]>
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/ktime.h> | |
14 | #include <linux/rbtree.h> | |
15 | #include <linux/ioprio.h> | |
16 | #include <linux/blktrace_api.h> | |
17 | #include <linux/blk-cgroup.h> | |
18 | #include "blk.h" | |
19 | ||
20 | /* | |
21 | * tunables | |
22 | */ | |
23 | /* max queue in one round of service */ | |
24 | static const int cfq_quantum = 8; | |
25 | static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 }; | |
26 | /* maximum backwards seek, in KiB */ | |
27 | static const int cfq_back_max = 16 * 1024; | |
28 | /* penalty of a backwards seek */ | |
29 | static const int cfq_back_penalty = 2; | |
30 | static const u64 cfq_slice_sync = NSEC_PER_SEC / 10; | |
31 | static u64 cfq_slice_async = NSEC_PER_SEC / 25; | |
32 | static const int cfq_slice_async_rq = 2; | |
33 | static u64 cfq_slice_idle = NSEC_PER_SEC / 125; | |
34 | static u64 cfq_group_idle = NSEC_PER_SEC / 125; | |
35 | static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */ | |
36 | static const int cfq_hist_divisor = 4; | |
37 | ||
38 | /* | |
39 | * offset from end of service tree | |
40 | */ | |
41 | #define CFQ_IDLE_DELAY (NSEC_PER_SEC / 5) | |
42 | ||
43 | /* | |
44 | * below this threshold, we consider thinktime immediate | |
45 | */ | |
46 | #define CFQ_MIN_TT (2 * NSEC_PER_SEC / HZ) | |
47 | ||
48 | #define CFQ_SLICE_SCALE (5) | |
49 | #define CFQ_HW_QUEUE_MIN (5) | |
50 | #define CFQ_SERVICE_SHIFT 12 | |
51 | ||
52 | #define CFQQ_SEEK_THR (sector_t)(8 * 100) | |
53 | #define CFQQ_CLOSE_THR (sector_t)(8 * 1024) | |
54 | #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32) | |
55 | #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8) | |
56 | ||
57 | #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq) | |
58 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0]) | |
59 | #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1]) | |
60 | ||
61 | static struct kmem_cache *cfq_pool; | |
62 | ||
63 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR | |
64 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
65 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
66 | ||
67 | #define sample_valid(samples) ((samples) > 80) | |
68 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) | |
69 | ||
70 | /* blkio-related constants */ | |
71 | #define CFQ_WEIGHT_LEGACY_MIN 10 | |
72 | #define CFQ_WEIGHT_LEGACY_DFL 500 | |
73 | #define CFQ_WEIGHT_LEGACY_MAX 1000 | |
74 | ||
75 | struct cfq_ttime { | |
76 | u64 last_end_request; | |
77 | ||
78 | u64 ttime_total; | |
79 | u64 ttime_mean; | |
80 | unsigned long ttime_samples; | |
81 | }; | |
82 | ||
83 | /* | |
84 | * Most of our rbtree usage is for sorting with min extraction, so | |
85 | * if we cache the leftmost node we don't have to walk down the tree | |
86 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
87 | * move this into the elevator for the rq sorting as well. | |
88 | */ | |
89 | struct cfq_rb_root { | |
90 | struct rb_root rb; | |
91 | struct rb_node *left; | |
92 | unsigned count; | |
93 | u64 min_vdisktime; | |
94 | struct cfq_ttime ttime; | |
95 | }; | |
96 | #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \ | |
97 | .ttime = {.last_end_request = ktime_get_ns(),},} | |
98 | ||
99 | /* | |
100 | * Per process-grouping structure | |
101 | */ | |
102 | struct cfq_queue { | |
103 | /* reference count */ | |
104 | int ref; | |
105 | /* various state flags, see below */ | |
106 | unsigned int flags; | |
107 | /* parent cfq_data */ | |
108 | struct cfq_data *cfqd; | |
109 | /* service_tree member */ | |
110 | struct rb_node rb_node; | |
111 | /* service_tree key */ | |
112 | u64 rb_key; | |
113 | /* prio tree member */ | |
114 | struct rb_node p_node; | |
115 | /* prio tree root we belong to, if any */ | |
116 | struct rb_root *p_root; | |
117 | /* sorted list of pending requests */ | |
118 | struct rb_root sort_list; | |
119 | /* if fifo isn't expired, next request to serve */ | |
120 | struct request *next_rq; | |
121 | /* requests queued in sort_list */ | |
122 | int queued[2]; | |
123 | /* currently allocated requests */ | |
124 | int allocated[2]; | |
125 | /* fifo list of requests in sort_list */ | |
126 | struct list_head fifo; | |
127 | ||
128 | /* time when queue got scheduled in to dispatch first request. */ | |
129 | u64 dispatch_start; | |
130 | u64 allocated_slice; | |
131 | u64 slice_dispatch; | |
132 | /* time when first request from queue completed and slice started. */ | |
133 | u64 slice_start; | |
134 | u64 slice_end; | |
135 | u64 slice_resid; | |
136 | ||
137 | /* pending priority requests */ | |
138 | int prio_pending; | |
139 | /* number of requests that are on the dispatch list or inside driver */ | |
140 | int dispatched; | |
141 | ||
142 | /* io prio of this group */ | |
143 | unsigned short ioprio, org_ioprio; | |
144 | unsigned short ioprio_class; | |
145 | ||
146 | pid_t pid; | |
147 | ||
148 | u32 seek_history; | |
149 | sector_t last_request_pos; | |
150 | ||
151 | struct cfq_rb_root *service_tree; | |
152 | struct cfq_queue *new_cfqq; | |
153 | struct cfq_group *cfqg; | |
154 | /* Number of sectors dispatched from queue in single dispatch round */ | |
155 | unsigned long nr_sectors; | |
156 | }; | |
157 | ||
158 | /* | |
159 | * First index in the service_trees. | |
160 | * IDLE is handled separately, so it has negative index | |
161 | */ | |
162 | enum wl_class_t { | |
163 | BE_WORKLOAD = 0, | |
164 | RT_WORKLOAD = 1, | |
165 | IDLE_WORKLOAD = 2, | |
166 | CFQ_PRIO_NR, | |
167 | }; | |
168 | ||
169 | /* | |
170 | * Second index in the service_trees. | |
171 | */ | |
172 | enum wl_type_t { | |
173 | ASYNC_WORKLOAD = 0, | |
174 | SYNC_NOIDLE_WORKLOAD = 1, | |
175 | SYNC_WORKLOAD = 2 | |
176 | }; | |
177 | ||
178 | struct cfqg_stats { | |
179 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
180 | /* number of ios merged */ | |
181 | struct blkg_rwstat merged; | |
182 | /* total time spent on device in ns, may not be accurate w/ queueing */ | |
183 | struct blkg_rwstat service_time; | |
184 | /* total time spent waiting in scheduler queue in ns */ | |
185 | struct blkg_rwstat wait_time; | |
186 | /* number of IOs queued up */ | |
187 | struct blkg_rwstat queued; | |
188 | /* total disk time and nr sectors dispatched by this group */ | |
189 | struct blkg_stat time; | |
190 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
191 | /* time not charged to this cgroup */ | |
192 | struct blkg_stat unaccounted_time; | |
193 | /* sum of number of ios queued across all samples */ | |
194 | struct blkg_stat avg_queue_size_sum; | |
195 | /* count of samples taken for average */ | |
196 | struct blkg_stat avg_queue_size_samples; | |
197 | /* how many times this group has been removed from service tree */ | |
198 | struct blkg_stat dequeue; | |
199 | /* total time spent waiting for it to be assigned a timeslice. */ | |
200 | struct blkg_stat group_wait_time; | |
201 | /* time spent idling for this blkcg_gq */ | |
202 | struct blkg_stat idle_time; | |
203 | /* total time with empty current active q with other requests queued */ | |
204 | struct blkg_stat empty_time; | |
205 | /* fields after this shouldn't be cleared on stat reset */ | |
206 | uint64_t start_group_wait_time; | |
207 | uint64_t start_idle_time; | |
208 | uint64_t start_empty_time; | |
209 | uint16_t flags; | |
210 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
211 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
212 | }; | |
213 | ||
214 | /* Per-cgroup data */ | |
215 | struct cfq_group_data { | |
216 | /* must be the first member */ | |
217 | struct blkcg_policy_data cpd; | |
218 | ||
219 | unsigned int weight; | |
220 | unsigned int leaf_weight; | |
221 | }; | |
222 | ||
223 | /* This is per cgroup per device grouping structure */ | |
224 | struct cfq_group { | |
225 | /* must be the first member */ | |
226 | struct blkg_policy_data pd; | |
227 | ||
228 | /* group service_tree member */ | |
229 | struct rb_node rb_node; | |
230 | ||
231 | /* group service_tree key */ | |
232 | u64 vdisktime; | |
233 | ||
234 | /* | |
235 | * The number of active cfqgs and sum of their weights under this | |
236 | * cfqg. This covers this cfqg's leaf_weight and all children's | |
237 | * weights, but does not cover weights of further descendants. | |
238 | * | |
239 | * If a cfqg is on the service tree, it's active. An active cfqg | |
240 | * also activates its parent and contributes to the children_weight | |
241 | * of the parent. | |
242 | */ | |
243 | int nr_active; | |
244 | unsigned int children_weight; | |
245 | ||
246 | /* | |
247 | * vfraction is the fraction of vdisktime that the tasks in this | |
248 | * cfqg are entitled to. This is determined by compounding the | |
249 | * ratios walking up from this cfqg to the root. | |
250 | * | |
251 | * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all | |
252 | * vfractions on a service tree is approximately 1. The sum may | |
253 | * deviate a bit due to rounding errors and fluctuations caused by | |
254 | * cfqgs entering and leaving the service tree. | |
255 | */ | |
256 | unsigned int vfraction; | |
257 | ||
258 | /* | |
259 | * There are two weights - (internal) weight is the weight of this | |
260 | * cfqg against the sibling cfqgs. leaf_weight is the wight of | |
261 | * this cfqg against the child cfqgs. For the root cfqg, both | |
262 | * weights are kept in sync for backward compatibility. | |
263 | */ | |
264 | unsigned int weight; | |
265 | unsigned int new_weight; | |
266 | unsigned int dev_weight; | |
267 | ||
268 | unsigned int leaf_weight; | |
269 | unsigned int new_leaf_weight; | |
270 | unsigned int dev_leaf_weight; | |
271 | ||
272 | /* number of cfqq currently on this group */ | |
273 | int nr_cfqq; | |
274 | ||
275 | /* | |
276 | * Per group busy queues average. Useful for workload slice calc. We | |
277 | * create the array for each prio class but at run time it is used | |
278 | * only for RT and BE class and slot for IDLE class remains unused. | |
279 | * This is primarily done to avoid confusion and a gcc warning. | |
280 | */ | |
281 | unsigned int busy_queues_avg[CFQ_PRIO_NR]; | |
282 | /* | |
283 | * rr lists of queues with requests. We maintain service trees for | |
284 | * RT and BE classes. These trees are subdivided in subclasses | |
285 | * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE | |
286 | * class there is no subclassification and all the cfq queues go on | |
287 | * a single tree service_tree_idle. | |
288 | * Counts are embedded in the cfq_rb_root | |
289 | */ | |
290 | struct cfq_rb_root service_trees[2][3]; | |
291 | struct cfq_rb_root service_tree_idle; | |
292 | ||
293 | u64 saved_wl_slice; | |
294 | enum wl_type_t saved_wl_type; | |
295 | enum wl_class_t saved_wl_class; | |
296 | ||
297 | /* number of requests that are on the dispatch list or inside driver */ | |
298 | int dispatched; | |
299 | struct cfq_ttime ttime; | |
300 | struct cfqg_stats stats; /* stats for this cfqg */ | |
301 | ||
302 | /* async queue for each priority case */ | |
303 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
304 | struct cfq_queue *async_idle_cfqq; | |
305 | ||
306 | }; | |
307 | ||
308 | struct cfq_io_cq { | |
309 | struct io_cq icq; /* must be the first member */ | |
310 | struct cfq_queue *cfqq[2]; | |
311 | struct cfq_ttime ttime; | |
312 | int ioprio; /* the current ioprio */ | |
313 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
314 | uint64_t blkcg_serial_nr; /* the current blkcg serial */ | |
315 | #endif | |
316 | }; | |
317 | ||
318 | /* | |
319 | * Per block device queue structure | |
320 | */ | |
321 | struct cfq_data { | |
322 | struct request_queue *queue; | |
323 | /* Root service tree for cfq_groups */ | |
324 | struct cfq_rb_root grp_service_tree; | |
325 | struct cfq_group *root_group; | |
326 | ||
327 | /* | |
328 | * The priority currently being served | |
329 | */ | |
330 | enum wl_class_t serving_wl_class; | |
331 | enum wl_type_t serving_wl_type; | |
332 | u64 workload_expires; | |
333 | struct cfq_group *serving_group; | |
334 | ||
335 | /* | |
336 | * Each priority tree is sorted by next_request position. These | |
337 | * trees are used when determining if two or more queues are | |
338 | * interleaving requests (see cfq_close_cooperator). | |
339 | */ | |
340 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
341 | ||
342 | unsigned int busy_queues; | |
343 | unsigned int busy_sync_queues; | |
344 | ||
345 | int rq_in_driver; | |
346 | int rq_in_flight[2]; | |
347 | ||
348 | /* | |
349 | * queue-depth detection | |
350 | */ | |
351 | int rq_queued; | |
352 | int hw_tag; | |
353 | /* | |
354 | * hw_tag can be | |
355 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
356 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
357 | * 0 => no NCQ | |
358 | */ | |
359 | int hw_tag_est_depth; | |
360 | unsigned int hw_tag_samples; | |
361 | ||
362 | /* | |
363 | * idle window management | |
364 | */ | |
365 | struct timer_list idle_slice_timer; | |
366 | struct work_struct unplug_work; | |
367 | ||
368 | struct cfq_queue *active_queue; | |
369 | struct cfq_io_cq *active_cic; | |
370 | ||
371 | sector_t last_position; | |
372 | ||
373 | /* | |
374 | * tunables, see top of file | |
375 | */ | |
376 | unsigned int cfq_quantum; | |
377 | unsigned int cfq_back_penalty; | |
378 | unsigned int cfq_back_max; | |
379 | unsigned int cfq_slice_async_rq; | |
380 | unsigned int cfq_latency; | |
381 | u64 cfq_fifo_expire[2]; | |
382 | u64 cfq_slice[2]; | |
383 | u64 cfq_slice_idle; | |
384 | u64 cfq_group_idle; | |
385 | u64 cfq_target_latency; | |
386 | ||
387 | /* | |
388 | * Fallback dummy cfqq for extreme OOM conditions | |
389 | */ | |
390 | struct cfq_queue oom_cfqq; | |
391 | ||
392 | u64 last_delayed_sync; | |
393 | }; | |
394 | ||
395 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); | |
396 | static void cfq_put_queue(struct cfq_queue *cfqq); | |
397 | ||
398 | static struct cfq_rb_root *st_for(struct cfq_group *cfqg, | |
399 | enum wl_class_t class, | |
400 | enum wl_type_t type) | |
401 | { | |
402 | if (!cfqg) | |
403 | return NULL; | |
404 | ||
405 | if (class == IDLE_WORKLOAD) | |
406 | return &cfqg->service_tree_idle; | |
407 | ||
408 | return &cfqg->service_trees[class][type]; | |
409 | } | |
410 | ||
411 | enum cfqq_state_flags { | |
412 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ | |
413 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
414 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ | |
415 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
416 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ | |
417 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
418 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
419 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ | |
420 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ | |
421 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ | |
422 | CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */ | |
423 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ | |
424 | CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */ | |
425 | }; | |
426 | ||
427 | #define CFQ_CFQQ_FNS(name) \ | |
428 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
429 | { \ | |
430 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ | |
431 | } \ | |
432 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
433 | { \ | |
434 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ | |
435 | } \ | |
436 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
437 | { \ | |
438 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ | |
439 | } | |
440 | ||
441 | CFQ_CFQQ_FNS(on_rr); | |
442 | CFQ_CFQQ_FNS(wait_request); | |
443 | CFQ_CFQQ_FNS(must_dispatch); | |
444 | CFQ_CFQQ_FNS(must_alloc_slice); | |
445 | CFQ_CFQQ_FNS(fifo_expire); | |
446 | CFQ_CFQQ_FNS(idle_window); | |
447 | CFQ_CFQQ_FNS(prio_changed); | |
448 | CFQ_CFQQ_FNS(slice_new); | |
449 | CFQ_CFQQ_FNS(sync); | |
450 | CFQ_CFQQ_FNS(coop); | |
451 | CFQ_CFQQ_FNS(split_coop); | |
452 | CFQ_CFQQ_FNS(deep); | |
453 | CFQ_CFQQ_FNS(wait_busy); | |
454 | #undef CFQ_CFQQ_FNS | |
455 | ||
456 | #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP) | |
457 | ||
458 | /* cfqg stats flags */ | |
459 | enum cfqg_stats_flags { | |
460 | CFQG_stats_waiting = 0, | |
461 | CFQG_stats_idling, | |
462 | CFQG_stats_empty, | |
463 | }; | |
464 | ||
465 | #define CFQG_FLAG_FNS(name) \ | |
466 | static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats) \ | |
467 | { \ | |
468 | stats->flags |= (1 << CFQG_stats_##name); \ | |
469 | } \ | |
470 | static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats) \ | |
471 | { \ | |
472 | stats->flags &= ~(1 << CFQG_stats_##name); \ | |
473 | } \ | |
474 | static inline int cfqg_stats_##name(struct cfqg_stats *stats) \ | |
475 | { \ | |
476 | return (stats->flags & (1 << CFQG_stats_##name)) != 0; \ | |
477 | } \ | |
478 | ||
479 | CFQG_FLAG_FNS(waiting) | |
480 | CFQG_FLAG_FNS(idling) | |
481 | CFQG_FLAG_FNS(empty) | |
482 | #undef CFQG_FLAG_FNS | |
483 | ||
484 | /* This should be called with the queue_lock held. */ | |
485 | static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats) | |
486 | { | |
487 | unsigned long long now; | |
488 | ||
489 | if (!cfqg_stats_waiting(stats)) | |
490 | return; | |
491 | ||
492 | now = sched_clock(); | |
493 | if (time_after64(now, stats->start_group_wait_time)) | |
494 | blkg_stat_add(&stats->group_wait_time, | |
495 | now - stats->start_group_wait_time); | |
496 | cfqg_stats_clear_waiting(stats); | |
497 | } | |
498 | ||
499 | /* This should be called with the queue_lock held. */ | |
500 | static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, | |
501 | struct cfq_group *curr_cfqg) | |
502 | { | |
503 | struct cfqg_stats *stats = &cfqg->stats; | |
504 | ||
505 | if (cfqg_stats_waiting(stats)) | |
506 | return; | |
507 | if (cfqg == curr_cfqg) | |
508 | return; | |
509 | stats->start_group_wait_time = sched_clock(); | |
510 | cfqg_stats_mark_waiting(stats); | |
511 | } | |
512 | ||
513 | /* This should be called with the queue_lock held. */ | |
514 | static void cfqg_stats_end_empty_time(struct cfqg_stats *stats) | |
515 | { | |
516 | unsigned long long now; | |
517 | ||
518 | if (!cfqg_stats_empty(stats)) | |
519 | return; | |
520 | ||
521 | now = sched_clock(); | |
522 | if (time_after64(now, stats->start_empty_time)) | |
523 | blkg_stat_add(&stats->empty_time, | |
524 | now - stats->start_empty_time); | |
525 | cfqg_stats_clear_empty(stats); | |
526 | } | |
527 | ||
528 | static void cfqg_stats_update_dequeue(struct cfq_group *cfqg) | |
529 | { | |
530 | blkg_stat_add(&cfqg->stats.dequeue, 1); | |
531 | } | |
532 | ||
533 | static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) | |
534 | { | |
535 | struct cfqg_stats *stats = &cfqg->stats; | |
536 | ||
537 | if (blkg_rwstat_total(&stats->queued)) | |
538 | return; | |
539 | ||
540 | /* | |
541 | * group is already marked empty. This can happen if cfqq got new | |
542 | * request in parent group and moved to this group while being added | |
543 | * to service tree. Just ignore the event and move on. | |
544 | */ | |
545 | if (cfqg_stats_empty(stats)) | |
546 | return; | |
547 | ||
548 | stats->start_empty_time = sched_clock(); | |
549 | cfqg_stats_mark_empty(stats); | |
550 | } | |
551 | ||
552 | static void cfqg_stats_update_idle_time(struct cfq_group *cfqg) | |
553 | { | |
554 | struct cfqg_stats *stats = &cfqg->stats; | |
555 | ||
556 | if (cfqg_stats_idling(stats)) { | |
557 | unsigned long long now = sched_clock(); | |
558 | ||
559 | if (time_after64(now, stats->start_idle_time)) | |
560 | blkg_stat_add(&stats->idle_time, | |
561 | now - stats->start_idle_time); | |
562 | cfqg_stats_clear_idling(stats); | |
563 | } | |
564 | } | |
565 | ||
566 | static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) | |
567 | { | |
568 | struct cfqg_stats *stats = &cfqg->stats; | |
569 | ||
570 | BUG_ON(cfqg_stats_idling(stats)); | |
571 | ||
572 | stats->start_idle_time = sched_clock(); | |
573 | cfqg_stats_mark_idling(stats); | |
574 | } | |
575 | ||
576 | static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) | |
577 | { | |
578 | struct cfqg_stats *stats = &cfqg->stats; | |
579 | ||
580 | blkg_stat_add(&stats->avg_queue_size_sum, | |
581 | blkg_rwstat_total(&stats->queued)); | |
582 | blkg_stat_add(&stats->avg_queue_size_samples, 1); | |
583 | cfqg_stats_update_group_wait_time(stats); | |
584 | } | |
585 | ||
586 | #else /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */ | |
587 | ||
588 | static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { } | |
589 | static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { } | |
590 | static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { } | |
591 | static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { } | |
592 | static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { } | |
593 | static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { } | |
594 | static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { } | |
595 | ||
596 | #endif /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */ | |
597 | ||
598 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
599 | ||
600 | static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd) | |
601 | { | |
602 | return pd ? container_of(pd, struct cfq_group, pd) : NULL; | |
603 | } | |
604 | ||
605 | static struct cfq_group_data | |
606 | *cpd_to_cfqgd(struct blkcg_policy_data *cpd) | |
607 | { | |
608 | return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL; | |
609 | } | |
610 | ||
611 | static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg) | |
612 | { | |
613 | return pd_to_blkg(&cfqg->pd); | |
614 | } | |
615 | ||
616 | static struct blkcg_policy blkcg_policy_cfq; | |
617 | ||
618 | static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg) | |
619 | { | |
620 | return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq)); | |
621 | } | |
622 | ||
623 | static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg) | |
624 | { | |
625 | return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq)); | |
626 | } | |
627 | ||
628 | static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) | |
629 | { | |
630 | struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent; | |
631 | ||
632 | return pblkg ? blkg_to_cfqg(pblkg) : NULL; | |
633 | } | |
634 | ||
635 | static inline bool cfqg_is_descendant(struct cfq_group *cfqg, | |
636 | struct cfq_group *ancestor) | |
637 | { | |
638 | return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup, | |
639 | cfqg_to_blkg(ancestor)->blkcg->css.cgroup); | |
640 | } | |
641 | ||
642 | static inline void cfqg_get(struct cfq_group *cfqg) | |
643 | { | |
644 | return blkg_get(cfqg_to_blkg(cfqg)); | |
645 | } | |
646 | ||
647 | static inline void cfqg_put(struct cfq_group *cfqg) | |
648 | { | |
649 | return blkg_put(cfqg_to_blkg(cfqg)); | |
650 | } | |
651 | ||
652 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) do { \ | |
653 | char __pbuf[128]; \ | |
654 | \ | |
655 | blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf)); \ | |
656 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \ | |
657 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
658 | cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\ | |
659 | __pbuf, ##args); \ | |
660 | } while (0) | |
661 | ||
662 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do { \ | |
663 | char __pbuf[128]; \ | |
664 | \ | |
665 | blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf)); \ | |
666 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args); \ | |
667 | } while (0) | |
668 | ||
669 | static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg, | |
670 | struct cfq_group *curr_cfqg, int op, | |
671 | int op_flags) | |
672 | { | |
673 | blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, 1); | |
674 | cfqg_stats_end_empty_time(&cfqg->stats); | |
675 | cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg); | |
676 | } | |
677 | ||
678 | static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg, | |
679 | uint64_t time, unsigned long unaccounted_time) | |
680 | { | |
681 | blkg_stat_add(&cfqg->stats.time, time); | |
682 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
683 | blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time); | |
684 | #endif | |
685 | } | |
686 | ||
687 | static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op, | |
688 | int op_flags) | |
689 | { | |
690 | blkg_rwstat_add(&cfqg->stats.queued, op, op_flags, -1); | |
691 | } | |
692 | ||
693 | static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op, | |
694 | int op_flags) | |
695 | { | |
696 | blkg_rwstat_add(&cfqg->stats.merged, op, op_flags, 1); | |
697 | } | |
698 | ||
699 | static inline void cfqg_stats_update_completion(struct cfq_group *cfqg, | |
700 | uint64_t start_time, uint64_t io_start_time, int op, | |
701 | int op_flags) | |
702 | { | |
703 | struct cfqg_stats *stats = &cfqg->stats; | |
704 | unsigned long long now = sched_clock(); | |
705 | ||
706 | if (time_after64(now, io_start_time)) | |
707 | blkg_rwstat_add(&stats->service_time, op, op_flags, | |
708 | now - io_start_time); | |
709 | if (time_after64(io_start_time, start_time)) | |
710 | blkg_rwstat_add(&stats->wait_time, op, op_flags, | |
711 | io_start_time - start_time); | |
712 | } | |
713 | ||
714 | /* @stats = 0 */ | |
715 | static void cfqg_stats_reset(struct cfqg_stats *stats) | |
716 | { | |
717 | /* queued stats shouldn't be cleared */ | |
718 | blkg_rwstat_reset(&stats->merged); | |
719 | blkg_rwstat_reset(&stats->service_time); | |
720 | blkg_rwstat_reset(&stats->wait_time); | |
721 | blkg_stat_reset(&stats->time); | |
722 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
723 | blkg_stat_reset(&stats->unaccounted_time); | |
724 | blkg_stat_reset(&stats->avg_queue_size_sum); | |
725 | blkg_stat_reset(&stats->avg_queue_size_samples); | |
726 | blkg_stat_reset(&stats->dequeue); | |
727 | blkg_stat_reset(&stats->group_wait_time); | |
728 | blkg_stat_reset(&stats->idle_time); | |
729 | blkg_stat_reset(&stats->empty_time); | |
730 | #endif | |
731 | } | |
732 | ||
733 | /* @to += @from */ | |
734 | static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from) | |
735 | { | |
736 | /* queued stats shouldn't be cleared */ | |
737 | blkg_rwstat_add_aux(&to->merged, &from->merged); | |
738 | blkg_rwstat_add_aux(&to->service_time, &from->service_time); | |
739 | blkg_rwstat_add_aux(&to->wait_time, &from->wait_time); | |
740 | blkg_stat_add_aux(&from->time, &from->time); | |
741 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
742 | blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time); | |
743 | blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum); | |
744 | blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples); | |
745 | blkg_stat_add_aux(&to->dequeue, &from->dequeue); | |
746 | blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time); | |
747 | blkg_stat_add_aux(&to->idle_time, &from->idle_time); | |
748 | blkg_stat_add_aux(&to->empty_time, &from->empty_time); | |
749 | #endif | |
750 | } | |
751 | ||
752 | /* | |
753 | * Transfer @cfqg's stats to its parent's aux counts so that the ancestors' | |
754 | * recursive stats can still account for the amount used by this cfqg after | |
755 | * it's gone. | |
756 | */ | |
757 | static void cfqg_stats_xfer_dead(struct cfq_group *cfqg) | |
758 | { | |
759 | struct cfq_group *parent = cfqg_parent(cfqg); | |
760 | ||
761 | lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock); | |
762 | ||
763 | if (unlikely(!parent)) | |
764 | return; | |
765 | ||
766 | cfqg_stats_add_aux(&parent->stats, &cfqg->stats); | |
767 | cfqg_stats_reset(&cfqg->stats); | |
768 | } | |
769 | ||
770 | #else /* CONFIG_CFQ_GROUP_IOSCHED */ | |
771 | ||
772 | static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; } | |
773 | static inline bool cfqg_is_descendant(struct cfq_group *cfqg, | |
774 | struct cfq_group *ancestor) | |
775 | { | |
776 | return true; | |
777 | } | |
778 | static inline void cfqg_get(struct cfq_group *cfqg) { } | |
779 | static inline void cfqg_put(struct cfq_group *cfqg) { } | |
780 | ||
781 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
782 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \ | |
783 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
784 | cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\ | |
785 | ##args) | |
786 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0) | |
787 | ||
788 | static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg, | |
789 | struct cfq_group *curr_cfqg, int op, int op_flags) { } | |
790 | static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg, | |
791 | uint64_t time, unsigned long unaccounted_time) { } | |
792 | static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int op, | |
793 | int op_flags) { } | |
794 | static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int op, | |
795 | int op_flags) { } | |
796 | static inline void cfqg_stats_update_completion(struct cfq_group *cfqg, | |
797 | uint64_t start_time, uint64_t io_start_time, int op, | |
798 | int op_flags) { } | |
799 | ||
800 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
801 | ||
802 | #define cfq_log(cfqd, fmt, args...) \ | |
803 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
804 | ||
805 | /* Traverses through cfq group service trees */ | |
806 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
807 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
808 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
809 | : &cfqg->service_tree_idle; \ | |
810 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
811 | (i == IDLE_WORKLOAD && j == 0); \ | |
812 | j++, st = i < IDLE_WORKLOAD ? \ | |
813 | &cfqg->service_trees[i][j]: NULL) \ | |
814 | ||
815 | static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd, | |
816 | struct cfq_ttime *ttime, bool group_idle) | |
817 | { | |
818 | u64 slice; | |
819 | if (!sample_valid(ttime->ttime_samples)) | |
820 | return false; | |
821 | if (group_idle) | |
822 | slice = cfqd->cfq_group_idle; | |
823 | else | |
824 | slice = cfqd->cfq_slice_idle; | |
825 | return ttime->ttime_mean > slice; | |
826 | } | |
827 | ||
828 | static inline bool iops_mode(struct cfq_data *cfqd) | |
829 | { | |
830 | /* | |
831 | * If we are not idling on queues and it is a NCQ drive, parallel | |
832 | * execution of requests is on and measuring time is not possible | |
833 | * in most of the cases until and unless we drive shallower queue | |
834 | * depths and that becomes a performance bottleneck. In such cases | |
835 | * switch to start providing fairness in terms of number of IOs. | |
836 | */ | |
837 | if (!cfqd->cfq_slice_idle && cfqd->hw_tag) | |
838 | return true; | |
839 | else | |
840 | return false; | |
841 | } | |
842 | ||
843 | static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq) | |
844 | { | |
845 | if (cfq_class_idle(cfqq)) | |
846 | return IDLE_WORKLOAD; | |
847 | if (cfq_class_rt(cfqq)) | |
848 | return RT_WORKLOAD; | |
849 | return BE_WORKLOAD; | |
850 | } | |
851 | ||
852 | ||
853 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
854 | { | |
855 | if (!cfq_cfqq_sync(cfqq)) | |
856 | return ASYNC_WORKLOAD; | |
857 | if (!cfq_cfqq_idle_window(cfqq)) | |
858 | return SYNC_NOIDLE_WORKLOAD; | |
859 | return SYNC_WORKLOAD; | |
860 | } | |
861 | ||
862 | static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class, | |
863 | struct cfq_data *cfqd, | |
864 | struct cfq_group *cfqg) | |
865 | { | |
866 | if (wl_class == IDLE_WORKLOAD) | |
867 | return cfqg->service_tree_idle.count; | |
868 | ||
869 | return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count + | |
870 | cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count + | |
871 | cfqg->service_trees[wl_class][SYNC_WORKLOAD].count; | |
872 | } | |
873 | ||
874 | static inline int cfqg_busy_async_queues(struct cfq_data *cfqd, | |
875 | struct cfq_group *cfqg) | |
876 | { | |
877 | return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count + | |
878 | cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count; | |
879 | } | |
880 | ||
881 | static void cfq_dispatch_insert(struct request_queue *, struct request *); | |
882 | static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync, | |
883 | struct cfq_io_cq *cic, struct bio *bio); | |
884 | ||
885 | static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq) | |
886 | { | |
887 | /* cic->icq is the first member, %NULL will convert to %NULL */ | |
888 | return container_of(icq, struct cfq_io_cq, icq); | |
889 | } | |
890 | ||
891 | static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd, | |
892 | struct io_context *ioc) | |
893 | { | |
894 | if (ioc) | |
895 | return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue)); | |
896 | return NULL; | |
897 | } | |
898 | ||
899 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync) | |
900 | { | |
901 | return cic->cfqq[is_sync]; | |
902 | } | |
903 | ||
904 | static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq, | |
905 | bool is_sync) | |
906 | { | |
907 | cic->cfqq[is_sync] = cfqq; | |
908 | } | |
909 | ||
910 | static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic) | |
911 | { | |
912 | return cic->icq.q->elevator->elevator_data; | |
913 | } | |
914 | ||
915 | /* | |
916 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
917 | * set (in which case it could also be direct WRITE). | |
918 | */ | |
919 | static inline bool cfq_bio_sync(struct bio *bio) | |
920 | { | |
921 | return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC); | |
922 | } | |
923 | ||
924 | /* | |
925 | * scheduler run of queue, if there are requests pending and no one in the | |
926 | * driver that will restart queueing | |
927 | */ | |
928 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
929 | { | |
930 | if (cfqd->busy_queues) { | |
931 | cfq_log(cfqd, "schedule dispatch"); | |
932 | kblockd_schedule_work(&cfqd->unplug_work); | |
933 | } | |
934 | } | |
935 | ||
936 | /* | |
937 | * Scale schedule slice based on io priority. Use the sync time slice only | |
938 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
939 | * io only, should not get full sync slice length. | |
940 | */ | |
941 | static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync, | |
942 | unsigned short prio) | |
943 | { | |
944 | u64 base_slice = cfqd->cfq_slice[sync]; | |
945 | u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE); | |
946 | ||
947 | WARN_ON(prio >= IOPRIO_BE_NR); | |
948 | ||
949 | return base_slice + (slice * (4 - prio)); | |
950 | } | |
951 | ||
952 | static inline u64 | |
953 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
954 | { | |
955 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
956 | } | |
957 | ||
958 | /** | |
959 | * cfqg_scale_charge - scale disk time charge according to cfqg weight | |
960 | * @charge: disk time being charged | |
961 | * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT | |
962 | * | |
963 | * Scale @charge according to @vfraction, which is in range (0, 1]. The | |
964 | * scaling is inversely proportional. | |
965 | * | |
966 | * scaled = charge / vfraction | |
967 | * | |
968 | * The result is also in fixed point w/ CFQ_SERVICE_SHIFT. | |
969 | */ | |
970 | static inline u64 cfqg_scale_charge(u64 charge, | |
971 | unsigned int vfraction) | |
972 | { | |
973 | u64 c = charge << CFQ_SERVICE_SHIFT; /* make it fixed point */ | |
974 | ||
975 | /* charge / vfraction */ | |
976 | c <<= CFQ_SERVICE_SHIFT; | |
977 | return div_u64(c, vfraction); | |
978 | } | |
979 | ||
980 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
981 | { | |
982 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
983 | if (delta > 0) | |
984 | min_vdisktime = vdisktime; | |
985 | ||
986 | return min_vdisktime; | |
987 | } | |
988 | ||
989 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
990 | { | |
991 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
992 | if (delta < 0) | |
993 | min_vdisktime = vdisktime; | |
994 | ||
995 | return min_vdisktime; | |
996 | } | |
997 | ||
998 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
999 | { | |
1000 | struct cfq_group *cfqg; | |
1001 | ||
1002 | if (st->left) { | |
1003 | cfqg = rb_entry_cfqg(st->left); | |
1004 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, | |
1005 | cfqg->vdisktime); | |
1006 | } | |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * get averaged number of queues of RT/BE priority. | |
1011 | * average is updated, with a formula that gives more weight to higher numbers, | |
1012 | * to quickly follows sudden increases and decrease slowly | |
1013 | */ | |
1014 | ||
1015 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, | |
1016 | struct cfq_group *cfqg, bool rt) | |
1017 | { | |
1018 | unsigned min_q, max_q; | |
1019 | unsigned mult = cfq_hist_divisor - 1; | |
1020 | unsigned round = cfq_hist_divisor / 2; | |
1021 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); | |
1022 | ||
1023 | min_q = min(cfqg->busy_queues_avg[rt], busy); | |
1024 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
1025 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
1026 | cfq_hist_divisor; | |
1027 | return cfqg->busy_queues_avg[rt]; | |
1028 | } | |
1029 | ||
1030 | static inline u64 | |
1031 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1032 | { | |
1033 | return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT; | |
1034 | } | |
1035 | ||
1036 | static inline u64 | |
1037 | cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1038 | { | |
1039 | u64 slice = cfq_prio_to_slice(cfqd, cfqq); | |
1040 | if (cfqd->cfq_latency) { | |
1041 | /* | |
1042 | * interested queues (we consider only the ones with the same | |
1043 | * priority class in the cfq group) | |
1044 | */ | |
1045 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
1046 | cfq_class_rt(cfqq)); | |
1047 | u64 sync_slice = cfqd->cfq_slice[1]; | |
1048 | u64 expect_latency = sync_slice * iq; | |
1049 | u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg); | |
1050 | ||
1051 | if (expect_latency > group_slice) { | |
1052 | u64 base_low_slice = 2 * cfqd->cfq_slice_idle; | |
1053 | u64 low_slice; | |
1054 | ||
1055 | /* scale low_slice according to IO priority | |
1056 | * and sync vs async */ | |
1057 | low_slice = div64_u64(base_low_slice*slice, sync_slice); | |
1058 | low_slice = min(slice, low_slice); | |
1059 | /* the adapted slice value is scaled to fit all iqs | |
1060 | * into the target latency */ | |
1061 | slice = div64_u64(slice*group_slice, expect_latency); | |
1062 | slice = max(slice, low_slice); | |
1063 | } | |
1064 | } | |
1065 | return slice; | |
1066 | } | |
1067 | ||
1068 | static inline void | |
1069 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1070 | { | |
1071 | u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq); | |
1072 | u64 now = ktime_get_ns(); | |
1073 | ||
1074 | cfqq->slice_start = now; | |
1075 | cfqq->slice_end = now + slice; | |
1076 | cfqq->allocated_slice = slice; | |
1077 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now); | |
1078 | } | |
1079 | ||
1080 | /* | |
1081 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
1082 | * isn't valid until the first request from the dispatch is activated | |
1083 | * and the slice time set. | |
1084 | */ | |
1085 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) | |
1086 | { | |
1087 | if (cfq_cfqq_slice_new(cfqq)) | |
1088 | return false; | |
1089 | if (ktime_get_ns() < cfqq->slice_end) | |
1090 | return false; | |
1091 | ||
1092 | return true; | |
1093 | } | |
1094 | ||
1095 | /* | |
1096 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
1097 | * We choose the request that is closest to the head right now. Distance | |
1098 | * behind the head is penalized and only allowed to a certain extent. | |
1099 | */ | |
1100 | static struct request * | |
1101 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) | |
1102 | { | |
1103 | sector_t s1, s2, d1 = 0, d2 = 0; | |
1104 | unsigned long back_max; | |
1105 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
1106 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
1107 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
1108 | ||
1109 | if (rq1 == NULL || rq1 == rq2) | |
1110 | return rq2; | |
1111 | if (rq2 == NULL) | |
1112 | return rq1; | |
1113 | ||
1114 | if (rq_is_sync(rq1) != rq_is_sync(rq2)) | |
1115 | return rq_is_sync(rq1) ? rq1 : rq2; | |
1116 | ||
1117 | if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO) | |
1118 | return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2; | |
1119 | ||
1120 | s1 = blk_rq_pos(rq1); | |
1121 | s2 = blk_rq_pos(rq2); | |
1122 | ||
1123 | /* | |
1124 | * by definition, 1KiB is 2 sectors | |
1125 | */ | |
1126 | back_max = cfqd->cfq_back_max * 2; | |
1127 | ||
1128 | /* | |
1129 | * Strict one way elevator _except_ in the case where we allow | |
1130 | * short backward seeks which are biased as twice the cost of a | |
1131 | * similar forward seek. | |
1132 | */ | |
1133 | if (s1 >= last) | |
1134 | d1 = s1 - last; | |
1135 | else if (s1 + back_max >= last) | |
1136 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
1137 | else | |
1138 | wrap |= CFQ_RQ1_WRAP; | |
1139 | ||
1140 | if (s2 >= last) | |
1141 | d2 = s2 - last; | |
1142 | else if (s2 + back_max >= last) | |
1143 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
1144 | else | |
1145 | wrap |= CFQ_RQ2_WRAP; | |
1146 | ||
1147 | /* Found required data */ | |
1148 | ||
1149 | /* | |
1150 | * By doing switch() on the bit mask "wrap" we avoid having to | |
1151 | * check two variables for all permutations: --> faster! | |
1152 | */ | |
1153 | switch (wrap) { | |
1154 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
1155 | if (d1 < d2) | |
1156 | return rq1; | |
1157 | else if (d2 < d1) | |
1158 | return rq2; | |
1159 | else { | |
1160 | if (s1 >= s2) | |
1161 | return rq1; | |
1162 | else | |
1163 | return rq2; | |
1164 | } | |
1165 | ||
1166 | case CFQ_RQ2_WRAP: | |
1167 | return rq1; | |
1168 | case CFQ_RQ1_WRAP: | |
1169 | return rq2; | |
1170 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
1171 | default: | |
1172 | /* | |
1173 | * Since both rqs are wrapped, | |
1174 | * start with the one that's further behind head | |
1175 | * (--> only *one* back seek required), | |
1176 | * since back seek takes more time than forward. | |
1177 | */ | |
1178 | if (s1 <= s2) | |
1179 | return rq1; | |
1180 | else | |
1181 | return rq2; | |
1182 | } | |
1183 | } | |
1184 | ||
1185 | /* | |
1186 | * The below is leftmost cache rbtree addon | |
1187 | */ | |
1188 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) | |
1189 | { | |
1190 | /* Service tree is empty */ | |
1191 | if (!root->count) | |
1192 | return NULL; | |
1193 | ||
1194 | if (!root->left) | |
1195 | root->left = rb_first(&root->rb); | |
1196 | ||
1197 | if (root->left) | |
1198 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
1199 | ||
1200 | return NULL; | |
1201 | } | |
1202 | ||
1203 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) | |
1204 | { | |
1205 | if (!root->left) | |
1206 | root->left = rb_first(&root->rb); | |
1207 | ||
1208 | if (root->left) | |
1209 | return rb_entry_cfqg(root->left); | |
1210 | ||
1211 | return NULL; | |
1212 | } | |
1213 | ||
1214 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
1215 | { | |
1216 | rb_erase(n, root); | |
1217 | RB_CLEAR_NODE(n); | |
1218 | } | |
1219 | ||
1220 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) | |
1221 | { | |
1222 | if (root->left == n) | |
1223 | root->left = NULL; | |
1224 | rb_erase_init(n, &root->rb); | |
1225 | --root->count; | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * would be nice to take fifo expire time into account as well | |
1230 | */ | |
1231 | static struct request * | |
1232 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1233 | struct request *last) | |
1234 | { | |
1235 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
1236 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
1237 | struct request *next = NULL, *prev = NULL; | |
1238 | ||
1239 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); | |
1240 | ||
1241 | if (rbprev) | |
1242 | prev = rb_entry_rq(rbprev); | |
1243 | ||
1244 | if (rbnext) | |
1245 | next = rb_entry_rq(rbnext); | |
1246 | else { | |
1247 | rbnext = rb_first(&cfqq->sort_list); | |
1248 | if (rbnext && rbnext != &last->rb_node) | |
1249 | next = rb_entry_rq(rbnext); | |
1250 | } | |
1251 | ||
1252 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); | |
1253 | } | |
1254 | ||
1255 | static u64 cfq_slice_offset(struct cfq_data *cfqd, | |
1256 | struct cfq_queue *cfqq) | |
1257 | { | |
1258 | /* | |
1259 | * just an approximation, should be ok. | |
1260 | */ | |
1261 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - | |
1262 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
1263 | } | |
1264 | ||
1265 | static inline s64 | |
1266 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1267 | { | |
1268 | return cfqg->vdisktime - st->min_vdisktime; | |
1269 | } | |
1270 | ||
1271 | static void | |
1272 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1273 | { | |
1274 | struct rb_node **node = &st->rb.rb_node; | |
1275 | struct rb_node *parent = NULL; | |
1276 | struct cfq_group *__cfqg; | |
1277 | s64 key = cfqg_key(st, cfqg); | |
1278 | int left = 1; | |
1279 | ||
1280 | while (*node != NULL) { | |
1281 | parent = *node; | |
1282 | __cfqg = rb_entry_cfqg(parent); | |
1283 | ||
1284 | if (key < cfqg_key(st, __cfqg)) | |
1285 | node = &parent->rb_left; | |
1286 | else { | |
1287 | node = &parent->rb_right; | |
1288 | left = 0; | |
1289 | } | |
1290 | } | |
1291 | ||
1292 | if (left) | |
1293 | st->left = &cfqg->rb_node; | |
1294 | ||
1295 | rb_link_node(&cfqg->rb_node, parent, node); | |
1296 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
1297 | } | |
1298 | ||
1299 | /* | |
1300 | * This has to be called only on activation of cfqg | |
1301 | */ | |
1302 | static void | |
1303 | cfq_update_group_weight(struct cfq_group *cfqg) | |
1304 | { | |
1305 | if (cfqg->new_weight) { | |
1306 | cfqg->weight = cfqg->new_weight; | |
1307 | cfqg->new_weight = 0; | |
1308 | } | |
1309 | } | |
1310 | ||
1311 | static void | |
1312 | cfq_update_group_leaf_weight(struct cfq_group *cfqg) | |
1313 | { | |
1314 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
1315 | ||
1316 | if (cfqg->new_leaf_weight) { | |
1317 | cfqg->leaf_weight = cfqg->new_leaf_weight; | |
1318 | cfqg->new_leaf_weight = 0; | |
1319 | } | |
1320 | } | |
1321 | ||
1322 | static void | |
1323 | cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1324 | { | |
1325 | unsigned int vfr = 1 << CFQ_SERVICE_SHIFT; /* start with 1 */ | |
1326 | struct cfq_group *pos = cfqg; | |
1327 | struct cfq_group *parent; | |
1328 | bool propagate; | |
1329 | ||
1330 | /* add to the service tree */ | |
1331 | BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node)); | |
1332 | ||
1333 | /* | |
1334 | * Update leaf_weight. We cannot update weight at this point | |
1335 | * because cfqg might already have been activated and is | |
1336 | * contributing its current weight to the parent's child_weight. | |
1337 | */ | |
1338 | cfq_update_group_leaf_weight(cfqg); | |
1339 | __cfq_group_service_tree_add(st, cfqg); | |
1340 | ||
1341 | /* | |
1342 | * Activate @cfqg and calculate the portion of vfraction @cfqg is | |
1343 | * entitled to. vfraction is calculated by walking the tree | |
1344 | * towards the root calculating the fraction it has at each level. | |
1345 | * The compounded ratio is how much vfraction @cfqg owns. | |
1346 | * | |
1347 | * Start with the proportion tasks in this cfqg has against active | |
1348 | * children cfqgs - its leaf_weight against children_weight. | |
1349 | */ | |
1350 | propagate = !pos->nr_active++; | |
1351 | pos->children_weight += pos->leaf_weight; | |
1352 | vfr = vfr * pos->leaf_weight / pos->children_weight; | |
1353 | ||
1354 | /* | |
1355 | * Compound ->weight walking up the tree. Both activation and | |
1356 | * vfraction calculation are done in the same loop. Propagation | |
1357 | * stops once an already activated node is met. vfraction | |
1358 | * calculation should always continue to the root. | |
1359 | */ | |
1360 | while ((parent = cfqg_parent(pos))) { | |
1361 | if (propagate) { | |
1362 | cfq_update_group_weight(pos); | |
1363 | propagate = !parent->nr_active++; | |
1364 | parent->children_weight += pos->weight; | |
1365 | } | |
1366 | vfr = vfr * pos->weight / parent->children_weight; | |
1367 | pos = parent; | |
1368 | } | |
1369 | ||
1370 | cfqg->vfraction = max_t(unsigned, vfr, 1); | |
1371 | } | |
1372 | ||
1373 | static void | |
1374 | cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1375 | { | |
1376 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1377 | struct cfq_group *__cfqg; | |
1378 | struct rb_node *n; | |
1379 | ||
1380 | cfqg->nr_cfqq++; | |
1381 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
1382 | return; | |
1383 | ||
1384 | /* | |
1385 | * Currently put the group at the end. Later implement something | |
1386 | * so that groups get lesser vtime based on their weights, so that | |
1387 | * if group does not loose all if it was not continuously backlogged. | |
1388 | */ | |
1389 | n = rb_last(&st->rb); | |
1390 | if (n) { | |
1391 | __cfqg = rb_entry_cfqg(n); | |
1392 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
1393 | } else | |
1394 | cfqg->vdisktime = st->min_vdisktime; | |
1395 | cfq_group_service_tree_add(st, cfqg); | |
1396 | } | |
1397 | ||
1398 | static void | |
1399 | cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
1400 | { | |
1401 | struct cfq_group *pos = cfqg; | |
1402 | bool propagate; | |
1403 | ||
1404 | /* | |
1405 | * Undo activation from cfq_group_service_tree_add(). Deactivate | |
1406 | * @cfqg and propagate deactivation upwards. | |
1407 | */ | |
1408 | propagate = !--pos->nr_active; | |
1409 | pos->children_weight -= pos->leaf_weight; | |
1410 | ||
1411 | while (propagate) { | |
1412 | struct cfq_group *parent = cfqg_parent(pos); | |
1413 | ||
1414 | /* @pos has 0 nr_active at this point */ | |
1415 | WARN_ON_ONCE(pos->children_weight); | |
1416 | pos->vfraction = 0; | |
1417 | ||
1418 | if (!parent) | |
1419 | break; | |
1420 | ||
1421 | propagate = !--parent->nr_active; | |
1422 | parent->children_weight -= pos->weight; | |
1423 | pos = parent; | |
1424 | } | |
1425 | ||
1426 | /* remove from the service tree */ | |
1427 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
1428 | cfq_rb_erase(&cfqg->rb_node, st); | |
1429 | } | |
1430 | ||
1431 | static void | |
1432 | cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1433 | { | |
1434 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1435 | ||
1436 | BUG_ON(cfqg->nr_cfqq < 1); | |
1437 | cfqg->nr_cfqq--; | |
1438 | ||
1439 | /* If there are other cfq queues under this group, don't delete it */ | |
1440 | if (cfqg->nr_cfqq) | |
1441 | return; | |
1442 | ||
1443 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); | |
1444 | cfq_group_service_tree_del(st, cfqg); | |
1445 | cfqg->saved_wl_slice = 0; | |
1446 | cfqg_stats_update_dequeue(cfqg); | |
1447 | } | |
1448 | ||
1449 | static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq, | |
1450 | u64 *unaccounted_time) | |
1451 | { | |
1452 | u64 slice_used; | |
1453 | u64 now = ktime_get_ns(); | |
1454 | ||
1455 | /* | |
1456 | * Queue got expired before even a single request completed or | |
1457 | * got expired immediately after first request completion. | |
1458 | */ | |
1459 | if (!cfqq->slice_start || cfqq->slice_start == now) { | |
1460 | /* | |
1461 | * Also charge the seek time incurred to the group, otherwise | |
1462 | * if there are mutiple queues in the group, each can dispatch | |
1463 | * a single request on seeky media and cause lots of seek time | |
1464 | * and group will never know it. | |
1465 | */ | |
1466 | slice_used = max_t(u64, (now - cfqq->dispatch_start), 1); | |
1467 | } else { | |
1468 | slice_used = now - cfqq->slice_start; | |
1469 | if (slice_used > cfqq->allocated_slice) { | |
1470 | *unaccounted_time = slice_used - cfqq->allocated_slice; | |
1471 | slice_used = cfqq->allocated_slice; | |
1472 | } | |
1473 | if (cfqq->slice_start > cfqq->dispatch_start) | |
1474 | *unaccounted_time += cfqq->slice_start - | |
1475 | cfqq->dispatch_start; | |
1476 | } | |
1477 | ||
1478 | return slice_used; | |
1479 | } | |
1480 | ||
1481 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
1482 | struct cfq_queue *cfqq) | |
1483 | { | |
1484 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
1485 | u64 used_sl, charge, unaccounted_sl = 0; | |
1486 | int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg) | |
1487 | - cfqg->service_tree_idle.count; | |
1488 | unsigned int vfr; | |
1489 | u64 now = ktime_get_ns(); | |
1490 | ||
1491 | BUG_ON(nr_sync < 0); | |
1492 | used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl); | |
1493 | ||
1494 | if (iops_mode(cfqd)) | |
1495 | charge = cfqq->slice_dispatch; | |
1496 | else if (!cfq_cfqq_sync(cfqq) && !nr_sync) | |
1497 | charge = cfqq->allocated_slice; | |
1498 | ||
1499 | /* | |
1500 | * Can't update vdisktime while on service tree and cfqg->vfraction | |
1501 | * is valid only while on it. Cache vfr, leave the service tree, | |
1502 | * update vdisktime and go back on. The re-addition to the tree | |
1503 | * will also update the weights as necessary. | |
1504 | */ | |
1505 | vfr = cfqg->vfraction; | |
1506 | cfq_group_service_tree_del(st, cfqg); | |
1507 | cfqg->vdisktime += cfqg_scale_charge(charge, vfr); | |
1508 | cfq_group_service_tree_add(st, cfqg); | |
1509 | ||
1510 | /* This group is being expired. Save the context */ | |
1511 | if (cfqd->workload_expires > now) { | |
1512 | cfqg->saved_wl_slice = cfqd->workload_expires - now; | |
1513 | cfqg->saved_wl_type = cfqd->serving_wl_type; | |
1514 | cfqg->saved_wl_class = cfqd->serving_wl_class; | |
1515 | } else | |
1516 | cfqg->saved_wl_slice = 0; | |
1517 | ||
1518 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
1519 | st->min_vdisktime); | |
1520 | cfq_log_cfqq(cfqq->cfqd, cfqq, | |
1521 | "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu", | |
1522 | used_sl, cfqq->slice_dispatch, charge, | |
1523 | iops_mode(cfqd), cfqq->nr_sectors); | |
1524 | cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl); | |
1525 | cfqg_stats_set_start_empty_time(cfqg); | |
1526 | } | |
1527 | ||
1528 | /** | |
1529 | * cfq_init_cfqg_base - initialize base part of a cfq_group | |
1530 | * @cfqg: cfq_group to initialize | |
1531 | * | |
1532 | * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED | |
1533 | * is enabled or not. | |
1534 | */ | |
1535 | static void cfq_init_cfqg_base(struct cfq_group *cfqg) | |
1536 | { | |
1537 | struct cfq_rb_root *st; | |
1538 | int i, j; | |
1539 | ||
1540 | for_each_cfqg_st(cfqg, i, j, st) | |
1541 | *st = CFQ_RB_ROOT; | |
1542 | RB_CLEAR_NODE(&cfqg->rb_node); | |
1543 | ||
1544 | cfqg->ttime.last_end_request = ktime_get_ns(); | |
1545 | } | |
1546 | ||
1547 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
1548 | static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val, | |
1549 | bool on_dfl, bool reset_dev, bool is_leaf_weight); | |
1550 | ||
1551 | static void cfqg_stats_exit(struct cfqg_stats *stats) | |
1552 | { | |
1553 | blkg_rwstat_exit(&stats->merged); | |
1554 | blkg_rwstat_exit(&stats->service_time); | |
1555 | blkg_rwstat_exit(&stats->wait_time); | |
1556 | blkg_rwstat_exit(&stats->queued); | |
1557 | blkg_stat_exit(&stats->time); | |
1558 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
1559 | blkg_stat_exit(&stats->unaccounted_time); | |
1560 | blkg_stat_exit(&stats->avg_queue_size_sum); | |
1561 | blkg_stat_exit(&stats->avg_queue_size_samples); | |
1562 | blkg_stat_exit(&stats->dequeue); | |
1563 | blkg_stat_exit(&stats->group_wait_time); | |
1564 | blkg_stat_exit(&stats->idle_time); | |
1565 | blkg_stat_exit(&stats->empty_time); | |
1566 | #endif | |
1567 | } | |
1568 | ||
1569 | static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp) | |
1570 | { | |
1571 | if (blkg_rwstat_init(&stats->merged, gfp) || | |
1572 | blkg_rwstat_init(&stats->service_time, gfp) || | |
1573 | blkg_rwstat_init(&stats->wait_time, gfp) || | |
1574 | blkg_rwstat_init(&stats->queued, gfp) || | |
1575 | blkg_stat_init(&stats->time, gfp)) | |
1576 | goto err; | |
1577 | ||
1578 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
1579 | if (blkg_stat_init(&stats->unaccounted_time, gfp) || | |
1580 | blkg_stat_init(&stats->avg_queue_size_sum, gfp) || | |
1581 | blkg_stat_init(&stats->avg_queue_size_samples, gfp) || | |
1582 | blkg_stat_init(&stats->dequeue, gfp) || | |
1583 | blkg_stat_init(&stats->group_wait_time, gfp) || | |
1584 | blkg_stat_init(&stats->idle_time, gfp) || | |
1585 | blkg_stat_init(&stats->empty_time, gfp)) | |
1586 | goto err; | |
1587 | #endif | |
1588 | return 0; | |
1589 | err: | |
1590 | cfqg_stats_exit(stats); | |
1591 | return -ENOMEM; | |
1592 | } | |
1593 | ||
1594 | static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp) | |
1595 | { | |
1596 | struct cfq_group_data *cgd; | |
1597 | ||
1598 | cgd = kzalloc(sizeof(*cgd), GFP_KERNEL); | |
1599 | if (!cgd) | |
1600 | return NULL; | |
1601 | return &cgd->cpd; | |
1602 | } | |
1603 | ||
1604 | static void cfq_cpd_init(struct blkcg_policy_data *cpd) | |
1605 | { | |
1606 | struct cfq_group_data *cgd = cpd_to_cfqgd(cpd); | |
1607 | unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ? | |
1608 | CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL; | |
1609 | ||
1610 | if (cpd_to_blkcg(cpd) == &blkcg_root) | |
1611 | weight *= 2; | |
1612 | ||
1613 | cgd->weight = weight; | |
1614 | cgd->leaf_weight = weight; | |
1615 | } | |
1616 | ||
1617 | static void cfq_cpd_free(struct blkcg_policy_data *cpd) | |
1618 | { | |
1619 | kfree(cpd_to_cfqgd(cpd)); | |
1620 | } | |
1621 | ||
1622 | static void cfq_cpd_bind(struct blkcg_policy_data *cpd) | |
1623 | { | |
1624 | struct blkcg *blkcg = cpd_to_blkcg(cpd); | |
1625 | bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys); | |
1626 | unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL; | |
1627 | ||
1628 | if (blkcg == &blkcg_root) | |
1629 | weight *= 2; | |
1630 | ||
1631 | WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false)); | |
1632 | WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true)); | |
1633 | } | |
1634 | ||
1635 | static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node) | |
1636 | { | |
1637 | struct cfq_group *cfqg; | |
1638 | ||
1639 | cfqg = kzalloc_node(sizeof(*cfqg), gfp, node); | |
1640 | if (!cfqg) | |
1641 | return NULL; | |
1642 | ||
1643 | cfq_init_cfqg_base(cfqg); | |
1644 | if (cfqg_stats_init(&cfqg->stats, gfp)) { | |
1645 | kfree(cfqg); | |
1646 | return NULL; | |
1647 | } | |
1648 | ||
1649 | return &cfqg->pd; | |
1650 | } | |
1651 | ||
1652 | static void cfq_pd_init(struct blkg_policy_data *pd) | |
1653 | { | |
1654 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1655 | struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg); | |
1656 | ||
1657 | cfqg->weight = cgd->weight; | |
1658 | cfqg->leaf_weight = cgd->leaf_weight; | |
1659 | } | |
1660 | ||
1661 | static void cfq_pd_offline(struct blkg_policy_data *pd) | |
1662 | { | |
1663 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1664 | int i; | |
1665 | ||
1666 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
1667 | if (cfqg->async_cfqq[0][i]) | |
1668 | cfq_put_queue(cfqg->async_cfqq[0][i]); | |
1669 | if (cfqg->async_cfqq[1][i]) | |
1670 | cfq_put_queue(cfqg->async_cfqq[1][i]); | |
1671 | } | |
1672 | ||
1673 | if (cfqg->async_idle_cfqq) | |
1674 | cfq_put_queue(cfqg->async_idle_cfqq); | |
1675 | ||
1676 | /* | |
1677 | * @blkg is going offline and will be ignored by | |
1678 | * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so | |
1679 | * that they don't get lost. If IOs complete after this point, the | |
1680 | * stats for them will be lost. Oh well... | |
1681 | */ | |
1682 | cfqg_stats_xfer_dead(cfqg); | |
1683 | } | |
1684 | ||
1685 | static void cfq_pd_free(struct blkg_policy_data *pd) | |
1686 | { | |
1687 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1688 | ||
1689 | cfqg_stats_exit(&cfqg->stats); | |
1690 | return kfree(cfqg); | |
1691 | } | |
1692 | ||
1693 | static void cfq_pd_reset_stats(struct blkg_policy_data *pd) | |
1694 | { | |
1695 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1696 | ||
1697 | cfqg_stats_reset(&cfqg->stats); | |
1698 | } | |
1699 | ||
1700 | static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd, | |
1701 | struct blkcg *blkcg) | |
1702 | { | |
1703 | struct blkcg_gq *blkg; | |
1704 | ||
1705 | blkg = blkg_lookup(blkcg, cfqd->queue); | |
1706 | if (likely(blkg)) | |
1707 | return blkg_to_cfqg(blkg); | |
1708 | return NULL; | |
1709 | } | |
1710 | ||
1711 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) | |
1712 | { | |
1713 | cfqq->cfqg = cfqg; | |
1714 | /* cfqq reference on cfqg */ | |
1715 | cfqg_get(cfqg); | |
1716 | } | |
1717 | ||
1718 | static u64 cfqg_prfill_weight_device(struct seq_file *sf, | |
1719 | struct blkg_policy_data *pd, int off) | |
1720 | { | |
1721 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1722 | ||
1723 | if (!cfqg->dev_weight) | |
1724 | return 0; | |
1725 | return __blkg_prfill_u64(sf, pd, cfqg->dev_weight); | |
1726 | } | |
1727 | ||
1728 | static int cfqg_print_weight_device(struct seq_file *sf, void *v) | |
1729 | { | |
1730 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1731 | cfqg_prfill_weight_device, &blkcg_policy_cfq, | |
1732 | 0, false); | |
1733 | return 0; | |
1734 | } | |
1735 | ||
1736 | static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf, | |
1737 | struct blkg_policy_data *pd, int off) | |
1738 | { | |
1739 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1740 | ||
1741 | if (!cfqg->dev_leaf_weight) | |
1742 | return 0; | |
1743 | return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight); | |
1744 | } | |
1745 | ||
1746 | static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v) | |
1747 | { | |
1748 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1749 | cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, | |
1750 | 0, false); | |
1751 | return 0; | |
1752 | } | |
1753 | ||
1754 | static int cfq_print_weight(struct seq_file *sf, void *v) | |
1755 | { | |
1756 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
1757 | struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg); | |
1758 | unsigned int val = 0; | |
1759 | ||
1760 | if (cgd) | |
1761 | val = cgd->weight; | |
1762 | ||
1763 | seq_printf(sf, "%u\n", val); | |
1764 | return 0; | |
1765 | } | |
1766 | ||
1767 | static int cfq_print_leaf_weight(struct seq_file *sf, void *v) | |
1768 | { | |
1769 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
1770 | struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg); | |
1771 | unsigned int val = 0; | |
1772 | ||
1773 | if (cgd) | |
1774 | val = cgd->leaf_weight; | |
1775 | ||
1776 | seq_printf(sf, "%u\n", val); | |
1777 | return 0; | |
1778 | } | |
1779 | ||
1780 | static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of, | |
1781 | char *buf, size_t nbytes, loff_t off, | |
1782 | bool on_dfl, bool is_leaf_weight) | |
1783 | { | |
1784 | unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN; | |
1785 | unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX; | |
1786 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
1787 | struct blkg_conf_ctx ctx; | |
1788 | struct cfq_group *cfqg; | |
1789 | struct cfq_group_data *cfqgd; | |
1790 | int ret; | |
1791 | u64 v; | |
1792 | ||
1793 | ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx); | |
1794 | if (ret) | |
1795 | return ret; | |
1796 | ||
1797 | if (sscanf(ctx.body, "%llu", &v) == 1) { | |
1798 | /* require "default" on dfl */ | |
1799 | ret = -ERANGE; | |
1800 | if (!v && on_dfl) | |
1801 | goto out_finish; | |
1802 | } else if (!strcmp(strim(ctx.body), "default")) { | |
1803 | v = 0; | |
1804 | } else { | |
1805 | ret = -EINVAL; | |
1806 | goto out_finish; | |
1807 | } | |
1808 | ||
1809 | cfqg = blkg_to_cfqg(ctx.blkg); | |
1810 | cfqgd = blkcg_to_cfqgd(blkcg); | |
1811 | ||
1812 | ret = -ERANGE; | |
1813 | if (!v || (v >= min && v <= max)) { | |
1814 | if (!is_leaf_weight) { | |
1815 | cfqg->dev_weight = v; | |
1816 | cfqg->new_weight = v ?: cfqgd->weight; | |
1817 | } else { | |
1818 | cfqg->dev_leaf_weight = v; | |
1819 | cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight; | |
1820 | } | |
1821 | ret = 0; | |
1822 | } | |
1823 | out_finish: | |
1824 | blkg_conf_finish(&ctx); | |
1825 | return ret ?: nbytes; | |
1826 | } | |
1827 | ||
1828 | static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of, | |
1829 | char *buf, size_t nbytes, loff_t off) | |
1830 | { | |
1831 | return __cfqg_set_weight_device(of, buf, nbytes, off, false, false); | |
1832 | } | |
1833 | ||
1834 | static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of, | |
1835 | char *buf, size_t nbytes, loff_t off) | |
1836 | { | |
1837 | return __cfqg_set_weight_device(of, buf, nbytes, off, false, true); | |
1838 | } | |
1839 | ||
1840 | static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val, | |
1841 | bool on_dfl, bool reset_dev, bool is_leaf_weight) | |
1842 | { | |
1843 | unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN; | |
1844 | unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX; | |
1845 | struct blkcg *blkcg = css_to_blkcg(css); | |
1846 | struct blkcg_gq *blkg; | |
1847 | struct cfq_group_data *cfqgd; | |
1848 | int ret = 0; | |
1849 | ||
1850 | if (val < min || val > max) | |
1851 | return -ERANGE; | |
1852 | ||
1853 | spin_lock_irq(&blkcg->lock); | |
1854 | cfqgd = blkcg_to_cfqgd(blkcg); | |
1855 | if (!cfqgd) { | |
1856 | ret = -EINVAL; | |
1857 | goto out; | |
1858 | } | |
1859 | ||
1860 | if (!is_leaf_weight) | |
1861 | cfqgd->weight = val; | |
1862 | else | |
1863 | cfqgd->leaf_weight = val; | |
1864 | ||
1865 | hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { | |
1866 | struct cfq_group *cfqg = blkg_to_cfqg(blkg); | |
1867 | ||
1868 | if (!cfqg) | |
1869 | continue; | |
1870 | ||
1871 | if (!is_leaf_weight) { | |
1872 | if (reset_dev) | |
1873 | cfqg->dev_weight = 0; | |
1874 | if (!cfqg->dev_weight) | |
1875 | cfqg->new_weight = cfqgd->weight; | |
1876 | } else { | |
1877 | if (reset_dev) | |
1878 | cfqg->dev_leaf_weight = 0; | |
1879 | if (!cfqg->dev_leaf_weight) | |
1880 | cfqg->new_leaf_weight = cfqgd->leaf_weight; | |
1881 | } | |
1882 | } | |
1883 | ||
1884 | out: | |
1885 | spin_unlock_irq(&blkcg->lock); | |
1886 | return ret; | |
1887 | } | |
1888 | ||
1889 | static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft, | |
1890 | u64 val) | |
1891 | { | |
1892 | return __cfq_set_weight(css, val, false, false, false); | |
1893 | } | |
1894 | ||
1895 | static int cfq_set_leaf_weight(struct cgroup_subsys_state *css, | |
1896 | struct cftype *cft, u64 val) | |
1897 | { | |
1898 | return __cfq_set_weight(css, val, false, false, true); | |
1899 | } | |
1900 | ||
1901 | static int cfqg_print_stat(struct seq_file *sf, void *v) | |
1902 | { | |
1903 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat, | |
1904 | &blkcg_policy_cfq, seq_cft(sf)->private, false); | |
1905 | return 0; | |
1906 | } | |
1907 | ||
1908 | static int cfqg_print_rwstat(struct seq_file *sf, void *v) | |
1909 | { | |
1910 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat, | |
1911 | &blkcg_policy_cfq, seq_cft(sf)->private, true); | |
1912 | return 0; | |
1913 | } | |
1914 | ||
1915 | static u64 cfqg_prfill_stat_recursive(struct seq_file *sf, | |
1916 | struct blkg_policy_data *pd, int off) | |
1917 | { | |
1918 | u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd), | |
1919 | &blkcg_policy_cfq, off); | |
1920 | return __blkg_prfill_u64(sf, pd, sum); | |
1921 | } | |
1922 | ||
1923 | static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf, | |
1924 | struct blkg_policy_data *pd, int off) | |
1925 | { | |
1926 | struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd), | |
1927 | &blkcg_policy_cfq, off); | |
1928 | return __blkg_prfill_rwstat(sf, pd, &sum); | |
1929 | } | |
1930 | ||
1931 | static int cfqg_print_stat_recursive(struct seq_file *sf, void *v) | |
1932 | { | |
1933 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1934 | cfqg_prfill_stat_recursive, &blkcg_policy_cfq, | |
1935 | seq_cft(sf)->private, false); | |
1936 | return 0; | |
1937 | } | |
1938 | ||
1939 | static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v) | |
1940 | { | |
1941 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1942 | cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq, | |
1943 | seq_cft(sf)->private, true); | |
1944 | return 0; | |
1945 | } | |
1946 | ||
1947 | static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd, | |
1948 | int off) | |
1949 | { | |
1950 | u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes); | |
1951 | ||
1952 | return __blkg_prfill_u64(sf, pd, sum >> 9); | |
1953 | } | |
1954 | ||
1955 | static int cfqg_print_stat_sectors(struct seq_file *sf, void *v) | |
1956 | { | |
1957 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1958 | cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false); | |
1959 | return 0; | |
1960 | } | |
1961 | ||
1962 | static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf, | |
1963 | struct blkg_policy_data *pd, int off) | |
1964 | { | |
1965 | struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL, | |
1966 | offsetof(struct blkcg_gq, stat_bytes)); | |
1967 | u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) + | |
1968 | atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]); | |
1969 | ||
1970 | return __blkg_prfill_u64(sf, pd, sum >> 9); | |
1971 | } | |
1972 | ||
1973 | static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v) | |
1974 | { | |
1975 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
1976 | cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0, | |
1977 | false); | |
1978 | return 0; | |
1979 | } | |
1980 | ||
1981 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
1982 | static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf, | |
1983 | struct blkg_policy_data *pd, int off) | |
1984 | { | |
1985 | struct cfq_group *cfqg = pd_to_cfqg(pd); | |
1986 | u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples); | |
1987 | u64 v = 0; | |
1988 | ||
1989 | if (samples) { | |
1990 | v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum); | |
1991 | v = div64_u64(v, samples); | |
1992 | } | |
1993 | __blkg_prfill_u64(sf, pd, v); | |
1994 | return 0; | |
1995 | } | |
1996 | ||
1997 | /* print avg_queue_size */ | |
1998 | static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v) | |
1999 | { | |
2000 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
2001 | cfqg_prfill_avg_queue_size, &blkcg_policy_cfq, | |
2002 | 0, false); | |
2003 | return 0; | |
2004 | } | |
2005 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
2006 | ||
2007 | static struct cftype cfq_blkcg_legacy_files[] = { | |
2008 | /* on root, weight is mapped to leaf_weight */ | |
2009 | { | |
2010 | .name = "weight_device", | |
2011 | .flags = CFTYPE_ONLY_ON_ROOT, | |
2012 | .seq_show = cfqg_print_leaf_weight_device, | |
2013 | .write = cfqg_set_leaf_weight_device, | |
2014 | }, | |
2015 | { | |
2016 | .name = "weight", | |
2017 | .flags = CFTYPE_ONLY_ON_ROOT, | |
2018 | .seq_show = cfq_print_leaf_weight, | |
2019 | .write_u64 = cfq_set_leaf_weight, | |
2020 | }, | |
2021 | ||
2022 | /* no such mapping necessary for !roots */ | |
2023 | { | |
2024 | .name = "weight_device", | |
2025 | .flags = CFTYPE_NOT_ON_ROOT, | |
2026 | .seq_show = cfqg_print_weight_device, | |
2027 | .write = cfqg_set_weight_device, | |
2028 | }, | |
2029 | { | |
2030 | .name = "weight", | |
2031 | .flags = CFTYPE_NOT_ON_ROOT, | |
2032 | .seq_show = cfq_print_weight, | |
2033 | .write_u64 = cfq_set_weight, | |
2034 | }, | |
2035 | ||
2036 | { | |
2037 | .name = "leaf_weight_device", | |
2038 | .seq_show = cfqg_print_leaf_weight_device, | |
2039 | .write = cfqg_set_leaf_weight_device, | |
2040 | }, | |
2041 | { | |
2042 | .name = "leaf_weight", | |
2043 | .seq_show = cfq_print_leaf_weight, | |
2044 | .write_u64 = cfq_set_leaf_weight, | |
2045 | }, | |
2046 | ||
2047 | /* statistics, covers only the tasks in the cfqg */ | |
2048 | { | |
2049 | .name = "time", | |
2050 | .private = offsetof(struct cfq_group, stats.time), | |
2051 | .seq_show = cfqg_print_stat, | |
2052 | }, | |
2053 | { | |
2054 | .name = "sectors", | |
2055 | .seq_show = cfqg_print_stat_sectors, | |
2056 | }, | |
2057 | { | |
2058 | .name = "io_service_bytes", | |
2059 | .private = (unsigned long)&blkcg_policy_cfq, | |
2060 | .seq_show = blkg_print_stat_bytes, | |
2061 | }, | |
2062 | { | |
2063 | .name = "io_serviced", | |
2064 | .private = (unsigned long)&blkcg_policy_cfq, | |
2065 | .seq_show = blkg_print_stat_ios, | |
2066 | }, | |
2067 | { | |
2068 | .name = "io_service_time", | |
2069 | .private = offsetof(struct cfq_group, stats.service_time), | |
2070 | .seq_show = cfqg_print_rwstat, | |
2071 | }, | |
2072 | { | |
2073 | .name = "io_wait_time", | |
2074 | .private = offsetof(struct cfq_group, stats.wait_time), | |
2075 | .seq_show = cfqg_print_rwstat, | |
2076 | }, | |
2077 | { | |
2078 | .name = "io_merged", | |
2079 | .private = offsetof(struct cfq_group, stats.merged), | |
2080 | .seq_show = cfqg_print_rwstat, | |
2081 | }, | |
2082 | { | |
2083 | .name = "io_queued", | |
2084 | .private = offsetof(struct cfq_group, stats.queued), | |
2085 | .seq_show = cfqg_print_rwstat, | |
2086 | }, | |
2087 | ||
2088 | /* the same statictics which cover the cfqg and its descendants */ | |
2089 | { | |
2090 | .name = "time_recursive", | |
2091 | .private = offsetof(struct cfq_group, stats.time), | |
2092 | .seq_show = cfqg_print_stat_recursive, | |
2093 | }, | |
2094 | { | |
2095 | .name = "sectors_recursive", | |
2096 | .seq_show = cfqg_print_stat_sectors_recursive, | |
2097 | }, | |
2098 | { | |
2099 | .name = "io_service_bytes_recursive", | |
2100 | .private = (unsigned long)&blkcg_policy_cfq, | |
2101 | .seq_show = blkg_print_stat_bytes_recursive, | |
2102 | }, | |
2103 | { | |
2104 | .name = "io_serviced_recursive", | |
2105 | .private = (unsigned long)&blkcg_policy_cfq, | |
2106 | .seq_show = blkg_print_stat_ios_recursive, | |
2107 | }, | |
2108 | { | |
2109 | .name = "io_service_time_recursive", | |
2110 | .private = offsetof(struct cfq_group, stats.service_time), | |
2111 | .seq_show = cfqg_print_rwstat_recursive, | |
2112 | }, | |
2113 | { | |
2114 | .name = "io_wait_time_recursive", | |
2115 | .private = offsetof(struct cfq_group, stats.wait_time), | |
2116 | .seq_show = cfqg_print_rwstat_recursive, | |
2117 | }, | |
2118 | { | |
2119 | .name = "io_merged_recursive", | |
2120 | .private = offsetof(struct cfq_group, stats.merged), | |
2121 | .seq_show = cfqg_print_rwstat_recursive, | |
2122 | }, | |
2123 | { | |
2124 | .name = "io_queued_recursive", | |
2125 | .private = offsetof(struct cfq_group, stats.queued), | |
2126 | .seq_show = cfqg_print_rwstat_recursive, | |
2127 | }, | |
2128 | #ifdef CONFIG_DEBUG_BLK_CGROUP | |
2129 | { | |
2130 | .name = "avg_queue_size", | |
2131 | .seq_show = cfqg_print_avg_queue_size, | |
2132 | }, | |
2133 | { | |
2134 | .name = "group_wait_time", | |
2135 | .private = offsetof(struct cfq_group, stats.group_wait_time), | |
2136 | .seq_show = cfqg_print_stat, | |
2137 | }, | |
2138 | { | |
2139 | .name = "idle_time", | |
2140 | .private = offsetof(struct cfq_group, stats.idle_time), | |
2141 | .seq_show = cfqg_print_stat, | |
2142 | }, | |
2143 | { | |
2144 | .name = "empty_time", | |
2145 | .private = offsetof(struct cfq_group, stats.empty_time), | |
2146 | .seq_show = cfqg_print_stat, | |
2147 | }, | |
2148 | { | |
2149 | .name = "dequeue", | |
2150 | .private = offsetof(struct cfq_group, stats.dequeue), | |
2151 | .seq_show = cfqg_print_stat, | |
2152 | }, | |
2153 | { | |
2154 | .name = "unaccounted_time", | |
2155 | .private = offsetof(struct cfq_group, stats.unaccounted_time), | |
2156 | .seq_show = cfqg_print_stat, | |
2157 | }, | |
2158 | #endif /* CONFIG_DEBUG_BLK_CGROUP */ | |
2159 | { } /* terminate */ | |
2160 | }; | |
2161 | ||
2162 | static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v) | |
2163 | { | |
2164 | struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); | |
2165 | struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg); | |
2166 | ||
2167 | seq_printf(sf, "default %u\n", cgd->weight); | |
2168 | blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device, | |
2169 | &blkcg_policy_cfq, 0, false); | |
2170 | return 0; | |
2171 | } | |
2172 | ||
2173 | static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of, | |
2174 | char *buf, size_t nbytes, loff_t off) | |
2175 | { | |
2176 | char *endp; | |
2177 | int ret; | |
2178 | u64 v; | |
2179 | ||
2180 | buf = strim(buf); | |
2181 | ||
2182 | /* "WEIGHT" or "default WEIGHT" sets the default weight */ | |
2183 | v = simple_strtoull(buf, &endp, 0); | |
2184 | if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) { | |
2185 | ret = __cfq_set_weight(of_css(of), v, true, false, false); | |
2186 | return ret ?: nbytes; | |
2187 | } | |
2188 | ||
2189 | /* "MAJ:MIN WEIGHT" */ | |
2190 | return __cfqg_set_weight_device(of, buf, nbytes, off, true, false); | |
2191 | } | |
2192 | ||
2193 | static struct cftype cfq_blkcg_files[] = { | |
2194 | { | |
2195 | .name = "weight", | |
2196 | .flags = CFTYPE_NOT_ON_ROOT, | |
2197 | .seq_show = cfq_print_weight_on_dfl, | |
2198 | .write = cfq_set_weight_on_dfl, | |
2199 | }, | |
2200 | { } /* terminate */ | |
2201 | }; | |
2202 | ||
2203 | #else /* GROUP_IOSCHED */ | |
2204 | static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd, | |
2205 | struct blkcg *blkcg) | |
2206 | { | |
2207 | return cfqd->root_group; | |
2208 | } | |
2209 | ||
2210 | static inline void | |
2211 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
2212 | cfqq->cfqg = cfqg; | |
2213 | } | |
2214 | ||
2215 | #endif /* GROUP_IOSCHED */ | |
2216 | ||
2217 | /* | |
2218 | * The cfqd->service_trees holds all pending cfq_queue's that have | |
2219 | * requests waiting to be processed. It is sorted in the order that | |
2220 | * we will service the queues. | |
2221 | */ | |
2222 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2223 | bool add_front) | |
2224 | { | |
2225 | struct rb_node **p, *parent; | |
2226 | struct cfq_queue *__cfqq; | |
2227 | u64 rb_key; | |
2228 | struct cfq_rb_root *st; | |
2229 | int left; | |
2230 | int new_cfqq = 1; | |
2231 | u64 now = ktime_get_ns(); | |
2232 | ||
2233 | st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq)); | |
2234 | if (cfq_class_idle(cfqq)) { | |
2235 | rb_key = CFQ_IDLE_DELAY; | |
2236 | parent = rb_last(&st->rb); | |
2237 | if (parent && parent != &cfqq->rb_node) { | |
2238 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
2239 | rb_key += __cfqq->rb_key; | |
2240 | } else | |
2241 | rb_key += now; | |
2242 | } else if (!add_front) { | |
2243 | /* | |
2244 | * Get our rb key offset. Subtract any residual slice | |
2245 | * value carried from last service. A negative resid | |
2246 | * count indicates slice overrun, and this should position | |
2247 | * the next service time further away in the tree. | |
2248 | */ | |
2249 | rb_key = cfq_slice_offset(cfqd, cfqq) + now; | |
2250 | rb_key -= cfqq->slice_resid; | |
2251 | cfqq->slice_resid = 0; | |
2252 | } else { | |
2253 | rb_key = -NSEC_PER_SEC; | |
2254 | __cfqq = cfq_rb_first(st); | |
2255 | rb_key += __cfqq ? __cfqq->rb_key : now; | |
2256 | } | |
2257 | ||
2258 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
2259 | new_cfqq = 0; | |
2260 | /* | |
2261 | * same position, nothing more to do | |
2262 | */ | |
2263 | if (rb_key == cfqq->rb_key && cfqq->service_tree == st) | |
2264 | return; | |
2265 | ||
2266 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
2267 | cfqq->service_tree = NULL; | |
2268 | } | |
2269 | ||
2270 | left = 1; | |
2271 | parent = NULL; | |
2272 | cfqq->service_tree = st; | |
2273 | p = &st->rb.rb_node; | |
2274 | while (*p) { | |
2275 | parent = *p; | |
2276 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
2277 | ||
2278 | /* | |
2279 | * sort by key, that represents service time. | |
2280 | */ | |
2281 | if (rb_key < __cfqq->rb_key) | |
2282 | p = &parent->rb_left; | |
2283 | else { | |
2284 | p = &parent->rb_right; | |
2285 | left = 0; | |
2286 | } | |
2287 | } | |
2288 | ||
2289 | if (left) | |
2290 | st->left = &cfqq->rb_node; | |
2291 | ||
2292 | cfqq->rb_key = rb_key; | |
2293 | rb_link_node(&cfqq->rb_node, parent, p); | |
2294 | rb_insert_color(&cfqq->rb_node, &st->rb); | |
2295 | st->count++; | |
2296 | if (add_front || !new_cfqq) | |
2297 | return; | |
2298 | cfq_group_notify_queue_add(cfqd, cfqq->cfqg); | |
2299 | } | |
2300 | ||
2301 | static struct cfq_queue * | |
2302 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, | |
2303 | sector_t sector, struct rb_node **ret_parent, | |
2304 | struct rb_node ***rb_link) | |
2305 | { | |
2306 | struct rb_node **p, *parent; | |
2307 | struct cfq_queue *cfqq = NULL; | |
2308 | ||
2309 | parent = NULL; | |
2310 | p = &root->rb_node; | |
2311 | while (*p) { | |
2312 | struct rb_node **n; | |
2313 | ||
2314 | parent = *p; | |
2315 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
2316 | ||
2317 | /* | |
2318 | * Sort strictly based on sector. Smallest to the left, | |
2319 | * largest to the right. | |
2320 | */ | |
2321 | if (sector > blk_rq_pos(cfqq->next_rq)) | |
2322 | n = &(*p)->rb_right; | |
2323 | else if (sector < blk_rq_pos(cfqq->next_rq)) | |
2324 | n = &(*p)->rb_left; | |
2325 | else | |
2326 | break; | |
2327 | p = n; | |
2328 | cfqq = NULL; | |
2329 | } | |
2330 | ||
2331 | *ret_parent = parent; | |
2332 | if (rb_link) | |
2333 | *rb_link = p; | |
2334 | return cfqq; | |
2335 | } | |
2336 | ||
2337 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2338 | { | |
2339 | struct rb_node **p, *parent; | |
2340 | struct cfq_queue *__cfqq; | |
2341 | ||
2342 | if (cfqq->p_root) { | |
2343 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
2344 | cfqq->p_root = NULL; | |
2345 | } | |
2346 | ||
2347 | if (cfq_class_idle(cfqq)) | |
2348 | return; | |
2349 | if (!cfqq->next_rq) | |
2350 | return; | |
2351 | ||
2352 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; | |
2353 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, | |
2354 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
2355 | if (!__cfqq) { | |
2356 | rb_link_node(&cfqq->p_node, parent, p); | |
2357 | rb_insert_color(&cfqq->p_node, cfqq->p_root); | |
2358 | } else | |
2359 | cfqq->p_root = NULL; | |
2360 | } | |
2361 | ||
2362 | /* | |
2363 | * Update cfqq's position in the service tree. | |
2364 | */ | |
2365 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2366 | { | |
2367 | /* | |
2368 | * Resorting requires the cfqq to be on the RR list already. | |
2369 | */ | |
2370 | if (cfq_cfqq_on_rr(cfqq)) { | |
2371 | cfq_service_tree_add(cfqd, cfqq, 0); | |
2372 | cfq_prio_tree_add(cfqd, cfqq); | |
2373 | } | |
2374 | } | |
2375 | ||
2376 | /* | |
2377 | * add to busy list of queues for service, trying to be fair in ordering | |
2378 | * the pending list according to last request service | |
2379 | */ | |
2380 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2381 | { | |
2382 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); | |
2383 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
2384 | cfq_mark_cfqq_on_rr(cfqq); | |
2385 | cfqd->busy_queues++; | |
2386 | if (cfq_cfqq_sync(cfqq)) | |
2387 | cfqd->busy_sync_queues++; | |
2388 | ||
2389 | cfq_resort_rr_list(cfqd, cfqq); | |
2390 | } | |
2391 | ||
2392 | /* | |
2393 | * Called when the cfqq no longer has requests pending, remove it from | |
2394 | * the service tree. | |
2395 | */ | |
2396 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2397 | { | |
2398 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); | |
2399 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
2400 | cfq_clear_cfqq_on_rr(cfqq); | |
2401 | ||
2402 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
2403 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
2404 | cfqq->service_tree = NULL; | |
2405 | } | |
2406 | if (cfqq->p_root) { | |
2407 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
2408 | cfqq->p_root = NULL; | |
2409 | } | |
2410 | ||
2411 | cfq_group_notify_queue_del(cfqd, cfqq->cfqg); | |
2412 | BUG_ON(!cfqd->busy_queues); | |
2413 | cfqd->busy_queues--; | |
2414 | if (cfq_cfqq_sync(cfqq)) | |
2415 | cfqd->busy_sync_queues--; | |
2416 | } | |
2417 | ||
2418 | /* | |
2419 | * rb tree support functions | |
2420 | */ | |
2421 | static void cfq_del_rq_rb(struct request *rq) | |
2422 | { | |
2423 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
2424 | const int sync = rq_is_sync(rq); | |
2425 | ||
2426 | BUG_ON(!cfqq->queued[sync]); | |
2427 | cfqq->queued[sync]--; | |
2428 | ||
2429 | elv_rb_del(&cfqq->sort_list, rq); | |
2430 | ||
2431 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { | |
2432 | /* | |
2433 | * Queue will be deleted from service tree when we actually | |
2434 | * expire it later. Right now just remove it from prio tree | |
2435 | * as it is empty. | |
2436 | */ | |
2437 | if (cfqq->p_root) { | |
2438 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
2439 | cfqq->p_root = NULL; | |
2440 | } | |
2441 | } | |
2442 | } | |
2443 | ||
2444 | static void cfq_add_rq_rb(struct request *rq) | |
2445 | { | |
2446 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
2447 | struct cfq_data *cfqd = cfqq->cfqd; | |
2448 | struct request *prev; | |
2449 | ||
2450 | cfqq->queued[rq_is_sync(rq)]++; | |
2451 | ||
2452 | elv_rb_add(&cfqq->sort_list, rq); | |
2453 | ||
2454 | if (!cfq_cfqq_on_rr(cfqq)) | |
2455 | cfq_add_cfqq_rr(cfqd, cfqq); | |
2456 | ||
2457 | /* | |
2458 | * check if this request is a better next-serve candidate | |
2459 | */ | |
2460 | prev = cfqq->next_rq; | |
2461 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); | |
2462 | ||
2463 | /* | |
2464 | * adjust priority tree position, if ->next_rq changes | |
2465 | */ | |
2466 | if (prev != cfqq->next_rq) | |
2467 | cfq_prio_tree_add(cfqd, cfqq); | |
2468 | ||
2469 | BUG_ON(!cfqq->next_rq); | |
2470 | } | |
2471 | ||
2472 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) | |
2473 | { | |
2474 | elv_rb_del(&cfqq->sort_list, rq); | |
2475 | cfqq->queued[rq_is_sync(rq)]--; | |
2476 | cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags); | |
2477 | cfq_add_rq_rb(rq); | |
2478 | cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group, | |
2479 | req_op(rq), rq->cmd_flags); | |
2480 | } | |
2481 | ||
2482 | static struct request * | |
2483 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
2484 | { | |
2485 | struct task_struct *tsk = current; | |
2486 | struct cfq_io_cq *cic; | |
2487 | struct cfq_queue *cfqq; | |
2488 | ||
2489 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
2490 | if (!cic) | |
2491 | return NULL; | |
2492 | ||
2493 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
2494 | if (cfqq) | |
2495 | return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio)); | |
2496 | ||
2497 | return NULL; | |
2498 | } | |
2499 | ||
2500 | static void cfq_activate_request(struct request_queue *q, struct request *rq) | |
2501 | { | |
2502 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2503 | ||
2504 | cfqd->rq_in_driver++; | |
2505 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", | |
2506 | cfqd->rq_in_driver); | |
2507 | ||
2508 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
2509 | } | |
2510 | ||
2511 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) | |
2512 | { | |
2513 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2514 | ||
2515 | WARN_ON(!cfqd->rq_in_driver); | |
2516 | cfqd->rq_in_driver--; | |
2517 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", | |
2518 | cfqd->rq_in_driver); | |
2519 | } | |
2520 | ||
2521 | static void cfq_remove_request(struct request *rq) | |
2522 | { | |
2523 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
2524 | ||
2525 | if (cfqq->next_rq == rq) | |
2526 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
2527 | ||
2528 | list_del_init(&rq->queuelist); | |
2529 | cfq_del_rq_rb(rq); | |
2530 | ||
2531 | cfqq->cfqd->rq_queued--; | |
2532 | cfqg_stats_update_io_remove(RQ_CFQG(rq), req_op(rq), rq->cmd_flags); | |
2533 | if (rq->cmd_flags & REQ_PRIO) { | |
2534 | WARN_ON(!cfqq->prio_pending); | |
2535 | cfqq->prio_pending--; | |
2536 | } | |
2537 | } | |
2538 | ||
2539 | static int cfq_merge(struct request_queue *q, struct request **req, | |
2540 | struct bio *bio) | |
2541 | { | |
2542 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2543 | struct request *__rq; | |
2544 | ||
2545 | __rq = cfq_find_rq_fmerge(cfqd, bio); | |
2546 | if (__rq && elv_rq_merge_ok(__rq, bio)) { | |
2547 | *req = __rq; | |
2548 | return ELEVATOR_FRONT_MERGE; | |
2549 | } | |
2550 | ||
2551 | return ELEVATOR_NO_MERGE; | |
2552 | } | |
2553 | ||
2554 | static void cfq_merged_request(struct request_queue *q, struct request *req, | |
2555 | int type) | |
2556 | { | |
2557 | if (type == ELEVATOR_FRONT_MERGE) { | |
2558 | struct cfq_queue *cfqq = RQ_CFQQ(req); | |
2559 | ||
2560 | cfq_reposition_rq_rb(cfqq, req); | |
2561 | } | |
2562 | } | |
2563 | ||
2564 | static void cfq_bio_merged(struct request_queue *q, struct request *req, | |
2565 | struct bio *bio) | |
2566 | { | |
2567 | cfqg_stats_update_io_merged(RQ_CFQG(req), bio_op(bio), bio->bi_rw); | |
2568 | } | |
2569 | ||
2570 | static void | |
2571 | cfq_merged_requests(struct request_queue *q, struct request *rq, | |
2572 | struct request *next) | |
2573 | { | |
2574 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
2575 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2576 | ||
2577 | /* | |
2578 | * reposition in fifo if next is older than rq | |
2579 | */ | |
2580 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
2581 | next->fifo_time < rq->fifo_time && | |
2582 | cfqq == RQ_CFQQ(next)) { | |
2583 | list_move(&rq->queuelist, &next->queuelist); | |
2584 | rq->fifo_time = next->fifo_time; | |
2585 | } | |
2586 | ||
2587 | if (cfqq->next_rq == next) | |
2588 | cfqq->next_rq = rq; | |
2589 | cfq_remove_request(next); | |
2590 | cfqg_stats_update_io_merged(RQ_CFQG(rq), req_op(next), next->cmd_flags); | |
2591 | ||
2592 | cfqq = RQ_CFQQ(next); | |
2593 | /* | |
2594 | * all requests of this queue are merged to other queues, delete it | |
2595 | * from the service tree. If it's the active_queue, | |
2596 | * cfq_dispatch_requests() will choose to expire it or do idle | |
2597 | */ | |
2598 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) && | |
2599 | cfqq != cfqd->active_queue) | |
2600 | cfq_del_cfqq_rr(cfqd, cfqq); | |
2601 | } | |
2602 | ||
2603 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, | |
2604 | struct bio *bio) | |
2605 | { | |
2606 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2607 | struct cfq_io_cq *cic; | |
2608 | struct cfq_queue *cfqq; | |
2609 | ||
2610 | /* | |
2611 | * Disallow merge of a sync bio into an async request. | |
2612 | */ | |
2613 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) | |
2614 | return false; | |
2615 | ||
2616 | /* | |
2617 | * Lookup the cfqq that this bio will be queued with and allow | |
2618 | * merge only if rq is queued there. | |
2619 | */ | |
2620 | cic = cfq_cic_lookup(cfqd, current->io_context); | |
2621 | if (!cic) | |
2622 | return false; | |
2623 | ||
2624 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
2625 | return cfqq == RQ_CFQQ(rq); | |
2626 | } | |
2627 | ||
2628 | static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2629 | { | |
2630 | del_timer(&cfqd->idle_slice_timer); | |
2631 | cfqg_stats_update_idle_time(cfqq->cfqg); | |
2632 | } | |
2633 | ||
2634 | static void __cfq_set_active_queue(struct cfq_data *cfqd, | |
2635 | struct cfq_queue *cfqq) | |
2636 | { | |
2637 | if (cfqq) { | |
2638 | cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d", | |
2639 | cfqd->serving_wl_class, cfqd->serving_wl_type); | |
2640 | cfqg_stats_update_avg_queue_size(cfqq->cfqg); | |
2641 | cfqq->slice_start = 0; | |
2642 | cfqq->dispatch_start = ktime_get_ns(); | |
2643 | cfqq->allocated_slice = 0; | |
2644 | cfqq->slice_end = 0; | |
2645 | cfqq->slice_dispatch = 0; | |
2646 | cfqq->nr_sectors = 0; | |
2647 | ||
2648 | cfq_clear_cfqq_wait_request(cfqq); | |
2649 | cfq_clear_cfqq_must_dispatch(cfqq); | |
2650 | cfq_clear_cfqq_must_alloc_slice(cfqq); | |
2651 | cfq_clear_cfqq_fifo_expire(cfqq); | |
2652 | cfq_mark_cfqq_slice_new(cfqq); | |
2653 | ||
2654 | cfq_del_timer(cfqd, cfqq); | |
2655 | } | |
2656 | ||
2657 | cfqd->active_queue = cfqq; | |
2658 | } | |
2659 | ||
2660 | /* | |
2661 | * current cfqq expired its slice (or was too idle), select new one | |
2662 | */ | |
2663 | static void | |
2664 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2665 | bool timed_out) | |
2666 | { | |
2667 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); | |
2668 | ||
2669 | if (cfq_cfqq_wait_request(cfqq)) | |
2670 | cfq_del_timer(cfqd, cfqq); | |
2671 | ||
2672 | cfq_clear_cfqq_wait_request(cfqq); | |
2673 | cfq_clear_cfqq_wait_busy(cfqq); | |
2674 | ||
2675 | /* | |
2676 | * If this cfqq is shared between multiple processes, check to | |
2677 | * make sure that those processes are still issuing I/Os within | |
2678 | * the mean seek distance. If not, it may be time to break the | |
2679 | * queues apart again. | |
2680 | */ | |
2681 | if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq)) | |
2682 | cfq_mark_cfqq_split_coop(cfqq); | |
2683 | ||
2684 | /* | |
2685 | * store what was left of this slice, if the queue idled/timed out | |
2686 | */ | |
2687 | if (timed_out) { | |
2688 | if (cfq_cfqq_slice_new(cfqq)) | |
2689 | cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq); | |
2690 | else | |
2691 | cfqq->slice_resid = cfqq->slice_end - ktime_get_ns(); | |
2692 | cfq_log_cfqq(cfqd, cfqq, "resid=%llu", cfqq->slice_resid); | |
2693 | } | |
2694 | ||
2695 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); | |
2696 | ||
2697 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2698 | cfq_del_cfqq_rr(cfqd, cfqq); | |
2699 | ||
2700 | cfq_resort_rr_list(cfqd, cfqq); | |
2701 | ||
2702 | if (cfqq == cfqd->active_queue) | |
2703 | cfqd->active_queue = NULL; | |
2704 | ||
2705 | if (cfqd->active_cic) { | |
2706 | put_io_context(cfqd->active_cic->icq.ioc); | |
2707 | cfqd->active_cic = NULL; | |
2708 | } | |
2709 | } | |
2710 | ||
2711 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) | |
2712 | { | |
2713 | struct cfq_queue *cfqq = cfqd->active_queue; | |
2714 | ||
2715 | if (cfqq) | |
2716 | __cfq_slice_expired(cfqd, cfqq, timed_out); | |
2717 | } | |
2718 | ||
2719 | /* | |
2720 | * Get next queue for service. Unless we have a queue preemption, | |
2721 | * we'll simply select the first cfqq in the service tree. | |
2722 | */ | |
2723 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) | |
2724 | { | |
2725 | struct cfq_rb_root *st = st_for(cfqd->serving_group, | |
2726 | cfqd->serving_wl_class, cfqd->serving_wl_type); | |
2727 | ||
2728 | if (!cfqd->rq_queued) | |
2729 | return NULL; | |
2730 | ||
2731 | /* There is nothing to dispatch */ | |
2732 | if (!st) | |
2733 | return NULL; | |
2734 | if (RB_EMPTY_ROOT(&st->rb)) | |
2735 | return NULL; | |
2736 | return cfq_rb_first(st); | |
2737 | } | |
2738 | ||
2739 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) | |
2740 | { | |
2741 | struct cfq_group *cfqg; | |
2742 | struct cfq_queue *cfqq; | |
2743 | int i, j; | |
2744 | struct cfq_rb_root *st; | |
2745 | ||
2746 | if (!cfqd->rq_queued) | |
2747 | return NULL; | |
2748 | ||
2749 | cfqg = cfq_get_next_cfqg(cfqd); | |
2750 | if (!cfqg) | |
2751 | return NULL; | |
2752 | ||
2753 | for_each_cfqg_st(cfqg, i, j, st) | |
2754 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
2755 | return cfqq; | |
2756 | return NULL; | |
2757 | } | |
2758 | ||
2759 | /* | |
2760 | * Get and set a new active queue for service. | |
2761 | */ | |
2762 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, | |
2763 | struct cfq_queue *cfqq) | |
2764 | { | |
2765 | if (!cfqq) | |
2766 | cfqq = cfq_get_next_queue(cfqd); | |
2767 | ||
2768 | __cfq_set_active_queue(cfqd, cfqq); | |
2769 | return cfqq; | |
2770 | } | |
2771 | ||
2772 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, | |
2773 | struct request *rq) | |
2774 | { | |
2775 | if (blk_rq_pos(rq) >= cfqd->last_position) | |
2776 | return blk_rq_pos(rq) - cfqd->last_position; | |
2777 | else | |
2778 | return cfqd->last_position - blk_rq_pos(rq); | |
2779 | } | |
2780 | ||
2781 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2782 | struct request *rq) | |
2783 | { | |
2784 | return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR; | |
2785 | } | |
2786 | ||
2787 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, | |
2788 | struct cfq_queue *cur_cfqq) | |
2789 | { | |
2790 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; | |
2791 | struct rb_node *parent, *node; | |
2792 | struct cfq_queue *__cfqq; | |
2793 | sector_t sector = cfqd->last_position; | |
2794 | ||
2795 | if (RB_EMPTY_ROOT(root)) | |
2796 | return NULL; | |
2797 | ||
2798 | /* | |
2799 | * First, if we find a request starting at the end of the last | |
2800 | * request, choose it. | |
2801 | */ | |
2802 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); | |
2803 | if (__cfqq) | |
2804 | return __cfqq; | |
2805 | ||
2806 | /* | |
2807 | * If the exact sector wasn't found, the parent of the NULL leaf | |
2808 | * will contain the closest sector. | |
2809 | */ | |
2810 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
2811 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
2812 | return __cfqq; | |
2813 | ||
2814 | if (blk_rq_pos(__cfqq->next_rq) < sector) | |
2815 | node = rb_next(&__cfqq->p_node); | |
2816 | else | |
2817 | node = rb_prev(&__cfqq->p_node); | |
2818 | if (!node) | |
2819 | return NULL; | |
2820 | ||
2821 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
2822 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
2823 | return __cfqq; | |
2824 | ||
2825 | return NULL; | |
2826 | } | |
2827 | ||
2828 | /* | |
2829 | * cfqd - obvious | |
2830 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
2831 | * closely cooperating with itself. | |
2832 | * | |
2833 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
2834 | * one request, and that cfqd->last_position reflects a position on the disk | |
2835 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
2836 | * assumption. | |
2837 | */ | |
2838 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
2839 | struct cfq_queue *cur_cfqq) | |
2840 | { | |
2841 | struct cfq_queue *cfqq; | |
2842 | ||
2843 | if (cfq_class_idle(cur_cfqq)) | |
2844 | return NULL; | |
2845 | if (!cfq_cfqq_sync(cur_cfqq)) | |
2846 | return NULL; | |
2847 | if (CFQQ_SEEKY(cur_cfqq)) | |
2848 | return NULL; | |
2849 | ||
2850 | /* | |
2851 | * Don't search priority tree if it's the only queue in the group. | |
2852 | */ | |
2853 | if (cur_cfqq->cfqg->nr_cfqq == 1) | |
2854 | return NULL; | |
2855 | ||
2856 | /* | |
2857 | * We should notice if some of the queues are cooperating, eg | |
2858 | * working closely on the same area of the disk. In that case, | |
2859 | * we can group them together and don't waste time idling. | |
2860 | */ | |
2861 | cfqq = cfqq_close(cfqd, cur_cfqq); | |
2862 | if (!cfqq) | |
2863 | return NULL; | |
2864 | ||
2865 | /* If new queue belongs to different cfq_group, don't choose it */ | |
2866 | if (cur_cfqq->cfqg != cfqq->cfqg) | |
2867 | return NULL; | |
2868 | ||
2869 | /* | |
2870 | * It only makes sense to merge sync queues. | |
2871 | */ | |
2872 | if (!cfq_cfqq_sync(cfqq)) | |
2873 | return NULL; | |
2874 | if (CFQQ_SEEKY(cfqq)) | |
2875 | return NULL; | |
2876 | ||
2877 | /* | |
2878 | * Do not merge queues of different priority classes | |
2879 | */ | |
2880 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
2881 | return NULL; | |
2882 | ||
2883 | return cfqq; | |
2884 | } | |
2885 | ||
2886 | /* | |
2887 | * Determine whether we should enforce idle window for this queue. | |
2888 | */ | |
2889 | ||
2890 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2891 | { | |
2892 | enum wl_class_t wl_class = cfqq_class(cfqq); | |
2893 | struct cfq_rb_root *st = cfqq->service_tree; | |
2894 | ||
2895 | BUG_ON(!st); | |
2896 | BUG_ON(!st->count); | |
2897 | ||
2898 | if (!cfqd->cfq_slice_idle) | |
2899 | return false; | |
2900 | ||
2901 | /* We never do for idle class queues. */ | |
2902 | if (wl_class == IDLE_WORKLOAD) | |
2903 | return false; | |
2904 | ||
2905 | /* We do for queues that were marked with idle window flag. */ | |
2906 | if (cfq_cfqq_idle_window(cfqq) && | |
2907 | !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)) | |
2908 | return true; | |
2909 | ||
2910 | /* | |
2911 | * Otherwise, we do only if they are the last ones | |
2912 | * in their service tree. | |
2913 | */ | |
2914 | if (st->count == 1 && cfq_cfqq_sync(cfqq) && | |
2915 | !cfq_io_thinktime_big(cfqd, &st->ttime, false)) | |
2916 | return true; | |
2917 | cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count); | |
2918 | return false; | |
2919 | } | |
2920 | ||
2921 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) | |
2922 | { | |
2923 | struct cfq_queue *cfqq = cfqd->active_queue; | |
2924 | struct cfq_rb_root *st = cfqq->service_tree; | |
2925 | struct cfq_io_cq *cic; | |
2926 | u64 sl, group_idle = 0; | |
2927 | u64 now = ktime_get_ns(); | |
2928 | ||
2929 | /* | |
2930 | * SSD device without seek penalty, disable idling. But only do so | |
2931 | * for devices that support queuing, otherwise we still have a problem | |
2932 | * with sync vs async workloads. | |
2933 | */ | |
2934 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) | |
2935 | return; | |
2936 | ||
2937 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2938 | WARN_ON(cfq_cfqq_slice_new(cfqq)); | |
2939 | ||
2940 | /* | |
2941 | * idle is disabled, either manually or by past process history | |
2942 | */ | |
2943 | if (!cfq_should_idle(cfqd, cfqq)) { | |
2944 | /* no queue idling. Check for group idling */ | |
2945 | if (cfqd->cfq_group_idle) | |
2946 | group_idle = cfqd->cfq_group_idle; | |
2947 | else | |
2948 | return; | |
2949 | } | |
2950 | ||
2951 | /* | |
2952 | * still active requests from this queue, don't idle | |
2953 | */ | |
2954 | if (cfqq->dispatched) | |
2955 | return; | |
2956 | ||
2957 | /* | |
2958 | * task has exited, don't wait | |
2959 | */ | |
2960 | cic = cfqd->active_cic; | |
2961 | if (!cic || !atomic_read(&cic->icq.ioc->active_ref)) | |
2962 | return; | |
2963 | ||
2964 | /* | |
2965 | * If our average think time is larger than the remaining time | |
2966 | * slice, then don't idle. This avoids overrunning the allotted | |
2967 | * time slice. | |
2968 | */ | |
2969 | if (sample_valid(cic->ttime.ttime_samples) && | |
2970 | (cfqq->slice_end - now < cic->ttime.ttime_mean)) { | |
2971 | cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu", | |
2972 | cic->ttime.ttime_mean); | |
2973 | return; | |
2974 | } | |
2975 | ||
2976 | /* | |
2977 | * There are other queues in the group or this is the only group and | |
2978 | * it has too big thinktime, don't do group idle. | |
2979 | */ | |
2980 | if (group_idle && | |
2981 | (cfqq->cfqg->nr_cfqq > 1 || | |
2982 | cfq_io_thinktime_big(cfqd, &st->ttime, true))) | |
2983 | return; | |
2984 | ||
2985 | cfq_mark_cfqq_wait_request(cfqq); | |
2986 | ||
2987 | if (group_idle) | |
2988 | sl = cfqd->cfq_group_idle; | |
2989 | else | |
2990 | sl = cfqd->cfq_slice_idle; | |
2991 | ||
2992 | mod_timer(&cfqd->idle_slice_timer, now + sl); | |
2993 | cfqg_stats_set_start_idle_time(cfqq->cfqg); | |
2994 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl, | |
2995 | group_idle ? 1 : 0); | |
2996 | } | |
2997 | ||
2998 | /* | |
2999 | * Move request from internal lists to the request queue dispatch list. | |
3000 | */ | |
3001 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) | |
3002 | { | |
3003 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3004 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3005 | ||
3006 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); | |
3007 | ||
3008 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); | |
3009 | cfq_remove_request(rq); | |
3010 | cfqq->dispatched++; | |
3011 | (RQ_CFQG(rq))->dispatched++; | |
3012 | elv_dispatch_sort(q, rq); | |
3013 | ||
3014 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++; | |
3015 | cfqq->nr_sectors += blk_rq_sectors(rq); | |
3016 | } | |
3017 | ||
3018 | /* | |
3019 | * return expired entry, or NULL to just start from scratch in rbtree | |
3020 | */ | |
3021 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) | |
3022 | { | |
3023 | struct request *rq = NULL; | |
3024 | ||
3025 | if (cfq_cfqq_fifo_expire(cfqq)) | |
3026 | return NULL; | |
3027 | ||
3028 | cfq_mark_cfqq_fifo_expire(cfqq); | |
3029 | ||
3030 | if (list_empty(&cfqq->fifo)) | |
3031 | return NULL; | |
3032 | ||
3033 | rq = rq_entry_fifo(cfqq->fifo.next); | |
3034 | if (ktime_get_ns() < rq->fifo_time) | |
3035 | rq = NULL; | |
3036 | ||
3037 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); | |
3038 | return rq; | |
3039 | } | |
3040 | ||
3041 | static inline int | |
3042 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3043 | { | |
3044 | const int base_rq = cfqd->cfq_slice_async_rq; | |
3045 | ||
3046 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); | |
3047 | ||
3048 | return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio); | |
3049 | } | |
3050 | ||
3051 | /* | |
3052 | * Must be called with the queue_lock held. | |
3053 | */ | |
3054 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
3055 | { | |
3056 | int process_refs, io_refs; | |
3057 | ||
3058 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
3059 | process_refs = cfqq->ref - io_refs; | |
3060 | BUG_ON(process_refs < 0); | |
3061 | return process_refs; | |
3062 | } | |
3063 | ||
3064 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
3065 | { | |
3066 | int process_refs, new_process_refs; | |
3067 | struct cfq_queue *__cfqq; | |
3068 | ||
3069 | /* | |
3070 | * If there are no process references on the new_cfqq, then it is | |
3071 | * unsafe to follow the ->new_cfqq chain as other cfqq's in the | |
3072 | * chain may have dropped their last reference (not just their | |
3073 | * last process reference). | |
3074 | */ | |
3075 | if (!cfqq_process_refs(new_cfqq)) | |
3076 | return; | |
3077 | ||
3078 | /* Avoid a circular list and skip interim queue merges */ | |
3079 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
3080 | if (__cfqq == cfqq) | |
3081 | return; | |
3082 | new_cfqq = __cfqq; | |
3083 | } | |
3084 | ||
3085 | process_refs = cfqq_process_refs(cfqq); | |
3086 | new_process_refs = cfqq_process_refs(new_cfqq); | |
3087 | /* | |
3088 | * If the process for the cfqq has gone away, there is no | |
3089 | * sense in merging the queues. | |
3090 | */ | |
3091 | if (process_refs == 0 || new_process_refs == 0) | |
3092 | return; | |
3093 | ||
3094 | /* | |
3095 | * Merge in the direction of the lesser amount of work. | |
3096 | */ | |
3097 | if (new_process_refs >= process_refs) { | |
3098 | cfqq->new_cfqq = new_cfqq; | |
3099 | new_cfqq->ref += process_refs; | |
3100 | } else { | |
3101 | new_cfqq->new_cfqq = cfqq; | |
3102 | cfqq->ref += new_process_refs; | |
3103 | } | |
3104 | } | |
3105 | ||
3106 | static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd, | |
3107 | struct cfq_group *cfqg, enum wl_class_t wl_class) | |
3108 | { | |
3109 | struct cfq_queue *queue; | |
3110 | int i; | |
3111 | bool key_valid = false; | |
3112 | u64 lowest_key = 0; | |
3113 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
3114 | ||
3115 | for (i = 0; i <= SYNC_WORKLOAD; ++i) { | |
3116 | /* select the one with lowest rb_key */ | |
3117 | queue = cfq_rb_first(st_for(cfqg, wl_class, i)); | |
3118 | if (queue && | |
3119 | (!key_valid || queue->rb_key < lowest_key)) { | |
3120 | lowest_key = queue->rb_key; | |
3121 | cur_best = i; | |
3122 | key_valid = true; | |
3123 | } | |
3124 | } | |
3125 | ||
3126 | return cur_best; | |
3127 | } | |
3128 | ||
3129 | static void | |
3130 | choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
3131 | { | |
3132 | u64 slice; | |
3133 | unsigned count; | |
3134 | struct cfq_rb_root *st; | |
3135 | u64 group_slice; | |
3136 | enum wl_class_t original_class = cfqd->serving_wl_class; | |
3137 | u64 now = ktime_get_ns(); | |
3138 | ||
3139 | /* Choose next priority. RT > BE > IDLE */ | |
3140 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) | |
3141 | cfqd->serving_wl_class = RT_WORKLOAD; | |
3142 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) | |
3143 | cfqd->serving_wl_class = BE_WORKLOAD; | |
3144 | else { | |
3145 | cfqd->serving_wl_class = IDLE_WORKLOAD; | |
3146 | cfqd->workload_expires = now + jiffies_to_nsecs(1); | |
3147 | return; | |
3148 | } | |
3149 | ||
3150 | if (original_class != cfqd->serving_wl_class) | |
3151 | goto new_workload; | |
3152 | ||
3153 | /* | |
3154 | * For RT and BE, we have to choose also the type | |
3155 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
3156 | * expiration time | |
3157 | */ | |
3158 | st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type); | |
3159 | count = st->count; | |
3160 | ||
3161 | /* | |
3162 | * check workload expiration, and that we still have other queues ready | |
3163 | */ | |
3164 | if (count && !(now > cfqd->workload_expires)) | |
3165 | return; | |
3166 | ||
3167 | new_workload: | |
3168 | /* otherwise select new workload type */ | |
3169 | cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg, | |
3170 | cfqd->serving_wl_class); | |
3171 | st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type); | |
3172 | count = st->count; | |
3173 | ||
3174 | /* | |
3175 | * the workload slice is computed as a fraction of target latency | |
3176 | * proportional to the number of queues in that workload, over | |
3177 | * all the queues in the same priority class | |
3178 | */ | |
3179 | group_slice = cfq_group_slice(cfqd, cfqg); | |
3180 | ||
3181 | slice = div_u64(group_slice * count, | |
3182 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class], | |
3183 | cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd, | |
3184 | cfqg))); | |
3185 | ||
3186 | if (cfqd->serving_wl_type == ASYNC_WORKLOAD) { | |
3187 | u64 tmp; | |
3188 | ||
3189 | /* | |
3190 | * Async queues are currently system wide. Just taking | |
3191 | * proportion of queues with-in same group will lead to higher | |
3192 | * async ratio system wide as generally root group is going | |
3193 | * to have higher weight. A more accurate thing would be to | |
3194 | * calculate system wide asnc/sync ratio. | |
3195 | */ | |
3196 | tmp = cfqd->cfq_target_latency * | |
3197 | cfqg_busy_async_queues(cfqd, cfqg); | |
3198 | tmp = div_u64(tmp, cfqd->busy_queues); | |
3199 | slice = min_t(u64, slice, tmp); | |
3200 | ||
3201 | /* async workload slice is scaled down according to | |
3202 | * the sync/async slice ratio. */ | |
3203 | slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]); | |
3204 | } else | |
3205 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
3206 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
3207 | ||
3208 | slice = max_t(u64, slice, CFQ_MIN_TT); | |
3209 | cfq_log(cfqd, "workload slice:%llu", slice); | |
3210 | cfqd->workload_expires = now + slice; | |
3211 | } | |
3212 | ||
3213 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) | |
3214 | { | |
3215 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
3216 | struct cfq_group *cfqg; | |
3217 | ||
3218 | if (RB_EMPTY_ROOT(&st->rb)) | |
3219 | return NULL; | |
3220 | cfqg = cfq_rb_first_group(st); | |
3221 | update_min_vdisktime(st); | |
3222 | return cfqg; | |
3223 | } | |
3224 | ||
3225 | static void cfq_choose_cfqg(struct cfq_data *cfqd) | |
3226 | { | |
3227 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); | |
3228 | u64 now = ktime_get_ns(); | |
3229 | ||
3230 | cfqd->serving_group = cfqg; | |
3231 | ||
3232 | /* Restore the workload type data */ | |
3233 | if (cfqg->saved_wl_slice) { | |
3234 | cfqd->workload_expires = now + cfqg->saved_wl_slice; | |
3235 | cfqd->serving_wl_type = cfqg->saved_wl_type; | |
3236 | cfqd->serving_wl_class = cfqg->saved_wl_class; | |
3237 | } else | |
3238 | cfqd->workload_expires = now - 1; | |
3239 | ||
3240 | choose_wl_class_and_type(cfqd, cfqg); | |
3241 | } | |
3242 | ||
3243 | /* | |
3244 | * Select a queue for service. If we have a current active queue, | |
3245 | * check whether to continue servicing it, or retrieve and set a new one. | |
3246 | */ | |
3247 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) | |
3248 | { | |
3249 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
3250 | u64 now = ktime_get_ns(); | |
3251 | ||
3252 | cfqq = cfqd->active_queue; | |
3253 | if (!cfqq) | |
3254 | goto new_queue; | |
3255 | ||
3256 | if (!cfqd->rq_queued) | |
3257 | return NULL; | |
3258 | ||
3259 | /* | |
3260 | * We were waiting for group to get backlogged. Expire the queue | |
3261 | */ | |
3262 | if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3263 | goto expire; | |
3264 | ||
3265 | /* | |
3266 | * The active queue has run out of time, expire it and select new. | |
3267 | */ | |
3268 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) { | |
3269 | /* | |
3270 | * If slice had not expired at the completion of last request | |
3271 | * we might not have turned on wait_busy flag. Don't expire | |
3272 | * the queue yet. Allow the group to get backlogged. | |
3273 | * | |
3274 | * The very fact that we have used the slice, that means we | |
3275 | * have been idling all along on this queue and it should be | |
3276 | * ok to wait for this request to complete. | |
3277 | */ | |
3278 | if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list) | |
3279 | && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
3280 | cfqq = NULL; | |
3281 | goto keep_queue; | |
3282 | } else | |
3283 | goto check_group_idle; | |
3284 | } | |
3285 | ||
3286 | /* | |
3287 | * The active queue has requests and isn't expired, allow it to | |
3288 | * dispatch. | |
3289 | */ | |
3290 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3291 | goto keep_queue; | |
3292 | ||
3293 | /* | |
3294 | * If another queue has a request waiting within our mean seek | |
3295 | * distance, let it run. The expire code will check for close | |
3296 | * cooperators and put the close queue at the front of the service | |
3297 | * tree. If possible, merge the expiring queue with the new cfqq. | |
3298 | */ | |
3299 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); | |
3300 | if (new_cfqq) { | |
3301 | if (!cfqq->new_cfqq) | |
3302 | cfq_setup_merge(cfqq, new_cfqq); | |
3303 | goto expire; | |
3304 | } | |
3305 | ||
3306 | /* | |
3307 | * No requests pending. If the active queue still has requests in | |
3308 | * flight or is idling for a new request, allow either of these | |
3309 | * conditions to happen (or time out) before selecting a new queue. | |
3310 | */ | |
3311 | if (timer_pending(&cfqd->idle_slice_timer)) { | |
3312 | cfqq = NULL; | |
3313 | goto keep_queue; | |
3314 | } | |
3315 | ||
3316 | /* | |
3317 | * This is a deep seek queue, but the device is much faster than | |
3318 | * the queue can deliver, don't idle | |
3319 | **/ | |
3320 | if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) && | |
3321 | (cfq_cfqq_slice_new(cfqq) || | |
3322 | (cfqq->slice_end - now > now - cfqq->slice_start))) { | |
3323 | cfq_clear_cfqq_deep(cfqq); | |
3324 | cfq_clear_cfqq_idle_window(cfqq); | |
3325 | } | |
3326 | ||
3327 | if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) { | |
3328 | cfqq = NULL; | |
3329 | goto keep_queue; | |
3330 | } | |
3331 | ||
3332 | /* | |
3333 | * If group idle is enabled and there are requests dispatched from | |
3334 | * this group, wait for requests to complete. | |
3335 | */ | |
3336 | check_group_idle: | |
3337 | if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 && | |
3338 | cfqq->cfqg->dispatched && | |
3339 | !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) { | |
3340 | cfqq = NULL; | |
3341 | goto keep_queue; | |
3342 | } | |
3343 | ||
3344 | expire: | |
3345 | cfq_slice_expired(cfqd, 0); | |
3346 | new_queue: | |
3347 | /* | |
3348 | * Current queue expired. Check if we have to switch to a new | |
3349 | * service tree | |
3350 | */ | |
3351 | if (!new_cfqq) | |
3352 | cfq_choose_cfqg(cfqd); | |
3353 | ||
3354 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); | |
3355 | keep_queue: | |
3356 | return cfqq; | |
3357 | } | |
3358 | ||
3359 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) | |
3360 | { | |
3361 | int dispatched = 0; | |
3362 | ||
3363 | while (cfqq->next_rq) { | |
3364 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
3365 | dispatched++; | |
3366 | } | |
3367 | ||
3368 | BUG_ON(!list_empty(&cfqq->fifo)); | |
3369 | ||
3370 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
3371 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
3372 | return dispatched; | |
3373 | } | |
3374 | ||
3375 | /* | |
3376 | * Drain our current requests. Used for barriers and when switching | |
3377 | * io schedulers on-the-fly. | |
3378 | */ | |
3379 | static int cfq_forced_dispatch(struct cfq_data *cfqd) | |
3380 | { | |
3381 | struct cfq_queue *cfqq; | |
3382 | int dispatched = 0; | |
3383 | ||
3384 | /* Expire the timeslice of the current active queue first */ | |
3385 | cfq_slice_expired(cfqd, 0); | |
3386 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) { | |
3387 | __cfq_set_active_queue(cfqd, cfqq); | |
3388 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
3389 | } | |
3390 | ||
3391 | BUG_ON(cfqd->busy_queues); | |
3392 | ||
3393 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); | |
3394 | return dispatched; | |
3395 | } | |
3396 | ||
3397 | static inline bool cfq_slice_used_soon(struct cfq_data *cfqd, | |
3398 | struct cfq_queue *cfqq) | |
3399 | { | |
3400 | u64 now = ktime_get_ns(); | |
3401 | ||
3402 | /* the queue hasn't finished any request, can't estimate */ | |
3403 | if (cfq_cfqq_slice_new(cfqq)) | |
3404 | return true; | |
3405 | if (now + cfqd->cfq_slice_idle * cfqq->dispatched > cfqq->slice_end) | |
3406 | return true; | |
3407 | ||
3408 | return false; | |
3409 | } | |
3410 | ||
3411 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3412 | { | |
3413 | unsigned int max_dispatch; | |
3414 | ||
3415 | /* | |
3416 | * Drain async requests before we start sync IO | |
3417 | */ | |
3418 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC]) | |
3419 | return false; | |
3420 | ||
3421 | /* | |
3422 | * If this is an async queue and we have sync IO in flight, let it wait | |
3423 | */ | |
3424 | if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq)) | |
3425 | return false; | |
3426 | ||
3427 | max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1); | |
3428 | if (cfq_class_idle(cfqq)) | |
3429 | max_dispatch = 1; | |
3430 | ||
3431 | /* | |
3432 | * Does this cfqq already have too much IO in flight? | |
3433 | */ | |
3434 | if (cfqq->dispatched >= max_dispatch) { | |
3435 | bool promote_sync = false; | |
3436 | /* | |
3437 | * idle queue must always only have a single IO in flight | |
3438 | */ | |
3439 | if (cfq_class_idle(cfqq)) | |
3440 | return false; | |
3441 | ||
3442 | /* | |
3443 | * If there is only one sync queue | |
3444 | * we can ignore async queue here and give the sync | |
3445 | * queue no dispatch limit. The reason is a sync queue can | |
3446 | * preempt async queue, limiting the sync queue doesn't make | |
3447 | * sense. This is useful for aiostress test. | |
3448 | */ | |
3449 | if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1) | |
3450 | promote_sync = true; | |
3451 | ||
3452 | /* | |
3453 | * We have other queues, don't allow more IO from this one | |
3454 | */ | |
3455 | if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) && | |
3456 | !promote_sync) | |
3457 | return false; | |
3458 | ||
3459 | /* | |
3460 | * Sole queue user, no limit | |
3461 | */ | |
3462 | if (cfqd->busy_queues == 1 || promote_sync) | |
3463 | max_dispatch = -1; | |
3464 | else | |
3465 | /* | |
3466 | * Normally we start throttling cfqq when cfq_quantum/2 | |
3467 | * requests have been dispatched. But we can drive | |
3468 | * deeper queue depths at the beginning of slice | |
3469 | * subjected to upper limit of cfq_quantum. | |
3470 | * */ | |
3471 | max_dispatch = cfqd->cfq_quantum; | |
3472 | } | |
3473 | ||
3474 | /* | |
3475 | * Async queues must wait a bit before being allowed dispatch. | |
3476 | * We also ramp up the dispatch depth gradually for async IO, | |
3477 | * based on the last sync IO we serviced | |
3478 | */ | |
3479 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { | |
3480 | u64 last_sync = ktime_get_ns() - cfqd->last_delayed_sync; | |
3481 | unsigned int depth; | |
3482 | ||
3483 | depth = div64_u64(last_sync, cfqd->cfq_slice[1]); | |
3484 | if (!depth && !cfqq->dispatched) | |
3485 | depth = 1; | |
3486 | if (depth < max_dispatch) | |
3487 | max_dispatch = depth; | |
3488 | } | |
3489 | ||
3490 | /* | |
3491 | * If we're below the current max, allow a dispatch | |
3492 | */ | |
3493 | return cfqq->dispatched < max_dispatch; | |
3494 | } | |
3495 | ||
3496 | /* | |
3497 | * Dispatch a request from cfqq, moving them to the request queue | |
3498 | * dispatch list. | |
3499 | */ | |
3500 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3501 | { | |
3502 | struct request *rq; | |
3503 | ||
3504 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
3505 | ||
3506 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
3507 | return false; | |
3508 | ||
3509 | /* | |
3510 | * follow expired path, else get first next available | |
3511 | */ | |
3512 | rq = cfq_check_fifo(cfqq); | |
3513 | if (!rq) | |
3514 | rq = cfqq->next_rq; | |
3515 | ||
3516 | /* | |
3517 | * insert request into driver dispatch list | |
3518 | */ | |
3519 | cfq_dispatch_insert(cfqd->queue, rq); | |
3520 | ||
3521 | if (!cfqd->active_cic) { | |
3522 | struct cfq_io_cq *cic = RQ_CIC(rq); | |
3523 | ||
3524 | atomic_long_inc(&cic->icq.ioc->refcount); | |
3525 | cfqd->active_cic = cic; | |
3526 | } | |
3527 | ||
3528 | return true; | |
3529 | } | |
3530 | ||
3531 | /* | |
3532 | * Find the cfqq that we need to service and move a request from that to the | |
3533 | * dispatch list | |
3534 | */ | |
3535 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
3536 | { | |
3537 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3538 | struct cfq_queue *cfqq; | |
3539 | ||
3540 | if (!cfqd->busy_queues) | |
3541 | return 0; | |
3542 | ||
3543 | if (unlikely(force)) | |
3544 | return cfq_forced_dispatch(cfqd); | |
3545 | ||
3546 | cfqq = cfq_select_queue(cfqd); | |
3547 | if (!cfqq) | |
3548 | return 0; | |
3549 | ||
3550 | /* | |
3551 | * Dispatch a request from this cfqq, if it is allowed | |
3552 | */ | |
3553 | if (!cfq_dispatch_request(cfqd, cfqq)) | |
3554 | return 0; | |
3555 | ||
3556 | cfqq->slice_dispatch++; | |
3557 | cfq_clear_cfqq_must_dispatch(cfqq); | |
3558 | ||
3559 | /* | |
3560 | * expire an async queue immediately if it has used up its slice. idle | |
3561 | * queue always expire after 1 dispatch round. | |
3562 | */ | |
3563 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
3564 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
3565 | cfq_class_idle(cfqq))) { | |
3566 | cfqq->slice_end = ktime_get_ns() + 1; | |
3567 | cfq_slice_expired(cfqd, 0); | |
3568 | } | |
3569 | ||
3570 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); | |
3571 | return 1; | |
3572 | } | |
3573 | ||
3574 | /* | |
3575 | * task holds one reference to the queue, dropped when task exits. each rq | |
3576 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
3577 | * | |
3578 | * Each cfq queue took a reference on the parent group. Drop it now. | |
3579 | * queue lock must be held here. | |
3580 | */ | |
3581 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
3582 | { | |
3583 | struct cfq_data *cfqd = cfqq->cfqd; | |
3584 | struct cfq_group *cfqg; | |
3585 | ||
3586 | BUG_ON(cfqq->ref <= 0); | |
3587 | ||
3588 | cfqq->ref--; | |
3589 | if (cfqq->ref) | |
3590 | return; | |
3591 | ||
3592 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); | |
3593 | BUG_ON(rb_first(&cfqq->sort_list)); | |
3594 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); | |
3595 | cfqg = cfqq->cfqg; | |
3596 | ||
3597 | if (unlikely(cfqd->active_queue == cfqq)) { | |
3598 | __cfq_slice_expired(cfqd, cfqq, 0); | |
3599 | cfq_schedule_dispatch(cfqd); | |
3600 | } | |
3601 | ||
3602 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
3603 | kmem_cache_free(cfq_pool, cfqq); | |
3604 | cfqg_put(cfqg); | |
3605 | } | |
3606 | ||
3607 | static void cfq_put_cooperator(struct cfq_queue *cfqq) | |
3608 | { | |
3609 | struct cfq_queue *__cfqq, *next; | |
3610 | ||
3611 | /* | |
3612 | * If this queue was scheduled to merge with another queue, be | |
3613 | * sure to drop the reference taken on that queue (and others in | |
3614 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
3615 | */ | |
3616 | __cfqq = cfqq->new_cfqq; | |
3617 | while (__cfqq) { | |
3618 | if (__cfqq == cfqq) { | |
3619 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
3620 | break; | |
3621 | } | |
3622 | next = __cfqq->new_cfqq; | |
3623 | cfq_put_queue(__cfqq); | |
3624 | __cfqq = next; | |
3625 | } | |
3626 | } | |
3627 | ||
3628 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3629 | { | |
3630 | if (unlikely(cfqq == cfqd->active_queue)) { | |
3631 | __cfq_slice_expired(cfqd, cfqq, 0); | |
3632 | cfq_schedule_dispatch(cfqd); | |
3633 | } | |
3634 | ||
3635 | cfq_put_cooperator(cfqq); | |
3636 | ||
3637 | cfq_put_queue(cfqq); | |
3638 | } | |
3639 | ||
3640 | static void cfq_init_icq(struct io_cq *icq) | |
3641 | { | |
3642 | struct cfq_io_cq *cic = icq_to_cic(icq); | |
3643 | ||
3644 | cic->ttime.last_end_request = ktime_get_ns(); | |
3645 | } | |
3646 | ||
3647 | static void cfq_exit_icq(struct io_cq *icq) | |
3648 | { | |
3649 | struct cfq_io_cq *cic = icq_to_cic(icq); | |
3650 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3651 | ||
3652 | if (cic_to_cfqq(cic, false)) { | |
3653 | cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, false)); | |
3654 | cic_set_cfqq(cic, NULL, false); | |
3655 | } | |
3656 | ||
3657 | if (cic_to_cfqq(cic, true)) { | |
3658 | cfq_exit_cfqq(cfqd, cic_to_cfqq(cic, true)); | |
3659 | cic_set_cfqq(cic, NULL, true); | |
3660 | } | |
3661 | } | |
3662 | ||
3663 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic) | |
3664 | { | |
3665 | struct task_struct *tsk = current; | |
3666 | int ioprio_class; | |
3667 | ||
3668 | if (!cfq_cfqq_prio_changed(cfqq)) | |
3669 | return; | |
3670 | ||
3671 | ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio); | |
3672 | switch (ioprio_class) { | |
3673 | default: | |
3674 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
3675 | case IOPRIO_CLASS_NONE: | |
3676 | /* | |
3677 | * no prio set, inherit CPU scheduling settings | |
3678 | */ | |
3679 | cfqq->ioprio = task_nice_ioprio(tsk); | |
3680 | cfqq->ioprio_class = task_nice_ioclass(tsk); | |
3681 | break; | |
3682 | case IOPRIO_CLASS_RT: | |
3683 | cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3684 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
3685 | break; | |
3686 | case IOPRIO_CLASS_BE: | |
3687 | cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3688 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3689 | break; | |
3690 | case IOPRIO_CLASS_IDLE: | |
3691 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
3692 | cfqq->ioprio = 7; | |
3693 | cfq_clear_cfqq_idle_window(cfqq); | |
3694 | break; | |
3695 | } | |
3696 | ||
3697 | /* | |
3698 | * keep track of original prio settings in case we have to temporarily | |
3699 | * elevate the priority of this queue | |
3700 | */ | |
3701 | cfqq->org_ioprio = cfqq->ioprio; | |
3702 | cfq_clear_cfqq_prio_changed(cfqq); | |
3703 | } | |
3704 | ||
3705 | static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio) | |
3706 | { | |
3707 | int ioprio = cic->icq.ioc->ioprio; | |
3708 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3709 | struct cfq_queue *cfqq; | |
3710 | ||
3711 | /* | |
3712 | * Check whether ioprio has changed. The condition may trigger | |
3713 | * spuriously on a newly created cic but there's no harm. | |
3714 | */ | |
3715 | if (unlikely(!cfqd) || likely(cic->ioprio == ioprio)) | |
3716 | return; | |
3717 | ||
3718 | cfqq = cic_to_cfqq(cic, false); | |
3719 | if (cfqq) { | |
3720 | cfq_put_queue(cfqq); | |
3721 | cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio); | |
3722 | cic_set_cfqq(cic, cfqq, false); | |
3723 | } | |
3724 | ||
3725 | cfqq = cic_to_cfqq(cic, true); | |
3726 | if (cfqq) | |
3727 | cfq_mark_cfqq_prio_changed(cfqq); | |
3728 | ||
3729 | cic->ioprio = ioprio; | |
3730 | } | |
3731 | ||
3732 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3733 | pid_t pid, bool is_sync) | |
3734 | { | |
3735 | RB_CLEAR_NODE(&cfqq->rb_node); | |
3736 | RB_CLEAR_NODE(&cfqq->p_node); | |
3737 | INIT_LIST_HEAD(&cfqq->fifo); | |
3738 | ||
3739 | cfqq->ref = 0; | |
3740 | cfqq->cfqd = cfqd; | |
3741 | ||
3742 | cfq_mark_cfqq_prio_changed(cfqq); | |
3743 | ||
3744 | if (is_sync) { | |
3745 | if (!cfq_class_idle(cfqq)) | |
3746 | cfq_mark_cfqq_idle_window(cfqq); | |
3747 | cfq_mark_cfqq_sync(cfqq); | |
3748 | } | |
3749 | cfqq->pid = pid; | |
3750 | } | |
3751 | ||
3752 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3753 | static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) | |
3754 | { | |
3755 | struct cfq_data *cfqd = cic_to_cfqd(cic); | |
3756 | struct cfq_queue *cfqq; | |
3757 | uint64_t serial_nr; | |
3758 | ||
3759 | rcu_read_lock(); | |
3760 | serial_nr = bio_blkcg(bio)->css.serial_nr; | |
3761 | rcu_read_unlock(); | |
3762 | ||
3763 | /* | |
3764 | * Check whether blkcg has changed. The condition may trigger | |
3765 | * spuriously on a newly created cic but there's no harm. | |
3766 | */ | |
3767 | if (unlikely(!cfqd) || likely(cic->blkcg_serial_nr == serial_nr)) | |
3768 | return; | |
3769 | ||
3770 | /* | |
3771 | * Drop reference to queues. New queues will be assigned in new | |
3772 | * group upon arrival of fresh requests. | |
3773 | */ | |
3774 | cfqq = cic_to_cfqq(cic, false); | |
3775 | if (cfqq) { | |
3776 | cfq_log_cfqq(cfqd, cfqq, "changed cgroup"); | |
3777 | cic_set_cfqq(cic, NULL, false); | |
3778 | cfq_put_queue(cfqq); | |
3779 | } | |
3780 | ||
3781 | cfqq = cic_to_cfqq(cic, true); | |
3782 | if (cfqq) { | |
3783 | cfq_log_cfqq(cfqd, cfqq, "changed cgroup"); | |
3784 | cic_set_cfqq(cic, NULL, true); | |
3785 | cfq_put_queue(cfqq); | |
3786 | } | |
3787 | ||
3788 | cic->blkcg_serial_nr = serial_nr; | |
3789 | } | |
3790 | #else | |
3791 | static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { } | |
3792 | #endif /* CONFIG_CFQ_GROUP_IOSCHED */ | |
3793 | ||
3794 | static struct cfq_queue ** | |
3795 | cfq_async_queue_prio(struct cfq_group *cfqg, int ioprio_class, int ioprio) | |
3796 | { | |
3797 | switch (ioprio_class) { | |
3798 | case IOPRIO_CLASS_RT: | |
3799 | return &cfqg->async_cfqq[0][ioprio]; | |
3800 | case IOPRIO_CLASS_NONE: | |
3801 | ioprio = IOPRIO_NORM; | |
3802 | /* fall through */ | |
3803 | case IOPRIO_CLASS_BE: | |
3804 | return &cfqg->async_cfqq[1][ioprio]; | |
3805 | case IOPRIO_CLASS_IDLE: | |
3806 | return &cfqg->async_idle_cfqq; | |
3807 | default: | |
3808 | BUG(); | |
3809 | } | |
3810 | } | |
3811 | ||
3812 | static struct cfq_queue * | |
3813 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic, | |
3814 | struct bio *bio) | |
3815 | { | |
3816 | int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio); | |
3817 | int ioprio = IOPRIO_PRIO_DATA(cic->ioprio); | |
3818 | struct cfq_queue **async_cfqq = NULL; | |
3819 | struct cfq_queue *cfqq; | |
3820 | struct cfq_group *cfqg; | |
3821 | ||
3822 | rcu_read_lock(); | |
3823 | cfqg = cfq_lookup_cfqg(cfqd, bio_blkcg(bio)); | |
3824 | if (!cfqg) { | |
3825 | cfqq = &cfqd->oom_cfqq; | |
3826 | goto out; | |
3827 | } | |
3828 | ||
3829 | if (!is_sync) { | |
3830 | if (!ioprio_valid(cic->ioprio)) { | |
3831 | struct task_struct *tsk = current; | |
3832 | ioprio = task_nice_ioprio(tsk); | |
3833 | ioprio_class = task_nice_ioclass(tsk); | |
3834 | } | |
3835 | async_cfqq = cfq_async_queue_prio(cfqg, ioprio_class, ioprio); | |
3836 | cfqq = *async_cfqq; | |
3837 | if (cfqq) | |
3838 | goto out; | |
3839 | } | |
3840 | ||
3841 | cfqq = kmem_cache_alloc_node(cfq_pool, GFP_NOWAIT | __GFP_ZERO, | |
3842 | cfqd->queue->node); | |
3843 | if (!cfqq) { | |
3844 | cfqq = &cfqd->oom_cfqq; | |
3845 | goto out; | |
3846 | } | |
3847 | ||
3848 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
3849 | cfq_init_prio_data(cfqq, cic); | |
3850 | cfq_link_cfqq_cfqg(cfqq, cfqg); | |
3851 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
3852 | ||
3853 | if (async_cfqq) { | |
3854 | /* a new async queue is created, pin and remember */ | |
3855 | cfqq->ref++; | |
3856 | *async_cfqq = cfqq; | |
3857 | } | |
3858 | out: | |
3859 | cfqq->ref++; | |
3860 | rcu_read_unlock(); | |
3861 | return cfqq; | |
3862 | } | |
3863 | ||
3864 | static void | |
3865 | __cfq_update_io_thinktime(struct cfq_ttime *ttime, u64 slice_idle) | |
3866 | { | |
3867 | u64 elapsed = ktime_get_ns() - ttime->last_end_request; | |
3868 | elapsed = min(elapsed, 2UL * slice_idle); | |
3869 | ||
3870 | ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8; | |
3871 | ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed, 8); | |
3872 | ttime->ttime_mean = div64_ul(ttime->ttime_total + 128, | |
3873 | ttime->ttime_samples); | |
3874 | } | |
3875 | ||
3876 | static void | |
3877 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3878 | struct cfq_io_cq *cic) | |
3879 | { | |
3880 | if (cfq_cfqq_sync(cfqq)) { | |
3881 | __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle); | |
3882 | __cfq_update_io_thinktime(&cfqq->service_tree->ttime, | |
3883 | cfqd->cfq_slice_idle); | |
3884 | } | |
3885 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3886 | __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle); | |
3887 | #endif | |
3888 | } | |
3889 | ||
3890 | static void | |
3891 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3892 | struct request *rq) | |
3893 | { | |
3894 | sector_t sdist = 0; | |
3895 | sector_t n_sec = blk_rq_sectors(rq); | |
3896 | if (cfqq->last_request_pos) { | |
3897 | if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
3898 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
3899 | else | |
3900 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
3901 | } | |
3902 | ||
3903 | cfqq->seek_history <<= 1; | |
3904 | if (blk_queue_nonrot(cfqd->queue)) | |
3905 | cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT); | |
3906 | else | |
3907 | cfqq->seek_history |= (sdist > CFQQ_SEEK_THR); | |
3908 | } | |
3909 | ||
3910 | /* | |
3911 | * Disable idle window if the process thinks too long or seeks so much that | |
3912 | * it doesn't matter | |
3913 | */ | |
3914 | static void | |
3915 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3916 | struct cfq_io_cq *cic) | |
3917 | { | |
3918 | int old_idle, enable_idle; | |
3919 | ||
3920 | /* | |
3921 | * Don't idle for async or idle io prio class | |
3922 | */ | |
3923 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
3924 | return; | |
3925 | ||
3926 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); | |
3927 | ||
3928 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) | |
3929 | cfq_mark_cfqq_deep(cfqq); | |
3930 | ||
3931 | if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE)) | |
3932 | enable_idle = 0; | |
3933 | else if (!atomic_read(&cic->icq.ioc->active_ref) || | |
3934 | !cfqd->cfq_slice_idle || | |
3935 | (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq))) | |
3936 | enable_idle = 0; | |
3937 | else if (sample_valid(cic->ttime.ttime_samples)) { | |
3938 | if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle) | |
3939 | enable_idle = 0; | |
3940 | else | |
3941 | enable_idle = 1; | |
3942 | } | |
3943 | ||
3944 | if (old_idle != enable_idle) { | |
3945 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
3946 | if (enable_idle) | |
3947 | cfq_mark_cfqq_idle_window(cfqq); | |
3948 | else | |
3949 | cfq_clear_cfqq_idle_window(cfqq); | |
3950 | } | |
3951 | } | |
3952 | ||
3953 | /* | |
3954 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
3955 | * no or if we aren't sure, a 1 will cause a preempt. | |
3956 | */ | |
3957 | static bool | |
3958 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
3959 | struct request *rq) | |
3960 | { | |
3961 | struct cfq_queue *cfqq; | |
3962 | ||
3963 | cfqq = cfqd->active_queue; | |
3964 | if (!cfqq) | |
3965 | return false; | |
3966 | ||
3967 | if (cfq_class_idle(new_cfqq)) | |
3968 | return false; | |
3969 | ||
3970 | if (cfq_class_idle(cfqq)) | |
3971 | return true; | |
3972 | ||
3973 | /* | |
3974 | * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice. | |
3975 | */ | |
3976 | if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq)) | |
3977 | return false; | |
3978 | ||
3979 | /* | |
3980 | * if the new request is sync, but the currently running queue is | |
3981 | * not, let the sync request have priority. | |
3982 | */ | |
3983 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) | |
3984 | return true; | |
3985 | ||
3986 | /* | |
3987 | * Treat ancestors of current cgroup the same way as current cgroup. | |
3988 | * For anybody else we disallow preemption to guarantee service | |
3989 | * fairness among cgroups. | |
3990 | */ | |
3991 | if (!cfqg_is_descendant(cfqq->cfqg, new_cfqq->cfqg)) | |
3992 | return false; | |
3993 | ||
3994 | if (cfq_slice_used(cfqq)) | |
3995 | return true; | |
3996 | ||
3997 | /* | |
3998 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
3999 | */ | |
4000 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
4001 | return true; | |
4002 | ||
4003 | WARN_ON_ONCE(cfqq->ioprio_class != new_cfqq->ioprio_class); | |
4004 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
4005 | if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD && | |
4006 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
4007 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
4008 | return true; | |
4009 | ||
4010 | /* | |
4011 | * So both queues are sync. Let the new request get disk time if | |
4012 | * it's a metadata request and the current queue is doing regular IO. | |
4013 | */ | |
4014 | if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending) | |
4015 | return true; | |
4016 | ||
4017 | /* An idle queue should not be idle now for some reason */ | |
4018 | if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq)) | |
4019 | return true; | |
4020 | ||
4021 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) | |
4022 | return false; | |
4023 | ||
4024 | /* | |
4025 | * if this request is as-good as one we would expect from the | |
4026 | * current cfqq, let it preempt | |
4027 | */ | |
4028 | if (cfq_rq_close(cfqd, cfqq, rq)) | |
4029 | return true; | |
4030 | ||
4031 | return false; | |
4032 | } | |
4033 | ||
4034 | /* | |
4035 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
4036 | * let it have half of its nominal slice. | |
4037 | */ | |
4038 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
4039 | { | |
4040 | enum wl_type_t old_type = cfqq_type(cfqd->active_queue); | |
4041 | ||
4042 | cfq_log_cfqq(cfqd, cfqq, "preempt"); | |
4043 | cfq_slice_expired(cfqd, 1); | |
4044 | ||
4045 | /* | |
4046 | * workload type is changed, don't save slice, otherwise preempt | |
4047 | * doesn't happen | |
4048 | */ | |
4049 | if (old_type != cfqq_type(cfqq)) | |
4050 | cfqq->cfqg->saved_wl_slice = 0; | |
4051 | ||
4052 | /* | |
4053 | * Put the new queue at the front of the of the current list, | |
4054 | * so we know that it will be selected next. | |
4055 | */ | |
4056 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
4057 | ||
4058 | cfq_service_tree_add(cfqd, cfqq, 1); | |
4059 | ||
4060 | cfqq->slice_end = 0; | |
4061 | cfq_mark_cfqq_slice_new(cfqq); | |
4062 | } | |
4063 | ||
4064 | /* | |
4065 | * Called when a new fs request (rq) is added (to cfqq). Check if there's | |
4066 | * something we should do about it | |
4067 | */ | |
4068 | static void | |
4069 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
4070 | struct request *rq) | |
4071 | { | |
4072 | struct cfq_io_cq *cic = RQ_CIC(rq); | |
4073 | ||
4074 | cfqd->rq_queued++; | |
4075 | if (rq->cmd_flags & REQ_PRIO) | |
4076 | cfqq->prio_pending++; | |
4077 | ||
4078 | cfq_update_io_thinktime(cfqd, cfqq, cic); | |
4079 | cfq_update_io_seektime(cfqd, cfqq, rq); | |
4080 | cfq_update_idle_window(cfqd, cfqq, cic); | |
4081 | ||
4082 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
4083 | ||
4084 | if (cfqq == cfqd->active_queue) { | |
4085 | /* | |
4086 | * Remember that we saw a request from this process, but | |
4087 | * don't start queuing just yet. Otherwise we risk seeing lots | |
4088 | * of tiny requests, because we disrupt the normal plugging | |
4089 | * and merging. If the request is already larger than a single | |
4090 | * page, let it rip immediately. For that case we assume that | |
4091 | * merging is already done. Ditto for a busy system that | |
4092 | * has other work pending, don't risk delaying until the | |
4093 | * idle timer unplug to continue working. | |
4094 | */ | |
4095 | if (cfq_cfqq_wait_request(cfqq)) { | |
4096 | if (blk_rq_bytes(rq) > PAGE_SIZE || | |
4097 | cfqd->busy_queues > 1) { | |
4098 | cfq_del_timer(cfqd, cfqq); | |
4099 | cfq_clear_cfqq_wait_request(cfqq); | |
4100 | __blk_run_queue(cfqd->queue); | |
4101 | } else { | |
4102 | cfqg_stats_update_idle_time(cfqq->cfqg); | |
4103 | cfq_mark_cfqq_must_dispatch(cfqq); | |
4104 | } | |
4105 | } | |
4106 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { | |
4107 | /* | |
4108 | * not the active queue - expire current slice if it is | |
4109 | * idle and has expired it's mean thinktime or this new queue | |
4110 | * has some old slice time left and is of higher priority or | |
4111 | * this new queue is RT and the current one is BE | |
4112 | */ | |
4113 | cfq_preempt_queue(cfqd, cfqq); | |
4114 | __blk_run_queue(cfqd->queue); | |
4115 | } | |
4116 | } | |
4117 | ||
4118 | static void cfq_insert_request(struct request_queue *q, struct request *rq) | |
4119 | { | |
4120 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
4121 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
4122 | ||
4123 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); | |
4124 | cfq_init_prio_data(cfqq, RQ_CIC(rq)); | |
4125 | ||
4126 | rq->fifo_time = ktime_get_ns() + cfqd->cfq_fifo_expire[rq_is_sync(rq)]; | |
4127 | list_add_tail(&rq->queuelist, &cfqq->fifo); | |
4128 | cfq_add_rq_rb(rq); | |
4129 | cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group, req_op(rq), | |
4130 | rq->cmd_flags); | |
4131 | cfq_rq_enqueued(cfqd, cfqq, rq); | |
4132 | } | |
4133 | ||
4134 | /* | |
4135 | * Update hw_tag based on peak queue depth over 50 samples under | |
4136 | * sufficient load. | |
4137 | */ | |
4138 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
4139 | { | |
4140 | struct cfq_queue *cfqq = cfqd->active_queue; | |
4141 | ||
4142 | if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth) | |
4143 | cfqd->hw_tag_est_depth = cfqd->rq_in_driver; | |
4144 | ||
4145 | if (cfqd->hw_tag == 1) | |
4146 | return; | |
4147 | ||
4148 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
4149 | cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN) | |
4150 | return; | |
4151 | ||
4152 | /* | |
4153 | * If active queue hasn't enough requests and can idle, cfq might not | |
4154 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
4155 | * case | |
4156 | */ | |
4157 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
4158 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
4159 | CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN) | |
4160 | return; | |
4161 | ||
4162 | if (cfqd->hw_tag_samples++ < 50) | |
4163 | return; | |
4164 | ||
4165 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) | |
4166 | cfqd->hw_tag = 1; | |
4167 | else | |
4168 | cfqd->hw_tag = 0; | |
4169 | } | |
4170 | ||
4171 | static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
4172 | { | |
4173 | struct cfq_io_cq *cic = cfqd->active_cic; | |
4174 | u64 now = ktime_get_ns(); | |
4175 | ||
4176 | /* If the queue already has requests, don't wait */ | |
4177 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
4178 | return false; | |
4179 | ||
4180 | /* If there are other queues in the group, don't wait */ | |
4181 | if (cfqq->cfqg->nr_cfqq > 1) | |
4182 | return false; | |
4183 | ||
4184 | /* the only queue in the group, but think time is big */ | |
4185 | if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) | |
4186 | return false; | |
4187 | ||
4188 | if (cfq_slice_used(cfqq)) | |
4189 | return true; | |
4190 | ||
4191 | /* if slice left is less than think time, wait busy */ | |
4192 | if (cic && sample_valid(cic->ttime.ttime_samples) | |
4193 | && (cfqq->slice_end - now < cic->ttime.ttime_mean)) | |
4194 | return true; | |
4195 | ||
4196 | /* | |
4197 | * If think times is less than a jiffy than ttime_mean=0 and above | |
4198 | * will not be true. It might happen that slice has not expired yet | |
4199 | * but will expire soon (4-5 ns) during select_queue(). To cover the | |
4200 | * case where think time is less than a jiffy, mark the queue wait | |
4201 | * busy if only 1 jiffy is left in the slice. | |
4202 | */ | |
4203 | if (cfqq->slice_end - now <= jiffies_to_nsecs(1)) | |
4204 | return true; | |
4205 | ||
4206 | return false; | |
4207 | } | |
4208 | ||
4209 | static void cfq_completed_request(struct request_queue *q, struct request *rq) | |
4210 | { | |
4211 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
4212 | struct cfq_data *cfqd = cfqq->cfqd; | |
4213 | const int sync = rq_is_sync(rq); | |
4214 | u64 now = ktime_get_ns(); | |
4215 | ||
4216 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", | |
4217 | !!(rq->cmd_flags & REQ_NOIDLE)); | |
4218 | ||
4219 | cfq_update_hw_tag(cfqd); | |
4220 | ||
4221 | WARN_ON(!cfqd->rq_in_driver); | |
4222 | WARN_ON(!cfqq->dispatched); | |
4223 | cfqd->rq_in_driver--; | |
4224 | cfqq->dispatched--; | |
4225 | (RQ_CFQG(rq))->dispatched--; | |
4226 | cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq), | |
4227 | rq_io_start_time_ns(rq), req_op(rq), | |
4228 | rq->cmd_flags); | |
4229 | ||
4230 | cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--; | |
4231 | ||
4232 | if (sync) { | |
4233 | struct cfq_rb_root *st; | |
4234 | ||
4235 | RQ_CIC(rq)->ttime.last_end_request = now; | |
4236 | ||
4237 | if (cfq_cfqq_on_rr(cfqq)) | |
4238 | st = cfqq->service_tree; | |
4239 | else | |
4240 | st = st_for(cfqq->cfqg, cfqq_class(cfqq), | |
4241 | cfqq_type(cfqq)); | |
4242 | ||
4243 | st->ttime.last_end_request = now; | |
4244 | if (!(rq->start_time + cfqd->cfq_fifo_expire[1] > now)) | |
4245 | cfqd->last_delayed_sync = now; | |
4246 | } | |
4247 | ||
4248 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4249 | cfqq->cfqg->ttime.last_end_request = now; | |
4250 | #endif | |
4251 | ||
4252 | /* | |
4253 | * If this is the active queue, check if it needs to be expired, | |
4254 | * or if we want to idle in case it has no pending requests. | |
4255 | */ | |
4256 | if (cfqd->active_queue == cfqq) { | |
4257 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); | |
4258 | ||
4259 | if (cfq_cfqq_slice_new(cfqq)) { | |
4260 | cfq_set_prio_slice(cfqd, cfqq); | |
4261 | cfq_clear_cfqq_slice_new(cfqq); | |
4262 | } | |
4263 | ||
4264 | /* | |
4265 | * Should we wait for next request to come in before we expire | |
4266 | * the queue. | |
4267 | */ | |
4268 | if (cfq_should_wait_busy(cfqd, cfqq)) { | |
4269 | u64 extend_sl = cfqd->cfq_slice_idle; | |
4270 | if (!cfqd->cfq_slice_idle) | |
4271 | extend_sl = cfqd->cfq_group_idle; | |
4272 | cfqq->slice_end = now + extend_sl; | |
4273 | cfq_mark_cfqq_wait_busy(cfqq); | |
4274 | cfq_log_cfqq(cfqd, cfqq, "will busy wait"); | |
4275 | } | |
4276 | ||
4277 | /* | |
4278 | * Idling is not enabled on: | |
4279 | * - expired queues | |
4280 | * - idle-priority queues | |
4281 | * - async queues | |
4282 | * - queues with still some requests queued | |
4283 | * - when there is a close cooperator | |
4284 | */ | |
4285 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) | |
4286 | cfq_slice_expired(cfqd, 1); | |
4287 | else if (sync && cfqq_empty && | |
4288 | !cfq_close_cooperator(cfqd, cfqq)) { | |
4289 | cfq_arm_slice_timer(cfqd); | |
4290 | } | |
4291 | } | |
4292 | ||
4293 | if (!cfqd->rq_in_driver) | |
4294 | cfq_schedule_dispatch(cfqd); | |
4295 | } | |
4296 | ||
4297 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) | |
4298 | { | |
4299 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { | |
4300 | cfq_mark_cfqq_must_alloc_slice(cfqq); | |
4301 | return ELV_MQUEUE_MUST; | |
4302 | } | |
4303 | ||
4304 | return ELV_MQUEUE_MAY; | |
4305 | } | |
4306 | ||
4307 | static int cfq_may_queue(struct request_queue *q, int op, int op_flags) | |
4308 | { | |
4309 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
4310 | struct task_struct *tsk = current; | |
4311 | struct cfq_io_cq *cic; | |
4312 | struct cfq_queue *cfqq; | |
4313 | ||
4314 | /* | |
4315 | * don't force setup of a queue from here, as a call to may_queue | |
4316 | * does not necessarily imply that a request actually will be queued. | |
4317 | * so just lookup a possibly existing queue, or return 'may queue' | |
4318 | * if that fails | |
4319 | */ | |
4320 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
4321 | if (!cic) | |
4322 | return ELV_MQUEUE_MAY; | |
4323 | ||
4324 | cfqq = cic_to_cfqq(cic, rw_is_sync(op, op_flags)); | |
4325 | if (cfqq) { | |
4326 | cfq_init_prio_data(cfqq, cic); | |
4327 | ||
4328 | return __cfq_may_queue(cfqq); | |
4329 | } | |
4330 | ||
4331 | return ELV_MQUEUE_MAY; | |
4332 | } | |
4333 | ||
4334 | /* | |
4335 | * queue lock held here | |
4336 | */ | |
4337 | static void cfq_put_request(struct request *rq) | |
4338 | { | |
4339 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
4340 | ||
4341 | if (cfqq) { | |
4342 | const int rw = rq_data_dir(rq); | |
4343 | ||
4344 | BUG_ON(!cfqq->allocated[rw]); | |
4345 | cfqq->allocated[rw]--; | |
4346 | ||
4347 | /* Put down rq reference on cfqg */ | |
4348 | cfqg_put(RQ_CFQG(rq)); | |
4349 | rq->elv.priv[0] = NULL; | |
4350 | rq->elv.priv[1] = NULL; | |
4351 | ||
4352 | cfq_put_queue(cfqq); | |
4353 | } | |
4354 | } | |
4355 | ||
4356 | static struct cfq_queue * | |
4357 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic, | |
4358 | struct cfq_queue *cfqq) | |
4359 | { | |
4360 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
4361 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
4362 | cfq_mark_cfqq_coop(cfqq->new_cfqq); | |
4363 | cfq_put_queue(cfqq); | |
4364 | return cic_to_cfqq(cic, 1); | |
4365 | } | |
4366 | ||
4367 | /* | |
4368 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
4369 | * was the last process referring to said cfqq. | |
4370 | */ | |
4371 | static struct cfq_queue * | |
4372 | split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq) | |
4373 | { | |
4374 | if (cfqq_process_refs(cfqq) == 1) { | |
4375 | cfqq->pid = current->pid; | |
4376 | cfq_clear_cfqq_coop(cfqq); | |
4377 | cfq_clear_cfqq_split_coop(cfqq); | |
4378 | return cfqq; | |
4379 | } | |
4380 | ||
4381 | cic_set_cfqq(cic, NULL, 1); | |
4382 | ||
4383 | cfq_put_cooperator(cfqq); | |
4384 | ||
4385 | cfq_put_queue(cfqq); | |
4386 | return NULL; | |
4387 | } | |
4388 | /* | |
4389 | * Allocate cfq data structures associated with this request. | |
4390 | */ | |
4391 | static int | |
4392 | cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio, | |
4393 | gfp_t gfp_mask) | |
4394 | { | |
4395 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
4396 | struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq); | |
4397 | const int rw = rq_data_dir(rq); | |
4398 | const bool is_sync = rq_is_sync(rq); | |
4399 | struct cfq_queue *cfqq; | |
4400 | ||
4401 | spin_lock_irq(q->queue_lock); | |
4402 | ||
4403 | check_ioprio_changed(cic, bio); | |
4404 | check_blkcg_changed(cic, bio); | |
4405 | new_queue: | |
4406 | cfqq = cic_to_cfqq(cic, is_sync); | |
4407 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
4408 | if (cfqq) | |
4409 | cfq_put_queue(cfqq); | |
4410 | cfqq = cfq_get_queue(cfqd, is_sync, cic, bio); | |
4411 | cic_set_cfqq(cic, cfqq, is_sync); | |
4412 | } else { | |
4413 | /* | |
4414 | * If the queue was seeky for too long, break it apart. | |
4415 | */ | |
4416 | if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) { | |
4417 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
4418 | cfqq = split_cfqq(cic, cfqq); | |
4419 | if (!cfqq) | |
4420 | goto new_queue; | |
4421 | } | |
4422 | ||
4423 | /* | |
4424 | * Check to see if this queue is scheduled to merge with | |
4425 | * another, closely cooperating queue. The merging of | |
4426 | * queues happens here as it must be done in process context. | |
4427 | * The reference on new_cfqq was taken in merge_cfqqs. | |
4428 | */ | |
4429 | if (cfqq->new_cfqq) | |
4430 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
4431 | } | |
4432 | ||
4433 | cfqq->allocated[rw]++; | |
4434 | ||
4435 | cfqq->ref++; | |
4436 | cfqg_get(cfqq->cfqg); | |
4437 | rq->elv.priv[0] = cfqq; | |
4438 | rq->elv.priv[1] = cfqq->cfqg; | |
4439 | spin_unlock_irq(q->queue_lock); | |
4440 | return 0; | |
4441 | } | |
4442 | ||
4443 | static void cfq_kick_queue(struct work_struct *work) | |
4444 | { | |
4445 | struct cfq_data *cfqd = | |
4446 | container_of(work, struct cfq_data, unplug_work); | |
4447 | struct request_queue *q = cfqd->queue; | |
4448 | ||
4449 | spin_lock_irq(q->queue_lock); | |
4450 | __blk_run_queue(cfqd->queue); | |
4451 | spin_unlock_irq(q->queue_lock); | |
4452 | } | |
4453 | ||
4454 | /* | |
4455 | * Timer running if the active_queue is currently idling inside its time slice | |
4456 | */ | |
4457 | static void cfq_idle_slice_timer(unsigned long data) | |
4458 | { | |
4459 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
4460 | struct cfq_queue *cfqq; | |
4461 | unsigned long flags; | |
4462 | int timed_out = 1; | |
4463 | ||
4464 | cfq_log(cfqd, "idle timer fired"); | |
4465 | ||
4466 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
4467 | ||
4468 | cfqq = cfqd->active_queue; | |
4469 | if (cfqq) { | |
4470 | timed_out = 0; | |
4471 | ||
4472 | /* | |
4473 | * We saw a request before the queue expired, let it through | |
4474 | */ | |
4475 | if (cfq_cfqq_must_dispatch(cfqq)) | |
4476 | goto out_kick; | |
4477 | ||
4478 | /* | |
4479 | * expired | |
4480 | */ | |
4481 | if (cfq_slice_used(cfqq)) | |
4482 | goto expire; | |
4483 | ||
4484 | /* | |
4485 | * only expire and reinvoke request handler, if there are | |
4486 | * other queues with pending requests | |
4487 | */ | |
4488 | if (!cfqd->busy_queues) | |
4489 | goto out_cont; | |
4490 | ||
4491 | /* | |
4492 | * not expired and it has a request pending, let it dispatch | |
4493 | */ | |
4494 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
4495 | goto out_kick; | |
4496 | ||
4497 | /* | |
4498 | * Queue depth flag is reset only when the idle didn't succeed | |
4499 | */ | |
4500 | cfq_clear_cfqq_deep(cfqq); | |
4501 | } | |
4502 | expire: | |
4503 | cfq_slice_expired(cfqd, timed_out); | |
4504 | out_kick: | |
4505 | cfq_schedule_dispatch(cfqd); | |
4506 | out_cont: | |
4507 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
4508 | } | |
4509 | ||
4510 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) | |
4511 | { | |
4512 | del_timer_sync(&cfqd->idle_slice_timer); | |
4513 | cancel_work_sync(&cfqd->unplug_work); | |
4514 | } | |
4515 | ||
4516 | static void cfq_exit_queue(struct elevator_queue *e) | |
4517 | { | |
4518 | struct cfq_data *cfqd = e->elevator_data; | |
4519 | struct request_queue *q = cfqd->queue; | |
4520 | ||
4521 | cfq_shutdown_timer_wq(cfqd); | |
4522 | ||
4523 | spin_lock_irq(q->queue_lock); | |
4524 | ||
4525 | if (cfqd->active_queue) | |
4526 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); | |
4527 | ||
4528 | spin_unlock_irq(q->queue_lock); | |
4529 | ||
4530 | cfq_shutdown_timer_wq(cfqd); | |
4531 | ||
4532 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4533 | blkcg_deactivate_policy(q, &blkcg_policy_cfq); | |
4534 | #else | |
4535 | kfree(cfqd->root_group); | |
4536 | #endif | |
4537 | kfree(cfqd); | |
4538 | } | |
4539 | ||
4540 | static int cfq_init_queue(struct request_queue *q, struct elevator_type *e) | |
4541 | { | |
4542 | struct cfq_data *cfqd; | |
4543 | struct blkcg_gq *blkg __maybe_unused; | |
4544 | int i, ret; | |
4545 | struct elevator_queue *eq; | |
4546 | ||
4547 | eq = elevator_alloc(q, e); | |
4548 | if (!eq) | |
4549 | return -ENOMEM; | |
4550 | ||
4551 | cfqd = kzalloc_node(sizeof(*cfqd), GFP_KERNEL, q->node); | |
4552 | if (!cfqd) { | |
4553 | kobject_put(&eq->kobj); | |
4554 | return -ENOMEM; | |
4555 | } | |
4556 | eq->elevator_data = cfqd; | |
4557 | ||
4558 | cfqd->queue = q; | |
4559 | spin_lock_irq(q->queue_lock); | |
4560 | q->elevator = eq; | |
4561 | spin_unlock_irq(q->queue_lock); | |
4562 | ||
4563 | /* Init root service tree */ | |
4564 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
4565 | ||
4566 | /* Init root group and prefer root group over other groups by default */ | |
4567 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4568 | ret = blkcg_activate_policy(q, &blkcg_policy_cfq); | |
4569 | if (ret) | |
4570 | goto out_free; | |
4571 | ||
4572 | cfqd->root_group = blkg_to_cfqg(q->root_blkg); | |
4573 | #else | |
4574 | ret = -ENOMEM; | |
4575 | cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group), | |
4576 | GFP_KERNEL, cfqd->queue->node); | |
4577 | if (!cfqd->root_group) | |
4578 | goto out_free; | |
4579 | ||
4580 | cfq_init_cfqg_base(cfqd->root_group); | |
4581 | cfqd->root_group->weight = 2 * CFQ_WEIGHT_LEGACY_DFL; | |
4582 | cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_LEGACY_DFL; | |
4583 | #endif | |
4584 | ||
4585 | /* | |
4586 | * Not strictly needed (since RB_ROOT just clears the node and we | |
4587 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
4588 | * to add magic to the rb code | |
4589 | */ | |
4590 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
4591 | cfqd->prio_trees[i] = RB_ROOT; | |
4592 | ||
4593 | /* | |
4594 | * Our fallback cfqq if cfq_get_queue() runs into OOM issues. | |
4595 | * Grab a permanent reference to it, so that the normal code flow | |
4596 | * will not attempt to free it. oom_cfqq is linked to root_group | |
4597 | * but shouldn't hold a reference as it'll never be unlinked. Lose | |
4598 | * the reference from linking right away. | |
4599 | */ | |
4600 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
4601 | cfqd->oom_cfqq.ref++; | |
4602 | ||
4603 | spin_lock_irq(q->queue_lock); | |
4604 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group); | |
4605 | cfqg_put(cfqd->root_group); | |
4606 | spin_unlock_irq(q->queue_lock); | |
4607 | ||
4608 | init_timer(&cfqd->idle_slice_timer); | |
4609 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
4610 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
4611 | ||
4612 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); | |
4613 | ||
4614 | cfqd->cfq_quantum = cfq_quantum; | |
4615 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; | |
4616 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
4617 | cfqd->cfq_back_max = cfq_back_max; | |
4618 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
4619 | cfqd->cfq_slice[0] = cfq_slice_async; | |
4620 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
4621 | cfqd->cfq_target_latency = cfq_target_latency; | |
4622 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
4623 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
4624 | cfqd->cfq_group_idle = cfq_group_idle; | |
4625 | cfqd->cfq_latency = 1; | |
4626 | cfqd->hw_tag = -1; | |
4627 | /* | |
4628 | * we optimistically start assuming sync ops weren't delayed in last | |
4629 | * second, in order to have larger depth for async operations. | |
4630 | */ | |
4631 | cfqd->last_delayed_sync = ktime_get_ns() - NSEC_PER_SEC; | |
4632 | return 0; | |
4633 | ||
4634 | out_free: | |
4635 | kfree(cfqd); | |
4636 | kobject_put(&eq->kobj); | |
4637 | return ret; | |
4638 | } | |
4639 | ||
4640 | static void cfq_registered_queue(struct request_queue *q) | |
4641 | { | |
4642 | struct elevator_queue *e = q->elevator; | |
4643 | struct cfq_data *cfqd = e->elevator_data; | |
4644 | ||
4645 | /* | |
4646 | * Default to IOPS mode with no idling for SSDs | |
4647 | */ | |
4648 | if (blk_queue_nonrot(q)) | |
4649 | cfqd->cfq_slice_idle = 0; | |
4650 | } | |
4651 | ||
4652 | /* | |
4653 | * sysfs parts below --> | |
4654 | */ | |
4655 | static ssize_t | |
4656 | cfq_var_show(unsigned int var, char *page) | |
4657 | { | |
4658 | return sprintf(page, "%u\n", var); | |
4659 | } | |
4660 | ||
4661 | static ssize_t | |
4662 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
4663 | { | |
4664 | char *p = (char *) page; | |
4665 | ||
4666 | *var = simple_strtoul(p, &p, 10); | |
4667 | return count; | |
4668 | } | |
4669 | ||
4670 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
4671 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
4672 | { \ | |
4673 | struct cfq_data *cfqd = e->elevator_data; \ | |
4674 | u64 __data = __VAR; \ | |
4675 | if (__CONV) \ | |
4676 | __data = div_u64(__data, NSEC_PER_MSEC); \ | |
4677 | return cfq_var_show(__data, (page)); \ | |
4678 | } | |
4679 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
4680 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); | |
4681 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
4682 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); | |
4683 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
4684 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); | |
4685 | SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1); | |
4686 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
4687 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
4688 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
4689 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); | |
4690 | SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1); | |
4691 | #undef SHOW_FUNCTION | |
4692 | ||
4693 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
4694 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
4695 | { \ | |
4696 | struct cfq_data *cfqd = e->elevator_data; \ | |
4697 | unsigned int __data; \ | |
4698 | int ret = cfq_var_store(&__data, (page), count); \ | |
4699 | if (__data < (MIN)) \ | |
4700 | __data = (MIN); \ | |
4701 | else if (__data > (MAX)) \ | |
4702 | __data = (MAX); \ | |
4703 | if (__CONV) \ | |
4704 | *(__PTR) = (u64)__data * NSEC_PER_MSEC; \ | |
4705 | else \ | |
4706 | *(__PTR) = __data; \ | |
4707 | return ret; \ | |
4708 | } | |
4709 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
4710 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, | |
4711 | UINT_MAX, 1); | |
4712 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
4713 | UINT_MAX, 1); | |
4714 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); | |
4715 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, | |
4716 | UINT_MAX, 0); | |
4717 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); | |
4718 | STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1); | |
4719 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
4720 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
4721 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, | |
4722 | UINT_MAX, 0); | |
4723 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); | |
4724 | STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1); | |
4725 | #undef STORE_FUNCTION | |
4726 | ||
4727 | #define CFQ_ATTR(name) \ | |
4728 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
4729 | ||
4730 | static struct elv_fs_entry cfq_attrs[] = { | |
4731 | CFQ_ATTR(quantum), | |
4732 | CFQ_ATTR(fifo_expire_sync), | |
4733 | CFQ_ATTR(fifo_expire_async), | |
4734 | CFQ_ATTR(back_seek_max), | |
4735 | CFQ_ATTR(back_seek_penalty), | |
4736 | CFQ_ATTR(slice_sync), | |
4737 | CFQ_ATTR(slice_async), | |
4738 | CFQ_ATTR(slice_async_rq), | |
4739 | CFQ_ATTR(slice_idle), | |
4740 | CFQ_ATTR(group_idle), | |
4741 | CFQ_ATTR(low_latency), | |
4742 | CFQ_ATTR(target_latency), | |
4743 | __ATTR_NULL | |
4744 | }; | |
4745 | ||
4746 | static struct elevator_type iosched_cfq = { | |
4747 | .ops = { | |
4748 | .elevator_merge_fn = cfq_merge, | |
4749 | .elevator_merged_fn = cfq_merged_request, | |
4750 | .elevator_merge_req_fn = cfq_merged_requests, | |
4751 | .elevator_allow_merge_fn = cfq_allow_merge, | |
4752 | .elevator_bio_merged_fn = cfq_bio_merged, | |
4753 | .elevator_dispatch_fn = cfq_dispatch_requests, | |
4754 | .elevator_add_req_fn = cfq_insert_request, | |
4755 | .elevator_activate_req_fn = cfq_activate_request, | |
4756 | .elevator_deactivate_req_fn = cfq_deactivate_request, | |
4757 | .elevator_completed_req_fn = cfq_completed_request, | |
4758 | .elevator_former_req_fn = elv_rb_former_request, | |
4759 | .elevator_latter_req_fn = elv_rb_latter_request, | |
4760 | .elevator_init_icq_fn = cfq_init_icq, | |
4761 | .elevator_exit_icq_fn = cfq_exit_icq, | |
4762 | .elevator_set_req_fn = cfq_set_request, | |
4763 | .elevator_put_req_fn = cfq_put_request, | |
4764 | .elevator_may_queue_fn = cfq_may_queue, | |
4765 | .elevator_init_fn = cfq_init_queue, | |
4766 | .elevator_exit_fn = cfq_exit_queue, | |
4767 | .elevator_registered_fn = cfq_registered_queue, | |
4768 | }, | |
4769 | .icq_size = sizeof(struct cfq_io_cq), | |
4770 | .icq_align = __alignof__(struct cfq_io_cq), | |
4771 | .elevator_attrs = cfq_attrs, | |
4772 | .elevator_name = "cfq", | |
4773 | .elevator_owner = THIS_MODULE, | |
4774 | }; | |
4775 | ||
4776 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4777 | static struct blkcg_policy blkcg_policy_cfq = { | |
4778 | .dfl_cftypes = cfq_blkcg_files, | |
4779 | .legacy_cftypes = cfq_blkcg_legacy_files, | |
4780 | ||
4781 | .cpd_alloc_fn = cfq_cpd_alloc, | |
4782 | .cpd_init_fn = cfq_cpd_init, | |
4783 | .cpd_free_fn = cfq_cpd_free, | |
4784 | .cpd_bind_fn = cfq_cpd_bind, | |
4785 | ||
4786 | .pd_alloc_fn = cfq_pd_alloc, | |
4787 | .pd_init_fn = cfq_pd_init, | |
4788 | .pd_offline_fn = cfq_pd_offline, | |
4789 | .pd_free_fn = cfq_pd_free, | |
4790 | .pd_reset_stats_fn = cfq_pd_reset_stats, | |
4791 | }; | |
4792 | #endif | |
4793 | ||
4794 | static int __init cfq_init(void) | |
4795 | { | |
4796 | int ret; | |
4797 | ||
4798 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4799 | ret = blkcg_policy_register(&blkcg_policy_cfq); | |
4800 | if (ret) | |
4801 | return ret; | |
4802 | #else | |
4803 | cfq_group_idle = 0; | |
4804 | #endif | |
4805 | ||
4806 | ret = -ENOMEM; | |
4807 | cfq_pool = KMEM_CACHE(cfq_queue, 0); | |
4808 | if (!cfq_pool) | |
4809 | goto err_pol_unreg; | |
4810 | ||
4811 | ret = elv_register(&iosched_cfq); | |
4812 | if (ret) | |
4813 | goto err_free_pool; | |
4814 | ||
4815 | return 0; | |
4816 | ||
4817 | err_free_pool: | |
4818 | kmem_cache_destroy(cfq_pool); | |
4819 | err_pol_unreg: | |
4820 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4821 | blkcg_policy_unregister(&blkcg_policy_cfq); | |
4822 | #endif | |
4823 | return ret; | |
4824 | } | |
4825 | ||
4826 | static void __exit cfq_exit(void) | |
4827 | { | |
4828 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
4829 | blkcg_policy_unregister(&blkcg_policy_cfq); | |
4830 | #endif | |
4831 | elv_unregister(&iosched_cfq); | |
4832 | kmem_cache_destroy(cfq_pool); | |
4833 | } | |
4834 | ||
4835 | module_init(cfq_init); | |
4836 | module_exit(cfq_exit); | |
4837 | ||
4838 | MODULE_AUTHOR("Jens Axboe"); | |
4839 | MODULE_LICENSE("GPL"); | |
4840 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |