]>
Commit | Line | Data |
---|---|---|
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/slab.h> | |
11 | #include <linux/blkdev.h> | |
12 | #include <linux/elevator.h> | |
13 | #include <linux/jiffies.h> | |
14 | #include <linux/rbtree.h> | |
15 | #include <linux/ioprio.h> | |
16 | #include <linux/blktrace_api.h> | |
17 | #include "blk.h" | |
18 | #include "blk-cgroup.h" | |
19 | ||
20 | static struct blkio_policy_type blkio_policy_cfq; | |
21 | ||
22 | /* | |
23 | * tunables | |
24 | */ | |
25 | /* max queue in one round of service */ | |
26 | static const int cfq_quantum = 8; | |
27 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; | |
28 | /* maximum backwards seek, in KiB */ | |
29 | static const int cfq_back_max = 16 * 1024; | |
30 | /* penalty of a backwards seek */ | |
31 | static const int cfq_back_penalty = 2; | |
32 | static const int cfq_slice_sync = HZ / 10; | |
33 | static int cfq_slice_async = HZ / 25; | |
34 | static const int cfq_slice_async_rq = 2; | |
35 | static int cfq_slice_idle = HZ / 125; | |
36 | static int cfq_group_idle = HZ / 125; | |
37 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ | |
38 | static const int cfq_hist_divisor = 4; | |
39 | ||
40 | /* | |
41 | * offset from end of service tree | |
42 | */ | |
43 | #define CFQ_IDLE_DELAY (HZ / 5) | |
44 | ||
45 | /* | |
46 | * below this threshold, we consider thinktime immediate | |
47 | */ | |
48 | #define CFQ_MIN_TT (2) | |
49 | ||
50 | #define CFQ_SLICE_SCALE (5) | |
51 | #define CFQ_HW_QUEUE_MIN (5) | |
52 | #define CFQ_SERVICE_SHIFT 12 | |
53 | ||
54 | #define CFQQ_SEEK_THR (sector_t)(8 * 100) | |
55 | #define CFQQ_CLOSE_THR (sector_t)(8 * 1024) | |
56 | #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
57 | #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8) | |
58 | ||
59 | #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq) | |
60 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0]) | |
61 | #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1]) | |
62 | ||
63 | static struct kmem_cache *cfq_pool; | |
64 | ||
65 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR | |
66 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
67 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
68 | ||
69 | #define sample_valid(samples) ((samples) > 80) | |
70 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) | |
71 | ||
72 | struct cfq_ttime { | |
73 | unsigned long last_end_request; | |
74 | ||
75 | unsigned long ttime_total; | |
76 | unsigned long ttime_samples; | |
77 | unsigned long ttime_mean; | |
78 | }; | |
79 | ||
80 | /* | |
81 | * Most of our rbtree usage is for sorting with min extraction, so | |
82 | * if we cache the leftmost node we don't have to walk down the tree | |
83 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
84 | * move this into the elevator for the rq sorting as well. | |
85 | */ | |
86 | struct cfq_rb_root { | |
87 | struct rb_root rb; | |
88 | struct rb_node *left; | |
89 | unsigned count; | |
90 | unsigned total_weight; | |
91 | u64 min_vdisktime; | |
92 | struct cfq_ttime ttime; | |
93 | }; | |
94 | #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \ | |
95 | .ttime = {.last_end_request = jiffies,},} | |
96 | ||
97 | /* | |
98 | * Per process-grouping structure | |
99 | */ | |
100 | struct cfq_queue { | |
101 | /* reference count */ | |
102 | int ref; | |
103 | /* various state flags, see below */ | |
104 | unsigned int flags; | |
105 | /* parent cfq_data */ | |
106 | struct cfq_data *cfqd; | |
107 | /* service_tree member */ | |
108 | struct rb_node rb_node; | |
109 | /* service_tree key */ | |
110 | unsigned long rb_key; | |
111 | /* prio tree member */ | |
112 | struct rb_node p_node; | |
113 | /* prio tree root we belong to, if any */ | |
114 | struct rb_root *p_root; | |
115 | /* sorted list of pending requests */ | |
116 | struct rb_root sort_list; | |
117 | /* if fifo isn't expired, next request to serve */ | |
118 | struct request *next_rq; | |
119 | /* requests queued in sort_list */ | |
120 | int queued[2]; | |
121 | /* currently allocated requests */ | |
122 | int allocated[2]; | |
123 | /* fifo list of requests in sort_list */ | |
124 | struct list_head fifo; | |
125 | ||
126 | /* time when queue got scheduled in to dispatch first request. */ | |
127 | unsigned long dispatch_start; | |
128 | unsigned int allocated_slice; | |
129 | unsigned int slice_dispatch; | |
130 | /* time when first request from queue completed and slice started. */ | |
131 | unsigned long slice_start; | |
132 | unsigned long slice_end; | |
133 | long slice_resid; | |
134 | ||
135 | /* pending priority requests */ | |
136 | int prio_pending; | |
137 | /* number of requests that are on the dispatch list or inside driver */ | |
138 | int dispatched; | |
139 | ||
140 | /* io prio of this group */ | |
141 | unsigned short ioprio, org_ioprio; | |
142 | unsigned short ioprio_class; | |
143 | ||
144 | pid_t pid; | |
145 | ||
146 | u32 seek_history; | |
147 | sector_t last_request_pos; | |
148 | ||
149 | struct cfq_rb_root *service_tree; | |
150 | struct cfq_queue *new_cfqq; | |
151 | struct cfq_group *cfqg; | |
152 | /* Number of sectors dispatched from queue in single dispatch round */ | |
153 | unsigned long nr_sectors; | |
154 | }; | |
155 | ||
156 | /* | |
157 | * First index in the service_trees. | |
158 | * IDLE is handled separately, so it has negative index | |
159 | */ | |
160 | enum wl_prio_t { | |
161 | BE_WORKLOAD = 0, | |
162 | RT_WORKLOAD = 1, | |
163 | IDLE_WORKLOAD = 2, | |
164 | CFQ_PRIO_NR, | |
165 | }; | |
166 | ||
167 | /* | |
168 | * Second index in the service_trees. | |
169 | */ | |
170 | enum wl_type_t { | |
171 | ASYNC_WORKLOAD = 0, | |
172 | SYNC_NOIDLE_WORKLOAD = 1, | |
173 | SYNC_WORKLOAD = 2 | |
174 | }; | |
175 | ||
176 | struct cfqg_stats { | |
177 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
178 | /* total bytes transferred */ | |
179 | struct blkg_rwstat service_bytes; | |
180 | /* total IOs serviced, post merge */ | |
181 | struct blkg_rwstat serviced; | |
182 | /* number of ios merged */ | |
183 | struct blkg_rwstat merged; | |
184 | /* total time spent on device in ns, may not be accurate w/ queueing */ | |
185 | struct blkg_rwstat service_time; | |
186 | /* total time spent waiting in scheduler queue in ns */ | |
187 | struct blkg_rwstat wait_time; | |
188 | /* number of IOs queued up */ | |
189 | struct blkg_rwstat queued; | |
190 | /* total sectors transferred */ | |
191 | struct blkg_stat sectors; | |
192 | /* total disk time and nr sectors dispatched by this group */ | |
193 | struct blkg_stat time; | |
194 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
195 | /* time not charged to this cgroup */ | |
196 | struct blkg_stat unaccounted_time; | |
197 | /* sum of number of ios queued across all samples */ | |
198 | struct blkg_stat avg_queue_size_sum; | |
199 | /* count of samples taken for average */ | |
200 | struct blkg_stat avg_queue_size_samples; | |
201 | /* how many times this group has been removed from service tree */ | |
202 | struct blkg_stat dequeue; | |
203 | /* total time spent waiting for it to be assigned a timeslice. */ | |
204 | struct blkg_stat group_wait_time; | |
205 | /* time spent idling for this blkio_group */ | |
206 | struct blkg_stat idle_time; | |
207 | /* total time with empty current active q with other requests queued */ | |
208 | struct blkg_stat empty_time; | |
209 | /* fields after this shouldn't be cleared on stat reset */ | |
210 | uint64_t start_group_wait_time; | |
211 | uint64_t start_idle_time; | |
212 | uint64_t start_empty_time; | |
213 | uint16_t flags; | |
214 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
215 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
216 | }; | |
217 | ||
218 | /* This is per cgroup per device grouping structure */ | |
219 | struct cfq_group { | |
220 | /* group service_tree member */ | |
221 | struct rb_node rb_node; | |
222 | ||
223 | /* group service_tree key */ | |
224 | u64 vdisktime; | |
225 | unsigned int weight; | |
226 | unsigned int new_weight; | |
227 | bool needs_update; | |
228 | ||
229 | /* number of cfqq currently on this group */ | |
230 | int nr_cfqq; | |
231 | ||
232 | /* | |
233 | * Per group busy queues average. Useful for workload slice calc. We | |
234 | * create the array for each prio class but at run time it is used | |
235 | * only for RT and BE class and slot for IDLE class remains unused. | |
236 | * This is primarily done to avoid confusion and a gcc warning. | |
237 | */ | |
238 | unsigned int busy_queues_avg[CFQ_PRIO_NR]; | |
239 | /* | |
240 | * rr lists of queues with requests. We maintain service trees for | |
241 | * RT and BE classes. These trees are subdivided in subclasses | |
242 | * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE | |
243 | * class there is no subclassification and all the cfq queues go on | |
244 | * a single tree service_tree_idle. | |
245 | * Counts are embedded in the cfq_rb_root | |
246 | */ | |
247 | struct cfq_rb_root service_trees[2][3]; | |
248 | struct cfq_rb_root service_tree_idle; | |
249 | ||
250 | unsigned long saved_workload_slice; | |
251 | enum wl_type_t saved_workload; | |
252 | enum wl_prio_t saved_serving_prio; | |
253 | ||
254 | /* number of requests that are on the dispatch list or inside driver */ | |
255 | int dispatched; | |
256 | struct cfq_ttime ttime; | |
257 | struct cfqg_stats stats; | |
258 | }; | |
259 | ||
260 | struct cfq_io_cq { | |
261 | struct io_cq icq; /* must be the first member */ | |
262 | struct cfq_queue *cfqq[2]; | |
263 | struct cfq_ttime ttime; | |
264 | int ioprio; /* the current ioprio */ | |
265 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
266 | uint64_t blkcg_id; /* the current blkcg ID */ | |
267 | #endif | |
268 | }; | |
269 | ||
270 | /* | |
271 | * Per block device queue structure | |
272 | */ | |
273 | struct cfq_data { | |
274 | struct request_queue *queue; | |
275 | /* Root service tree for cfq_groups */ | |
276 | struct cfq_rb_root grp_service_tree; | |
277 | struct cfq_group *root_group; | |
278 | ||
279 | /* | |
280 | * The priority currently being served | |
281 | */ | |
282 | enum wl_prio_t serving_prio; | |
283 | enum wl_type_t serving_type; | |
284 | unsigned long workload_expires; | |
285 | struct cfq_group *serving_group; | |
286 | ||
287 | /* | |
288 | * Each priority tree is sorted by next_request position. These | |
289 | * trees are used when determining if two or more queues are | |
290 | * interleaving requests (see cfq_close_cooperator). | |
291 | */ | |
292 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
293 | ||
294 | unsigned int busy_queues; | |
295 | unsigned int busy_sync_queues; | |
296 | ||
297 | int rq_in_driver; | |
298 | int rq_in_flight[2]; | |
299 | ||
300 | /* | |
301 | * queue-depth detection | |
302 | */ | |
303 | int rq_queued; | |
304 | int hw_tag; | |
305 | /* | |
306 | * hw_tag can be | |
307 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
308 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
309 | * 0 => no NCQ | |
310 | */ | |
311 | int hw_tag_est_depth; | |
312 | unsigned int hw_tag_samples; | |
313 | ||
314 | /* | |
315 | * idle window management | |
316 | */ | |
317 | struct timer_list idle_slice_timer; | |
318 | struct work_struct unplug_work; | |
319 | ||
320 | struct cfq_queue *active_queue; | |
321 | struct cfq_io_cq *active_cic; | |
322 | ||
323 | /* | |
324 | * async queue for each priority case | |
325 | */ | |
326 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
327 | struct cfq_queue *async_idle_cfqq; | |
328 | ||
329 | sector_t last_position; | |
330 | ||
331 | /* | |
332 | * tunables, see top of file | |
333 | */ | |
334 | unsigned int cfq_quantum; | |
335 | unsigned int cfq_fifo_expire[2]; | |
336 | unsigned int cfq_back_penalty; | |
337 | unsigned int cfq_back_max; | |
338 | unsigned int cfq_slice[2]; | |
339 | unsigned int cfq_slice_async_rq; | |
340 | unsigned int cfq_slice_idle; | |
341 | unsigned int cfq_group_idle; | |
342 | unsigned int cfq_latency; | |
343 | ||
344 | /* | |
345 | * Fallback dummy cfqq for extreme OOM conditions | |
346 | */ | |
347 | struct cfq_queue oom_cfqq; | |
348 | ||
349 | unsigned long last_delayed_sync; | |
350 | }; | |
351 | ||
352 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); | |
353 | ||
354 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, | |
355 | enum wl_prio_t prio, | |
356 | enum wl_type_t type) | |
357 | { | |
358 | if (!cfqg) | |
359 | return NULL; | |
360 | ||
361 | if (prio == IDLE_WORKLOAD) | |
362 | return &cfqg->service_tree_idle; | |
363 | ||
364 | return &cfqg->service_trees[prio][type]; | |
365 | } | |
366 | ||
367 | enum cfqq_state_flags { | |
368 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ | |
369 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
370 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ | |
371 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
372 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ | |
373 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
374 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
375 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ | |
376 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ | |
377 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ | |
378 | CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ | |
379 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ | |
380 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ | |
381 | }; | |
382 | ||
383 | #define CFQ_CFQQ_FNS(name) \ | |
384 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
385 | { \ | |
386 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ | |
387 | } \ | |
388 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
389 | { \ | |
390 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ | |
391 | } \ | |
392 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
393 | { \ | |
394 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ | |
395 | } | |
396 | ||
397 | CFQ_CFQQ_FNS(on_rr); | |
398 | CFQ_CFQQ_FNS(wait_request); | |
399 | CFQ_CFQQ_FNS(must_dispatch); | |
400 | CFQ_CFQQ_FNS(must_alloc_slice); | |
401 | CFQ_CFQQ_FNS(fifo_expire); | |
402 | CFQ_CFQQ_FNS(idle_window); | |
403 | CFQ_CFQQ_FNS(prio_changed); | |
404 | CFQ_CFQQ_FNS(slice_new); | |
405 | CFQ_CFQQ_FNS(sync); | |
406 | CFQ_CFQQ_FNS(coop); | |
407 | CFQ_CFQQ_FNS(split_coop); | |
408 | CFQ_CFQQ_FNS(deep); | |
409 | CFQ_CFQQ_FNS(wait_busy); | |
410 | #undef CFQ_CFQQ_FNS | |
411 | ||
412 | #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP) | |
413 | ||
414 | /* cfqg stats flags */ | |
415 | enum cfqg_stats_flags { | |
416 | CFQG_stats_waiting = 0, | |
417 | CFQG_stats_idling, | |
418 | CFQG_stats_empty, | |
419 | }; | |
420 | ||
421 | #define CFQG_FLAG_FNS(name) \ | |
422 | static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \ | |
423 | { \ | |
424 | stats->flags |= (1 << CFQG_stats_##name); \ | |
425 | } \ | |
426 | static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \ | |
427 | { \ | |
428 | stats->flags &= ~(1 << CFQG_stats_##name); \ | |
429 | } \ | |
430 | static inline int cfqg_stats_##name(struct cfqg_stats *stats) \ | |
431 | { \ | |
432 | return (stats->flags & (1 << CFQG_stats_##name)) != 0; \ | |
433 | } \ | |
434 | ||
435 | CFQG_FLAG_FNS(waiting) | |
436 | CFQG_FLAG_FNS(idling) | |
437 | CFQG_FLAG_FNS(empty) | |
438 | #undef CFQG_FLAG_FNS | |
439 | ||
440 | /* This should be called with the queue_lock held. */ | |
441 | static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats) | |
442 | { | |
443 | unsigned long long now; | |
444 | ||
445 | if (!cfqg_stats_waiting(stats)) | |
446 | return; | |
447 | ||
448 | now = sched_clock(); | |
449 | if (time_after64(now, stats->start_group_wait_time)) | |
450 | blkg_stat_add(&stats->group_wait_time, | |
451 | now - stats->start_group_wait_time); | |
452 | cfqg_stats_clear_waiting(stats); | |
453 | } | |
454 | ||
455 | /* This should be called with the queue_lock held. */ | |
456 | static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, | |
457 | struct cfq_group *curr_cfqg) | |
458 | { | |
459 | struct cfqg_stats *stats = &cfqg->stats; | |
460 | ||
461 | if (cfqg_stats_waiting(stats)) | |
462 | return; | |
463 | if (cfqg == curr_cfqg) | |
464 | return; | |
465 | stats->start_group_wait_time = sched_clock(); | |
466 | cfqg_stats_mark_waiting(stats); | |
467 | } | |
468 | ||
469 | /* This should be called with the queue_lock held. */ | |
470 | static void cfqg_stats_end_empty_time(struct cfqg_stats *stats) | |
471 | { | |
472 | unsigned long long now; | |
473 | ||
474 | if (!cfqg_stats_empty(stats)) | |
475 | return; | |
476 | ||
477 | now = sched_clock(); | |
478 | if (time_after64(now, stats->start_empty_time)) | |
479 | blkg_stat_add(&stats->empty_time, | |
480 | now - stats->start_empty_time); | |
481 | cfqg_stats_clear_empty(stats); | |
482 | } | |
483 | ||
484 | static void cfqg_stats_update_dequeue(struct cfq_group *cfqg) | |
485 | { | |
486 | blkg_stat_add(&cfqg->stats.dequeue, 1); | |
487 | } | |
488 | ||
489 | static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) | |
490 | { | |
491 | struct cfqg_stats *stats = &cfqg->stats; | |
492 | ||
493 | if (blkg_rwstat_sum(&stats->queued)) | |
494 | return; | |
495 | ||
496 | /* | |
497 | * group is already marked empty. This can happen if cfqq got new | |
498 | * request in parent group and moved to this group while being added | |
499 | * to service tree. Just ignore the event and move on. | |
500 | */ | |
501 | if (cfqg_stats_empty(stats)) | |
502 | return; | |
503 | ||
504 | stats->start_empty_time = sched_clock(); | |
505 | cfqg_stats_mark_empty(stats); | |
506 | } | |
507 | ||
508 | static void cfqg_stats_update_idle_time(struct cfq_group *cfqg) | |
509 | { | |
510 | struct cfqg_stats *stats = &cfqg->stats; | |
511 | ||
512 | if (cfqg_stats_idling(stats)) { | |
513 | unsigned long long now = sched_clock(); | |
514 | ||
515 | if (time_after64(now, stats->start_idle_time)) | |
516 | blkg_stat_add(&stats->idle_time, | |
517 | now - stats->start_idle_time); | |
518 | cfqg_stats_clear_idling(stats); | |
519 | } | |
520 | } | |
521 | ||
522 | static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) | |
523 | { | |
524 | struct cfqg_stats *stats = &cfqg->stats; | |
