0);
1346         return process_refs;
1347 }
1348
1349 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1350 {
1351         int process_refs, new_process_refs;
1352         struct cfq_queue *__cfqq;
1353
1354         /* Avoid a circular list and skip interim queue merges */
1355         while ((__cfqq = new_cfqq->new_cfqq)) {
1356                 if (__cfqq == cfqq)
1357                         return;
1358                 new_cfqq = __cfqq;
1359         }
1360
1361         process_refs = cfqq_process_refs(cfqq);
1362         /*
1363          * If the process for the cfqq has gone away, there is no
1364          * sense in merging the queues.
1365          */
1366         if (process_refs == 0)
1367                 return;
1368
1369         /*
1370          * Merge in the direction of the lesser amount of work.
1371          */
1372         new_process_refs = cfqq_process_refs(new_cfqq);
1373         if (new_process_refs >= process_refs) {
1374                 cfqq->new_cfqq = new_cfqq;
1375                 atomic_add(process_refs, &new_cfqq->ref);
1376         } else {
1377                 new_cfqq->new_cfqq = cfqq;
1378                 atomic_add(new_process_refs, &cfqq->ref);
1379         }
1380 }
1381
1382 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, enum wl_prio_t prio,
1383                                     bool prio_changed)
1384 {
1385         struct cfq_queue *queue;
1386         int i;
1387         bool key_valid = false;
1388         unsigned long lowest_key = 0;
1389         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1390
1391         if (prio_changed) {
1392                 /*
1393                  * When priorities switched, we prefer starting
1394                  * from SYNC_NOIDLE (first choice), or just SYNC
1395                  * over ASYNC
1396                  */
1397                 if (service_tree_for(prio, cur_best, cfqd)->count)
1398                         return cur_best;
1399                 cur_best = SYNC_WORKLOAD;
1400                 if (service_tree_for(prio, cur_best, cfqd)->count)
1401                         return cur_best;
1402
1403                 return ASYNC_WORKLOAD;
1404         }
1405
1406         for (i = 0; i < 3; ++i) {
1407                 /* otherwise, select the one with lowest rb_key */
1408                 queue = cfq_rb_first(service_tree_for(prio, i, cfqd));
1409                 if (queue &&
1410                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
1411                         lowest_key = queue->rb_key;
1412                         cur_best = i;
1413                         key_valid = true;
1414                 }
1415         }
1416
1417         return cur_best;
1418 }
1419
1420 static void choose_service_tree(struct cfq_data *cfqd)
1421 {
1422         enum wl_prio_t previous_prio = cfqd->serving_prio;
1423         bool prio_changed;
1424         unsigned slice;
1425         unsigned count;
1426
1427         /* Choose next priority. RT > BE > IDLE */
1428         if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1429                 cfqd->serving_prio = RT_WORKLOAD;
1430         else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1431                 cfqd->serving_prio = BE_WORKLOAD;
1432         else {
1433                 cfqd->serving_prio = IDLE_WORKLOAD;
1434                 cfqd->workload_expires = jiffies + 1;
1435                 return;
1436         }
1437
1438         /*
1439          * For RT and BE, we have to choose also the type
1440          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1441          * expiration time
1442          */
1443         prio_changed = (cfqd->serving_prio != previous_prio);
1444         count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1445                 ->count;
1446
1447         /*
1448          * If priority didn't change, check workload expiration,
1449          * and that we still have other queues ready
1450          */
1451         if (!prio_changed && count &&
1452             !time_after(jiffies, cfqd->workload_expires))
1453                 return;
1454
1455         /* otherwise select new workload type */
1456         cfqd->serving_type =
1457                 cfq_choose_wl(cfqd, cfqd->serving_prio, prio_changed);
1458         count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1459                 ->count;
1460
1461         /*
1462          * the workload slice is computed as a fraction of target latency
1463          * proportional to the number of queues in that workload, over
1464          * all the queues in the same priority class
1465          */
1466         slice = cfq_target_latency * count /
1467                 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1468                       cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1469
1470         if (cfqd->serving_type == ASYNC_WORKLOAD)
1471                 /* async workload slice is scaled down according to
1472                  * the sync/async slice ratio. */
1473                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1474         else
1475                 /* sync workload slice is at least 2 * cfq_slice_idle */
1476                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1477
1478         slice = max_t(unsigned, slice, CFQ_MIN_TT);
1479         cfqd->workload_expires = jiffies + slice;
1480 }
1481
1482 /*
1483  * Select a queue for service. If we have a current active queue,
1484  * check whether to continue servicing it, or retrieve and set a new one.
1485  */
1486 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1487 {
1488         struct cfq_queue *cfqq, *new_cfqq = NULL;
1489
1490         cfqq = cfqd->active_queue;
1491         if (!cfqq)
1492                 goto new_queue;
1493
1494         /*
1495          * The active queue has run out of time, expire it and select new.
1496          */
1497         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1498                 goto expire;
1499
1500         /*
1501          * The active queue has requests and isn't expired, allow it to
1502          * dispatch.
1503          */
1504         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1505                 goto keep_queue;
1506
1507         /*
1508          * If another queue has a request waiting within our mean seek
1509          * distance, let it run.  The expire code will check for close
1510          * cooperators and put the close queue at the front of the service
1511          * tree.  If possible, merge the expiring queue with the new cfqq.
1512          */
1513         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1514         if (new_cfqq) {
1515                 if (!cfqq->new_cfqq)
1516                         cfq_setup_merge(cfqq, new_cfqq);
1517                 goto expire;
1518         }
1519
1520         /*
1521          * No requests pending. If the active queue still has requests in
1522          * flight or is idling for a new request, allow either of these
1523          * conditions to happen (or time out) before selecting a new queue.
1524          */
1525         if (timer_pending(&cfqd->idle_slice_timer) ||
1526             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1527                 cfqq = NULL;
1528                 goto keep_queue;
1529         }
1530
1531 expire:
1532         cfq_slice_expired(cfqd, 0);
1533 new_queue:
1534         /*
1535          * Current queue expired. Check if we have to switch to a new
1536          * service tree
1537          */
1538         if (!new_cfqq)
1539                 choose_service_tree(cfqd);
1540
1541         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1542 keep_queue:
1543         return cfqq;
1544 }
1545
1546 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1547 {
1548         int dispatched = 0;
1549
1550         while (cfqq->next_rq) {
1551                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1552                 dispatched++;
1553         }
1554
1555         BUG_ON(!list_empty(&cfqq->fifo));
1556         return dispatched;
1557 }
1558
1559 /*
1560  * Drain our current requests. Used for barriers and when switching
1561  * io schedulers on-the-fly.
1562  */
1563 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1564 {
1565         struct cfq_queue *cfqq;
1566         int dispatched = 0;
1567         int i, j;
1568         for (i = 0; i < 2; ++i)
1569                 for (j = 0; j < 3; ++j)
1570                         while ((cfqq = cfq_rb_first(&cfqd->service_trees[i][j]))
1571                                 != NULL)
1572                                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1573
1574         while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
1575                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1576
1577         cfq_slice_expired(cfqd, 0);
1578
1579         BUG_ON(cfqd->busy_queues);
1580
1581         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1582         return dispatched;
1583 }
1584
1585 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1586 {
1587         unsigned int max_dispatch;
1588
1589         /*
1590          * Drain async requests before we start sync IO
1591          */
1592         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1593                 return false;
1594
1595         /*
1596          * If this is an async queue and we have sync IO in flight, let it wait
1597          */
1598         if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1599                 return false;
1600
1601         max_dispatch = cfqd->cfq_quantum;
1602         if (cfq_class_idle(cfqq))
1603                 max_dispatch = 1;
1604
1605         /*
1606          * Does this cfqq already have too much IO in flight?
1607          */
1608         if (cfqq->dispatched >= max_dispatch) {
1609                 /*
1610                  * idle queue must always only have a single IO in flight
1611                  */
1612                 if (cfq_class_idle(cfqq))
1613                         return false;
1614
1615                 /*
1616                  * We have other queues, don't allow more IO from this one
1617                  */
1618                 if (cfqd->busy_queues > 1)
1619                         return false;
1620
1621                 /*
1622                  * Sole queue user, allow bigger slice
1623                  */
1624                 max_dispatch *= 4;
1625         }
1626
1627         /*
1628          * Async queues must wait a bit before being allowed dispatch.
1629          * We also ramp up the dispatch depth gradually for async IO,
1630          * based on the last sync IO we serviced
1631          */
1632         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1633                 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1634                 unsigned int depth;
1635
1636                 depth = last_sync / cfqd->cfq_slice[1];
1637                 if (!depth && !cfqq->dispatched)
1638                         depth = 1;
1639                 if (depth < max_dispatch)
1640                         max_dispatch = depth;
1641         }
1642
1643         /*
1644          * If we're below the current max, allow a dispatch
1645          */
1646         return cfqq->dispatched < max_dispatch;
1647 }
1648
1649 /*
1650  * Dispatch a request from cfqq, moving them to the request queue
1651  * dispatch list.
1652  */
1653 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1654 {
1655         struct request *rq;
1656
1657         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1658
1659         if (!cfq_may_dispatch(cfqd, cfqq))
1660                 return false;
1661
1662         /*
1663          * follow expired path, else get first next available
1664          */
1665         rq = cfq_check_fifo(cfqq);
1666         if (!rq)
1667                 rq = cfqq->next_rq;
1668
1669         /*
1670          * insert request into driver dispatch list
1671          */
1672         cfq_dispatch_insert(cfqd->queue, rq);
1673
1674         if (!cfqd->active_cic) {
1675                 struct cfq_io_context *cic = RQ_CIC(rq);
1676
1677                 atomic_long_inc(&cic->ioc->refcount);
1678                 cfqd->active_cic = cic;
1679         }
1680
1681         return true;
1682 }
1683
1684 /*
1685  * Find the cfqq that we need to service and move a request from that to the
1686  * dispatch list
1687  */
1688 static int cfq_dispatch_requests(struct request_queue *q, int force)
1689 {
1690         struct cfq_data *cfqd = q->elevator->elevator_data;
1691         struct cfq_queue *cfqq;
1692
1693         if (!cfqd->busy_queues)
1694                 return 0;
1695
1696         if (unlikely(force))
1697                 return cfq_forced_dispatch(cfqd);
1698
1699         cfqq = cfq_select_queue(cfqd);
1700         if (!cfqq)
1701                 return 0;
1702
1703         /*
1704          * Dispatch a request from this cfqq, if it is allowed
1705          */
1706         if (!cfq_dispatch_request(cfqd, cfqq))
1707                 return 0;
1708
1709         cfqq->slice_dispatch++;
1710         cfq_clear_cfqq_must_dispatch(cfqq);
1711
1712         /*
1713          * expire an async queue immediately if it has used up its slice. idle
1714          * queue always expire after 1 dispatch round.
1715          */
1716         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1717             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1718             cfq_class_idle(cfqq))) {
1719                 cfqq->slice_end = jiffies + 1;
1720                 cfq_slice_expired(cfqd, 0);
1721         }
1722
1723         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1724         return 1;
1725 }
1726
1727 /*
1728  * task holds one reference to the queue, dropped when task exits. each rq
1729  * in-flight on this queue also holds a reference, dropped when rq is freed.
1730  *
1731  * queue lock must be held here.
1732  */
1733 static void cfq_put_queue(struct cfq_queue *cfqq)
1734 {
1735         struct cfq_data *cfqd = cfqq->cfqd;
1736
1737         BUG_ON(atomic_read(&cfqq->ref) <= 0);
1738
1739         if (!atomic_dec_and_test(&cfqq->ref))
1740                 return;
1741
1742         cfq_log_cfqq(cfqd, cfqq, "put_queue");
1743         BUG_ON(rb_first(&cfqq->sort_list));
1744         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1745         BUG_ON(cfq_cfqq_on_rr(cfqq));
1746
1747         if (unlikely(cfqd->active_queue == cfqq)) {
1748                 __cfq_slice_expired(cfqd, cfqq, 0);
1749                 cfq_schedule_dispatch(cfqd);
1750         }
1751
1752         kmem_cache_free(cfq_pool, cfqq);
1753 }
1754
1755 /*
1756  * Must always be called with the rcu_read_lock() held
1757  */
1758 static void
1759 __call_for_each_cic(struct io_context *ioc,
1760                     void (*func)(struct io_context *, struct cfq_io_context *))
1761 {
1762         struct cfq_io_context *cic;
1763         struct hlist_node *n;
1764
1765         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1766                 func(ioc, cic);
1767 }
1768
1769 /*
1770  * Call func for each cic attached to this ioc.
1771  */
1772 static void
1773 call_for_each_cic(struct io_context *ioc,
1774                   void (*func)(struct io_context *, struct cfq_io_context *))
1775 {
1776         rcu_read_lock();
1777         __call_for_each_cic(ioc, func);
1778         rcu_read_unlock();
1779 }
1780
1781 static void cfq_cic_free_rcu(struct rcu_head *head)
1782 {
1783         struct cfq_io_context *cic;
1784
1785         cic = container_of(head, struct cfq_io_context, rcu_head);
1786
1787         kmem_cache_free(cfq_ioc_pool, cic);
1788         elv_ioc_count_dec(cfq_ioc_count);
1789
1790         if (ioc_gone) {
1791                 /*
1792                  * CFQ scheduler is exiting, grab exit lock and check
1793                  * the pending io context count. If it hits zero,
1794                  * complete ioc_gone and set it back to NULL
1795                  */
1796                 spin_lock(&ioc_gone_lock);
1797                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1798                         complete(ioc_gone);
1799                         ioc_gone = NULL;
1800                 }
1801                 spin_unlock(&ioc_gone_lock);
1802         }
1803 }
1804
1805 static void cfq_cic_free(struct cfq_io_context *cic)
1806 {
1807         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1808 }
1809
1810 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1811 {
1812         unsigned long flags;
1813
1814         BUG_ON(!cic->dead_key);
1815
1816         spin_lock_irqsave(&ioc->lock, flags);
1817         radix_tree_delete(&ioc->radix_root, cic->dead_key);
1818         hlist_del_rcu(&cic->cic_list);
1819         spin_unlock_irqrestore(&ioc->lock, flags);
1820
1821         cfq_cic_free(cic);
1822 }
1823
1824 /*
1825  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1826  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1827  * and ->trim() which is called with the task lock held
1828  */
1829 static void cfq_free_io_context(struct io_context *ioc)
1830 {
1831         /*
1832          * ioc->refcount is zero here, or we are called from elv_unregister(),
1833          * so no more cic's are allowed to be linked into this ioc.  So it
1834          * should be ok to iterate over the known list, we will see all cic's
1835          * since no new ones are added.
1836          */
1837         __call_for_each_cic(ioc, cic_free_func);
1838 }
1839
1840 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1841 {
1842         struct cfq_queue *__cfqq, *next;
1843
1844         if (unlikely(cfqq == cfqd->active_queue)) {
1845                 __cfq_slice_expired(cfqd, cfqq, 0);
1846                 cfq_schedule_dispatch(cfqd);
1847         }
1848
1849         /*
1850          * If this queue was scheduled to merge with another queue, be
1851          * sure to drop the reference taken on that queue (and others in
1852          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
1853          */
1854         __cfqq = cfqq->new_cfqq;
1855         while (__cfqq) {
1856                 if (__cfqq == cfqq) {
1857                         WARN(1, "cfqq->new_cfqq loop detected\n");
1858                         break;
1859                 }
1860                 next = __cfqq->new_cfqq;
1861                 cfq_put_queue(__cfqq);
1862                 __cfqq = next;
1863         }
1864
1865         cfq_put_queue(cfqq);
1866 }
1867
1868 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1869                                          struct cfq_io_context *cic)
1870 {
1871         struct io_context *ioc = cic->ioc;
1872
1873         list_del_init(&cic->queue_list);
1874
1875         /*
1876          * Make sure key == NULL is seen for dead queues
1877          */
1878         smp_wmb();
1879         cic->dead_key = (unsigned long) cic->key;
1880         cic->key = NULL;
1881
1882         if (ioc->ioc_data == cic)
1883                 rcu_assign_pointer(ioc->ioc_data, NULL);
1884
1885         if (cic->cfqq[BLK_RW_ASYNC]) {
1886                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1887                 cic->cfqq[BLK_RW_ASYNC] = NULL;
1888         }
1889
1890         if (cic->cfqq[BLK_RW_SYNC]) {
1891                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1892                 cic->cfqq[BLK_RW_SYNC] = NULL;
1893         }
1894 }
1895
1896 static void cfq_exit_single_io_context(struct io_context *ioc,
1897                                        struct cfq_io_context *cic)
1898 {
1899         struct cfq_data *cfqd = cic->key;
1900
1901         if (cfqd) {
1902                 struct request_queue *q = cfqd->queue;
1903                 unsigned long flags;
1904
1905                 spin_lock_irqsave(q->queue_lock, flags);
1906
1907                 /*
1908                  * Ensure we get a fresh copy of the ->key to prevent
1909                  * race between exiting task and queue
1910                  */
1911                 smp_read_barrier_depends();
1912                 if (cic->key)
1913                         __cfq_exit_single_io_context(cfqd, cic);
1914
1915                 spin_unlock_irqrestore(q->queue_lock, flags);
1916         }
1917 }
1918
1919 /*
1920  * The process that ioc belongs to has exited, we need to clean up
1921  * and put the internal structures we have that belongs to that process.
1922  */
1923 static void cfq_exit_io_context(struct io_context *ioc)
1924 {
1925         call_for_each_cic(ioc, cfq_exit_single_io_context);
1926 }
1927
1928 static struct cfq_io_context *
1929 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1930 {
1931         struct cfq_io_context *cic;
1932
1933         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1934                                                         cfqd->queue->node);
1935         if (cic) {
1936                 cic->last_end_request = jiffies;
1937                 INIT_LIST_HEAD(&cic->queue_list);
1938                 INIT_HLIST_NODE(&cic->cic_list);
1939                 cic->dtor = cfq_free_io_context;
1940                 cic->exit = cfq_exit_io_context;
1941                 elv_ioc_count_inc(cfq_ioc_count);
1942         }
1943
1944         return cic;
1945 }
1946
1947 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1948 {
1949         struct task_struct *tsk = current;
1950         int ioprio_class;
1951
1952         if (!cfq_cfqq_prio_changed(cfqq))
1953                 return;
1954
1955         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1956         switch (ioprio_class) {
1957         default:
1958                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1959         case IOPRIO_CLASS_NONE:
1960                 /*
1961                  * no prio set, inherit CPU scheduling settings
1962                  */
1963                 cfqq->ioprio = task_nice_ioprio(tsk);
1964                 cfqq->ioprio_class = task_nice_ioclass(tsk);
1965                 break;
1966         case IOPRIO_CLASS_RT:
1967                 cfqq->ioprio = task_ioprio(ioc);
1968                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1969                 break;
1970         case IOPRIO_CLASS_BE:
1971                 cfqq->ioprio = task_ioprio(ioc);
1972                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1973                 break;
1974         case IOPRIO_CLASS_IDLE:
1975                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1976                 cfqq->ioprio = 7;
1977                 cfq_clear_cfqq_idle_window(cfqq);
1978                 break;
1979         }
1980
1981         /*
1982          * keep track of original prio settings in case we have to temporarily
1983          * elevate the priority of this queue
1984          */
1985         cfqq->org_ioprio = cfqq->ioprio;
1986         cfqq->org_ioprio_class = cfqq->ioprio_class;
1987         cfq_clear_cfqq_prio_changed(cfqq);
1988 }
1989
1990 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
1991 {
1992         struct cfq_data *cfqd = cic->key;
1993         struct cfq_queue *cfqq;
1994         unsigned long flags;
1995
1996         if (unlikely(!cfqd))
1997                 return;
1998
1999         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2000
2001         cfqq = cic->cfqq[BLK_RW_ASYNC];
2002         if (cfqq) {
2003                 struct cfq_queue *new_cfqq;
2004                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2005                                                 GFP_ATOMIC);
2006                 if (new_cfqq) {
2007                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2008                         cfq_put_queue(cfqq);
2009                 }
2010         }
2011
2012         cfqq = cic->cfqq[BLK_RW_SYNC];
2013         if (cfqq)
2014                 cfq_mark_cfqq_prio_changed(cfqq);
2015
2016         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2017 }
2018
2019 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2020 {
2021         call_for_each_cic(ioc, changed_ioprio);
2022         ioc->ioprio_changed = 0;
2023 }
2024
2025 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2026                           pid_t pid, bool is_sync)
2027 {
2028         RB_CLEAR_NODE(&cfqq->rb_node);
2029         RB_CLEAR_NODE(&cfqq->p_node);
2030         INIT_LIST_HEAD(&cfqq->fifo);
2031
2032         atomic_set(&cfqq->ref, 0);
2033         cfqq->cfqd = cfqd;
2034
2035         cfq_mark_cfqq_prio_changed(cfqq);
2036
2037         if (is_sync) {
2038                 if (!cfq_class_idle(cfqq))
2039                         cfq_mark_cfqq_idle_window(cfqq);
2040                 cfq_mark_cfqq_sync(cfqq);
2041         }
2042         cfqq->pid = pid;
2043 }
2044
2045 static struct cfq_queue *
2046 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2047                      struct io_context *ioc, gfp_t gfp_mask)
2048 {
2049         struct cfq_queue *cfqq, *new_cfqq = NULL;
2050         struct cfq_io_context *cic;
2051
2052 retry:
2053         cic = cfq_cic_lookup(cfqd, ioc);
2054         /* cic always exists here */
2055         cfqq = cic_to_cfqq(cic, is_sync);
2056
2057         /*
2058          * Always try a new alloc if we fell back to the OOM cfqq
2059          * originally, since it should just be a temporary situation.
2060          */
2061         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2062                 cfqq = NULL;
2063                 if (new_cfqq) {
2064                         cfqq = new_cfqq;
2065                         new_cfqq = NULL;
2066                 } else if (gfp_mask & __GFP_WAIT) {
2067                         spin_unlock_irq(cfqd->queue->queue_lock);
2068                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2069                                         gfp_mask | __GFP_ZERO,
2070                                         cfqd->queue->node);
2071                         spin_lock_irq(cfqd->queue->queue_lock);
2072                         if (new_cfqq)
2073                                 goto retry;
2074                 } else {
2075                         cfqq = kmem_cache_alloc_node(cfq_pool,
2076                                         gfp_mask | __GFP_ZERO,
2077                                         cfqd->queue->node);
2078                 }
2079
2080                 if (cfqq) {
2081                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2082                         cfq_init_prio_data(cfqq, ioc);
2083                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2084                 } else
2085                         cfqq = &cfqd->oom_cfqq;
2086         }
2087
2088         if (new_cfqq)
2089                 kmem_cache_free(cfq_pool, new_cfqq);
2090
2091         return cfqq;
2092 }
2093
2094 static struct cfq_queue **
2095 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2096 {
2097         switch (ioprio_class) {
2098         case IOPRIO_CLASS_RT:
2099                 return &cfqd->async_cfqq[0][ioprio];
2100         case IOPRIO_CLASS_BE:
2101                 return &cfqd->async_cfqq[1][ioprio];
2102         case IOPRIO_CLASS_IDLE:
2103                 return &cfqd->async_idle_cfqq;
2104         default:
2105                 BUG();
2106         }
2107 }
2108
2109 static struct cfq_queue *
2110 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2111               gfp_t gfp_mask)
2112 {
2113         const int ioprio = task_ioprio(ioc);
2114         const int ioprio_class = task_ioprio_class(ioc);
2115         struct cfq_queue **async_cfqq = NULL;
2116         struct cfq_queue *cfqq = NULL;
2117
2118         if (!is_sync) {
2119                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2120                 cfqq = *async_cfqq;
2121         }
2122
2123         if (!cfqq)
2124                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2125
2126         /*
2127          * pin the queue now that it's allocated, scheduler exit will prune it
2128          */
2129         if (!is_sync && !(*async_cfqq)) {
2130                 atomic_inc(&cfqq->ref);
2131                 *async_cfqq = cfqq;
2132         }
2133
2134         atomic_inc(&cfqq->ref);
2135         return cfqq;
2136 }
2137
2138 /*
2139  * We drop cfq io contexts lazily, so we may find a dead one.
2140  */
2141 static void
2142 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2143                   struct cfq_io_context *cic)
2144 {
2145         unsigned long flags;
2146
2147         WARN_ON(!list_empty(&cic->queue_list));
2148
2149         spin_lock_irqsave(&ioc->lock, flags);
2150
2151         BUG_ON(ioc->ioc_data == cic);
2152
2153         radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2154         hlist_del_rcu(&cic->cic_list);
2155         spin_unlock_irqrestore(&ioc->lock, flags);
2156
2157         cfq_cic_free(cic);
2158 }
2159
2160 static struct cfq_io_context *
2161 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2162 {
2163         struct cfq_io_context *cic;
2164         unsigned long flags;
2165         void *k;
2166
2167         if (unlikely(!ioc))
2168                 return NULL;
2169
2170         rcu_read_lock();
2171
2172         /*
2173          * we maintain a last-hit cache, to avoid browsing over the tree
2174          */
2175         cic = rcu_dereference(ioc->ioc_data);
2176         if (cic && cic->key == cfqd) {
2177                 rcu_read_unlock();
2178                 return cic;
2179         }
2180
2181         do {
2182                 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2183                 rcu_read_unlock();
2184                 if (!cic)
2185                         break;
2186                 /* ->key must be copied to avoid race with cfq_exit_queue() */
2187                 k = cic->key;
2188                 if (unlikely(!k)) {
2189                         cfq_drop_dead_cic(cfqd, ioc, cic);
2190                         rcu_read_lock();
2191                         continue;
2192                 }
2193
2194                 spin_lock_irqsave(&ioc->lock, flags);
2195                 rcu_assign_pointer(ioc->ioc_data, cic);
2196                 spin_unlock_irqrestore(&ioc->lock, flags);
2197                 break;
2198         } while (1);
2199
2200         return cic;
2201 }
2202
2203 /*
2204  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2205  * the process specific cfq io context when entered from the block layer.
2206  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2207  */
2208 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2209                         struct cfq_io_context *cic, gfp_t gfp_mask)
2210 {
2211         unsigned long flags;
2212         int ret;
2213
2214         ret = radix_tree_preload(gfp_mask);
2215         if (!ret) {
2216                 cic->ioc = ioc;
2217                 cic->key = cfqd;
2218
2219                 spin_lock_irqsave(&ioc->lock, flags);
2220                 ret = radix_tree_insert(&ioc->radix_root,
2221                                                 (unsigned long) cfqd, cic);
2222                 if (!ret)
2223                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2224                 spin_unlock_irqrestore(&ioc->lock, flags);
2225
2226                 radix_tree_preload_end();
2227
2228                 if (!ret) {
2229                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2230                         list_add(&cic->queue_list, &cfqd->cic_list);
2231                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2232                 }
2233         }
2234
2235         if (ret)
2236                 printk(KERN_ERR "cfq: cic link failed!\n");
2237
2238         return ret;
2239 }
2240
2241 /*
2242  * Setup general io context and cfq io context. There can be several cfq
2243  * io contexts per general io context, if this process is doing io to more
2244  * than one device managed by cfq.
2245  */
2246 static struct cfq_io_context *
2247 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2248 {
2249         struct io_context *ioc = NULL;
2250         struct cfq_io_context *cic;
2251
2252         might_sleep_if(gfp_mask & __GFP_WAIT);
2253
2254         ioc = get_io_context(gfp_mask, cfqd->queue->node);
2255         if (!ioc)
2256                 return NULL;
2257
2258         cic = cfq_cic_lookup(cfqd, ioc);
2259         if (cic)
2260                 goto out;
2261
2262         cic = cfq_alloc_io_context(cfqd, gfp_mask);
2263         if (cic == NULL)
2264                 goto err;
2265
2266         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2267                 goto err_free;
2268
2269 out:
2270         smp_read_barrier_depends();
2271         if (unlikely(ioc->ioprio_changed))
2272                 cfq_ioc_set_ioprio(ioc);
2273
2274         return cic;
2275 err_free:
2276         cfq_cic_free(cic);
2277 err:
2278         put_io_context(ioc);
2279         return NULL;
2280 }
2281
2282 static void
2283 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2284 {
2285         unsigned long elapsed = jiffies - cic->last_end_request;
2286         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2287
2288         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2289         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2290         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2291 }
2292
2293 static void
2294 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2295                        struct request *rq)
2296 {
2297         sector_t sdist;
2298         u64 total;
2299
2300         if (!cfqq->last_request_pos)
2301                 sdist = 0;
2302         else if (cfqq->last_request_pos < blk_rq_pos(rq))
2303                 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2304         else
2305                 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2306
2307         /*
2308          * Don't allow the seek distance to get too large from the
2309          * odd fragment, pagein, etc
2310          */
2311         if (cfqq->seek_samples <= 60) /* second&third seek */
2312                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2313         else
2314                 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2315
2316         cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2317         cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2318         total = cfqq->seek_total + (cfqq->seek_samples/2);
2319         do_div(total, cfqq->seek_samples);
2320         cfqq->seek_mean = (sector_t)total;
2321
2322         /*
2323          * If this cfqq is shared between multiple processes, check to
2324          * make sure that those processes are still issuing I/Os within
2325          * the mean seek distance.  If not, it may be time to break the
2326          * queues apart again.
2327          */
2328         if (cfq_cfqq_coop(cfqq)) {
2329                 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2330                         cfqq->seeky_start = jiffies;
2331                 else if (!CFQQ_SEEKY(cfqq))
2332                         cfqq->seeky_start = 0;
2333         }
2334 }
2335
2336 /*
2337  * Disable idle window if the process thinks too long or seeks so much that
2338  * it doesn't matter
2339  */
2340 static void
2341 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2342                        struct cfq_io_context *cic)
2343 {
2344         int old_idle, enable_idle;
2345
2346         /*
2347          * Don't idle for async or idle io prio class
2348          */
2349         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2350                 return;
2351
2352         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2353
2354         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2355             (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
2356                 enable_idle = 0;
2357         else if (sample_valid(cic->ttime_samples)) {
2358                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2359                         enable_idle = 0;
2360                 else
2361                         enable_idle = 1;
2362         }
2363
2364         if (old_idle != enable_idle) {
2365                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2366                 if (enable_idle)
2367                         cfq_mark_cfqq_idle_window(cfqq);
2368                 else
2369                         cfq_clear_cfqq_idle_window(cfqq);
2370         }
2371 }
2372
2373 /*
2374  * Check if new_cfqq should preempt the currently active queue. Return 0 for
2375  * no or if we aren't sure, a 1 will cause a preempt.
2376  */
2377 static bool
2378 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2379                    struct request *rq)
2380 {
2381         struct cfq_queue *cfqq;
2382
2383         cfqq = cfqd->active_queue;
2384         if (!cfqq)
2385                 return false;
2386
2387         if (cfq_slice_used(cfqq))
2388                 return true;
2389
2390         if (cfq_class_idle(new_cfqq))
2391                 return false;
2392
2393         if (cfq_class_idle(cfqq))
2394                 return true;
2395
2396         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD
2397             && new_cfqq->service_tree == cfqq->service_tree)
2398                 return true;
2399
2400         /*
2401          * if the new request is sync, but the currently running queue is
2402          * not, let the sync request have priority.
2403          */
2404         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2405                 return true;
2406
2407         /*
2408          * So both queues are sync. Let the new request get disk time if
2409          * it's a metadata request and the current queue is doing regular IO.
2410          */
2411         if (rq_is_meta(rq) && !cfqq->meta_pending)
2412                 return false;
2413
2414         /*
2415          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2416          */
2417         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2418                 return true;
2419
2420         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2421                 return false;
2422
2423         /*
2424          * if this request is as-good as one we would expect from the
2425          * current cfqq, let it preempt
2426          */
2427         if (cfq_rq_close(cfqd, cfqq, rq))
2428                 return true;
2429
2430         return false;
2431 }
2432
2433 /*
2434  * cfqq preempts the active queue. if we allowed preempt with no slice left,
2435  * let it have half of its nominal slice.
2436  */
2437 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2438 {
2439         cfq_log_cfqq(cfqd, cfqq, "preempt");
2440         cfq_slice_expired(cfqd, 1);
2441
2442         /*
2443          * Put the new queue at the front of the of the current list,
2444          * so we know that it will be selected next.
2445          */
2446         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2447
2448         cfq_service_tree_add(cfqd, cfqq, 1);
2449
2450         cfqq->slice_end = 0;
2451         cfq_mark_cfqq_slice_new(cfqq);
2452 }
2453
2454 /*
2455  * Called when a new fs request (rq) is added (to cfqq). Check if there's
2456  * something we should do about it
2457  */
2458 static void
2459 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2460                 struct request *rq)
2461 {
2462         struct cfq_io_context *cic = RQ_CIC(rq);
2463
2464         cfqd->rq_queued++;
2465         if (rq_is_meta(rq))
2466                 cfqq->meta_pending++;
2467
2468         cfq_update_io_thinktime(cfqd, cic);
2469         cfq_update_io_seektime(cfqd, cfqq, rq);
2470         cfq_update_idle_window(cfqd, cfqq, cic);
2471
2472         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2473
2474         if (cfqq == cfqd->active_queue) {
2475                 /*
2476                  * Remember that we saw a request from this process, but
2477                  * don't start queuing just yet. Otherwise we risk seeing lots
2478                  * of tiny requests, because we disrupt the normal plugging
2479                  * and merging. If the request is already larger than a single
2480                  * page, let it rip immediately. For that case we assume that
2481                  * merging is already done. Ditto for a busy system that
2482                  * has other work pending, don't risk delaying until the
2483                  * idle timer unplug to continue working.
2484                  */
2485                 if (cfq_cfqq_wait_request(cfqq)) {
2486                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2487                             cfqd->busy_queues > 1) {
2488                                 del_timer(&cfqd->idle_slice_timer);
2489                         __blk_run_queue(cfqd->queue);
2490                         }
2491                         cfq_mark_cfqq_must_dispatch(cfqq);
2492                 }
2493         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2494                 /*
2495                  * not the active queue - expire current slice if it is
2496                  * idle and has expired it's mean thinktime or this new queue
2497                  * has some old slice time left and is of higher priority or
2498                  * this new queue is RT and the current one is BE
2499                  */
2500                 cfq_preempt_queue(cfqd, cfqq);
2501                 __blk_run_queue(cfqd->queue);
2502         }
2503 }
2504
2505 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2506 {
2507         struct cfq_data *cfqd = q->elevator->elevator_data;
2508         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2509
2510         cfq_log_cfqq(cfqd, cfqq, "insert_request");
2511         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2512
2513         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2514         list_add_tail(&rq->queuelist, &cfqq->fifo);
2515         cfq_add_rq_rb(rq);
2516
2517         cfq_rq_enqueued(cfqd, cfqq, rq);
2518 }
2519
2520 /*
2521  * Update hw_tag based on peak queue depth over 50 samples under
2522  * sufficient load.
2523  */
2524 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2525 {
2526         struct cfq_queue *cfqq = cfqd->active_queue;
2527
2528         if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2529                 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2530
2531         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2532             rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2533                 return;
2534
2535         /*
2536          * If active queue hasn't enough requests and can idle, cfq might not
2537          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2538          * case
2539          */
2540         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2541             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2542             CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2543                 return;
2544
2545         if (cfqd->hw_tag_samples++ < 50)
2546                 return;
2547
2548         if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2549                 cfqd->hw_tag = 1;
2550         else
2551                 cfqd->hw_tag = 0;
2552
2553         cfqd->hw_tag_samples = 0;
2554         cfqd->rq_in_driver_peak = 0;
2555 }
2556
2557 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2558 {
2559         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2560         struct cfq_data *cfqd = cfqq->cfqd;
2561         const int sync = rq_is_sync(rq);
2562         unsigned long now;
2563
2564         now = jiffies;
2565         cfq_log_cfqq(cfqd, cfqq, "complete");
2566
2567         cfq_update_hw_tag(cfqd);
2568
2569         WARN_ON(!cfqd->rq_in_driver[sync]);
2570         WARN_ON(!cfqq->dispatched);
2571         cfqd->rq_in_driver[sync]--;
2572         cfqq->dispatched--;
2573
2574         if (cfq_cfqq_sync(cfqq))
2575                 cfqd->sync_flight--;
2576
2577         if (sync) {
2578                 RQ_CIC(rq)->last_end_request = now;
2579                 cfqd->last_end_sync_rq = now;
2580         }
2581
2582         /*
2583          * If this is the active queue, check if it needs to be expired,
2584          * or if we want to idle in case it has no pending requests.
2585          */
2586         if (cfqd->active_queue == cfqq) {
2587                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2588
2589                 if (cfq_cfqq_slice_new(cfqq)) {
2590                         cfq_set_prio_slice(cfqd, cfqq);
2591                         cfq_clear_cfqq_slice_new(cfqq);
2592                 }
2593                 /*
2594                  * If there are no requests waiting in this queue, and
2595                  * there are other queues ready to issue requests, AND
2596                  * those other queues are issuing requests within our
2597                  * mean seek distance, give them a chance to run instead
2598                  * of idling.
2599                  */
2600                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2601                         cfq_slice_expired(cfqd, 1);
2602                 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2603                          sync && !rq_noidle(rq))
2604                         cfq_arm_slice_timer(cfqd);
2605         }
2606
2607         if (!rq_in_driver(cfqd))
2608                 cfq_schedule_dispatch(cfqd);
2609 }
2610
2611 /*
2612  * we temporarily boost lower priority queues if they are holding fs exclusive
2613  * resources. they are boosted to normal prio (CLASS_BE/4)
2614  */
2615 static void cfq_prio_boost(struct cfq_queue *cfqq)
2616 {
2617         if (has_fs_excl()) {
2618                 /*
2619                  * boost idle prio on transactions that would lock out other
2620                  * users of the filesystem
2621                  */
2622                 if (cfq_class_idle(cfqq))
2623                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
2624                 if (cfqq->ioprio > IOPRIO_NORM)
2625                         cfqq->ioprio = IOPRIO_NORM;
2626         } else {
2627                 /*
2628                  * unboost the queue (if needed)
2629                  */
2630                 cfqq->ioprio_class = cfqq->org_ioprio_class;
2631                 cfqq->ioprio = cfqq->org_ioprio;
2632         }
2633 }
2634
2635 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2636 {
2637         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2638                 cfq_mark_cfqq_must_alloc_slice(cfqq);
2639                 return ELV_MQUEUE_MUST;
2640         }
2641
2642         return ELV_MQUEUE_MAY;
2643 }
2644
2645 static int cfq_may_queue(struct request_queue *q, int rw)
2646 {
2647         struct cfq_data *cfqd = q->elevator->elevator_data;
2648         struct task_struct *tsk = current;
2649         struct cfq_io_context *cic;
2650         struct cfq_queue *cfqq;
2651
2652         /*
2653          * don't force setup of a queue from here, as a call to may_queue
2654          * does not necessarily imply that a request actually will be queued.
2655          * so just lookup a possibly existing queue, or return 'may queue'
2656          * if that fails
2657          */
2658         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2659         if (!cic)
2660                 return ELV_MQUEUE_MAY;
2661
2662         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2663         if (cfqq) {
2664                 cfq_init_prio_data(cfqq, cic->ioc);
2665                 cfq_prio_boost(cfqq);
2666
2667                 return __cfq_may_queue(cfqq);
2668         }
2669
2670         return ELV_MQUEUE_MAY;
2671 }
2672
2673 /*
2674  * queue lock held here
2675  */
2676 static void cfq_put_request(struct request *rq)
2677 {
2678         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2679
2680         if (cfqq) {
2681                 const int rw = rq_data_dir(rq);
2682
2683                 BUG_ON(!cfqq->allocated[rw]);
2684                 cfqq->allocated[rw]--;
2685
2686                 put_io_context(RQ_CIC(rq)->ioc);
2687
2688                 rq->elevator_private = NULL;
2689                 rq->elevator_private2 = NULL;
2690
2691                 cfq_put_queue(cfqq);
2692         }
2693 }
2694
2695 static struct cfq_queue *
2696 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2697                 struct cfq_queue *cfqq)
2698 {
2699         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2700         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2701         cfq_mark_cfqq_coop(cfqq->new_cfqq);
2702         cfq_put_queue(cfqq);
2703         return cic_to_cfqq(cic, 1);
2704 }
2705
2706 static int should_split_cfqq(struct cfq_queue *cfqq)
2707 {
2708         if (cfqq->seeky_start &&
2709             time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2710                 return 1;
2711         return 0;
2712 }
2713
2714 /*
2715  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2716  * was the last process referring to said cfqq.
2717  */
2718 static struct cfq_queue *
2719 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2720 {
2721         if (cfqq_process_refs(cfqq) == 1) {
2722                 cfqq->seeky_start = 0;
2723                 cfqq->pid = current->pid;
2724                 cfq_clear_cfqq_coop(cfqq);
2725                 return cfqq;
2726         }
2727
2728         cic_set_cfqq(cic, NULL, 1);
2729         cfq_put_queue(cfqq);
2730         return NULL;
2731 }
2732 /*
2733  * Allocate cfq data structures associated with this request.
2734  */
2735 static int
2736 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2737 {
2738         struct cfq_data *cfqd = q->elevator->elevator_data;
2739         struct cfq_io_context *cic;
2740         const int rw = rq_data_dir(rq);
2741         const bool is_sync = rq_is_sync(rq);
2742         struct cfq_queue *cfqq;
2743         unsigned long flags;
2744
2745         might_sleep_if(gfp_mask & __GFP_WAIT);
2746
2747         cic = cfq_get_io_context(cfqd, gfp_mask);
2748
2749         spin_lock_irqsave(q->queue_lock, flags);
2750
2751         if (!cic)
2752                 goto queue_fail;
2753
2754 new_queue:
2755         cfqq = cic_to_cfqq(cic, is_sync);
2756         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2757                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2758                 cic_set_cfqq(cic, cfqq, is_sync);
2759         } else {
2760                 /*
2761                  * If the queue was seeky for too long, break it apart.
2762                  */
2763                 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2764                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2765                         cfqq = split_cfqq(cic, cfqq);
2766                         if (!cfqq)
2767                                 goto new_queue;
2768                 }
2769
2770                 /*
2771                  * Check to see if this queue is scheduled to merge with
2772                  * another, closely cooperating queue.  The merging of
2773                  * queues happens here as it must be done in process context.
2774                  * The reference on new_cfqq was taken in merge_cfqqs.
2775                  */
2776                 if (cfqq->new_cfqq)
2777                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2778         }
2779
2780         cfqq->allocated[rw]++;
2781         atomic_inc(&cfqq->ref);
2782
2783         spin_unlock_irqrestore(q->queue_lock, flags);
2784
2785         rq->elevator_private = cic;
2786         rq->elevator_private2 = cfqq;
2787         return 0;
2788
2789 queue_fail:
2790         if (cic)
2791                 put_io_context(cic->ioc);
2792
2793         cfq_schedule_dispatch(cfqd);
2794         spin_unlock_irqrestore(q->queue_lock, flags);
2795         cfq_log(cfqd, "set_request fail");
2796         return 1;
2797 }
2798
2799 static void cfq_kick_queue(struct work_struct *work)
2800 {
2801         struct cfq_data *cfqd =
2802                 container_of(work, struct cfq_data, unplug_work);
2803         struct request_queue *q = cfqd->queue;
2804
2805         spin_lock_irq(q->queue_lock);
2806         __blk_run_queue(cfqd->queue);
2807         spin_unlock_irq(q->queue_lock);
2808 }
2809
2810 /*
2811  * Timer running if the active_queue is currently idling inside its time slice
2812  */
2813 static void cfq_idle_slice_timer(unsigned long data)
2814 {
2815         struct cfq_data *cfqd = (struct cfq_data *) data;
2816         struct cfq_queue *cfqq;
2817         unsigned long flags;
2818         int timed_out = 1;
2819
2820         cfq_log(cfqd, "idle timer fired");
2821
2822         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2823
2824         cfqq = cfqd->active_queue;
2825         if (cfqq) {
2826                 timed_out = 0;
2827
2828                 /*
2829                  * We saw a request before the queue expired, let it through
2830                  */
2831                 if (cfq_cfqq_must_dispatch(cfqq))
2832                         goto out_kick;
2833
2834                 /*
2835                  * expired
2836                  */
2837                 if (cfq_slice_used(cfqq))
2838                         goto expire;
2839
2840                 /*
2841                  * only expire and reinvoke request handler, if there are
2842                  * other queues with pending requests
2843                  */
2844                 if (!cfqd->busy_queues)
2845                         goto out_cont;
2846
2847                 /*
2848                  * not expired and it has a request pending, let it dispatch
2849                  */
2850                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2851                         goto out_kick;
2852         }
2853 expire:
2854         cfq_slice_expired(cfqd, timed_out);
2855 out_kick:
2856         cfq_schedule_dispatch(cfqd);
2857 out_cont:
2858         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2859 }
2860
2861 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2862 {
2863         del_timer_sync(&cfqd->idle_slice_timer);
2864         cancel_work_sync(&cfqd->unplug_work);
2865 }
2866
2867 static void cfq_put_async_queues(struct cfq_data *cfqd)
2868 {
2869         int i;
2870
2871         for (i = 0; i < IOPRIO_BE_NR; i++) {
2872                 if (cfqd->async_cfqq[0][i])
2873                         cfq_put_queue(cfqd->async_cfqq[0][i]);
2874                 if (cfqd->async_cfqq[1][i])
2875                         cfq_put_queue(cfqd->async_cfqq[1][i]);
2876         }
2877
2878         if (cfqd->async_idle_cfqq)
2879                 cfq_put_queue(cfqd->async_idle_cfqq);
2880 }
2881
2882 static void cfq_exit_queue(struct elevator_queue *e)
2883 {
2884         struct cfq_data *cfqd = e->elevator_data;
2885         struct request_queue *q = cfqd->queue;
2886
2887         cfq_shutdown_timer_wq(cfqd);
2888
2889         spin_lock_irq(q->queue_lock);
2890
2891         if (cfqd->active_queue)
2892                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2893
2894         while (!list_empty(&cfqd->cic_list)) {
2895                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2896                                                         struct cfq_io_context,
2897                                                         queue_list);
2898
2899                 __cfq_exit_single_io_context(cfqd, cic);
2900         }
2901
2902         cfq_put_async_queues(cfqd);
2903
2904         spin_unlock_irq(q->queue_lock);
2905
2906         cfq_shutdown_timer_wq(cfqd);
2907
2908         kfree(cfqd);
2909 }
2910
2911 static void *cfq_init_queue(struct request_queue *q)
2912 {
2913         struct cfq_data *cfqd;
2914         int i, j;
2915
2916         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2917         if (!cfqd)
2918                 return NULL;
2919
2920         for (i = 0; i < 2; ++i)
2921                 for (j = 0; j < 3; ++j)
2922                         cfqd->service_trees[i][j] = CFQ_RB_ROOT;
2923         cfqd->service_tree_idle = CFQ_RB_ROOT;
2924
2925         /*
2926          * Not strictly needed (since RB_ROOT just clears the node and we
2927          * zeroed cfqd on alloc), but better be safe in case someone decides
2928          * to add magic to the rb code
2929          */
2930         for (i = 0; i < CFQ_PRIO_LISTS; i++)
2931                 cfqd->prio_trees[i] = RB_ROOT;
2932
2933         /*
2934          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2935          * Grab a permanent reference to it, so that the normal code flow
2936          * will not attempt to free it.
2937          */
2938         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2939         atomic_inc(&cfqd->oom_cfqq.ref);
2940
2941         INIT_LIST_HEAD(&cfqd->cic_list);
2942
2943         cfqd->queue = q;
2944
2945         init_timer(&cfqd->idle_slice_timer);
2946         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2947         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2948
2949         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2950
2951         cfqd->cfq_quantum = cfq_quantum;
2952         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2953         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2954         cfqd->cfq_back_max = cfq_back_max;
2955         cfqd->cfq_back_penalty = cfq_back_penalty;
2956         cfqd->cfq_slice[0] = cfq_slice_async;
2957         cfqd->cfq_slice[1] = cfq_slice_sync;
2958         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2959         cfqd->cfq_slice_idle = cfq_slice_idle;
2960         cfqd->cfq_latency = 1;
2961         cfqd->hw_tag = 1;
2962         cfqd->last_end_sync_rq = jiffies;
2963         return cfqd;
2964 }
2965
2966 static void cfq_slab_kill(void)
2967 {
2968         /*
2969          * Caller already ensured that pending RCU callbacks are completed,
2970          * so we should have no busy allocations at this point.
2971          */
2972         if (cfq_pool)
2973                 kmem_cache_destroy(cfq_pool);
2974         if (cfq_ioc_pool)
2975                 kmem_cache_destroy(cfq_ioc_pool);
2976 }
2977
2978 static int __init cfq_slab_setup(void)
2979 {
2980         cfq_pool = KMEM_CACHE(cfq_queue, 0);
2981         if (!cfq_pool)
2982                 goto fail;
2983
2984         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
2985         if (!cfq_ioc_pool)
2986                 goto fail;
2987
2988         return 0;
2989 fail:
2990         cfq_slab_kill();
2991         return -ENOMEM;
2992 }
2993
2994 /*
2995  * sysfs parts below -->
2996  */
2997 static ssize_t
2998 cfq_var_show(unsigned int var, char *page)
2999 {
3000         return sprintf(page, "%d\n", var);
3001 }
3002
3003 static ssize_t
3004 cfq_var_store(unsigned int *var, const char *page, size_t count)
3005 {
3006         char *p = (char *) page;
3007
3008         *var = simple_strtoul(p, &p, 10);
3009         return count;
3010 }
3011
3012 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3013 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3014 {                                                                       \
3015         struct cfq_data *cfqd = e->elevator_data;                       \
3016         unsigned int __data = __VAR;                                    \
3017         if (__CONV)                                                     \
3018                 __data = jiffies_to_msecs(__data);                      \
3019         return cfq_var_show(__data, (page));                            \
3020 }
3021 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3022 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3023 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3024 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3025 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3026 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3027 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3028 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3029 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3030 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3031 #undef SHOW_FUNCTION
3032
3033 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3034 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3035 {                                                                       \
3036         struct cfq_data *cfqd = e->elevator_data;                       \
3037         unsigned int __data;                                            \
3038         int ret = cfq_var_store(&__data, (page), count);                \
3039         if (__data < (MIN))                                             \
3040                 __data = (MIN);                                         \
3041         else if (__data > (MAX))                                        \
3042                 __data = (MAX);                                         \
3043         if (__CONV)                                                     \
3044                 *(__PTR) = msecs_to_jiffies(__data);                    \
3045         else                                                            \
3046                 *(__PTR) = __data;                                      \
3047         return ret;                                                     \
3048 }
3049 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3050 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3051                 UINT_MAX, 1);
3052 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3053                 UINT_MAX, 1);
3054 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3055 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3056                 UINT_MAX, 0);
3057 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3058 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3059 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3060 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3061                 UINT_MAX, 0);
3062 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3063 #undef STORE_FUNCTION
3064
3065 #define CFQ_ATTR(name) \
3066         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3067
3068 static struct elv_fs_entry cfq_attrs[] = {
3069         CFQ_ATTR(quantum),
3070         CFQ_ATTR(fifo_expire_sync),
3071         CFQ_ATTR(fifo_expire_async),
3072         CFQ_ATTR(back_seek_max),
3073         CFQ_ATTR(back_seek_penalty),
3074         CFQ_ATTR(slice_sync),
3075         CFQ_ATTR(slice_async),
3076         CFQ_ATTR(slice_async_rq),
3077         CFQ_ATTR(slice_idle),
3078         CFQ_ATTR(low_latency),
3079         __ATTR_NULL
3080 };
3081
3082 static struct elevator_type iosched_cfq = {
3083         .ops = {
3084                 .elevator_merge_fn =            cfq_merge,
3085                 .elevator_merged_fn =           cfq_merged_request,
3086                 .elevator_merge_req_fn =        cfq_merged_requests,
3087                 .elevator_allow_merge_fn =      cfq_allow_merge,
3088                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3089                 .elevator_add_req_fn =          cfq_insert_request,
3090                 .elevator_activate_req_fn =     cfq_activate_request,
3091                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3092                 .elevator_queue_empty_fn =      cfq_queue_empty,
3093                 .elevator_completed_req_fn =    cfq_completed_request,
3094                 .elevator_former_req_fn =       elv_rb_former_request,
3095                 .elevator_latter_req_fn =       elv_rb_latter_request,
3096                 .elevator_set_req_fn =          cfq_set_request,
3097                 .elevator_put_req_fn =          cfq_put_request,
3098                 .elevator_may_queue_fn =        cfq_may_queue,
3099                 .elevator_init_fn =             cfq_init_queue,
3100                 .elevator_exit_fn =             cfq_exit_queue,
3101                 .trim =                         cfq_free_io_context,
3102         },
3103         .elevator_attrs =       cfq_attrs,
3104         .elevator_name =        "cfq",
3105         .elevator_owner =       THIS_MODULE,
3106 };
3107
3108 static int __init cfq_init(void)
3109 {
3110         /*
3111          * could be 0 on HZ < 1000 setups
3112          */
3113         if (!cfq_slice_async)
3114                 cfq_slice_async = 1;
3115         if (!cfq_slice_idle)
3116                 cfq_slice_idle = 1;
3117
3118         if (cfq_slab_setup())
3119                 return -ENOMEM;
3120
3121         elv_register(&iosched_cfq);
3122
3123         return 0;
3124 }
3125
3126 static void __exit cfq_exit(void)
3127 {
3128         DECLARE_COMPLETION_ONSTACK(all_gone);
3129         elv_unregister(&iosched_cfq);
3130         ioc_gone = &all_gone;
3131         /* ioc_gone's update must be visible before reading ioc_count */
3132         smp_wmb();
3133
3134         /*
3135          * this also protects us from entering cfq_slab_kill() with
3136          * pending RCU callbacks
3137          */
3138         if (elv_ioc_count_read(cfq_ioc_count))
3139                 wait_for_completion(&all_gone);
3140         cfq_slab_kill();
3141 }
3142
3143 module_init(cfq_init);
3144 module_exit(cfq_exit);
3145
3146 MODULE_AUTHOR("Jens Axboe");
3147 MODULE_LICENSE("GPL");
3148 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");