525 | ||
526 | BUG_ON(cfqg_stats_idling(stats)); | |
527 | ||
528 | stats->start_idle_time = sched_clock(); | |
529 | cfqg_stats_mark_idling(stats); | |
530 | } | |
531 | ||
532 | static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) | |
533 | { | |
534 | struct cfqg_stats *stats = &cfqg->stats; | |
535 | ||
536 | blkg_stat_add(&stats->avg_queue_size_sum, | |
537 | blkg_rwstat_sum(&stats->queued)); | |
538 | blkg_stat_add(&stats->avg_queue_size_samples, 1); | |
539 | cfqg_stats_update_group_wait_time(stats); | |
540 | } | |
541 | ||
542 | #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */ | |
543 | ||
544 | static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, | |
545 | struct cfq_group *curr_cfqg) { } | |
546 | static void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { } | |
547 | static void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { } | |
548 | static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { } | |
549 | static void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { } | |
550 | static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { } | |
551 | static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { } | |
552 | ||
553 | #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */ | |
554 | ||
555 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
556 | ||
557 | static inline struct cfq_group *blkg_to_cfqg(struct blkio_group *blkg) | |
558 | { | |
559 | return blkg_to_pdata(blkg, &blkio_policy_cfq); | |
560 | } | |
561 | ||
562 | static inline struct blkio_group *cfqg_to_blkg(struct cfq_group *cfqg) | |
563 | { | |
564 | return pdata_to_blkg(cfqg); | |
565 | } | |
566 | ||
567 | static inline void cfqg_get(struct cfq_group *cfqg) | |
568 | { | |
569 | return blkg_get(cfqg_to_blkg(cfqg)); | |
570 | } | |
571 | ||
572 | static inline void cfqg_put(struct cfq_group *cfqg) | |
573 | { | |
574 | return blkg_put(cfqg_to_blkg(cfqg)); | |
575 | } | |
576 | ||
577 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
578 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
579 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
580 | blkg_path(cfqg_to_blkg((cfqq)->cfqg)), ##args) | |
581 | ||
582 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
583 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
584 | blkg_path(cfqg_to_blkg((cfqg))), ##args) \ | |
585 | ||
586 | static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg, | |
587 | struct cfq_group *curr_cfqg, int rw) | |
588 | { | |
589 | blkg_rwstat_add(&cfqg->stats.queued, rw, 1); | |
590 | cfqg_stats_end_empty_time(&cfqg->stats); | |
591 | cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg); | |
592 | } | |
593 | ||
594 | static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg, | |
595 | unsigned long time, unsigned long unaccounted_time) | |
596 | { | |
597 | blkg_stat_add(&cfqg->stats.time, time); | |
598 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
599 | blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time); | |
600 | #endif | |
601 | } | |
602 | ||
603 | static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) | |
604 | { | |
605 | blkg_rwstat_add(&cfqg->stats.queued, rw, -1); | |
606 | } | |
607 | ||
608 | static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) | |
609 | { | |
610 | blkg_rwstat_add(&cfqg->stats.merged, rw, 1); | |
611 | } | |
612 | ||
613 | static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg, | |
614 | uint64_t bytes, int rw) | |
615 | { | |
616 | blkg_stat_add(&cfqg->stats.sectors, bytes >> 9); | |
617 | blkg_rwstat_add(&cfqg->stats.serviced, rw, 1); | |
618 | blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes); | |
619 | } | |
620 | ||
621 | static inline void cfqg_stats_update_completion(struct cfq_group *cfqg, | |
622 | uint64_t start_time, uint64_t io_start_time, int rw) | |
623 | { | |
624 | struct cfqg_stats *stats = &cfqg->stats; | |
625 | unsigned long long now = sched_clock(); | |
626 | ||
627 | if (time_after64(now, io_start_time)) | |
628 | blkg_rwstat_add(&stats->service_time, rw, now - io_start_time); | |
629 | if (time_after64(io_start_time, start_time)) | |
630 | blkg_rwstat_add(&stats->wait_time, rw, | |
631 | io_start_time - start_time); | |
632 | } | |
633 | ||
634 | static void cfqg_stats_reset(struct blkio_group *blkg) | |
635 | { | |
636 | struct cfq_group *cfqg = blkg_to_cfqg(blkg); | |
637 | struct cfqg_stats *stats = &cfqg->stats; | |
638 | ||
639 | /* queued stats shouldn't be cleared */ | |
640 | blkg_rwstat_reset(&stats->service_bytes); | |
641 | blkg_rwstat_reset(&stats->serviced); | |
642 | blkg_rwstat_reset(&stats->merged); | |
643 | blkg_rwstat_reset(&stats->service_time); | |
644 | blkg_rwstat_reset(&stats->wait_time); | |
645 | blkg_stat_reset(&stats->time); | |
646 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
647 | blkg_stat_reset(&stats->unaccounted_time); | |
648 | blkg_stat_reset(&stats->avg_queue_size_sum); | |
649 | blkg_stat_reset(&stats->avg_queue_size_samples); | |
650 | blkg_stat_reset(&stats->dequeue); | |
651 | blkg_stat_reset(&stats->group_wait_time); | |
652 | blkg_stat_reset(&stats->idle_time); | |
653 | blkg_stat_reset(&stats->empty_time); | |
654 | #endif | |
655 | } | |
656 | ||
657 | #else /* CONFIG_CFQ_GROUP_IOSCHED */ | |
658 | ||
659 | static inline struct cfq_group *blkg_to_cfqg(struct blkio_group *blkg) { return NULL; } | |
660 | static inline struct blkio_group *cfqg_to_blkg(struct cfq_group *cfqg) { return NULL; } | |
661 | static inline void cfqg_get(struct cfq_group *cfqg) { } | |
662 | static inline void cfqg_put(struct cfq_group *cfqg) { } | |
663 | ||
664 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
665 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
666 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0) | |
667 | ||
668 | static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg, | |
669 | struct cfq_group *curr_cfqg, int rw) { } | |
670 | static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg, | |
671 | unsigned long time, unsigned long unaccounted_time) { } | |
672 | static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { } | |
673 | static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { } | |
674 | static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg, | |
675 | uint64_t bytes, int rw) { } | |
676 | static inline void cfqg_stats_update_completion(struct cfq_group *cfqg, | |
677 | uint64_t start_time, uint64_t io_start_time, int rw) { } | |
678 | ||
679 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
680 | ||
681 | #define cfq_log(cfqd, fmt, args...) \ | |
682 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
683 | ||
684 | /* Traverses through cfq group service trees */ | |
685 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
686 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
687 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
688 | : &cfqg->service_tree_idle; \ | |
689 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
690 | (i == IDLE_WORKLOAD && j == 0); \ | |
691 | j++, st = i < IDLE_WORKLOAD ? \ | |
692 | &cfqg->service_trees[i][j]: NULL) \ | |
693 | ||
694 | static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd, | |
695 | struct cfq_ttime *ttime, bool group_idle) | |
696 | { | |
697 | unsigned long slice; | |
698 | if (!sample_valid(ttime->ttime_samples)) | |
699 | return false; | |
700 | if (group_idle) | |
701 | slice = cfqd->cfq_group_idle; | |
702 | else | |
703 | slice = cfqd->cfq_slice_idle; | |
704 | return ttime->ttime_mean > slice; | |
705 | } | |
706 | ||
707 | static inline bool iops_mode(struct cfq_data *cfqd) | |
708 | { | |
709 | /* | |
710 | * If we are not idling on queues and it is a NCQ drive, parallel | |
711 | * execution of requests is on and measuring time is not possible | |
712 | * in most of the cases until and unless we drive shallower queue | |
713 | * depths and that becomes a performance bottleneck. In such cases | |
714 | * switch to start providing fairness in terms of number of IOs. | |
715 | */ | |
716 | if (!cfqd->cfq_slice_idle && cfqd->hw_tag) | |
717 | return true; | |
718 | else | |
719 | return false; | |
720 | } | |
721 | ||
722 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) | |
723 | { | |
724 | if (cfq_class_idle(cfqq)) | |
725 | return IDLE_WORKLOAD; | |
726 | if (cfq_class_rt(cfqq)) | |
727 | return RT_WORKLOAD; | |
728 | return BE_WORKLOAD; | |
729 | } | |
730 | ||
731 | ||
732 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
733 | { | |
734 | if (!cfq_cfqq_sync(cfqq)) | |
735 | return ASYNC_WORKLOAD; | |
736 | if (!cfq_cfqq_idle_window(cfqq)) | |
737 | return SYNC_NOIDLE_WORKLOAD; | |
738 | return SYNC_WORKLOAD; | |
739 | } | |
740 | ||
741 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, | |
742 | struct cfq_data *cfqd, | |
743 | struct cfq_group *cfqg) | |
744 | { | |
745 | if (wl == IDLE_WORKLOAD) | |
746 | return cfqg->service_tree_idle.count; | |
747 | ||
748 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count | |
749 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
750 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
751 | } | |
752 | ||
753 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, | |
754 | struct cfq_group *cfqg) | |
755 | { | |
756 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count | |
757 | + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
758 | } | |
759 | ||
760 | static void cfq_dispatch_insert(struct request_queue *, struct request *); | |
761 | static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync, | |
762 | struct cfq_io_cq *cic, struct bio *bio, | |
763 | gfp_t gfp_mask); | |
764 | ||
765 | static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq) | |
766 | { | |
767 | /* cic->icq is the first member, %NULL will convert to %NULL */ | |
768 | return container_of(icq, struct cfq_io_cq, icq); | |
769 | } | |
770 | ||
771 | static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd, | |
772 | struct io_context *ioc) | |
773 | { | |
774 | if (ioc) | |
775 | return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue)); | |
776 | return NULL; | |
777 | } | |
778 | ||
779 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync) | |
780 | { | |
781 | return cic->cfqq[is_sync]; | |
782 | } | |
783 | ||
784 | static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq, | |
785 | bool is_sync) | |
786 | { | |
787 | cic->cfqq[is_sync] = cfqq; | |
788 | } | |
789 | ||
790 | static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic) | |
791 | { | |
792 | return cic->icq.q->elevator->elevator_data; | |
793 | } | |
794 | ||
795 | /* | |
796 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
797 | * set (in which case it could also be direct WRITE). | |
798 | */ | |
799 | static inline bool cfq_bio_sync(struct bio *bio) | |
800 | { | |
801 | return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); | |
802 | } | |
803 | ||
804 | /* | |
805 | * scheduler run of queue, if there are requests pending and no one in the | |
806 | * driver that will restart queueing | |
807 | */ | |
808 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
809 | { | |
810 | if (cfqd->busy_queues) { | |
811 | cfq_log(cfqd, "schedule dispatch"); | |
812 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); | |
813 | } | |
814 | } | |
815 | ||
816 | /* | |
817 | * Scale schedule slice based on io priority. Use the sync time slice only | |
818 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
819 | * io only, should not get full sync slice length. | |
820 | */ | |
821 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, | |
822 | unsigned short prio) | |
823 | { | |
824 | const int base_slice = cfqd->cfq_slice[sync]; | |
825 | ||
826 | WARN_ON(prio >= IOPRIO_BE_NR); | |
827 | ||
828 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
829 | } | |
830 | ||
831 | static inline int | |
832 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
833 | { | |
834 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
835 | } | |
836 | ||
837 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) | |
838 | { | |
839 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
840 | ||
841 | d = d * BLKIO_WEIGHT_DEFAULT; | |
842 | do_div(d, cfqg->weight); | |
843 | return d; | |
844 | } | |
845 | ||
846 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
847 | { | |
848 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
849 | if (delta > 0) | |
850 | min_vdisktime = vdisktime; | |
851 | ||
852 | return min_vdisktime; | |
853 | } | |
854 | ||
855 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
856 | { | |
857 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
858 | if (delta < 0) | |
859 | min_vdisktime = vdisktime; | |
860 | ||
861 | return min_vdisktime; | |
862 | } | |
863 | ||
864 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
865 | { | |
866 | struct cfq_group *cfqg; | |
867 | ||
868 | if (st->left) { | |
869 | cfqg = rb_entry_cfqg(st->left); | |
870 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, | |
871 | cfqg->vdisktime); | |
872 | } | |
873 | } | |
874 | ||
875 | /* | |
876 | * get averaged number of queues of RT/BE priority. | |
877 | * average is updated, with a formula that gives more weight to higher numbers, | |
878 | * to quickly follows sudden increases and decrease slowly | |
879 | */ | |
880 | ||
881 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, | |
882 | struct cfq_group *cfqg, bool rt) | |
883 | { | |
884 | unsigned min_q, max_q; | |
885 | unsigned mult = cfq_hist_divisor - 1; | |
886 | unsigned round = cfq_hist_divisor / 2; | |
887 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); | |
888 | ||
889 | min_q = min(cfqg->busy_queues_avg[rt], busy); | |
890 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
891 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
892 | cfq_hist_divisor; | |
893 | return cfqg->busy_queues_avg[rt]; | |
894 | } | |
895 | ||
896 | static inline unsigned | |
897 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
898 | { | |
899 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
900 | ||
901 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
902 | } | |
903 | ||
904 | static inline unsigned | |
905 | cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
906 | { | |
907 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); | |
908 | if (cfqd->cfq_latency) { | |
909 | /* | |
910 | * interested queues (we consider only the ones with the same | |
911 | * priority class in the cfq group) | |
912 | */ | |
913 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
914 | cfq_class_rt(cfqq)); | |
915 | unsigned sync_slice = cfqd->cfq_slice[1]; | |
916 | unsigned expect_latency = sync_slice * iq; | |
917 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); | |
918 | ||
919 | if (expect_latency > group_slice) { | |
920 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; | |
921 | /* scale low_slice according to IO priority | |
922 | * and sync vs async */ | |
923 | unsigned low_slice = | |
924 | min(slice, base_low_slice * slice / sync_slice); | |
925 | /* the adapted slice value is scaled to fit all iqs | |
926 | * into the target latency */ | |
927 | slice = max(slice * group_slice / expect_latency, | |
928 | low_slice); | |
929 | } | |
930 | } | |
931 | return slice; | |
932 | } | |
933 | ||
934 | static inline void | |
935 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
936 | { | |
937 | unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq); | |
938 | ||
939 | cfqq->slice_start = jiffies; | |
940 | cfqq->slice_end = jiffies + slice; | |
941 | cfqq->allocated_slice = slice; | |
942 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); | |
943 | } | |
944 | ||
945 | /* | |
946 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
947 | * isn't valid until the first request from the dispatch is activated | |
948 | * and the slice time set. | |
949 | */ | |
950 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) | |
951 | { | |
952 | if (cfq_cfqq_slice_new(cfqq)) | |
953 | return false; | |
954 | if (time_before(jiffies, cfqq->slice_end)) | |
955 | return false; | |
956 | ||
957 | return true; | |
958 | } | |
959 | ||
960 | /* | |
961 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
962 | * We choose the request that is closest to the head right now. Distance | |
963 | * behind the head is penalized and only allowed to a certain extent. | |
964 | */ | |
965 | static struct request * | |
966 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) | |
967 | { | |
968 | sector_t s1, s2, d1 = 0, d2 = 0; | |
969 | unsigned long back_max; | |
970 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
971 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
972 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
973 | ||
974 | if (rq1 == NULL || rq1 == rq2) | |
975 | return rq2; | |
976 | if (rq2 == NULL) | |
977 | return rq1; | |
978 | ||
979 | if (rq_is_sync(rq1) != rq_is_sync(rq2)) | |
980 | return rq_is_sync(rq1) ? rq1 : rq2; | |
981 | ||
982 | if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO) | |
983 | return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2; | |
984 | ||
985 | s1 = blk_rq_pos(rq1); | |
986 | s2 = blk_rq_pos(rq2); | |
987 | ||
988 | /* | |
989 | * by definition, 1KiB is 2 sectors | |
990 | */ | |
991 | back_max = cfqd->cfq_back_max * 2; | |
992 | ||
993 | /* | |
994 | * Strict one way elevator _except_ in the case where we allow | |
995 | * short backward seeks which are biased as twice the cost of a | |
996 | * similar forward seek. | |
997 | */ | |
998 | if (s1 >= last) | |
999 | d1 = s1 - last; | |
1000 | else if (s1 + back_max >= last) | |
1001 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
1002 | else | |
1003 | wrap |= CFQ_RQ1_WRAP; | |
1004 | ||
1005 | if (s2 >= last) | |
1006 | d2 = s2 - last; | |
1007 | else if (s2 + back_max >= last) | |
1008 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
1009 | else | |
1010 | wrap |= CFQ_RQ2_WRAP; | |
1011 | ||
1012 | /* Found required data */ | |
1013 | ||
1014 | /* | |
1015 | * By doing switch() on the bit mask "wrap" we avoid having to | |
1016 | * check two variables for all permutations: --> faster! | |
1017 | */ | |
1018 | switch (wrap) { | |
1019 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
1020 | if (d1 < d2) | |
1021 | return rq1; | |
1022 | else if (d2 < d1) | |
1023 | return rq2; | |
1024 | else { | |
1025 | if (s1 >= s2) | |
1026 | return rq1; | |
1027 | else | |
1028 | return rq2; | |
1029 | } | |
1030 | ||
1031 | case CFQ_RQ2_WRAP: | |
1032 | return rq1; | |
1033 | case CFQ_RQ1_WRAP: | |
1034 | return rq2; | |
1035 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
1036 | default: | |
1037 | /* | |
1038 | * Since both rqs are wrapped, | |
1039 | * start with the one that's further behind head | |
1040 | * (--> only *one* back seek required), | |
1041 | * since back seek takes more time than forward. | |
1042 | */ | |
1043 | if (s1 <= s2) | |
1044 | return rq1; | |
1045 | else | |
1046 | return rq2; | |
1047 | } | |
1048 | } | |
1049 | ||
1050 | /* | |
1051 | * The below is leftmost cache rbtree addon | |
1052 | */ | |
1053 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) | |
1054 | { | |
1055 | /* Service tree is empty */ | |
1056 | if (!root->count) | |
1057 | return NULL; | |
1058 | ||
1059 | if (!root->left) | |
1060 | root->left = rb_first(&root->rb); | |
1061 | ||
1062 | if (root->left) | |
1063 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
1064 | ||
1065 | return NULL; | |
1066 | } | |
1067 | ||
1068 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) | |
1069 | { | |
1070 | if (!root->left) | |
1071 | root->left = rb_first(&root->rb); | |
1072 | ||
1073 | if (root->left) | |
1074 | return rb_entry_cfqg(root->left); | |
1075 | ||
1076 | return NULL; | |
1077 | } | |
1078 | ||
1079 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
1080 | { | |
1081 | rb_erase(n, root); | |
1082 | RB_CLEAR_NODE(n); | |
1083 | } | |
1084 | ||
1085 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) | |
1086 | { | |
1087 | if (root->left == n) | |
1088 | root->left = NULL; | |
1089 | rb_erase_init(n, &root->rb); | |
1090 | --root->count; | |
1091 | } | |
1092 | ||
1093 | /* | |
1094 | * would be nice to take fifo expire time into account as well | |
1095 | */ | |
1096 | static struct request * | |
1097 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1098 | struct request *last) | |
1099 | { | |
1100 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
1101 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
1102 | struct request *next = NULL, *prev = NULL; | |
1103 | ||
1104 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); | |
1105 | ||
1106 | if (rbprev) | |
1107 | prev = rb_entry_rq(rbprev); | |
1108 | ||
1109 | if (rbnext) | |
1110 | next = rb_entry_rq(rbnext); | |
1111 | else { | |
1112 | rbnext = rb_first(&cfqq->sort_list); | |
1113 | if (rbnext && rbnext != &last->rb_node) | |
1114 | next = rb_entry_rq(rbnext); | |
1115 | } | |
1116 | ||
1117 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); | |
1118 | } | |
1119 | ||
1120 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, | |
1121 | struct cfq_queue *cfqq) | |
1122 | { | |
1123 | /* | |
1124 | * just an approximation, should be ok. | |
1125 | */ | |
1126 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - | |
1127 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
1128 | } | |
1129 | ||
1130 | static inline s64 | |
1131 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1132 | { | |
1133 | return cfqg->vdisktime - st->min_vdisktime; | |
1134 | } | |
1135 | ||
1136 | static void | |
1137 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1138 | { | |
1139 | struct rb_node **node = &st->rb.rb_node; | |
1140 | struct rb_node *parent = NULL; | |
1141 | struct cfq_group *__cfqg; | |
1142 | s64 key = cfqg_key(st, cfqg); | |
1143 | int left = 1; | |
1144 | ||
1145 | while (*node != NULL) { | |
1146 | parent = *node; | |
1147 | __cfqg = rb_entry_cfqg(parent); | |
1148 | ||
1149 | if (key < cfqg_key(st, __cfqg)) | |
1150 | node = &parent->rb_left; | |
1151 | else { | |
1152 | node = &parent->rb_right; | |
1153 | left = 0; | |
1154 | } | |
1155 | } | |
1156 | ||
1157 | if (left) | |
1158 | st->left = &cfqg->rb_node; | |
1159 | ||
1160 | rb_link_node(&cfqg->rb_node, parent, node); | |
1161 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
1162 | } | |
1163 | ||
1164 | static void | |
1165 | cfq_update_group_weight(struct cfq_group *cfqg) | |
1166 | { | |
1167 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
1168 | if (cfqg->needs_update) { | |
1169 | cfqg->weight = cfqg->new_weight; | |
1170 | cfqg->needs_update = false; | |
1171 | } | |
1172 | } | |
1173 | ||
1174 | static void | |
1175 | cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1176 | { | |
1177 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
1178 | ||
1179 | cfq_update_group_weight(cfqg); | |
1180 | __cfq_group_service_tree_add(st, cfqg); | |
1181 | st->total_weight += cfqg->weight; | |
1182 | } | |
1183 | ||
1184 | static void | |
1185 | cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1186 | { | |
1187 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1188 | struct cfq_group *__cfqg; | |
1189 | struct rb_node *n; | |
1190 | ||
1191 | cfqg->nr_cfqq++; | |
1192 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
1193 | return; | |
1194 | ||
1195 | /* | |
1196 | * Currently put the group at the end. Later implement something | |
1197 | * so that groups get lesser vtime based on their weights, so that | |
1198 | * if group does not loose all if it was not continuously backlogged. | |
1199 | */ | |
1200 | n = rb_last(&st->rb); | |
1201 | if (n) { | |
1202 | __cfqg = rb_entry_cfqg(n); | |
1203 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
1204 | } else | |
1205 | cfqg->vdisktime = st->min_vdisktime; | |
1206 | cfq_group_service_tree_add(st, cfqg); | |
1207 | } | |
1208 | ||
1209 | static void | |
1210 | cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1211 | { | |
1212 | st->total_weight -= cfqg->weight; | |
1213 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
1214 | cfq_rb_erase(&cfqg->rb_node, st); | |
1215 | } | |
1216 | ||
1217 | static void | |
1218 | cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1219 | { | |
1220 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1221 | ||
1222 | BUG_ON(cfqg->nr_cfqq < 1); | |
1223 | cfqg->nr_cfqq--; | |
1224 | ||
1225 | /* If there are other cfq queues under this group, don't delete it */ | |
1226 | if (cfqg->nr_cfqq) | |
1227 | return; | |
1228 | ||
1229 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); | |
1230 | cfq_group_service_tree_del(st, cfqg); | |
1231 | cfqg->saved_workload_slice = 0; | |
1232 | cfqg_stats_update_dequeue(cfqg); | |
1233 | } | |
1234 | ||
1235 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq, | |
1236 | unsigned int *unaccounted_time) | |
1237 | { | |
1238 | unsigned int slice_used; | |
1239 | ||
1240 | /* | |
1241 | * Queue got expired before even a single request completed or | |
1242 | * got expired immediately after first request completion. | |
1243 | */ | |
1244 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
1245 | /* | |
1246 | * Also charge the seek time incurred to the group, otherwise | |
1247 | * if there are mutiple queues in the group, each can dispatch | |
1248 | * a single request on seeky media and cause lots of seek time | |
1249 | * and group will never know it. | |
1250 | */ | |
1251 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
1252 | 1); | |
1253 | } else { | |
1254 | slice_used = jiffies - cfqq->slice_start; | |
1255 | if (slice_used > cfqq->allocated_slice) { | |
1256 | *unaccounted_time = slice_used - cfqq->allocated_slice; | |
1257 | slice_used = cfqq->allocated_slice; | |
1258 | } | |
1259 | if (time_after(cfqq->slice_start, cfqq->dispatch_start)) | |
1260 | *unaccounted_time += cfqq->slice_start - | |
1261 | cfqq->dispatch_start; | |
1262 | } | |
1263 | ||
1264 | return slice_used; | |
1265 | } | |
1266 | ||
1267 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
1268 | struct cfq_queue *cfqq) | |
1269 | { | |
1270 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1271 | unsigned int used_sl, charge, unaccounted_sl = 0; | |
1272 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) | |
1273 | - cfqg->service_tree_idle.count; | |
1274 | ||
1275 | BUG_ON(nr_sync < 0); | |
1276 | used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); | |
1277 | ||
1278 | if (iops_mode(cfqd)) | |
1279 | charge = cfqq->slice_dispatch; | |
1280 | else if (!cfq_cfqq_sync(cfqq) && !nr_sync) | |
1281 | charge = cfqq->allocated_slice; | |
1282 | ||
1283 | /* Can't update vdisktime while group is on service tree */ | |
1284 | cfq_group_service_tree_del(st, cfqg); | |
1285 | cfqg->vdisktime += cfq_scale_slice(charge, cfqg); | |
1286 | /* If a new weight was requested, update now, off tree */ | |
1287 | cfq_group_service_tree_add(st, cfqg); | |
1288 | ||
1289 | /* This group is being expired. Save the context */ | |
1290 | if (time_after(cfqd->workload_expires, jiffies)) { | |
1291 | cfqg->saved_workload_slice = cfqd->workload_expires | |
1292 | - jiffies; | |
1293 | cfqg->saved_workload = cfqd->serving_type; | |
1294 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
1295 | } else | |
1296 | cfqg->saved_workload_slice = 0; | |
1297 | ||
1298 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
1299 | st->min_vdisktime); | |
1300 | cfq_log_cfqq(cfqq->cfqd, cfqq, | |
1301 | "sl_used=%u disp=%u charge=%u iops=%u sect=%lu", | |
1302 | used_sl, cfqq->slice_dispatch, charge, | |
1303 | iops_mode(cfqd), cfqq->nr_sectors); | |
1304 | cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl); | |
1305 | cfqg_stats_set_start_empty_time(cfqg); | |
1306 | } | |
1307 | ||
1308 | /** | |
1309 | * cfq_init_cfqg_base - initialize base part of a cfq_group | |
1310 | * @cfqg: cfq_group to initialize | |
1311 | * | |
1312 | * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED | |
1313 | * is enabled or not. | |
1314 | */ | |
1315 | static void cfq_init_cfqg_base(struct cfq_group *cfqg) | |
1316 | { | |
1317 | struct cfq_rb_root *st; | |
1318 | int i, j; | |
1319 | ||
1320 | for_each_cfqg_st(cfqg, i, j, st) | |
1321 | *st = CFQ_RB_ROOT; | |
1322 | RB_CLEAR_NODE(&cfqg->rb_node); | |
1323 | ||
1324 | cfqg->ttime.last_end_request = jiffies; | |
1325 | } | |
1326 | ||
1327 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
1328 | static void cfq_update_blkio_group_weight(struct blkio_group *blkg, | |
1329 | unsigned int weight) | |
1330 | { | |
1331 | struct cfq_group *cfqg = blkg_to_cfqg(blkg); | |
1332 | ||
1333 | cfqg->new_weight = weight; | |
1334 | cfqg->needs_update = true; | |
1335 | } | |
1336 | ||
1337 | static void cfq_init_blkio_group(struct blkio_group *blkg) | |
1338 | { | |
1339 | struct cfq_group *cfqg = blkg_to_cfqg(blkg); | |
1340 | ||
1341 | cfq_init_cfqg_base(cfqg); | |
1342 | cfqg->weight = blkg->blkcg->weight; | |
1343 | } | |
1344 | ||
1345 | /* | |
1346 | * Search for the cfq group current task belongs to. request_queue lock must | |
1347 | * be held. | |
1348 | */ | |
1349 | static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd, | |
1350 | struct blkio_cgroup *blkcg) | |
1351 | { | |
1352 | struct request_queue *q = cfqd->queue; | |
1353 | struct cfq_group *cfqg = NULL; | |
1354 | ||
1355 | /* avoid lookup for the common case where there's no blkio cgroup */ | |
1356 | if (blkcg == &blkio_root_cgroup) { | |
1357 | cfqg = cfqd->root_group; | |
1358 | } else { | |
1359 | struct blkio_group *blkg; | |
1360 | ||
1361 | blkg = blkg_lookup_create(blkcg, q, false); | |
1362 | if (!IS_ERR(blkg)) | |
1363 | cfqg = blkg_to_cfqg(blkg); | |
1364 | } | |
1365 | ||
1366 | return cfqg; | |
1367 | } | |
1368 | ||
1369 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) | |
1370 | { | |
1371 | /* Currently, all async queues are mapped to root group */ | |
1372 | if (!cfq_cfqq_sync(cfqq)) | |
1373 | cfqg = cfqq->cfqd->root_group; | |
1374 | ||
1375 | cfqq->cfqg = cfqg; | |
1376 | /* cfqq reference on cfqg */ | |
1377 | cfqg_get(cfqg); | |
1378 | } | |
1379 | ||
1380 | static u64 blkg_prfill_weight_device(struct seq_file *sf, | |
1381 | struct blkg_policy_data *pd, int off) | |
1382 | { | |
1383 | if (!pd->conf.weight) | |
1384 | return 0; | |
1385 | return __blkg_prfill_u64(sf, pd, pd->conf.weight); | |
1386 | } | |
1387 | ||
1388 | static int blkcg_print_weight_device(struct cgroup *cgrp, struct cftype *cft, | |
1389 | struct seq_file *sf) | |
1390 | { | |
1391 | blkcg_print_blkgs(sf, cgroup_to_blkio_cgroup(cgrp), | |
1392 | blkg_prfill_weight_device, BLKIO_POLICY_PROP, 0, | |
1393 | false); | |
1394 | return 0; | |
1395 | } | |
1396 | ||
1397 | static int blkcg_print_weight(struct cgroup *cgrp, struct cftype *cft, | |
1398 | struct seq_file *sf) | |
1399 | { | |
1400 | seq_printf(sf, "%u\n", cgroup_to_blkio_cgroup(cgrp)->weight); | |
1401 | return 0; | |
1402 | } | |
1403 | ||
1404 | static int blkcg_set_weight_device(struct cgroup *cgrp, struct cftype *cft, | |
1405 | const char *buf) | |
1406 | { | |
1407 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
1408 | struct blkg_policy_data *pd; | |
1409 | struct blkg_conf_ctx ctx; | |
1410 | int ret; | |
1411 | ||
1412 | ret = blkg_conf_prep(blkcg, buf, &ctx); | |
1413 | if (ret) | |
1414 | return ret; | |
1415 | ||
1416 | ret = -EINVAL; | |
1417 | pd = ctx.blkg->pd[BLKIO_POLICY_PROP]; | |
1418 | if (pd && (!ctx.v || (ctx.v >= BLKIO_WEIGHT_MIN && | |
1419 | ctx.v <= BLKIO_WEIGHT_MAX))) { | |
1420 | pd->conf.weight = ctx.v; | |
1421 | cfq_update_blkio_group_weight(ctx.blkg, ctx.v ?: blkcg->weight); | |
1422 | ret = 0; | |
1423 | } | |
1424 | ||
1425 | blkg_conf_finish(&ctx); | |
1426 | return ret; | |
1427 | } | |
1428 | ||
1429 | static int blkcg_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val) | |
1430 | { | |
1431 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
1432 | struct blkio_group *blkg; | |
1433 | struct hlist_node *n; | |
1434 | ||
1435 | if (val < BLKIO_WEIGHT_MIN || val > BLKIO_WEIGHT_MAX) | |
1436 | return -EINVAL; | |
1437 | ||
1438 | spin_lock_irq(&blkcg->lock); | |
1439 | blkcg->weight = (unsigned int)val; | |
1440 | ||
1441 | hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) { | |
1442 | struct blkg_policy_data *pd = blkg->pd[BLKIO_POLICY_PROP]; | |
1443 | ||
1444 | if (pd && !pd->conf.weight) | |
1445 | cfq_update_blkio_group_weight(blkg, blkcg->weight); | |
1446 | } | |
1447 | ||
1448 | spin_unlock_irq(&blkcg->lock); | |
1449 | return 0; | |
1450 | } | |
1451 | ||
1452 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
1453 | static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf, | |
1454 | struct blkg_policy_data *pd, int off) | |
1455 | { | |
1456 | struct cfq_group *cfqg = (void *)pd->pdata; | |
1457 | u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples); | |
1458 | u64 v = 0; | |
1459 | ||
1460 | if (samples) { | |
1461 | v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum); | |
1462 | do_div(v, samples); | |
1463 | } | |
1464 | __blkg_prfill_u64(sf, pd, v); | |
1465 | return 0; | |
1466 | } | |
1467 | ||
1468 | /* print avg_queue_size */ | |
1469 | static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft, | |
1470 | struct seq_file *sf) | |
1471 | { | |
1472 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgrp); | |
1473 | ||
1474 | blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size, | |
1475 | BLKIO_POLICY_PROP, 0, false); | |
1476 | return 0; | |
1477 | } | |
1478 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
1479 | ||
1480 | static struct cftype cfq_blkcg_files[] = { | |
1481 | { | |
1482 | .name = "weight_device", | |
1483 | .read_seq_string = blkcg_print_weight_device, | |
1484 | .write_string = blkcg_set_weight_device, | |
1485 | .max_write_len = 256, | |
1486 | }, | |
1487 | { | |
1488 | .name = "weight", | |
1489 | .read_seq_string = blkcg_print_weight, | |
1490 | .write_u64 = blkcg_set_weight, | |
1491 | }, | |
1492 | { | |
1493 | .name = "time", | |
1494 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1495 | offsetof(struct cfq_group, stats.time)), | |
1496 | .read_seq_string = blkcg_print_stat, | |
1497 | }, | |
1498 | { | |
1499 | .name = "sectors", | |
1500 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1501 | offsetof(struct cfq_group, stats.sectors)), | |
1502 | .read_seq_string = blkcg_print_stat, | |
1503 | }, | |
1504 | { | |
1505 | .name = "io_service_bytes", | |
1506 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1507 | offsetof(struct cfq_group, stats.service_bytes)), | |
1508 | .read_seq_string = blkcg_print_rwstat, | |
1509 | }, | |
1510 | { | |
1511 | .name = "io_serviced", | |
1512 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1513 | offsetof(struct cfq_group, stats.serviced)), | |
1514 | .read_seq_string = blkcg_print_rwstat, | |
1515 | }, | |
1516 | { | |
1517 | .name = "io_service_time", | |
1518 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1519 | offsetof(struct cfq_group, stats.service_time)), | |
1520 | .read_seq_string = blkcg_print_rwstat, | |
1521 | }, | |
1522 | { | |
1523 | .name = "io_wait_time", | |
1524 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1525 | offsetof(struct cfq_group, stats.wait_time)), | |
1526 | .read_seq_string = blkcg_print_rwstat, | |
1527 | }, | |
1528 | { | |
1529 | .name = "io_merged", | |
1530 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1531 | offsetof(struct cfq_group, stats.merged)), | |
1532 | .read_seq_string = blkcg_print_rwstat, | |
1533 | }, | |
1534 | { | |
1535 | .name = "io_queued", | |
1536 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1537 | offsetof(struct cfq_group, stats.queued)), | |
1538 | .read_seq_string = blkcg_print_rwstat, | |
1539 | }, | |
1540 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
1541 | { | |
1542 | .name = "avg_queue_size", | |
1543 | .read_seq_string = cfqg_print_avg_queue_size, | |
1544 | }, | |
1545 | { | |
1546 | .name = "group_wait_time", | |
1547 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1548 | offsetof(struct cfq_group, stats.group_wait_time)), | |
1549 | .read_seq_string = blkcg_print_stat, | |
1550 | }, | |
1551 | { | |
1552 | .name = "idle_time", | |
1553 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1554 | offsetof(struct cfq_group, stats.idle_time)), | |
1555 | .read_seq_string = blkcg_print_stat, | |
1556 | }, | |
1557 | { | |
1558 | .name = "empty_time", | |
1559 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1560 | offsetof(struct cfq_group, stats.empty_time)), | |
1561 | .read_seq_string = blkcg_print_stat, | |
1562 | }, | |
1563 | { | |
1564 | .name = "dequeue", | |
1565 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1566 | offsetof(struct cfq_group, stats.dequeue)), | |
1567 | .read_seq_string = blkcg_print_stat, | |
1568 | }, | |
1569 | { | |
1570 | .name = "unaccounted_time", | |
1571 | .private = BLKCG_STAT_PRIV(BLKIO_POLICY_PROP, | |
1572 | offsetof(struct cfq_group, stats.unaccounted_time)), | |
1573 | .read_seq_string = blkcg_print_stat, | |
1574 | }, | |
1575 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
1576 | { } /* terminate */ | |
1577 | }; | |
1578 | #else /* GROUP_IOSCHED */ | |
1579 | static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd, | |
1580 | struct blkio_cgroup *blkcg) | |
1581 | { | |
1582 | return cfqd->root_group; | |
1583 | } | |
1584 | ||
1585 | static inline void | |
1586 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1587 | cfqq->cfqg = cfqg; | |
1588 | } | |
1589 | ||
1590 | #endif /* GROUP_IOSCHED */ | |
1591 | ||
1592 | /* | |
1593 | * The cfqd->service_trees holds all pending cfq_queue's that have | |
1594 | * requests waiting to be processed. It is sorted in the order that | |
1595 | * we will service the queues. | |
1596 | */ | |
1597 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1598 | bool add_front) | |
1599 | { | |
1600 | struct rb_node **p, *parent; | |
1601 | struct cfq_queue *__cfqq; | |
1602 | unsigned long rb_key; | |
1603 | struct cfq_rb_root *service_tree; | |
1604 | int left; | |
1605 | int new_cfqq = 1; | |
1606 | ||
1607 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), | |
1608 | cfqq_type(cfqq)); | |
1609 | if (cfq_class_idle(cfqq)) { | |
1610 | rb_key = CFQ_IDLE_DELAY; | |
1611 | parent = rb_last(&service_tree->rb); | |
1612 | if (parent && parent != &cfqq->rb_node) { | |
1613 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1614 | rb_key += __cfqq->rb_key; | |
1615 | } else | |
1616 | rb_key += jiffies; | |
1617 | } else if (!add_front) { | |
1618 | /* | |
1619 | * Get our rb key offset. Subtract any residual slice | |
1620 | * value carried from last service. A negative resid | |
1621 | * count indicates slice overrun, and this should position | |
1622 | * the next service time further away in the tree. | |
1623 | */ | |
1624 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; | |
1625 | rb_key -= cfqq->slice_resid; | |
1626 | cfqq->slice_resid = 0; | |
1627 | } else { | |
1628 | rb_key = -HZ; | |
1629 | __cfqq = cfq_rb_first(service_tree); | |
1630 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; | |
1631 | } | |
1632 | ||
1633 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
1634 | new_cfqq = 0; | |
1635 | /* | |
1636 | * same position, nothing more to do | |
1637 | */ | |
1638 | if (rb_key == cfqq->rb_key && | |
1639 | cfqq->service_tree == service_tree) | |
1640 | return; | |
1641 | ||
1642 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1643 | cfqq->service_tree = NULL; | |
1644 | } | |
1645 | ||
1646 | left = 1; | |
1647 | parent = NULL; | |
1648 | cfqq->service_tree = service_tree; | |
1649 | p = &service_tree->rb.rb_node; | |
1650 | while (*p) { | |
1651 | struct rb_node **n; | |
1652 | ||
1653 | parent = *p; | |
1654 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1655 | ||
1656 | /* | |
1657 | * sort by key, that represents service time. | |
1658 | */ | |
1659 | if (time_before(rb_key, __cfqq->rb_key)) | |
1660 | n = &(*p)->rb_left; | |
1661 | else { | |
1662 | n = &(*p)->rb_right; | |
1663 | left = 0; | |
1664 | } | |
1665 | ||
1666 | p = n; | |
1667 | } | |
1668 | ||
1669 | if (left) | |
1670 | service_tree->left = &cfqq->rb_node; | |
1671 | ||
1672 | cfqq->rb_key = rb_key; | |
1673 | rb_link_node(&cfqq->rb_node, parent, p); | |
1674 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); | |
1675 | service_tree->count++; | |
1676 | if (add_front || !new_cfqq) | |
1677 | return; | |
1678 | cfq_group_notify_queue_add(cfqd, cfqq->cfqg); | |
1679 | } | |
1680 | ||
1681 | static struct cfq_queue * | |
1682 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, | |
1683 | sector_t sector, struct rb_node **ret_parent, | |
1684 | struct rb_node ***rb_link) | |
1685 | { | |
1686 | struct rb_node **p, *parent; | |
1687 | struct cfq_queue *cfqq = NULL; | |
1688 | ||
1689 | parent = NULL; | |
1690 | p = &root->rb_node; | |
1691 | while (*p) { | |
1692 | struct rb_node **n; | |
1693 | ||
1694 | parent = *p; | |
1695 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1696 | ||
1697 | /* | |
1698 | * Sort strictly based on sector. Smallest to the left, | |
1699 | * largest to the right. | |
1700 | */ | |
1701 | if (sector > blk_rq_pos(cfqq->next_rq)) | |
1702 | n = &(*p)->rb_right; | |
1703 | else if (sector < blk_rq_pos(cfqq->next_rq)) | |
1704 | n = &(*p)->rb_left; | |
1705 | else | |
1706 | break; | |
1707 | p = n; | |
1708 | cfqq = NULL; | |
1709 | } | |
1710 | ||
1711 | *ret_parent = parent; | |
1712 | if (rb_link) | |
1713 | *rb_link = p; | |
1714 | return cfqq; | |
1715 | } | |
1716 | ||
1717 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1718 | { | |
1719 | struct rb_node **p, *parent; | |
1720 | struct cfq_queue *__cfqq; | |
1721 | ||
1722 | if (cfqq->p_root) { | |
1723 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1724 | cfqq->p_root = NULL; | |
1725 | } | |
1726 | ||
1727 | if (cfq_class_idle(cfqq)) | |
1728 | return; | |
1729 | if (!cfqq->next_rq) | |
1730 | return; | |
1731 | ||
1732 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; | |
1733 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, | |
1734 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
1735 | if (!__cfqq) { | |
1736 | rb_link_node(&cfqq->p_node, parent, p); | |
1737 | rb_insert_color(&cfqq->p_node, cfqq->p_root); | |
1738 | } else | |
1739 | cfqq->p_root = NULL; | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * Update cfqq's position in the service tree. | |
1744 | */ | |
1745 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1746 | { | |
1747 | /* | |
1748 | * Resorting requires the cfqq to be on the RR list already. | |
1749 | */ | |
1750 | if (cfq_cfqq_on_rr(cfqq)) { | |
1751 | cfq_service_tree_add(cfqd, cfqq, 0); | |
1752 | cfq_prio_tree_add(cfqd, cfqq); | |
1753 | } | |
1754 | } | |
1755 | ||
1756 | /* | |
1757 | * add to busy list of queues for service, trying to be fair in ordering | |
1758 | * the pending list according to last request service | |
1759 | */ | |
1760 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1761 | { | |
1762 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); | |
1763 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
1764 | cfq_mark_cfqq_on_rr(cfqq); | |
1765 | cfqd->busy_queues++; | |
1766 | if (cfq_cfqq_sync(cfqq)) | |
1767 | cfqd->busy_sync_queues++; | |
1768 | ||
1769 | cfq_resort_rr_list(cfqd, cfqq); | |
1770 | } | |
1771 | ||
1772 | /* | |
1773 | * Called when the cfqq no longer has requests pending, remove it from | |
1774 | * the service tree. | |
1775 | */ | |
1776 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1777 | { | |
1778 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); | |
1779 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
1780 | cfq_clear_cfqq_on_rr(cfqq); | |
1781 | ||
1782 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
1783 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1784 | cfqq->service_tree = NULL; | |
1785 | } | |
1786 | if (cfqq->p_root) { | |
1787 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1788 | cfqq->p_root = NULL; | |
1789 | } | |
1790 | ||
1791 | cfq_group_notify_queue_del(cfqd, cfqq->cfqg); | |
1792 | BUG_ON(!cfqd->busy_queues); | |
1793 | cfqd->busy_queues--; | |
1794 | if (cfq_cfqq_sync(cfqq)) | |
1795 | cfqd->busy_sync_queues--; | |
1796 | } | |
1797 | ||
1798 | /* | |
1799 | * rb tree support functions | |
1800 | */ | |
1801 | static void cfq_del_rq_rb(struct request *rq) | |
1802 | { | |
1803 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1804 | const int sync = rq_is_sync(rq); | |
1805 | ||
1806 | BUG_ON(!cfqq->queued[sync]); | |
1807 | cfqq->queued[sync]--; | |
1808 | ||
1809 | elv_rb_del(&cfqq->sort_list, rq); | |
1810 | ||
1811 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { | |
1812 | /* | |
1813 | * Queue will be deleted from service tree when we actually | |
1814 | * expire it later. Right now just remove it from prio tree | |
1815 | * as it is empty. | |
1816 | */ | |
1817 | if (cfqq->p_root) { | |
1818 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1819 | cfqq->p_root = NULL; | |
1820 | } | |
1821 | } | |
1822 | } | |
1823 | ||
1824 | static void cfq_add_rq_rb(struct request *rq) | |
1825 | { | |
1826 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1827 | struct cfq_data *cfqd = cfqq->cfqd; | |
1828 | struct request *prev; | |
1829 | ||
1830 | cfqq->queued[rq_is_sync(rq)]++; | |
1831 | ||
1832 | elv_rb_add(&cfqq->sort_list, rq); | |
1833 | ||
1834 | if (!cfq_cfqq_on_rr(cfqq)) | |
1835 | cfq_add_cfqq_rr(cfqd, cfqq); | |
1836 | ||
1837 | /* | |
1838 | * check if this request is a better next-serve candidate | |
1839 | */ | |
1840 | prev = cfqq->next_rq; | |
1841 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); | |
1842 | ||
1843 | /* | |
1844 | * adjust priority tree position, if ->next_rq changes | |
1845 | */ | |
1846 | if (prev != cfqq->next_rq) | |
1847 | cfq_prio_tree_add(cfqd, cfqq); | |
1848 | ||
1849 | BUG_ON(!cfqq->next_rq); | |
1850 | } | |
1851 | ||
1852 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) | |
1853 | { | |
1854 | elv_rb_del(&cfqq->sort_list, rq); | |
1855 | cfqq->queued[rq_is_sync(rq)]--; | |
1856 | cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags); | |
1857 | cfq_add_rq_rb(rq); | |
1858 | cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group, | |
1859 | rq->cmd_flags); | |
1860 | } | |
1861 | ||
1862 | static struct request * | |
1863 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1864 | { | |
1865 | struct task_struct *tsk = current; | |
1866 | struct cfq_io_cq *cic; | |
1867 | struct cfq_queue *cfqq; | |
1868 | ||
1869 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
1870 | if (!cic) | |
1871 | return NULL; | |
1872 | ||
1873 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
1874 | if (cfqq) { | |
1875 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1876 | ||
1877 | return elv_rb_find(&cfqq->sort_list, sector); | |
1878 | } | |
1879 | ||
1880 | return NULL; | |
1881 | } | |
1882 | ||
1883 | static void cfq_activate_request(struct request_queue *q, struct request *rq) | |
1884 | { | |
1885 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1886 | ||
1887 | cfqd->rq_in_driver++; | |
1888 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", | |
1889 | cfqd->rq_in_driver); | |
1890 | ||
1891 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
1892 | } | |
1893 | ||
1894 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) | |
1895 | { | |
1896 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1897 | ||
1898 | WARN_ON(!cfqd->rq_in_driver); | |
1899 | cfqd->rq_in_driver--; | |
1900 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", | |
1901 | cfqd->rq_in_driver); | |
1902 | } | |
1903 | ||
1904 | static void cfq_remove_request(struct request *rq) | |
1905 | { | |
1906 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1907 | ||
1908 | if (cfqq->next_rq == rq) | |
1909 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1910 | ||
1911 | list_del_init(&rq->queuelist); | |
1912 | cfq_del_rq_rb(rq); | |
1913 | ||
1914 | cfqq->cfqd->rq_queued--; | |
1915 | cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags); | |
1916 | if (rq->cmd_flags & REQ_PRIO) { | |
1917 | WARN_ON(!cfqq->prio_pending); | |
1918 | cfqq->prio_pending--; | |
1919 | } | |
1920 | } | |
1921 | ||
1922 | static int cfq_merge(struct request_queue *q, struct request **req, | |
1923 | struct bio *bio) | |
1924 | { | |
1925 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1926 | struct request *__rq; | |
1927 | ||
1928 | __rq = cfq_find_rq_fmerge(cfqd, bio); | |
1929 | if (__rq && elv_rq_merge_ok(__rq, bio)) { | |
1930 | *req = __rq; | |
1931 | return ELEVATOR_FRONT_MERGE; | |
1932 | } | |
1933 | ||
1934 | return ELEVATOR_NO_MERGE; | |
1935 | } | |
1936 | ||
1937 | static void cfq_merged_request(struct request_queue *q, struct request *req, | |
1938 | int type) | |
1939 | { | |
1940 | if (type == ELEVATOR_FRONT_MERGE) { | |
1941 | struct cfq_queue *cfqq = RQ_CFQQ(req); | |
1942 | ||
1943 | cfq_reposition_rq_rb(cfqq, req); | |
1944 | } | |
1945 | } | |
1946 | ||
1947 | static void cfq_bio_merged(struct request_queue *q, struct request *req, | |
1948 | struct bio *bio) | |
1949 | { | |
1950 | cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw); | |
1951 | } | |
1952 | ||
1953 | static void | |
1954 | cfq_merged_requests(struct request_queue *q, struct request *rq, | |
1955 | struct request *next) | |
1956 | { | |
1957 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1958 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1959 | ||
1960 | /* | |
1961 | * reposition in fifo if next is older than rq | |
1962 | */ | |
1963 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
1964 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { | |
1965 | list_move(&rq->queuelist, &next->queuelist); | |
1966 | rq_set_fifo_time(rq, rq_fifo_time(next)); | |
1967 | } | |
1968 | ||
1969 | if (cfqq->next_rq == next) | |
1970 | cfqq->next_rq = rq; | |
1971 | cfq_remove_request(next); | |
1972 | cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags); | |
1973 | ||
1974 | cfqq = RQ_CFQQ(next); | |
1975 | /* | |
1976 | * all requests of this queue are merged to other queues, delete it | |
1977 | * from the service tree. If it's the active_queue, | |
1978 | * cfq_dispatch_requests() will choose to expire it or do idle | |
1979 | */ | |
1980 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) && | |
1981 | cfqq != cfqd->active_queue) | |
1982 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1983 | } | |
1984 | ||
1985 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, | |
1986 | struct bio *bio) | |
1987 | { | |
1988 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1989 | struct cfq_io_cq *cic; | |
1990 | struct cfq_queue *cfqq; | |
1991 | ||
1992 | /* | |
1993 | * Disallow merge of a sync bio into an async request. | |
1994 | */ | |
1995 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) | |
1996 | return false; | |
1997 | ||
1998 | /* | |
1999 | * Lookup the cfqq that this bio will be queued with and allow | |
2000 | * merge only if rq is queued there. | |
2001 | */ | |
2002 | cic = cfq_cic_lookup(cfqd, current->io_context); | |
2003 | if (!cic) | |
2004 | return false; | |
2005 | ||
2006 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
2007 | return cfqq == RQ_CFQQ(rq); | |
2008 | } | |
2009 | ||
2010 | static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2011 | { | |
2012 | del_timer(&cfqd->idle_slice_timer); | |
2013 | cfqg_stats_update_idle_time(cfqq->cfqg); | |
2014 | } | |
2015 | ||
2016 | static void __cfq_set_active_queue(struct cfq_data *cfqd, | |
2017 | struct cfq_queue *cfqq) | |
2018 | { | |
2019 | if (cfqq) { | |
2020 | cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d", | |
2021 | cfqd->serving_prio, cfqd->serving_type); | |
2022 | cfqg_stats_update_avg_queue_size(cfqq->cfqg); | |
2023 | cfqq->slice_start = 0; | |
2024 | cfqq->dispatch_start = jiffies; | |
2025 | cfqq->allocated_slice = 0; | |
2026 | cfqq->slice_end = 0; | |
2027 | cfqq->slice_dispatch = 0; | |
2028 | cfqq->nr_sectors = 0; | |
2029 | ||
2030 | cfq_clear_cfqq_wait_request(cfqq); | |
2031 | cfq_clear_cfqq_must_dispatch(cfqq); | |
2032 | cfq_clear_cfqq_must_alloc_slice(cfqq); | |
2033 | cfq_clear_cfqq_fifo_expire(cfqq); | |
2034 | cfq_mark_cfqq_slice_new(cfqq); | |
2035 | ||
2036 | cfq_del_timer(cfqd, cfqq); | |
2037 | } | |
2038 | ||
2039 | cfqd->active_queue = cfqq; | |
2040 | } | |
2041 | ||
2042 | /* | |
2043 | * current cfqq expired its slice (or was too idle), select new one | |
2044 | */ | |
2045 | static void | |
2046 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2047 | bool timed_out) | |
2048 | { | |
2049 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); | |
2050 | ||
2051 | if (cfq_cfqq_wait_request(cfqq)) | |
2052 | cfq_del_timer(cfqd, cfqq); | |
2053 | ||
2054 | cfq_clear_cfqq_wait_request(cfqq); | |
2055 | cfq_clear_cfqq_wait_busy(cfqq); | |
2056 | ||
2057 | /* | |
2058 | * If this cfqq is shared between multiple processes, check to | |
2059 | * make sure that those processes are still issuing I/Os within | |
2060 | * the mean seek distance. If not, it may be time to break the | |
2061 | * queues apart again. | |
2062 | */ | |
2063 | if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) | |
2064 | cfq_mark_cfqq_split_coop(cfqq); | |
2065 | ||
2066 | /* | |
2067 | * store what was left of this slice, if the queue idled/timed out | |
2068 | */ | |
2069 | if (timed_out) { | |
2070 | if (cfq_cfqq_slice_new(cfqq)) | |
2071 | cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); | |
2072 | else | |
2073 | cfqq->slice_resid = cfqq->slice_end - jiffies; | |
2074 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); | |
2075 | } | |
2076 | ||
2077 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); | |
2078 | ||
2079 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2080 | cfq_del_cfqq_rr(cfqd, cfqq); | |
2081 | ||
2082 | cfq_resort_rr_list(cfqd, cfqq); | |
2083 | ||
2084 | if (cfqq == cfqd->active_queue) | |
2085 | cfqd->active_queue = NULL; | |
2086 | ||
2087 | if (cfqd->active_cic) { | |
2088 | put_io_context(cfqd->active_cic->icq.ioc); | |
2089 | cfqd->active_cic = NULL; | |
2090 | } | |
2091 | } | |
2092 | ||
2093 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) | |
2094 | { | |
2095 | struct cfq_queue *cfqq = cfqd->active_queue; | |
2096 | ||
2097 | if (cfqq) | |
2098 | __cfq_slice_expired(cfqd, cfqq, timed_out); | |
2099 | } | |
2100 | ||
2101 | /* | |
2102 | * Get next queue for service. Unless we have a queue preemption, | |
2103 | * we'll simply select the first cfqq in the service tree. | |
2104 | */ | |
2105 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) | |
2106 | { | |
2107 | struct cfq_rb_root *service_tree = | |
2108 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, | |
2109 | cfqd->serving_type); | |
2110 | ||
2111 | if (!cfqd->rq_queued) | |
2112 | return NULL; | |
2113 | ||
2114 | /* There is nothing to dispatch */ | |
2115 | if (!service_tree) | |
2116 | return NULL; | |
2117 | if (RB_EMPTY_ROOT(&service_tree->rb)) | |
2118 | return NULL; | |
2119 | return cfq_rb_first(service_tree); | |
2120 | } | |
2121 | ||
2122 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) | |
2123 | { | |
2124 | struct cfq_group *cfqg; | |
2125 | struct cfq_queue *cfqq; | |
2126 | int i, j; | |
2127 | struct cfq_rb_root *st; | |
2128 | ||
2129 | if (!cfqd->rq_queued) | |
2130 | return NULL; | |
2131 | ||
2132 | cfqg = cfq_get_next_cfqg(cfqd); | |
2133 | if (!cfqg) | |
2134 | return NULL; | |
2135 | ||
2136 | for_each_cfqg_st(cfqg, i, j, st) | |
2137 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
2138 | return cfqq; | |
2139 | return NULL; | |
2140 | } | |
2141 | ||
2142 | /* | |
2143 | * Get and set a new active queue for service. | |
2144 | */ | |
2145 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, | |
2146 | struct cfq_queue *cfqq) | |
2147 | { | |
2148 | if (!cfqq) | |
2149 | cfqq = cfq_get_next_queue(cfqd); | |
2150 | ||
2151 | __cfq_set_active_queue(cfqd, cfqq); | |
2152 | return cfqq; | |
2153 | } | |
2154 | ||
2155 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, | |
2156 | struct request *rq) | |
2157 | { | |
2158 | if (blk_rq_pos(rq) >= cfqd->last_position) | |
2159 | return blk_rq_pos(rq) - cfqd->last_position; | |
2160 | else | |
2161 | return cfqd->last_position - blk_rq_pos(rq); | |
2162 | } | |
2163 | ||
2164 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2165 | struct request *rq) | |
2166 | { | |
2167 | return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; | |
2168 | } | |
2169 | ||
2170 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, | |
2171 | struct cfq_queue *cur_cfqq) | |
2172 | { | |
2173 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; | |
2174 | struct rb_node *parent, *node; | |
2175 | struct cfq_queue *__cfqq; | |
2176 | sector_t sector = cfqd->last_position; | |
2177 | ||
2178 | if (RB_EMPTY_ROOT(root)) | |
2179 | return NULL; | |
2180 | ||
2181 | /* | |
2182 | * First, if we find a request starting at the end of the last | |
2183 | * request, choose it. | |
2184 | */ | |
2185 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); | |
2186 | if (__cfqq) | |
2187 | return __cfqq; | |
2188 | ||
2189 | /* | |
2190 | * If the exact sector wasn't found, the parent of the NULL leaf | |
2191 | * will contain the closest sector. | |
2192 | */ | |
2193 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
2194 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
2195 | return __cfqq; | |
2196 | ||
2197 | if (blk_rq_pos(__cfqq->next_rq) < sector) | |
2198 | node = rb_next(&__cfqq->p_node); | |
2199 | else | |
2200 | node = rb_prev(&__cfqq->p_node); | |
2201 | if (!node) | |
2202 | return NULL; | |
2203 | ||
2204 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
2205 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
2206 | return __cfqq; | |
2207 | ||
2208 | return NULL; | |
2209 | } | |
2210 | ||
2211 | /* | |
2212 | * cfqd - obvious | |
2213 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
2214 | * closely cooperating with itself. | |
2215 | * | |
2216 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
2217 | * one request, and that cfqd->last_position reflects a position on the disk | |
2218 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
2219 | * assumption. | |
2220 | */ | |
2221 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
2222 | struct cfq_queue *cur_cfqq) | |
2223 | { | |
2224 | struct cfq_queue *cfqq; | |
2225 | ||
2226 | if (cfq_class_idle(cur_cfqq)) | |
2227 | return NULL; | |
2228 | if (!cfq_cfqq_sync(cur_cfqq)) | |
2229 | return NULL; | |
2230 | if (CFQQ_SEEKY(cur_cfqq)) | |
2231 | return NULL; | |
2232 | ||
2233 | /* | |
2234 | * Don't search priority tree if it's the only queue in the group. | |
2235 | */ | |
2236 | if (cur_cfqq->cfqg->nr_cfqq == 1) | |
2237 | return NULL; | |
2238 | ||
2239 | /* | |
2240 | * We should notice if some of the queues are cooperating, eg | |
2241 | * working closely on the same area of the disk. In that case, | |
2242 | * we can group them together and don't waste time idling. | |
2243 | */ | |
2244 | cfqq = cfqq_close(cfqd, cur_cfqq); | |
2245 | if (!cfqq) | |
2246 | return NULL; | |
2247 | ||
2248 | /* If new queue belongs to different cfq_group, don't choose it */ | |
2249 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
2250 | return NULL; | |
2251 | ||
2252 | /* | |
2253 | * It only makes sense to merge sync queues. | |
2254 | */ | |
2255 | if (!cfq_cfqq_sync(cfqq)) | |
2256 | return NULL; | |
2257 | if (CFQQ_SEEKY(cfqq)) | |
2258 | return NULL; | |
2259 | ||
2260 | /* | |
2261 | * Do not merge queues of different priority classes | |
2262 | */ | |
2263 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
2264 | return NULL; | |
2265 | ||
2266 | return cfqq; | |
2267 | } | |
2268 | ||
2269 | /* | |
2270 | * Determine whether we should enforce idle window for this queue. | |
2271 | */ | |
2272 | ||
2273 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2274 | { | |
2275 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
2276 | struct cfq_rb_root *service_tree = cfqq->service_tree; | |
2277 | ||
2278 | BUG_ON(!service_tree); | |
2279 | BUG_ON(!service_tree->count); | |
2280 | ||
2281 | if (!cfqd->cfq_slice_idle) | |
2282 | return false; | |
2283 | ||
2284 | /* We never do for idle class queues. */ | |
2285 | if (prio == IDLE_WORKLOAD) | |
2286 | return false; | |
2287 | ||
2288 | /* We do for queues that were marked with idle window flag. */ | |
2289 | if (cfq_cfqq_idle_window(cfqq) && | |
2290 | !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) | |
2291 | return true; | |
2292 | ||
2293 | /* | |
2294 | * Otherwise, we do only if they are the last ones | |
2295 | * in their service tree. | |
2296 | */ | |
2297 | if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) && | |
2298 | !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false)) | |
2299 | return true; | |
2300 | cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", | |
2301 | service_tree->count); | |
2302 | return false; | |
2303 | } | |
2304 | ||
2305 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) | |
2306 | { | |
2307 | struct cfq_queue *cfqq = cfqd->active_queue; | |
2308 | struct cfq_io_cq *cic; | |
2309 | unsigned long sl, group_idle = 0; | |
2310 | ||
2311 | /* | |
2312 | * SSD device without seek penalty, disable idling. But only do so | |
2313 | * for devices that support queuing, otherwise we still have a problem | |
2314 | * with sync vs async workloads. | |
2315 | */ | |
2316 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) | |
2317 | return; | |
2318 | ||
2319 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2320 | WARN_ON(cfq_cfqq_slice_new(cfqq)); | |
2321 | ||
2322 | /* | |
2323 | * idle is disabled, either manually or by past process history | |
2324 | */ | |
2325 | if (!cfq_should_idle(cfqd, cfqq)) { | |
2326 | /* no queue idling. Check for group idling */ | |
2327 | if (cfqd->cfq_group_idle) | |
2328 | group_idle = cfqd->cfq_group_idle; | |
2329 | else | |
2330 | return; | |
2331 | } | |
2332 | ||
2333 | /* | |
2334 | * still active requests from this queue, don't idle | |
2335 | */ | |
2336 | if (cfqq->dispatched) | |
2337 | return; | |
2338 | ||
2339 | /* | |
2340 | * task has exited, don't wait | |
2341 | */ | |
2342 | cic = cfqd->active_cic; | |
2343 | if (!cic || !atomic_read(&cic->icq.ioc->active_ref)) | |
2344 | return; | |
2345 | ||
2346 | /* | |
2347 | * If our average think time is larger than the remaining time | |
2348 | * slice, then don't idle. This avoids overrunning the allotted | |
2349 | * time slice. | |
2350 | */ | |
2351 | if (sample_valid(cic->ttime.ttime_samples) && | |
2352 | (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) { | |
2353 | cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu", | |
2354 | cic->ttime.ttime_mean); | |
2355 | return; | |
2356 | } | |
2357 | ||
2358 | /* There are other queues in the group, don't do group idle */ | |
2359 | if (group_idle && cfqq->cfqg->nr_cfqq > 1) | |
2360 | return; | |
2361 | ||
2362 | cfq_mark_cfqq_wait_request(cfqq); | |
2363 | ||
2364 | if (group_idle) | |
2365 | sl = cfqd->cfq_group_idle; | |
2366 | else | |
2367 | sl = cfqd->cfq_slice_idle; | |
2368 | ||
2369 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); | |
2370 | cfqg_stats_set_start_idle_time(cfqq->cfqg); | |
2371 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl, | |
2372 | group_idle ? 1 : 0); | |
2373 | } | |
2374 | ||
2375 | /* | |
2376 | * Move request from internal lists to the request queue dispatch list. | |
2377 | */ | |
2378 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) | |
2379 | { | |
2380 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2381 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
2382 | ||
2383 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); | |
2384 | ||
2385 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); | |
2386 | cfq_remove_request(rq); | |
2387 | cfqq->dispatched++; | |
2388 | (RQ_CFQG(rq))->dispatched++; | |
2389 | elv_dispatch_sort(q, rq); | |
2390 | ||
2391 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; | |
2392 | cfqq->nr_sectors += blk_rq_sectors(rq); | |
2393 | cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags); | |
2394 | } | |
2395 | ||
2396 | /* | |
2397 | * return expired entry, or NULL to just start from scratch in rbtree | |
2398 | */ | |
2399 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) | |
2400 | { | |
2401 | struct request *rq = NULL; | |
2402 | ||
2403 | if (cfq_cfqq_fifo_expire(cfqq)) | |
2404 | return NULL; | |
2405 | ||
2406 | cfq_mark_cfqq_fifo_expire(cfqq); | |
2407 | ||
2408 | if (list_empty(&cfqq->fifo)) | |
2409 | return NULL; | |
2410 | ||
2411 | rq = rq_entry_fifo(cfqq->fifo.next); | |
2412 | if (time_before(jiffies, rq_fifo_time(rq))) | |
2413 | rq = NULL; | |
2414 | ||
2415 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); | |
2416 | return rq; | |
2417 | } | |
2418 | ||
2419 | static inline int | |
2420 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2421 | { | |
2422 | const int base_rq = cfqd->cfq_slice_async_rq; | |
2423 | ||
2424 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); | |
2425 | ||
2426 | return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); | |
2427 | } | |
2428 | ||
2429 | /* | |
2430 | * Must be called with the queue_lock held. | |
2431 | */ | |
2432 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
2433 | { | |
2434 | int process_refs, io_refs; | |
2435 | ||
2436 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
2437 | process_refs = cfqq->ref - io_refs; | |
2438 | BUG_ON(process_refs < 0); | |
2439 | return process_refs; | |
2440 | } | |
2441 | ||
2442 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
2443 | { | |
2444 | int process_refs, new_process_refs; | |
2445 | struct cfq_queue *__cfqq; | |
2446 | ||
2447 | /* | |
2448 | * If there are no process references on the new_cfqq, then it is | |
2449 | * unsafe to follow the ->new_cfqq chain as other cfqq's in the | |
2450 | * chain may have dropped their last reference (not just their | |
2451 | * last process reference). | |
2452 | */ | |
2453 | if (!cfqq_process_refs(new_cfqq)) | |
2454 | return; | |
2455 | ||
2456 | /* Avoid a circular list and skip interim queue merges */ | |
2457 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
2458 | if (__cfqq == cfqq) | |
2459 | return; | |
2460 | new_cfqq = __cfqq; | |
2461 | } | |
2462 | ||
2463 | process_refs = cfqq_process_refs(cfqq); | |
2464 | new_process_refs = cfqq_process_refs(new_cfqq); | |
2465 | /* | |
2466 | * If the process for the cfqq has gone away, there is no | |
2467 | * sense in merging the queues. | |
2468 | */ | |
2469 | if (process_refs == 0 || new_process_refs == 0) | |
2470 | return; | |
2471 | ||
2472 | /* | |
2473 | * Merge in the direction of the lesser amount of work. | |
2474 | */ | |
2475 | if (new_process_refs >= process_refs) { | |
2476 | cfqq->new_cfqq = new_cfqq; | |
2477 | new_cfqq->ref += process_refs; | |
2478 | } else { | |
2479 | new_cfqq->new_cfqq = cfqq; | |
2480 | cfqq->ref += new_process_refs; | |
2481 | } | |
2482 | } | |
2483 | ||
2484 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, | |
2485 | struct cfq_group *cfqg, enum wl_prio_t prio) | |
2486 | { | |
2487 | struct cfq_queue *queue; | |
2488 | int i; | |
2489 | bool key_valid = false; | |
2490 | unsigned long lowest_key = 0; | |
2491 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
2492 | ||
2493 | for (i = 0; i <= SYNC_WORKLOAD; ++i) { | |
2494 | /* select the one with lowest rb_key */ | |
2495 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i)); | |
2496 | if (queue && | |
2497 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
2498 | lowest_key = queue->rb_key; | |
2499 | cur_best = i; | |
2500 | key_valid = true; | |
2501 | } | |
2502 | } | |
2503 | ||
2504 | return cur_best; | |
2505 | } | |
2506 | ||
2507 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
2508 | { | |
2509 | unsigned slice; | |
2510 | unsigned count; | |
2511 | struct cfq_rb_root *st; | |
2512 | unsigned group_slice; | |
2513 | enum wl_prio_t original_prio = cfqd->serving_prio; | |
2514 | ||
2515 | /* Choose next priority. RT > BE > IDLE */ | |
2516 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) | |
2517 | cfqd->serving_prio = RT_WORKLOAD; | |
2518 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) | |
2519 | cfqd->serving_prio = BE_WORKLOAD; | |
2520 | else { | |
2521 | cfqd->serving_prio = IDLE_WORKLOAD; | |
2522 | cfqd->workload_expires = jiffies + 1; | |
2523 | return; | |
2524 | } | |
2525 | ||
2526 | if (original_prio != cfqd->serving_prio) | |
2527 | goto new_workload; | |
2528 | ||
2529 | /* | |
2530 | * For RT and BE, we have to choose also the type | |
2531 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
2532 | * expiration time | |
2533 | */ | |
2534 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); | |
2535 | count = st->count; | |
2536 | ||
2537 | /* | |
2538 | * check workload expiration, and that we still have other queues ready | |
2539 | */ | |
2540 | if (count && !time_after(jiffies, cfqd->workload_expires)) | |
2541 | return; | |
2542 | ||
2543 | new_workload: | |
2544 | /* otherwise select new workload type */ | |
2545 | cfqd->serving_type = | |
2546 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio); | |
2547 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type); | |
2548 | count = st->count; | |
2549 | ||
2550 | /* | |
2551 | * the workload slice is computed as a fraction of target latency | |
2552 | * proportional to the number of queues in that workload, over | |
2553 | * all the queues in the same priority class | |
2554 | */ | |
2555 | group_slice = cfq_group_slice(cfqd, cfqg); | |
2556 | ||
2557 | slice = group_slice * count / | |
2558 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
2559 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
2560 | ||
2561 | if (cfqd->serving_type == ASYNC_WORKLOAD) { | |
2562 | unsigned int tmp; | |
2563 | ||
2564 | /* | |
2565 | * Async queues are currently system wide. Just taking | |
2566 | * proportion of queues with-in same group will lead to higher | |
2567 | * async ratio system wide as generally root group is going | |
2568 | * to have higher weight. A more accurate thing would be to | |
2569 | * calculate system wide asnc/sync ratio. | |
2570 | */ | |
2571 | tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg); | |
2572 | tmp = tmp/cfqd->busy_queues; | |
2573 | slice = min_t(unsigned, slice, tmp); | |
2574 | ||
2575 | /* async workload slice is scaled down according to | |
2576 | * the sync/async slice ratio. */ | |
2577 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
2578 | } else | |
2579 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
2580 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2581 | ||
2582 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
2583 | cfq_log(cfqd, "workload slice:%d", slice); | |
2584 | cfqd->workload_expires = jiffies + slice; | |
2585 | } | |
2586 | ||
2587 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) | |
2588 | { | |
2589 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
2590 | struct cfq_group *cfqg; | |
2591 | ||
2592 | if (RB_EMPTY_ROOT(&st->rb)) | |
2593 | return NULL; | |
2594 | cfqg = cfq_rb_first_group(st); | |
2595 | update_min_vdisktime(st); | |
2596 | return cfqg; | |
2597 | } | |
2598 | ||
2599 | static void cfq_choose_cfqg(struct cfq_data *cfqd) | |
2600 | { | |
2601 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); | |
2602 | ||
2603 | cfqd->serving_group = cfqg; | |
2604 | ||
2605 | /* Restore the workload type data */ | |
2606 | if (cfqg->saved_workload_slice) { | |
2607 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2608 | cfqd->serving_type = cfqg->saved_workload; | |
2609 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
2610 | } else | |
2611 | cfqd->workload_expires = jiffies - 1; | |
2612 | ||
2613 | choose_service_tree(cfqd, cfqg); | |
2614 | } | |
2615 | ||
2616 | /* | |
2617 | * Select a queue for service. If we have a current active queue, | |
2618 | * check whether to continue servicing it, or retrieve and set a new one. | |
2619 | */ | |
2620 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) | |
2621 | { | |
2622 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
2623 | ||
2624 | cfqq = cfqd->active_queue; | |
2625 | if (!cfqq) | |
2626 | goto new_queue; | |
2627 | ||
2628 | if (!cfqd->rq_queued) | |
2629 | return NULL; | |
2630 | ||
2631 | /* | |
2632 | * We were waiting for group to get backlogged. Expire the queue | |
2633 | */ | |
2634 | if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2635 | goto expire; | |
2636 | ||
2637 | /* | |
2638 | * The active queue has run out of time, expire it and select new. | |
2639 | */ | |
2640 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { | |
2641 | /* | |
2642 | * If slice had not expired at the completion of last request | |
2643 | * we might not have turned on wait_busy flag. Don't expire | |
2644 | * the queue yet. Allow the group to get backlogged. | |
2645 | * | |
2646 | * The very fact that we have used the slice, that means we | |
2647 | * have been idling all along on this queue and it should be | |
2648 | * ok to wait for this request to complete. | |
2649 | */ | |
2650 | if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) | |
2651 | && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
2652 | cfqq = NULL; | |
2653 | goto keep_queue; | |
2654 | } else | |
2655 | goto check_group_idle; | |
2656 | } | |
2657 | ||
2658 | /* | |
2659 | * The active queue has requests and isn't expired, allow it to | |
2660 | * dispatch. | |
2661 | */ | |
2662 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2663 | goto keep_queue; | |
2664 | ||
2665 | /* | |
2666 | * If another queue has a request waiting within our mean seek | |
2667 | * distance, let it run. The expire code will check for close | |
2668 | * cooperators and put the close queue at the front of the service | |
2669 | * tree. If possible, merge the expiring queue with the new cfqq. | |
2670 | */ | |
2671 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); | |
2672 | if (new_cfqq) { | |
2673 | if (!cfqq->new_cfqq) | |
2674 | cfq_setup_merge(cfqq, new_cfqq); | |
2675 | goto expire; | |
2676 | } | |
2677 | ||
2678 | /* | |
2679 | * No requests pending. If the active queue still has requests in | |
2680 | * flight or is idling for a new request, allow either of these | |
2681 | * conditions to happen (or time out) before selecting a new queue. | |
2682 | */ | |
2683 | if (timer_pending(&cfqd->idle_slice_timer)) { | |
2684 | cfqq = NULL; | |
2685 | goto keep_queue; | |
2686 | } | |
2687 | ||
2688 | /* | |
2689 | * This is a deep seek queue, but the device is much faster than | |
2690 | * the queue can deliver, don't idle | |
2691 | **/ | |
2692 | if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && | |
2693 | (cfq_cfqq_slice_new(cfqq) || | |
2694 | (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) { | |
2695 | cfq_clear_cfqq_deep(cfqq); | |
2696 | cfq_clear_cfqq_idle_window(cfqq); | |
2697 | } | |
2698 | ||
2699 | if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
2700 | cfqq = NULL; | |
2701 | goto keep_queue; | |
2702 | } | |
2703 | ||
2704 | /* | |
2705 | * If group idle is enabled and there are requests dispatched from | |
2706 | * this group, wait for requests to complete. | |
2707 | */ | |
2708 | check_group_idle: | |
2709 | if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && | |
2710 | cfqq->cfqg->dispatched && | |
2711 | !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { | |
2712 | cfqq = NULL; | |
2713 | goto keep_queue; | |
2714 | } | |
2715 | ||
2716 | expire: | |
2717 | cfq_slice_expired(cfqd, 0); | |
2718 | new_queue: | |
2719 | /* | |
2720 | * Current queue expired. Check if we have to switch to a new | |
2721 | * service tree | |
2722 | */ | |
2723 | if (!new_cfqq) | |
2724 | cfq_choose_cfqg(cfqd); | |
2725 | ||
2726 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); | |
2727 | keep_queue: | |
2728 | return cfqq; | |
2729 | } | |
2730 | ||
2731 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) | |
2732 | { | |
2733 | int dispatched = 0; | |
2734 | ||
2735 | while (cfqq->next_rq) { | |
2736 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2737 | dispatched++; | |
2738 | } | |
2739 | ||
2740 | BUG_ON(!list_empty(&cfqq->fifo)); | |
2741 | ||
2742 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
2743 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
2744 | return dispatched; | |
2745 | } | |
2746 | ||
2747 | /* | |
2748 | * Drain our current requests. Used for barriers and when switching | |
2749 | * io schedulers on-the-fly. | |
2750 | */ | |
2751 | static int cfq_forced_dispatch(struct cfq_data *cfqd) | |
2752 | { | |
2753 | struct cfq_queue *cfqq; | |
2754 | int dispatched = 0; | |
2755 | ||
2756 | /* Expire the timeslice of the current active queue first */ | |
2757 | cfq_slice_expired(cfqd, 0); | |
2758 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { | |
2759 | __cfq_set_active_queue(cfqd, cfqq); | |
2760 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
2761 | } | |
2762 | ||
2763 | BUG_ON(cfqd->busy_queues); | |
2764 | ||
2765 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); | |
2766 | return dispatched; | |
2767 | } | |
2768 | ||
2769 | static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, | |
2770 | struct cfq_queue *cfqq) | |
2771 | { | |
2772 | /* the queue hasn't finished any request, can't estimate */ | |
2773 | if (cfq_cfqq_slice_new(cfqq)) | |
2774 | return true; | |
2775 | if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched, | |
2776 | cfqq->slice_end)) | |
2777 | return true; | |
2778 | ||
2779 | return false; | |
2780 | } | |
2781 | ||
2782 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2783 | { | |
2784 | unsigned int max_dispatch; | |
2785 | ||
2786 | /* | |
2787 | * Drain async requests before we start sync IO | |
2788 | */ | |
2789 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) | |
2790 | return false; | |
2791 | ||
2792 | /* | |
2793 | * If this is an async queue and we have sync IO in flight, let it wait | |
2794 | */ | |
2795 | if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) | |
2796 | return false; | |
2797 | ||
2798 | max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); | |
2799 | if (cfq_class_idle(cfqq)) | |
2800 | max_dispatch = 1; | |
2801 | ||
2802 | /* | |
2803 | * Does this cfqq already have too much IO in flight? | |
2804 | */ | |
2805 | if (cfqq->dispatched >= max_dispatch) { | |
2806 | bool promote_sync = false; | |
2807 | /* | |
2808 | * idle queue must always only have a single IO in flight | |
2809 | */ | |
2810 | if (cfq_class_idle(cfqq)) | |
2811 | return false; | |
2812 | ||
2813 | /* | |
2814 | * If there is only one sync queue | |
2815 | * we can ignore async queue here and give the sync | |
2816 | * queue no dispatch limit. The reason is a sync queue can | |
2817 | * preempt async queue, limiting the sync queue doesn't make | |
2818 | * sense. This is useful for aiostress test. | |
2819 | */ | |
2820 | if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) | |
2821 | promote_sync = true; | |
2822 | ||
2823 | /* | |
2824 | * We have other queues, don't allow more IO from this one | |
2825 | */ | |
2826 | if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && | |
2827 | !promote_sync) | |
2828 | return false; | |
2829 | ||
2830 | /* | |
2831 | * Sole queue user, no limit | |
2832 | */ | |
2833 | if (cfqd->busy_queues == 1 || promote_sync) | |
2834 | max_dispatch = -1; | |
2835 | else | |
2836 | /* | |
2837 | * Normally we start throttling cfqq when cfq_quantum/2 | |
2838 | * requests have been dispatched. But we can drive | |
2839 | * deeper queue depths at the beginning of slice | |
2840 | * subjected to upper limit of cfq_quantum. | |
2841 | * */ | |
2842 | max_dispatch = cfqd->cfq_quantum; | |
2843 | } | |
2844 | ||
2845 | /* | |
2846 | * Async queues must wait a bit before being allowed dispatch. | |
2847 | * We also ramp up the dispatch depth gradually for async IO, | |
2848 | * based on the last sync IO we serviced | |
2849 | */ | |
2850 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { | |
2851 | unsigned long last_sync = jiffies - cfqd->last_delayed_sync; | |
2852 | unsigned int depth; | |
2853 | ||
2854 | depth = last_sync / cfqd->cfq_slice[1]; | |
2855 | if (!depth && !cfqq->dispatched) | |
2856 | depth = 1; | |
2857 | if (depth < max_dispatch) | |
2858 | max_dispatch = depth; | |
2859 | } | |
2860 | ||
2861 | /* | |
2862 | * If we're below the current max, allow a dispatch | |
2863 | */ | |
2864 | return cfqq->dispatched < max_dispatch; | |
2865 | } | |
2866 | ||
2867 | /* | |
2868 | * Dispatch a request from cfqq, moving them to the request queue | |
2869 | * dispatch list. | |
2870 | */ | |
2871 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2872 | { | |
2873 | struct request *rq; | |
2874 | ||
2875 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2876 | ||
2877 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2878 | return false; | |
2879 | ||
2880 | /* | |
2881 | * follow expired path, else get first next available | |
2882 | */ | |
2883 | rq = cfq_check_fifo(cfqq); | |
2884 | if (!rq) | |
2885 | rq = cfqq->next_rq; | |
2886 | ||
2887 | /* | |
2888 | * insert request into driver dispatch list | |
2889 | */ | |
2890 | cfq_dispatch_insert(cfqd->queue, rq); | |
2891 | ||
2892 | if (!cfqd->active_cic) { | |
2893 | struct cfq_io_cq *cic = RQ_CIC(rq); | |
2894 | ||
2895 | atomic_long_inc(&cic->icq.ioc->refcount); | |
2896 | cfqd->active_cic = cic; | |
2897 | } | |
2898 | ||
2899 | return true; | |
2900 | } | |
2901 | ||
2902 | /* | |
2903 | * Find the cfqq that we need to service and move a request from that to the | |
2904 | * dispatch list | |
2905 | */ | |
2906 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2907 | { | |
2908 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2909 | struct cfq_queue *cfqq; | |
2910 | ||
2911 | if (!cfqd->busy_queues) | |
2912 | return 0; | |
2913 | ||
2914 | if (unlikely(force)) | |
2915 | return cfq_forced_dispatch(cfqd); | |
2916 | ||
2917 | cfqq = cfq_select_queue(cfqd); | |
2918 | if (!cfqq) | |
2919 | return 0; | |
2920 | ||
2921 | /* | |
2922 | * Dispatch a request from this cfqq, if it is allowed | |
2923 | */ | |
2924 | if (!cfq_dispatch_request(cfqd, cfqq)) | |
2925 | return 0; | |
2926 | ||
2927 | cfqq->slice_dispatch++; | |
2928 | cfq_clear_cfqq_must_dispatch(cfqq); | |
2929 | ||
2930 | /* | |
2931 | * expire an async queue immediately if it has used up its slice. idle | |
2932 | * queue always expire after 1 dispatch round. | |
2933 | */ | |
2934 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2935 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2936 | cfq_class_idle(cfqq))) { | |
2937 | cfqq->slice_end = jiffies + 1; | |
2938 | cfq_slice_expired(cfqd, 0); | |
2939 | } | |
2940 | ||
2941 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); | |
2942 | return 1; | |
2943 | } | |
2944 | ||
2945 | /* | |
2946 | * task holds one reference to the queue, dropped when task exits. each rq | |
2947 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
2948 | * | |
2949 | * Each cfq queue took a reference on the parent group. Drop it now. | |
2950 | * queue lock must be held here. | |
2951 | */ | |
2952 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2953 | { | |
2954 | struct cfq_data *cfqd = cfqq->cfqd; | |
2955 | struct cfq_group *cfqg; | |
2956 | ||
2957 | BUG_ON(cfqq->ref <= 0); | |
2958 | ||
2959 | cfqq->ref--; | |
2960 | if (cfqq->ref) | |
2961 | return; | |
2962 | ||
2963 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); | |
2964 | BUG_ON(rb_first(&cfqq->sort_list)); | |
2965 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); | |
2966 | cfqg = cfqq->cfqg; | |
2967 | ||
2968 | if (unlikely(cfqd->active_queue == cfqq)) { | |
2969 | __cfq_slice_expired(cfqd, cfqq, 0); | |
2970 | cfq_schedule_dispatch(cfqd); | |
2971 | } | |
2972 | ||
2973 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
2974 | kmem_cache_free(cfq_pool, cfqq); | |
2975 | cfqg_put(cfqg); | |
2976 | } | |
2977 | ||
2978 | static void cfq_put_cooperator(struct cfq_queue *cfqq) | |
2979 | { | |
2980 | struct cfq_queue *__cfqq, *next; | |
2981 | ||
2982 | /* | |
2983 | * If this queue was scheduled to merge with another queue, be | |
2984 | * sure to drop the reference taken on that queue (and others in | |
2985 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2986 | */ | |
2987 | __cfqq = cfqq->new_cfqq; | |
2988 | while (__cfqq) { | |
2989 | if (__cfqq == cfqq) { | |
2990 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2991 | break; | |
2992 | } | |
2993 | next = __cfqq->new_cfqq; | |
2994 | cfq_put_queue(__cfqq); | |
2995 | __cfqq = next; | |
2996 | } | |
2997 | } | |
2998 | ||
2999 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3000 | { | |
3001 | if (unlikely(cfqq == cfqd->active_queue)) { | |
3002 | __cfq_slice_expired(cfqd, cfqq, 0); | |
3003 | cfq_schedule_dispatch(cfqd); | |
3004 | } | |
3005 | ||
3006 | cfq_put_cooperator(cfqq); | |
3007 | ||
3008 | cfq_put_queue(cfqq); | |
3009 | } | |
3010 | ||
3011 | static void cfq_init_icq(struct io_cq *icq) | |
3012 | { | |
3013 | struct cfq_io_cq *cic = icq_to_cic(icq); | |
3014 | ||
3015 | cic->ttime.last_end_request = jiffies; | |
3016 | } | |
3017 | ||
3018 | static void cfq_exit_icq(struct io_cq *icq) | |
3019 | { | |
3020 | struct cfq_io_cq *cic = icq_to_cic(icq); | |
3021 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3022 | ||
3023 | if (cic->cfqq[BLK_RW_ASYNC]) { | |
3024 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
3025 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
3026 | } | |
3027 | ||
3028 | if (cic->cfqq[BLK_RW_SYNC]) { | |
3029 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
3030 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
3031 | } | |
3032 | } | |
3033 | ||
3034 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic) | |
3035 | { | |
3036 | struct task_struct *tsk = current; | |
3037 | int ioprio_class; | |
3038 | ||
3039 | if (!cfq_cfqq_prio_changed(cfqq)) | |
3040 | return; | |
3041 | ||
3042 | ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio); | |
3043 | switch (ioprio_class) { | |
3044 | default: | |
3045 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
3046 | case IOPRIO_CLASS_NONE: | |
3047 | /* | |
3048 | * no prio set, inherit CPU scheduling settings | |
3049 | */ | |
3050 | cfqq->ioprio = task_nice_ioprio(tsk); | |
3051 | cfqq->ioprio_class = task_nice_ioclass(tsk); | |
3052 | break; | |
3053 | case IOPRIO_CLASS_RT: | |
3054 | cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3055 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
3056 | break; | |
3057 | case IOPRIO_CLASS_BE: | |
3058 | cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3059 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3060 | break; | |
3061 | case IOPRIO_CLASS_IDLE: | |
3062 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
3063 | cfqq->ioprio = 7; | |
3064 | cfq_clear_cfqq_idle_window(cfqq); | |
3065 | break; | |
3066 | } | |
3067 | ||
3068 | /* | |
3069 | * keep track of original prio settings in case we have to temporarily | |
3070 | * elevate the priority of this queue | |
3071 | */ | |
3072 | cfqq->org_ioprio = cfqq->ioprio; | |
3073 | cfq_clear_cfqq_prio_changed(cfqq); | |
3074 | } | |
3075 | ||
3076 | static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio) | |
3077 | { | |
3078 | int ioprio = cic->icq.ioc->ioprio; | |
3079 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3080 | struct cfq_queue *cfqq; | |
3081 | ||
3082 | /* | |
3083 | * Check whether ioprio has changed. The condition may trigger | |
3084 | * spuriously on a newly created cic but there's no harm. | |
3085 | */ | |
3086 | if (unlikely(!cfqd) || likely(cic->ioprio == ioprio)) | |
3087 | return; | |
3088 | ||
3089 | cfqq = cic->cfqq[BLK_RW_ASYNC]; | |
3090 | if (cfqq) { | |
3091 | struct cfq_queue *new_cfqq; | |
3092 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio, | |
3093 | GFP_ATOMIC); | |
3094 | if (new_cfqq) { | |
3095 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; | |
3096 | cfq_put_queue(cfqq); | |
3097 | } | |
3098 | } | |
3099 | ||
3100 | cfqq = cic->cfqq[BLK_RW_SYNC]; | |
3101 | if (cfqq) | |
3102 | cfq_mark_cfqq_prio_changed(cfqq); | |
3103 | ||
3104 | cic->ioprio = ioprio; | |
3105 | } | |
3106 | ||
3107 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3108 | pid_t pid, bool is_sync) | |
3109 | { | |
3110 | RB_CLEAR_NODE(&cfqq->rb_node); | |
3111 | RB_CLEAR_NODE(&cfqq->p_node); | |
3112 | INIT_LIST_HEAD(&cfqq->fifo); | |
3113 | ||
3114 | cfqq->ref = 0; | |
3115 | cfqq->cfqd = cfqd; | |
3116 | ||
3117 | cfq_mark_cfqq_prio_changed(cfqq); | |
3118 | ||
3119 | if (is_sync) { | |
3120 | if (!cfq_class_idle(cfqq)) | |
3121 | cfq_mark_cfqq_idle_window(cfqq); | |
3122 | cfq_mark_cfqq_sync(cfqq); | |
3123 | } | |
3124 | cfqq->pid = pid; | |
3125 | } | |
3126 | ||
3127 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3128 | static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) | |
3129 | { | |
3130 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3131 | struct cfq_queue *sync_cfqq; | |
3132 | uint64_t id; | |
3133 | ||
3134 | rcu_read_lock(); | |
3135 | id = bio_blkio_cgroup(bio)->id; | |
3136 | rcu_read_unlock(); | |
3137 | ||
3138 | /* | |
3139 | * Check whether blkcg has changed. The condition may trigger | |
3140 | * spuriously on a newly created cic but there's no harm. | |
3141 | */ | |
3142 | if (unlikely(!cfqd) || likely(cic->blkcg_id == id)) | |
3143 | return; | |
3144 | ||
3145 | sync_cfqq = cic_to_cfqq(cic, 1); | |
3146 | if (sync_cfqq) { | |
3147 | /* | |
3148 | * Drop reference to sync queue. A new sync queue will be | |
3149 | * assigned in new group upon arrival of a fresh request. | |
3150 | */ | |
3151 | cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup"); | |
3152 | cic_set_cfqq(cic, NULL, 1); | |
3153 | cfq_put_queue(sync_cfqq); | |
3154 | } | |
3155 | ||
3156 | cic->blkcg_id = id; | |
3157 | } | |
3158 | #else | |
3159 | static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { } | |
3160 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
3161 | ||
3162 | static struct cfq_queue * | |
3163 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic, | |
3164 | struct bio *bio, gfp_t gfp_mask) | |
3165 | { | |
3166 | struct blkio_cgroup *blkcg; | |
3167 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
3168 | struct cfq_group *cfqg; | |
3169 | ||
3170 | retry: | |
3171 | rcu_read_lock(); | |
3172 | ||
3173 | blkcg = bio_blkio_cgroup(bio); | |
3174 | cfqg = cfq_lookup_create_cfqg(cfqd, blkcg); | |
3175 | cfqq = cic_to_cfqq(cic, is_sync); | |
3176 | ||
3177 | /* | |
3178 | * Always try a new alloc if we fell back to the OOM cfqq | |
3179 | * originally, since it should just be a temporary situation. | |
3180 | */ | |
3181 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
3182 | cfqq = NULL; | |
3183 | if (new_cfqq) { | |
3184 | cfqq = new_cfqq; | |
3185 | new_cfqq = NULL; | |
3186 | } else if (gfp_mask & __GFP_WAIT) { | |
3187 | rcu_read_unlock(); | |
3188 | spin_unlock_irq(cfqd->queue->queue_lock); | |
3189 | new_cfqq = kmem_cache_alloc_node(cfq_pool, | |
3190 | gfp_mask | __GFP_ZERO, | |
3191 | cfqd->queue->node); | |
3192 | spin_lock_irq(cfqd->queue->queue_lock); | |
3193 | if (new_cfqq) | |
3194 | goto retry; | |
3195 | } else { | |
3196 | cfqq = kmem_cache_alloc_node(cfq_pool, | |
3197 | gfp_mask | __GFP_ZERO, | |
3198 | cfqd->queue->node); | |
3199 | } | |
3200 | ||
3201 | if (cfqq) { | |
3202 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
3203 | cfq_init_prio_data(cfqq, cic); | |
3204 | cfq_link_cfqq_cfqg(cfqq, cfqg); | |
3205 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
3206 | } else | |
3207 | cfqq = &cfqd->oom_cfqq; | |
3208 | } | |
3209 | ||
3210 | if (new_cfqq) | |
3211 | kmem_cache_free(cfq_pool, new_cfqq); | |
3212 | ||
3213 | rcu_read_unlock(); | |
3214 | return cfqq; | |
3215 | } | |
3216 | ||
3217 | static struct cfq_queue ** | |
3218 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
3219 | { | |
3220 | switch (ioprio_class) { | |
3221 | case IOPRIO_CLASS_RT: | |
3222 | return &cfqd->async_cfqq[0][ioprio]; | |
3223 | case IOPRIO_CLASS_NONE: | |
3224 | ioprio = IOPRIO_NORM; | |
3225 | /* fall through */ | |
3226 | case IOPRIO_CLASS_BE: | |
3227 | return &cfqd->async_cfqq[1][ioprio]; | |
3228 | case IOPRIO_CLASS_IDLE: | |
3229 | return &cfqd->async_idle_cfqq; | |
3230 | default: | |
3231 | BUG(); | |
3232 | } | |
3233 | } | |
3234 | ||
3235 | static struct cfq_queue * | |
3236 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic, | |
3237 | struct bio *bio, gfp_t gfp_mask) | |
3238 | { | |
3239 | const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio); | |
3240 | const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3241 | struct cfq_queue **async_cfqq = NULL; | |
3242 | struct cfq_queue *cfqq = NULL; | |
3243 | ||
3244 | if (!is_sync) { | |
3245 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
3246 | cfqq = *async_cfqq; | |
3247 | } | |
3248 | ||
3249 | if (!cfqq) | |
3250 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask); | |
3251 | ||
3252 | /* | |
3253 | * pin the queue now that it's allocated, scheduler exit will prune it | |
3254 | */ | |
3255 | if (!is_sync && !(*async_cfqq)) { | |
3256 | cfqq->ref++; | |
3257 | *async_cfqq = cfqq; | |
3258 | } | |
3259 | ||
3260 | cfqq->ref++; | |
3261 | return cfqq; | |
3262 | } | |
3263 | ||
3264 | static void | |
3265 | __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle) | |
3266 | { | |
3267 | unsigned long elapsed = jiffies - ttime->last_end_request; | |
3268 | elapsed = min(elapsed, 2UL * slice_idle); | |
3269 | ||
3270 | ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; | |
3271 | ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8; | |
3272 | ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples; | |
3273 | } | |
3274 | ||
3275 | static void | |
3276 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3277 | struct cfq_io_cq *cic) | |
3278 | { | |
3279 | if (cfq_cfqq_sync(cfqq)) { | |
3280 | __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); | |
3281 | __cfq_update_io_thinktime(&cfqq->service_tree->ttime, | |
3282 | cfqd->cfq_slice_idle); | |
3283 | } | |
3284 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3285 | __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); | |
3286 | #endif | |
3287 | } | |
3288 | ||
3289 | static void | |
3290 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3291 | struct request *rq) | |
3292 | { | |
3293 | sector_t sdist = 0; | |
3294 | sector_t n_sec = blk_rq_sectors(rq); | |
3295 | if (cfqq->last_request_pos) { | |
3296 | if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
3297 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
3298 | else | |
3299 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
3300 | } | |
3301 | ||
3302 | cfqq->seek_history <<= 1; | |
3303 | if (blk_queue_nonrot(cfqd->queue)) | |
3304 | cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); | |
3305 | else | |
3306 | cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); | |
3307 | } | |
3308 | ||
3309 | /* | |
3310 | * Disable idle window if the process thinks too long or seeks so much that | |
3311 | * it doesn't matter | |
3312 | */ | |
3313 | static void | |
3314 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3315 | struct cfq_io_cq *cic) | |
3316 | { | |
3317 | int old_idle, enable_idle; | |
3318 | ||
3319 | /* | |
3320 | * Don't idle for async or idle io prio class | |
3321 | */ | |
3322 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
3323 | return; | |
3324 | ||
3325 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); | |
3326 | ||
3327 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) | |
3328 | cfq_mark_cfqq_deep(cfqq); | |
3329 | ||
3330 | if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) | |
3331 | enable_idle = 0; | |
3332 | else if (!atomic_read(&cic->icq.ioc->active_ref) || | |
3333 | !cfqd->cfq_slice_idle || | |
3334 | (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) | |
3335 | enable_idle = 0; | |
3336 | else if (sample_valid(cic->ttime.ttime_samples)) { | |
3337 | if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) | |
3338 | enable_idle = 0; | |
3339 | else | |
3340 | enable_idle = 1; | |
3341 | } | |
3342 | ||
3343 | if (old_idle != enable_idle) { | |
3344 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3345 | if (enable_idle) | |
3346 | cfq_mark_cfqq_idle_window(cfqq); | |
3347 | else | |
3348 | cfq_clear_cfqq_idle_window(cfqq); | |
3349 | } | |
3350 | } | |
3351 | ||
3352 | /* | |
3353 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3354 | * no or if we aren't sure, a 1 will cause a preempt. | |
3355 | */ | |
3356 | static bool | |
3357 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
3358 | struct request *rq) | |
3359 | { | |
3360 | struct cfq_queue *cfqq; | |
3361 | ||
3362 | cfqq = cfqd->active_queue; | |
3363 | if (!cfqq) | |
3364 | return false; | |
3365 | ||
3366 | if (cfq_class_idle(new_cfqq)) | |
3367 | return false; | |
3368 | ||
3369 | if (cfq_class_idle(cfqq)) | |
3370 | return true; | |
3371 | ||
3372 | /* | |
3373 | * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. | |
3374 | */ | |
3375 | if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) | |
3376 | return false; | |
3377 | ||
3378 | /* | |
3379 | * if the new request is sync, but the currently running queue is | |
3380 | * not, let the sync request have priority. | |
3381 | */ | |
3382 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) | |
3383 | return true; | |
3384 | ||
3385 | if (new_cfqq->cfqg != cfqq->cfqg) | |
3386 | return false; | |
3387 | ||
3388 | if (cfq_slice_used(cfqq)) | |
3389 | return true; | |
3390 | ||
3391 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
3392 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
3393 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
3394 | new_cfqq->service_tree->count == 2 && | |
3395 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3396 | return true; | |
3397 | ||
3398 | /* | |
3399 | * So both queues are sync. Let the new request get disk time if | |
3400 | * it's a metadata request and the current queue is doing regular IO. | |
3401 | */ | |
3402 | if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) | |
3403 | return true; | |
3404 | ||
3405 | /* | |
3406 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3407 | */ | |
3408 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
3409 | return true; | |
3410 | ||
3411 | /* An idle queue should not be idle now for some reason */ | |
3412 | if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) | |
3413 | return true; | |
3414 | ||
3415 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) | |
3416 | return false; | |
3417 | ||
3418 | /* | |
3419 | * if this request is as-good as one we would expect from the | |
3420 | * current cfqq, let it preempt | |
3421 | */ | |
3422 | if (cfq_rq_close(cfqd, cfqq, rq)) | |
3423 | return true; | |
3424 | ||
3425 | return false; | |
3426 | } | |
3427 | ||
3428 | /* | |
3429 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3430 | * let it have half of its nominal slice. | |
3431 | */ | |
3432 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3433 | { | |
3434 | enum wl_type_t old_type = cfqq_type(cfqd->active_queue); | |
3435 | ||
3436 | cfq_log_cfqq(cfqd, cfqq, "preempt"); | |
3437 | cfq_slice_expired(cfqd, 1); | |
3438 | ||
3439 | /* | |
3440 | * workload type is changed, don't save slice, otherwise preempt | |
3441 | * doesn't happen | |
3442 | */ | |
3443 | if (old_type != cfqq_type(cfqq)) | |
3444 | cfqq->cfqg->saved_workload_slice = 0; | |
3445 | ||
3446 | /* | |
3447 | * Put the new queue at the front of the of the current list, | |
3448 | * so we know that it will be selected next. | |
3449 | */ | |
3450 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
3451 | ||
3452 | cfq_service_tree_add(cfqd, cfqq, 1); | |
3453 | ||
3454 | cfqq->slice_end = 0; | |
3455 | cfq_mark_cfqq_slice_new(cfqq); | |
3456 | } | |
3457 | ||
3458 | /* | |
3459 | * Called when a new fs request (rq) is added (to cfqq). Check if there's | |
3460 | * something we should do about it | |
3461 | */ | |
3462 | static void | |
3463 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3464 | struct request *rq) | |
3465 | { | |
3466 | struct cfq_io_cq *cic = RQ_CIC(rq); | |
3467 | ||
3468 | cfqd->rq_queued++; | |
3469 | if (rq->cmd_flags & REQ_PRIO) | |
3470 | cfqq->prio_pending++; | |
3471 | ||
3472 | cfq_update_io_thinktime(cfqd, cfqq, cic); | |
3473 | cfq_update_io_seektime(cfqd, cfqq, rq); | |
3474 | cfq_update_idle_window(cfqd, cfqq, cic); | |
3475 | ||
3476 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
3477 | ||
3478 | if (cfqq == cfqd->active_queue) { | |
3479 | /* | |
3480 | * Remember that we saw a request from this process, but | |
3481 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3482 | * of tiny requests, because we disrupt the normal plugging | |
3483 | * and merging. If the request is already larger than a single | |
3484 | * page, let it rip immediately. For that case we assume that | |
3485 | * merging is already done. Ditto for a busy system that | |
3486 | * has other work pending, don't risk delaying until the | |
3487 | * idle timer unplug to continue working. | |
3488 | */ | |
3489 | if (cfq_cfqq_wait_request(cfqq)) { | |
3490 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || | |
3491 | cfqd->busy_queues > 1) { | |
3492 | cfq_del_timer(cfqd, cfqq); | |
3493 | cfq_clear_cfqq_wait_request(cfqq); | |
3494 | __blk_run_queue(cfqd->queue); | |
3495 | } else { | |
3496 | cfqg_stats_update_idle_time(cfqq->cfqg); | |
3497 | cfq_mark_cfqq_must_dispatch(cfqq); | |
3498 | } | |
3499 | } | |
3500 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { | |
3501 | /* | |
3502 | * not the active queue - expire current slice if it is | |
3503 | * idle and has expired it's mean thinktime or this new queue | |
3504 | * has some old slice time left and is of higher priority or | |
3505 | * this new queue is RT and the current one is BE | |
3506 | */ | |
3507 | cfq_preempt_queue(cfqd, cfqq); | |
3508 | __blk_run_queue(cfqd->queue); | |
3509 | } | |
3510 | } | |
3511 | ||
3512 | static void cfq_insert_request(struct request_queue *q, struct request *rq) | |
3513 | { | |
3514 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3515 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3516 | ||
3517 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); | |
3518 | cfq_init_prio_data(cfqq, RQ_CIC(rq)); | |
3519 | ||
3520 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); | |
3521 | list_add_tail(&rq->queuelist, &cfqq->fifo); | |
3522 | cfq_add_rq_rb(rq); | |
3523 | cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, | |
3524 | rq->cmd_flags); | |
3525 | cfq_rq_enqueued(cfqd, cfqq, rq); | |
3526 | } | |
3527 | ||
3528 | /* | |
3529 | * Update hw_tag based on peak queue depth over 50 samples under | |
3530 | * sufficient load. | |
3531 | */ | |
3532 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3533 | { | |
3534 | struct cfq_queue *cfqq = cfqd->active_queue; | |
3535 | ||
3536 | if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) | |
3537 | cfqd->hw_tag_est_depth = cfqd->rq_in_driver; | |
3538 | ||
3539 | if (cfqd->hw_tag == 1) | |
3540 | return; | |
3541 | ||
3542 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
3543 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) | |
3544 | return; | |
3545 | ||
3546 | /* | |
3547 | * If active queue hasn't enough requests and can idle, cfq might not | |
3548 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3549 | * case | |
3550 | */ | |
3551 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3552 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
3553 | CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) | |
3554 | return; | |
3555 | ||
3556 | if (cfqd->hw_tag_samples++ < 50) | |
3557 | return; | |
3558 | ||
3559 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) | |
3560 | cfqd->hw_tag = 1; | |
3561 | else | |
3562 | cfqd->hw_tag = 0; | |
3563 | } | |
3564 | ||
3565 | static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3566 | { | |
3567 | struct cfq_io_cq *cic = cfqd->active_cic; | |
3568 | ||
3569 | /* If the queue already has requests, don't wait */ | |
3570 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3571 | return false; | |
3572 | ||
3573 | /* If there are other queues in the group, don't wait */ | |
3574 | if (cfqq->cfqg->nr_cfqq > 1) | |
3575 | return false; | |
3576 | ||
3577 | /* the only queue in the group, but think time is big */ | |
3578 | if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) | |
3579 | return false; | |
3580 | ||
3581 | if (cfq_slice_used(cfqq)) | |
3582 | return true; | |
3583 | ||
3584 | /* if slice left is less than think time, wait busy */ | |
3585 | if (cic && sample_valid(cic->ttime.ttime_samples) | |
3586 | && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) | |
3587 | return true; | |
3588 | ||
3589 | /* | |
3590 | * If think times is less than a jiffy than ttime_mean=0 and above | |
3591 | * will not be true. It might happen that slice has not expired yet | |
3592 | * but will expire soon (4-5 ns) during select_queue(). To cover the | |
3593 | * case where think time is less than a jiffy, mark the queue wait | |
3594 | * busy if only 1 jiffy is left in the slice. | |
3595 | */ | |
3596 | if (cfqq->slice_end - jiffies == 1) | |
3597 | return true; | |
3598 | ||
3599 | return false; | |
3600 | } | |
3601 | ||
3602 | static void cfq_completed_request(struct request_queue *q, struct request *rq) | |
3603 | { | |
3604 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3605 | struct cfq_data *cfqd = cfqq->cfqd; | |
3606 | const int sync = rq_is_sync(rq); | |
3607 | unsigned long now; | |
3608 | ||
3609 | now = jiffies; | |
3610 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", | |
3611 | !!(rq->cmd_flags & REQ_NOIDLE)); | |
3612 | ||
3613 | cfq_update_hw_tag(cfqd); | |
3614 | ||
3615 | WARN_ON(!cfqd->rq_in_driver); | |
3616 | WARN_ON(!cfqq->dispatched); | |
3617 | cfqd->rq_in_driver--; | |
3618 | cfqq->dispatched--; | |
3619 | (RQ_CFQG(rq))->dispatched--; | |
3620 | cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq), | |
3621 | rq_io_start_time_ns(rq), rq->cmd_flags); | |
3622 | ||
3623 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; | |
3624 | ||
3625 | if (sync) { | |
3626 | struct cfq_rb_root *service_tree; | |
3627 | ||
3628 | RQ_CIC(rq)->ttime.last_end_request = now; | |
3629 | ||
3630 | if (cfq_cfqq_on_rr(cfqq)) | |
3631 | service_tree = cfqq->service_tree; | |
3632 | else | |
3633 | service_tree = service_tree_for(cfqq->cfqg, | |
3634 | cfqq_prio(cfqq), cfqq_type(cfqq)); | |
3635 | service_tree->ttime.last_end_request = now; | |
3636 | if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now)) | |
3637 | cfqd->last_delayed_sync = now; | |
3638 | } | |
3639 | ||
3640 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3641 | cfqq->cfqg->ttime.last_end_request = now; | |
3642 | #endif | |
3643 | ||
3644 | /* | |
3645 | * If this is the active queue, check if it needs to be expired, | |
3646 | * or if we want to idle in case it has no pending requests. | |
3647 | */ | |
3648 | if (cfqd->active_queue == cfqq) { | |
3649 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); | |
3650 | ||
3651 | if (cfq_cfqq_slice_new(cfqq)) { | |
3652 | cfq_set_prio_slice(cfqd, cfqq); | |
3653 | cfq_clear_cfqq_slice_new(cfqq); | |
3654 | } | |
3655 | ||
3656 | /* | |
3657 | * Should we wait for next request to come in before we expire | |
3658 | * the queue. | |
3659 | */ | |
3660 | if (cfq_should_wait_busy(cfqd, cfqq)) { | |
3661 | unsigned long extend_sl = cfqd->cfq_slice_idle; | |
3662 | if (!cfqd->cfq_slice_idle) | |
3663 | extend_sl = cfqd->cfq_group_idle; | |
3664 | cfqq->slice_end = jiffies + extend_sl; | |
3665 | cfq_mark_cfqq_wait_busy(cfqq); | |
3666 | cfq_log_cfqq(cfqd, cfqq, "will busy wait"); | |
3667 | } | |
3668 | ||
3669 | /* | |
3670 | * Idling is not enabled on: | |
3671 | * - expired queues | |
3672 | * - idle-priority queues | |
3673 | * - async queues | |
3674 | * - queues with still some requests queued | |
3675 | * - when there is a close cooperator | |
3676 | */ | |
3677 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) | |
3678 | cfq_slice_expired(cfqd, 1); | |
3679 | else if (sync && cfqq_empty && | |
3680 | !cfq_close_cooperator(cfqd, cfqq)) { | |
3681 | cfq_arm_slice_timer(cfqd); | |
3682 | } | |
3683 | } | |
3684 | ||
3685 | if (!cfqd->rq_in_driver) | |
3686 | cfq_schedule_dispatch(cfqd); | |
3687 | } | |
3688 | ||
3689 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) | |
3690 | { | |
3691 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { | |
3692 | cfq_mark_cfqq_must_alloc_slice(cfqq); | |
3693 | return ELV_MQUEUE_MUST; | |
3694 | } | |
3695 | ||
3696 | return ELV_MQUEUE_MAY; | |
3697 | } | |
3698 | ||
3699 | static int cfq_may_queue(struct request_queue *q, int rw) | |
3700 | { | |
3701 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3702 | struct task_struct *tsk = current; | |
3703 | struct cfq_io_cq *cic; | |
3704 | struct cfq_queue *cfqq; | |
3705 | ||
3706 | /* | |
3707 | * don't force setup of a queue from here, as a call to may_queue | |
3708 | * does not necessarily imply that a request actually will be queued. | |
3709 | * so just lookup a possibly existing queue, or return 'may queue' | |
3710 | * if that fails | |
3711 | */ | |
3712 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
3713 | if (!cic) | |
3714 | return ELV_MQUEUE_MAY; | |
3715 | ||
3716 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); | |
3717 | if (cfqq) { | |
3718 | cfq_init_prio_data(cfqq, cic); | |
3719 | ||
3720 | return __cfq_may_queue(cfqq); | |
3721 | } | |
3722 | ||
3723 | return ELV_MQUEUE_MAY; | |
3724 | } | |
3725 | ||
3726 | /* | |
3727 | * queue lock held here | |
3728 | */ | |
3729 | static void cfq_put_request(struct request *rq) | |
3730 | { | |
3731 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3732 | ||
3733 | if (cfqq) { | |
3734 | const int rw = rq_data_dir(rq); | |
3735 | ||
3736 | BUG_ON(!cfqq->allocated[rw]); | |
3737 | cfqq->allocated[rw]--; | |
3738 | ||
3739 | /* Put down rq reference on cfqg */ | |
3740 | cfqg_put(RQ_CFQG(rq)); | |
3741 | rq->elv.priv[0] = NULL; | |
3742 | rq->elv.priv[1] = NULL; | |
3743 | ||
3744 | cfq_put_queue(cfqq); | |
3745 | } | |
3746 | } | |
3747 | ||
3748 | static struct cfq_queue * | |
3749 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic, | |
3750 | struct cfq_queue *cfqq) | |
3751 | { | |
3752 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3753 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
3754 | cfq_mark_cfqq_coop(cfqq->new_cfqq); | |
3755 | cfq_put_queue(cfqq); | |
3756 | return cic_to_cfqq(cic, 1); | |
3757 | } | |
3758 | ||
3759 | /* | |
3760 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3761 | * was the last process referring to said cfqq. | |
3762 | */ | |
3763 | static struct cfq_queue * | |
3764 | split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq) | |
3765 | { | |
3766 | if (cfqq_process_refs(cfqq) == 1) { | |
3767 | cfqq->pid = current->pid; | |
3768 | cfq_clear_cfqq_coop(cfqq); | |
3769 | cfq_clear_cfqq_split_coop(cfqq); | |
3770 | return cfqq; | |
3771 | } | |
3772 | ||
3773 | cic_set_cfqq(cic, NULL, 1); | |
3774 | ||
3775 | cfq_put_cooperator(cfqq); | |
3776 | ||
3777 | cfq_put_queue(cfqq); | |
3778 | return NULL; | |
3779 | } | |
3780 | /* | |
3781 | * Allocate cfq data structures associated with this request. | |
3782 | */ | |
3783 | static int | |
3784 | cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio, | |
3785 | gfp_t gfp_mask) | |
3786 | { | |
3787 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3788 | struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq); | |
3789 | const int rw = rq_data_dir(rq); | |
3790 | const bool is_sync = rq_is_sync(rq); | |
3791 | struct cfq_queue *cfqq; | |
3792 | ||
3793 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3794 | ||
3795 | spin_lock_irq(q->queue_lock); | |
3796 | ||
3797 | check_ioprio_changed(cic, bio); | |
3798 | check_blkcg_changed(cic, bio); | |
3799 | new_queue: | |
3800 | cfqq = cic_to_cfqq(cic, is_sync); | |
3801 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
3802 | cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask); | |
3803 | cic_set_cfqq(cic, cfqq, is_sync); | |
3804 | } else { | |
3805 | /* | |
3806 | * If the queue was seeky for too long, break it apart. | |
3807 | */ | |
3808 | if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { | |
3809 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
3810 | cfqq = split_cfqq(cic, cfqq); | |
3811 | if (!cfqq) | |
3812 | goto new_queue; | |
3813 | } | |
3814 | ||
3815 | /* | |
3816 | * Check to see if this queue is scheduled to merge with | |
3817 | * another, closely cooperating queue. The merging of | |
3818 | * queues happens here as it must be done in process context. | |
3819 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3820 | */ | |
3821 | if (cfqq->new_cfqq) | |
3822 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
3823 | } | |
3824 | ||
3825 | cfqq->allocated[rw]++; | |
3826 | ||
3827 | cfqq->ref++; | |
3828 | cfqg_get(cfqq->cfqg); | |
3829 | rq->elv.priv[0] = cfqq; | |
3830 | rq->elv.priv[1] = cfqq->cfqg; | |
3831 | spin_unlock_irq(q->queue_lock); | |
3832 | return 0; | |
3833 | } | |
3834 | ||
3835 | static void cfq_kick_queue(struct work_struct *work) | |
3836 | { | |
3837 | struct cfq_data *cfqd = | |
3838 | container_of(work, struct cfq_data, unplug_work); | |
3839 | struct request_queue *q = cfqd->queue; | |
3840 | ||
3841 | spin_lock_irq(q->queue_lock); | |
3842 | __blk_run_queue(cfqd->queue); | |
3843 | spin_unlock_irq(q->queue_lock); | |
3844 | } | |
3845 | ||
3846 | /* | |
3847 | * Timer running if the active_queue is currently idling inside its time slice | |
3848 | */ | |
3849 | static void cfq_idle_slice_timer(unsigned long data) | |
3850 | { | |
3851 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3852 | struct cfq_queue *cfqq; | |
3853 | unsigned long flags; | |
3854 | int timed_out = 1; | |
3855 | ||
3856 | cfq_log(cfqd, "idle timer fired"); | |
3857 | ||
3858 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
3859 | ||
3860 | cfqq = cfqd->active_queue; | |
3861 | if (cfqq) { | |
3862 | timed_out = 0; | |
3863 | ||
3864 | /* | |
3865 | * We saw a request before the queue expired, let it through | |
3866 | */ | |
3867 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3868 | goto out_kick; | |
3869 | ||
3870 | /* | |
3871 | * expired | |
3872 | */ | |
3873 | if (cfq_slice_used(cfqq)) | |
3874 | goto expire; | |
3875 | ||
3876 | /* | |
3877 | * only expire and reinvoke request handler, if there are | |
3878 | * other queues with pending requests | |
3879 | */ | |
3880 | if (!cfqd->busy_queues) | |
3881 | goto out_cont; | |
3882 | ||
3883 | /* | |
3884 | * not expired and it has a request pending, let it dispatch | |
3885 | */ | |
3886 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3887 | goto out_kick; | |
3888 | ||
3889 | /* | |
3890 | * Queue depth flag is reset only when the idle didn't succeed | |
3891 | */ | |
3892 | cfq_clear_cfqq_deep(cfqq); | |
3893 | } | |
3894 | expire: | |
3895 | cfq_slice_expired(cfqd, timed_out); | |
3896 | out_kick: | |
3897 | cfq_schedule_dispatch(cfqd); | |
3898 | out_cont: | |
3899 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3900 | } | |
3901 | ||
3902 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) | |
3903 | { | |
3904 | del_timer_sync(&cfqd->idle_slice_timer); | |
3905 | cancel_work_sync(&cfqd->unplug_work); | |
3906 | } | |
3907 | ||
3908 | static void cfq_put_async_queues(struct cfq_data *cfqd) | |
3909 | { | |
3910 | int i; | |
3911 | ||
3912 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3913 | if (cfqd->async_cfqq[0][i]) | |
3914 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3915 | if (cfqd->async_cfqq[1][i]) | |
3916 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
3917 | } | |
3918 | ||
3919 | if (cfqd->async_idle_cfqq) | |
3920 | cfq_put_queue(cfqd->async_idle_cfqq); | |
3921 | } | |
3922 | ||
3923 | static void cfq_exit_queue(struct elevator_queue *e) | |
3924 | { | |
3925 | struct cfq_data *cfqd = e->elevator_data; | |
3926 | struct request_queue *q = cfqd->queue; | |
3927 | ||
3928 | cfq_shutdown_timer_wq(cfqd); | |
3929 | ||
3930 | spin_lock_irq(q->queue_lock); | |
3931 | ||
3932 | if (cfqd->active_queue) | |
3933 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); | |
3934 | ||
3935 | cfq_put_async_queues(cfqd); | |
3936 | ||
3937 | spin_unlock_irq(q->queue_lock); | |
3938 | ||
3939 | cfq_shutdown_timer_wq(cfqd); | |
3940 | ||
3941 | #ifndef CONFIG_CFQ_GROUP_IOSCHED | |
3942 | kfree(cfqd->root_group); | |
3943 | #endif | |
3944 | update_root_blkg_pd(q, BLKIO_POLICY_PROP); | |
3945 | kfree(cfqd); | |
3946 | } | |
3947 | ||
3948 | static int cfq_init_queue(struct request_queue *q) | |
3949 | { | |
3950 | struct cfq_data *cfqd; | |
3951 | struct blkio_group *blkg __maybe_unused; | |
3952 | int i; | |
3953 | ||
3954 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); | |
3955 | if (!cfqd) | |
3956 | return -ENOMEM; | |
3957 | ||
3958 | cfqd->queue = q; | |
3959 | q->elevator->elevator_data = cfqd; | |
3960 | ||
3961 | /* Init root service tree */ | |
3962 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3963 | ||
3964 | /* Init root group and prefer root group over other groups by default */ | |
3965 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3966 | rcu_read_lock(); | |
3967 | spin_lock_irq(q->queue_lock); | |
3968 | ||
3969 | blkg = blkg_lookup_create(&blkio_root_cgroup, q, true); | |
3970 | if (!IS_ERR(blkg)) | |
3971 | cfqd->root_group = blkg_to_cfqg(blkg); | |
3972 | ||
3973 | spin_unlock_irq(q->queue_lock); | |
3974 | rcu_read_unlock(); | |
3975 | #else | |
3976 | cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group), | |
3977 | GFP_KERNEL, cfqd->queue->node); | |
3978 | if (cfqd->root_group) | |
3979 | cfq_init_cfqg_base(cfqd->root_group); | |
3980 | #endif | |
3981 | if (!cfqd->root_group) { | |
3982 | kfree(cfqd); | |
3983 | return -ENOMEM; | |
3984 | } | |
3985 | ||
3986 | cfqd->root_group->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3987 | ||
3988 | /* | |
3989 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3990 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3991 | * to add magic to the rb code | |
3992 | */ | |
3993 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3994 | cfqd->prio_trees[i] = RB_ROOT; | |
3995 | ||
3996 | /* | |
3997 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3998 | * Grab a permanent reference to it, so that the normal code flow | |
3999 | * will not attempt to free it. oom_cfqq is linked to root_group | |
4000 | * but shouldn't hold a reference as it'll never be unlinked. Lose | |
4001 | * the reference from linking right away. | |
4002 | */ | |
4003 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
4004 | cfqd->oom_cfqq.ref++; | |
4005 | ||
4006 | spin_lock_irq(q->queue_lock); | |
4007 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group); | |
4008 | cfqg_put(cfqd->root_group); | |
4009 | spin_unlock_irq(q->queue_lock); | |
4010 | ||
4011 | init_timer(&cfqd->idle_slice_timer); | |
4012 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
4013 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
4014 | ||
4015 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); | |
4016 | ||
4017 | cfqd->cfq_quantum = cfq_quantum; | |
4018 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; | |
4019 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
4020 | cfqd->cfq_back_max = cfq_back_max; | |
4021 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
4022 | cfqd->cfq_slice[0] = cfq_slice_async; | |
4023 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
4024 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
4025 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
4026 | cfqd->cfq_group_idle = cfq_group_idle; | |
4027 | cfqd->cfq_latency = 1; | |
4028 | cfqd->hw_tag = -1; | |
4029 | /* | |
4030 | * we optimistically start assuming sync ops weren't delayed in last | |
4031 | * second, in order to have larger depth for async operations. | |
4032 | */ | |
4033 | cfqd->last_delayed_sync = jiffies - HZ; | |
4034 | return 0; | |
4035 | } | |
4036 | ||
4037 | /* | |
4038 | * sysfs parts below --> | |
4039 | */ | |
4040 | static ssize_t | |
4041 | cfq_var_show(unsigned int var, char *page) | |
4042 | { | |
4043 | return sprintf(page, "%d\n", var); | |
4044 | } | |
4045 | ||
4046 | static ssize_t | |
4047 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
4048 | { | |
4049 | char *p = (char *) page; | |
4050 | ||
4051 | *var = simple_strtoul(p, &p, 10); | |
4052 | return count; | |
4053 | } | |
4054 | ||
4055 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
4056 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
4057 | { \ | |
4058 | struct cfq_data *cfqd = e->elevator_data; \ | |
4059 | unsigned int __data = __VAR; \ | |
4060 | if (__CONV) \ | |
4061 | __data = jiffies_to_msecs(__data); \ | |
4062 | return cfq_var_show(__data, (page)); \ | |
4063 | } | |
4064 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
4065 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); | |
4066 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
4067 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); | |
4068 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
4069 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); | |
4070 | SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); | |
4071 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
4072 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
4073 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
4074 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); | |
4075 | #undef SHOW_FUNCTION | |
4076 | ||
4077 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
4078 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
4079 | { \ | |
4080 | struct cfq_data *cfqd = e->elevator_data; \ | |
4081 | unsigned int __data; \ | |
4082 | int ret = cfq_var_store(&__data, (page), count); \ | |
4083 | if (__data < (MIN)) \ | |
4084 | __data = (MIN); \ | |
4085 | else if (__data > (MAX)) \ | |
4086 | __data = (MAX); \ | |
4087 | if (__CONV) \ | |
4088 | *(__PTR) = msecs_to_jiffies(__data); \ | |
4089 | else \ | |
4090 | *(__PTR) = __data; \ | |
4091 | return ret; \ | |
4092 | } | |
4093 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
4094 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, | |
4095 | UINT_MAX, 1); | |
4096 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
4097 | UINT_MAX, 1); | |
4098 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); | |
4099 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, | |
4100 | UINT_MAX, 0); | |
4101 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); | |
4102 | STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); | |
4103 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
4104 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
4105 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, | |
4106 | UINT_MAX, 0); | |
4107 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); | |
4108 | #undef STORE_FUNCTION | |
4109 | ||
4110 | #define CFQ_ATTR(name) \ | |
4111 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
4112 | ||
4113 | static struct elv_fs_entry cfq_attrs[] = { | |
4114 | CFQ_ATTR(quantum), | |
4115 | CFQ_ATTR(fifo_expire_sync), | |
4116 | CFQ_ATTR(fifo_expire_async), | |
4117 | CFQ_ATTR(back_seek_max), | |
4118 | CFQ_ATTR(back_seek_penalty), | |
4119 | CFQ_ATTR(slice_sync), | |
4120 | CFQ_ATTR(slice_async), | |
4121 | CFQ_ATTR(slice_async_rq), | |
4122 | CFQ_ATTR(slice_idle), | |
4123 | CFQ_ATTR(group_idle), | |
4124 | CFQ_ATTR(low_latency), | |
4125 | __ATTR_NULL | |
4126 | }; | |
4127 | ||
4128 | static struct elevator_type iosched_cfq = { | |
4129 | .ops = { | |
4130 | .elevator_merge_fn = cfq_merge, | |
4131 | .elevator_merged_fn = cfq_merged_request, | |
4132 | .elevator_merge_req_fn = cfq_merged_requests, | |
4133 | .elevator_allow_merge_fn = cfq_allow_merge, | |
4134 | .elevator_bio_merged_fn = cfq_bio_merged, | |
4135 | .elevator_dispatch_fn = cfq_dispatch_requests, | |
4136 | .elevator_add_req_fn = cfq_insert_request, | |
4137 | .elevator_activate_req_fn = cfq_activate_request, | |
4138 | .elevator_deactivate_req_fn = cfq_deactivate_request, | |
4139 | .elevator_completed_req_fn = cfq_completed_request, | |
4140 | .elevator_former_req_fn = elv_rb_former_request, | |
4141 | .elevator_latter_req_fn = elv_rb_latter_request, | |
4142 | .elevator_init_icq_fn = cfq_init_icq, | |
4143 | .elevator_exit_icq_fn = cfq_exit_icq, | |
4144 | .elevator_set_req_fn = cfq_set_request, | |
4145 | .elevator_put_req_fn = cfq_put_request, | |
4146 | .elevator_may_queue_fn = cfq_may_queue, | |
4147 | .elevator_init_fn = cfq_init_queue, | |
4148 | .elevator_exit_fn = cfq_exit_queue, | |
4149 | }, | |
4150 | .icq_size = sizeof(struct cfq_io_cq), | |
4151 | .icq_align = __alignof__(struct cfq_io_cq), | |
4152 | .elevator_attrs = cfq_attrs, | |
4153 | .elevator_name = "cfq", | |
4154 | .elevator_owner = THIS_MODULE, | |
4155 | }; | |
4156 | ||
4157 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4158 | static struct blkio_policy_type blkio_policy_cfq = { | |
4159 | .ops = { | |
4160 | .blkio_init_group_fn = cfq_init_blkio_group, | |
4161 | .blkio_reset_group_stats_fn = cfqg_stats_reset, | |
4162 | }, | |
4163 | .plid = BLKIO_POLICY_PROP, | |
4164 | .pdata_size = sizeof(struct cfq_group), | |
4165 | .cftypes = cfq_blkcg_files, | |
4166 | }; | |
4167 | #endif | |
4168 | ||
4169 | static int __init cfq_init(void) | |
4170 | { | |
4171 | int ret; | |
4172 | ||
4173 | /* | |
4174 | * could be 0 on HZ < 1000 setups | |
4175 | */ | |
4176 | if (!cfq_slice_async) | |
4177 | cfq_slice_async = 1; | |
4178 | if (!cfq_slice_idle) | |
4179 | cfq_slice_idle = 1; | |
4180 | ||
4181 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4182 | if (!cfq_group_idle) | |
4183 | cfq_group_idle = 1; | |
4184 | #else | |
4185 | cfq_group_idle = 0; | |
4186 | #endif | |
4187 | cfq_pool = KMEM_CACHE(cfq_queue, 0); | |
4188 | if (!cfq_pool) | |
4189 | return -ENOMEM; | |
4190 | ||
4191 | ret = elv_register(&iosched_cfq); | |
4192 | if (ret) { | |
4193 | kmem_cache_destroy(cfq_pool); | |
4194 | return ret; | |
4195 | } | |
4196 | ||
4197 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4198 | blkio_policy_register(&blkio_policy_cfq); | |
4199 | #endif | |
4200 | return 0; | |
4201 | } | |
4202 | ||
4203 | static void __exit cfq_exit(void) | |
4204 | { | |
4205 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4206 | blkio_policy_unregister(&blkio_policy_cfq); | |
4207 | #endif | |
4208 | elv_unregister(&iosched_cfq); | |
4209 | kmem_cache_destroy(cfq_pool); | |
4210 | } | |
4211 | ||
4212 | module_init(cfq_init); | |
4213 | module_exit(cfq_exit); | |
4214 | ||
4215 | MODULE_AUTHOR("Jens Axboe"); | |
4216 | MODULE_LICENSE("GPL"); | |
4217 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |