]>
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/blkdev.h> | |
11 | #include <linux/elevator.h> | |
12 | #include <linux/jiffies.h> | |
13 | #include <linux/rbtree.h> | |
14 | #include <linux/ioprio.h> | |
15 | #include <linux/blktrace_api.h> | |
16 | #include "blk-cgroup.h" | |
17 | ||
18 | /* | |
19 | * tunables | |
20 | */ | |
21 | /* max queue in one round of service */ | |
22 | static const int cfq_quantum = 4; | |
23 | static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; | |
24 | /* maximum backwards seek, in KiB */ | |
25 | static const int cfq_back_max = 16 * 1024; | |
26 | /* penalty of a backwards seek */ | |
27 | static const int cfq_back_penalty = 2; | |
28 | static const int cfq_slice_sync = HZ / 10; | |
29 | static int cfq_slice_async = HZ / 25; | |
30 | static const int cfq_slice_async_rq = 2; | |
31 | static int cfq_slice_idle = HZ / 125; | |
32 | static const int cfq_target_latency = HZ * 3/10; /* 300 ms */ | |
33 | static const int cfq_hist_divisor = 4; | |
34 | ||
35 | /* | |
36 | * offset from end of service tree | |
37 | */ | |
38 | #define CFQ_IDLE_DELAY (HZ / 5) | |
39 | ||
40 | /* | |
41 | * below this threshold, we consider thinktime immediate | |
42 | */ | |
43 | #define CFQ_MIN_TT (2) | |
44 | ||
45 | /* | |
46 | * Allow merged cfqqs to perform this amount of seeky I/O before | |
47 | * deciding to break the queues up again. | |
48 | */ | |
49 | #define CFQQ_COOP_TOUT (HZ) | |
50 | ||
51 | #define CFQ_SLICE_SCALE (5) | |
52 | #define CFQ_HW_QUEUE_MIN (5) | |
53 | #define CFQ_SERVICE_SHIFT 12 | |
54 | ||
55 | #define RQ_CIC(rq) \ | |
56 | ((struct cfq_io_context *) (rq)->elevator_private) | |
57 | #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2) | |
58 | ||
59 | static struct kmem_cache *cfq_pool; | |
60 | static struct kmem_cache *cfq_ioc_pool; | |
61 | ||
62 | static DEFINE_PER_CPU(unsigned long, cfq_ioc_count); | |
63 | static struct completion *ioc_gone; | |
64 | static DEFINE_SPINLOCK(ioc_gone_lock); | |
65 | ||
66 | #define CFQ_PRIO_LISTS IOPRIO_BE_NR | |
67 | #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE) | |
68 | #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT) | |
69 | ||
70 | #define sample_valid(samples) ((samples) > 80) | |
71 | #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node) | |
72 | ||
73 | /* | |
74 | * Most of our rbtree usage is for sorting with min extraction, so | |
75 | * if we cache the leftmost node we don't have to walk down the tree | |
76 | * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should | |
77 | * move this into the elevator for the rq sorting as well. | |
78 | */ | |
79 | struct cfq_rb_root { | |
80 | struct rb_root rb; | |
81 | struct rb_node *left; | |
82 | unsigned count; | |
83 | u64 min_vdisktime; | |
84 | struct rb_node *active; | |
85 | unsigned total_weight; | |
86 | }; | |
87 | #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, } | |
88 | ||
89 | /* | |
90 | * Per process-grouping structure | |
91 | */ | |
92 | struct cfq_queue { | |
93 | /* reference count */ | |
94 | atomic_t ref; | |
95 | /* various state flags, see below */ | |
96 | unsigned int flags; | |
97 | /* parent cfq_data */ | |
98 | struct cfq_data *cfqd; | |
99 | /* service_tree member */ | |
100 | struct rb_node rb_node; | |
101 | /* service_tree key */ | |
102 | unsigned long rb_key; | |
103 | /* prio tree member */ | |
104 | struct rb_node p_node; | |
105 | /* prio tree root we belong to, if any */ | |
106 | struct rb_root *p_root; | |
107 | /* sorted list of pending requests */ | |
108 | struct rb_root sort_list; | |
109 | /* if fifo isn't expired, next request to serve */ | |
110 | struct request *next_rq; | |
111 | /* requests queued in sort_list */ | |
112 | int queued[2]; | |
113 | /* currently allocated requests */ | |
114 | int allocated[2]; | |
115 | /* fifo list of requests in sort_list */ | |
116 | struct list_head fifo; | |
117 | ||
118 | /* time when queue got scheduled in to dispatch first request. */ | |
119 | unsigned long dispatch_start; | |
120 | /* time when first request from queue completed and slice started. */ | |
121 | unsigned long slice_start; | |
122 | unsigned long slice_end; | |
123 | long slice_resid; | |
124 | unsigned int slice_dispatch; | |
125 | ||
126 | /* pending metadata requests */ | |
127 | int meta_pending; | |
128 | /* number of requests that are on the dispatch list or inside driver */ | |
129 | int dispatched; | |
130 | ||
131 | /* io prio of this group */ | |
132 | unsigned short ioprio, org_ioprio; | |
133 | unsigned short ioprio_class, org_ioprio_class; | |
134 | ||
135 | unsigned int seek_samples; | |
136 | u64 seek_total; | |
137 | sector_t seek_mean; | |
138 | sector_t last_request_pos; | |
139 | unsigned long seeky_start; | |
140 | ||
141 | pid_t pid; | |
142 | ||
143 | struct cfq_rb_root *service_tree; | |
144 | struct cfq_queue *new_cfqq; | |
145 | struct cfq_group *cfqg; | |
146 | /* Sectors dispatched in current dispatch round */ | |
147 | unsigned long nr_sectors; | |
148 | }; | |
149 | ||
150 | /* | |
151 | * First index in the service_trees. | |
152 | * IDLE is handled separately, so it has negative index | |
153 | */ | |
154 | enum wl_prio_t { | |
155 | BE_WORKLOAD = 0, | |
156 | RT_WORKLOAD = 1, | |
157 | IDLE_WORKLOAD = 2, | |
158 | }; | |
159 | ||
160 | /* | |
161 | * Second index in the service_trees. | |
162 | */ | |
163 | enum wl_type_t { | |
164 | ASYNC_WORKLOAD = 0, | |
165 | SYNC_NOIDLE_WORKLOAD = 1, | |
166 | SYNC_WORKLOAD = 2 | |
167 | }; | |
168 | ||
169 | /* This is per cgroup per device grouping structure */ | |
170 | struct cfq_group { | |
171 | /* group service_tree member */ | |
172 | struct rb_node rb_node; | |
173 | ||
174 | /* group service_tree key */ | |
175 | u64 vdisktime; | |
176 | unsigned int weight; | |
177 | bool on_st; | |
178 | ||
179 | /* number of cfqq currently on this group */ | |
180 | int nr_cfqq; | |
181 | ||
182 | /* Per group busy queus average. Useful for workload slice calc. */ | |
183 | unsigned int busy_queues_avg[2]; | |
184 | /* | |
185 | * rr lists of queues with requests, onle rr for each priority class. | |
186 | * Counts are embedded in the cfq_rb_root | |
187 | */ | |
188 | struct cfq_rb_root service_trees[2][3]; | |
189 | struct cfq_rb_root service_tree_idle; | |
190 | ||
191 | unsigned long saved_workload_slice; | |
192 | enum wl_type_t saved_workload; | |
193 | enum wl_prio_t saved_serving_prio; | |
194 | struct blkio_group blkg; | |
195 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
196 | struct hlist_node cfqd_node; | |
197 | atomic_t ref; | |
198 | #endif | |
199 | }; | |
200 | ||
201 | /* | |
202 | * Per block device queue structure | |
203 | */ | |
204 | struct cfq_data { | |
205 | struct request_queue *queue; | |
206 | /* Root service tree for cfq_groups */ | |
207 | struct cfq_rb_root grp_service_tree; | |
208 | struct cfq_group root_group; | |
209 | /* Number of active cfq groups on group service tree */ | |
210 | int nr_groups; | |
211 | ||
212 | /* | |
213 | * The priority currently being served | |
214 | */ | |
215 | enum wl_prio_t serving_prio; | |
216 | enum wl_type_t serving_type; | |
217 | unsigned long workload_expires; | |
218 | struct cfq_group *serving_group; | |
219 | bool noidle_tree_requires_idle; | |
220 | ||
221 | /* | |
222 | * Each priority tree is sorted by next_request position. These | |
223 | * trees are used when determining if two or more queues are | |
224 | * interleaving requests (see cfq_close_cooperator). | |
225 | */ | |
226 | struct rb_root prio_trees[CFQ_PRIO_LISTS]; | |
227 | ||
228 | unsigned int busy_queues; | |
229 | ||
230 | int rq_in_driver[2]; | |
231 | int sync_flight; | |
232 | ||
233 | /* | |
234 | * queue-depth detection | |
235 | */ | |
236 | int rq_queued; | |
237 | int hw_tag; | |
238 | /* | |
239 | * hw_tag can be | |
240 | * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection) | |
241 | * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth) | |
242 | * 0 => no NCQ | |
243 | */ | |
244 | int hw_tag_est_depth; | |
245 | unsigned int hw_tag_samples; | |
246 | ||
247 | /* | |
248 | * idle window management | |
249 | */ | |
250 | struct timer_list idle_slice_timer; | |
251 | struct work_struct unplug_work; | |
252 | ||
253 | struct cfq_queue *active_queue; | |
254 | struct cfq_io_context *active_cic; | |
255 | ||
256 | /* | |
257 | * async queue for each priority case | |
258 | */ | |
259 | struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR]; | |
260 | struct cfq_queue *async_idle_cfqq; | |
261 | ||
262 | sector_t last_position; | |
263 | ||
264 | /* | |
265 | * tunables, see top of file | |
266 | */ | |
267 | unsigned int cfq_quantum; | |
268 | unsigned int cfq_fifo_expire[2]; | |
269 | unsigned int cfq_back_penalty; | |
270 | unsigned int cfq_back_max; | |
271 | unsigned int cfq_slice[2]; | |
272 | unsigned int cfq_slice_async_rq; | |
273 | unsigned int cfq_slice_idle; | |
274 | unsigned int cfq_latency; | |
275 | ||
276 | struct list_head cic_list; | |
277 | ||
278 | /* | |
279 | * Fallback dummy cfqq for extreme OOM conditions | |
280 | */ | |
281 | struct cfq_queue oom_cfqq; | |
282 | ||
283 | unsigned long last_end_sync_rq; | |
284 | ||
285 | /* List of cfq groups being managed on this device*/ | |
286 | struct hlist_head cfqg_list; | |
287 | }; | |
288 | ||
289 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd); | |
290 | ||
291 | static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg, | |
292 | enum wl_prio_t prio, | |
293 | enum wl_type_t type, | |
294 | struct cfq_data *cfqd) | |
295 | { | |
296 | if (!cfqg) | |
297 | return NULL; | |
298 | ||
299 | if (prio == IDLE_WORKLOAD) | |
300 | return &cfqg->service_tree_idle; | |
301 | ||
302 | return &cfqg->service_trees[prio][type]; | |
303 | } | |
304 | ||
305 | enum cfqq_state_flags { | |
306 | CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */ | |
307 | CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */ | |
308 | CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */ | |
309 | CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */ | |
310 | CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ | |
311 | CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */ | |
312 | CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */ | |
313 | CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */ | |
314 | CFQ_CFQQ_FLAG_sync, /* synchronous queue */ | |
315 | CFQ_CFQQ_FLAG_coop, /* cfqq is shared */ | |
316 | CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */ | |
317 | }; | |
318 | ||
319 | #define CFQ_CFQQ_FNS(name) \ | |
320 | static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \ | |
321 | { \ | |
322 | (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \ | |
323 | } \ | |
324 | static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \ | |
325 | { \ | |
326 | (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \ | |
327 | } \ | |
328 | static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \ | |
329 | { \ | |
330 | return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \ | |
331 | } | |
332 | ||
333 | CFQ_CFQQ_FNS(on_rr); | |
334 | CFQ_CFQQ_FNS(wait_request); | |
335 | CFQ_CFQQ_FNS(must_dispatch); | |
336 | CFQ_CFQQ_FNS(must_alloc_slice); | |
337 | CFQ_CFQQ_FNS(fifo_expire); | |
338 | CFQ_CFQQ_FNS(idle_window); | |
339 | CFQ_CFQQ_FNS(prio_changed); | |
340 | CFQ_CFQQ_FNS(slice_new); | |
341 | CFQ_CFQQ_FNS(sync); | |
342 | CFQ_CFQQ_FNS(coop); | |
343 | CFQ_CFQQ_FNS(deep); | |
344 | #undef CFQ_CFQQ_FNS | |
345 | ||
346 | #ifdef CONFIG_DEBUG_CFQ_IOSCHED | |
347 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
348 | blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \ | |
349 | cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \ | |
350 | blkg_path(&(cfqq)->cfqg->blkg), ##args); | |
351 | ||
352 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \ | |
353 | blk_add_trace_msg((cfqd)->queue, "%s " fmt, \ | |
354 | blkg_path(&(cfqg)->blkg), ##args); \ | |
355 | ||
356 | #else | |
357 | #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \ | |
358 | blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args) | |
359 | #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0); | |
360 | #endif | |
361 | #define cfq_log(cfqd, fmt, args...) \ | |
362 | blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args) | |
363 | ||
364 | /* Traverses through cfq group service trees */ | |
365 | #define for_each_cfqg_st(cfqg, i, j, st) \ | |
366 | for (i = 0; i <= IDLE_WORKLOAD; i++) \ | |
367 | for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\ | |
368 | : &cfqg->service_tree_idle; \ | |
369 | (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \ | |
370 | (i == IDLE_WORKLOAD && j == 0); \ | |
371 | j++, st = i < IDLE_WORKLOAD ? \ | |
372 | &cfqg->service_trees[i][j]: NULL) \ | |
373 | ||
374 | ||
375 | static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq) | |
376 | { | |
377 | if (cfq_class_idle(cfqq)) | |
378 | return IDLE_WORKLOAD; | |
379 | if (cfq_class_rt(cfqq)) | |
380 | return RT_WORKLOAD; | |
381 | return BE_WORKLOAD; | |
382 | } | |
383 | ||
384 | ||
385 | static enum wl_type_t cfqq_type(struct cfq_queue *cfqq) | |
386 | { | |
387 | if (!cfq_cfqq_sync(cfqq)) | |
388 | return ASYNC_WORKLOAD; | |
389 | if (!cfq_cfqq_idle_window(cfqq)) | |
390 | return SYNC_NOIDLE_WORKLOAD; | |
391 | return SYNC_WORKLOAD; | |
392 | } | |
393 | ||
394 | static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl, | |
395 | struct cfq_data *cfqd, | |
396 | struct cfq_group *cfqg) | |
397 | { | |
398 | if (wl == IDLE_WORKLOAD) | |
399 | return cfqg->service_tree_idle.count; | |
400 | ||
401 | return cfqg->service_trees[wl][ASYNC_WORKLOAD].count | |
402 | + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count | |
403 | + cfqg->service_trees[wl][SYNC_WORKLOAD].count; | |
404 | } | |
405 | ||
406 | static void cfq_dispatch_insert(struct request_queue *, struct request *); | |
407 | static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool, | |
408 | struct io_context *, gfp_t); | |
409 | static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *, | |
410 | struct io_context *); | |
411 | ||
412 | static inline int rq_in_driver(struct cfq_data *cfqd) | |
413 | { | |
414 | return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1]; | |
415 | } | |
416 | ||
417 | static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic, | |
418 | bool is_sync) | |
419 | { | |
420 | return cic->cfqq[is_sync]; | |
421 | } | |
422 | ||
423 | static inline void cic_set_cfqq(struct cfq_io_context *cic, | |
424 | struct cfq_queue *cfqq, bool is_sync) | |
425 | { | |
426 | cic->cfqq[is_sync] = cfqq; | |
427 | } | |
428 | ||
429 | /* | |
430 | * We regard a request as SYNC, if it's either a read or has the SYNC bit | |
431 | * set (in which case it could also be direct WRITE). | |
432 | */ | |
433 | static inline bool cfq_bio_sync(struct bio *bio) | |
434 | { | |
435 | return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO); | |
436 | } | |
437 | ||
438 | /* | |
439 | * scheduler run of queue, if there are requests pending and no one in the | |
440 | * driver that will restart queueing | |
441 | */ | |
442 | static inline void cfq_schedule_dispatch(struct cfq_data *cfqd) | |
443 | { | |
444 | if (cfqd->busy_queues) { | |
445 | cfq_log(cfqd, "schedule dispatch"); | |
446 | kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work); | |
447 | } | |
448 | } | |
449 | ||
450 | static int cfq_queue_empty(struct request_queue *q) | |
451 | { | |
452 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
453 | ||
454 | return !cfqd->rq_queued; | |
455 | } | |
456 | ||
457 | /* | |
458 | * Scale schedule slice based on io priority. Use the sync time slice only | |
459 | * if a queue is marked sync and has sync io queued. A sync queue with async | |
460 | * io only, should not get full sync slice length. | |
461 | */ | |
462 | static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync, | |
463 | unsigned short prio) | |
464 | { | |
465 | const int base_slice = cfqd->cfq_slice[sync]; | |
466 | ||
467 | WARN_ON(prio >= IOPRIO_BE_NR); | |
468 | ||
469 | return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio)); | |
470 | } | |
471 | ||
472 | static inline int | |
473 | cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
474 | { | |
475 | return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio); | |
476 | } | |
477 | ||
478 | static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg) | |
479 | { | |
480 | u64 d = delta << CFQ_SERVICE_SHIFT; | |
481 | ||
482 | d = d * BLKIO_WEIGHT_DEFAULT; | |
483 | do_div(d, cfqg->weight); | |
484 | return d; | |
485 | } | |
486 | ||
487 | static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
488 | { | |
489 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
490 | if (delta > 0) | |
491 | min_vdisktime = vdisktime; | |
492 | ||
493 | return min_vdisktime; | |
494 | } | |
495 | ||
496 | static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime) | |
497 | { | |
498 | s64 delta = (s64)(vdisktime - min_vdisktime); | |
499 | if (delta < 0) | |
500 | min_vdisktime = vdisktime; | |
501 | ||
502 | return min_vdisktime; | |
503 | } | |
504 | ||
505 | static void update_min_vdisktime(struct cfq_rb_root *st) | |
506 | { | |
507 | u64 vdisktime = st->min_vdisktime; | |
508 | struct cfq_group *cfqg; | |
509 | ||
510 | if (st->active) { | |
511 | cfqg = rb_entry_cfqg(st->active); | |
512 | vdisktime = cfqg->vdisktime; | |
513 | } | |
514 | ||
515 | if (st->left) { | |
516 | cfqg = rb_entry_cfqg(st->left); | |
517 | vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime); | |
518 | } | |
519 | ||
520 | st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime); | |
521 | } | |
522 | ||
523 | /* | |
524 | * get averaged number of queues of RT/BE priority. | |
525 | * average is updated, with a formula that gives more weight to higher numbers, | |
526 | * to quickly follows sudden increases and decrease slowly | |
527 | */ | |
528 | ||
529 | static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd, | |
530 | struct cfq_group *cfqg, bool rt) | |
531 | { | |
532 | unsigned min_q, max_q; | |
533 | unsigned mult = cfq_hist_divisor - 1; | |
534 | unsigned round = cfq_hist_divisor / 2; | |
535 | unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg); | |
536 | ||
537 | min_q = min(cfqg->busy_queues_avg[rt], busy); | |
538 | max_q = max(cfqg->busy_queues_avg[rt], busy); | |
539 | cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) / | |
540 | cfq_hist_divisor; | |
541 | return cfqg->busy_queues_avg[rt]; | |
542 | } | |
543 | ||
544 | static inline unsigned | |
545 | cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
546 | { | |
547 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
548 | ||
549 | return cfq_target_latency * cfqg->weight / st->total_weight; | |
550 | } | |
551 | ||
552 | static inline void | |
553 | cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
554 | { | |
555 | unsigned slice = cfq_prio_to_slice(cfqd, cfqq); | |
556 | if (cfqd->cfq_latency) { | |
557 | /* | |
558 | * interested queues (we consider only the ones with the same | |
559 | * priority class in the cfq group) | |
560 | */ | |
561 | unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg, | |
562 | cfq_class_rt(cfqq)); | |
563 | unsigned sync_slice = cfqd->cfq_slice[1]; | |
564 | unsigned expect_latency = sync_slice * iq; | |
565 | unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg); | |
566 | ||
567 | if (expect_latency > group_slice) { | |
568 | unsigned base_low_slice = 2 * cfqd->cfq_slice_idle; | |
569 | /* scale low_slice according to IO priority | |
570 | * and sync vs async */ | |
571 | unsigned low_slice = | |
572 | min(slice, base_low_slice * slice / sync_slice); | |
573 | /* the adapted slice value is scaled to fit all iqs | |
574 | * into the target latency */ | |
575 | slice = max(slice * group_slice / expect_latency, | |
576 | low_slice); | |
577 | } | |
578 | } | |
579 | cfqq->slice_start = jiffies; | |
580 | cfqq->slice_end = jiffies + slice; | |
581 | cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies); | |
582 | } | |
583 | ||
584 | /* | |
585 | * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end | |
586 | * isn't valid until the first request from the dispatch is activated | |
587 | * and the slice time set. | |
588 | */ | |
589 | static inline bool cfq_slice_used(struct cfq_queue *cfqq) | |
590 | { | |
591 | if (cfq_cfqq_slice_new(cfqq)) | |
592 | return 0; | |
593 | if (time_before(jiffies, cfqq->slice_end)) | |
594 | return 0; | |
595 | ||
596 | return 1; | |
597 | } | |
598 | ||
599 | /* | |
600 | * Lifted from AS - choose which of rq1 and rq2 that is best served now. | |
601 | * We choose the request that is closest to the head right now. Distance | |
602 | * behind the head is penalized and only allowed to a certain extent. | |
603 | */ | |
604 | static struct request * | |
605 | cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last) | |
606 | { | |
607 | sector_t s1, s2, d1 = 0, d2 = 0; | |
608 | unsigned long back_max; | |
609 | #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */ | |
610 | #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */ | |
611 | unsigned wrap = 0; /* bit mask: requests behind the disk head? */ | |
612 | ||
613 | if (rq1 == NULL || rq1 == rq2) | |
614 | return rq2; | |
615 | if (rq2 == NULL) | |
616 | return rq1; | |
617 | ||
618 | if (rq_is_sync(rq1) && !rq_is_sync(rq2)) | |
619 | return rq1; | |
620 | else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) | |
621 | return rq2; | |
622 | if (rq_is_meta(rq1) && !rq_is_meta(rq2)) | |
623 | return rq1; | |
624 | else if (rq_is_meta(rq2) && !rq_is_meta(rq1)) | |
625 | return rq2; | |
626 | ||
627 | s1 = blk_rq_pos(rq1); | |
628 | s2 = blk_rq_pos(rq2); | |
629 | ||
630 | /* | |
631 | * by definition, 1KiB is 2 sectors | |
632 | */ | |
633 | back_max = cfqd->cfq_back_max * 2; | |
634 | ||
635 | /* | |
636 | * Strict one way elevator _except_ in the case where we allow | |
637 | * short backward seeks which are biased as twice the cost of a | |
638 | * similar forward seek. | |
639 | */ | |
640 | if (s1 >= last) | |
641 | d1 = s1 - last; | |
642 | else if (s1 + back_max >= last) | |
643 | d1 = (last - s1) * cfqd->cfq_back_penalty; | |
644 | else | |
645 | wrap |= CFQ_RQ1_WRAP; | |
646 | ||
647 | if (s2 >= last) | |
648 | d2 = s2 - last; | |
649 | else if (s2 + back_max >= last) | |
650 | d2 = (last - s2) * cfqd->cfq_back_penalty; | |
651 | else | |
652 | wrap |= CFQ_RQ2_WRAP; | |
653 | ||
654 | /* Found required data */ | |
655 | ||
656 | /* | |
657 | * By doing switch() on the bit mask "wrap" we avoid having to | |
658 | * check two variables for all permutations: --> faster! | |
659 | */ | |
660 | switch (wrap) { | |
661 | case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ | |
662 | if (d1 < d2) | |
663 | return rq1; | |
664 | else if (d2 < d1) | |
665 | return rq2; | |
666 | else { | |
667 | if (s1 >= s2) | |
668 | return rq1; | |
669 | else | |
670 | return rq2; | |
671 | } | |
672 | ||
673 | case CFQ_RQ2_WRAP: | |
674 | return rq1; | |
675 | case CFQ_RQ1_WRAP: | |
676 | return rq2; | |
677 | case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */ | |
678 | default: | |
679 | /* | |
680 | * Since both rqs are wrapped, | |
681 | * start with the one that's further behind head | |
682 | * (--> only *one* back seek required), | |
683 | * since back seek takes more time than forward. | |
684 | */ | |
685 | if (s1 <= s2) | |
686 | return rq1; | |
687 | else | |
688 | return rq2; | |
689 | } | |
690 | } | |
691 | ||
692 | /* | |
693 | * The below is leftmost cache rbtree addon | |
694 | */ | |
695 | static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root) | |
696 | { | |
697 | /* Service tree is empty */ | |
698 | if (!root->count) | |
699 | return NULL; | |
700 | ||
701 | if (!root->left) | |
702 | root->left = rb_first(&root->rb); | |
703 | ||
704 | if (root->left) | |
705 | return rb_entry(root->left, struct cfq_queue, rb_node); | |
706 | ||
707 | return NULL; | |
708 | } | |
709 | ||
710 | static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root) | |
711 | { | |
712 | if (!root->left) | |
713 | root->left = rb_first(&root->rb); | |
714 | ||
715 | if (root->left) | |
716 | return rb_entry_cfqg(root->left); | |
717 | ||
718 | return NULL; | |
719 | } | |
720 | ||
721 | static void rb_erase_init(struct rb_node *n, struct rb_root *root) | |
722 | { | |
723 | rb_erase(n, root); | |
724 | RB_CLEAR_NODE(n); | |
725 | } | |
726 | ||
727 | static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root) | |
728 | { | |
729 | if (root->left == n) | |
730 | root->left = NULL; | |
731 | rb_erase_init(n, &root->rb); | |
732 | --root->count; | |
733 | } | |
734 | ||
735 | /* | |
736 | * would be nice to take fifo expire time into account as well | |
737 | */ | |
738 | static struct request * | |
739 | cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
740 | struct request *last) | |
741 | { | |
742 | struct rb_node *rbnext = rb_next(&last->rb_node); | |
743 | struct rb_node *rbprev = rb_prev(&last->rb_node); | |
744 | struct request *next = NULL, *prev = NULL; | |
745 | ||
746 | BUG_ON(RB_EMPTY_NODE(&last->rb_node)); | |
747 | ||
748 | if (rbprev) | |
749 | prev = rb_entry_rq(rbprev); | |
750 | ||
751 | if (rbnext) | |
752 | next = rb_entry_rq(rbnext); | |
753 | else { | |
754 | rbnext = rb_first(&cfqq->sort_list); | |
755 | if (rbnext && rbnext != &last->rb_node) | |
756 | next = rb_entry_rq(rbnext); | |
757 | } | |
758 | ||
759 | return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last)); | |
760 | } | |
761 | ||
762 | static unsigned long cfq_slice_offset(struct cfq_data *cfqd, | |
763 | struct cfq_queue *cfqq) | |
764 | { | |
765 | /* | |
766 | * just an approximation, should be ok. | |
767 | */ | |
768 | return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) - | |
769 | cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio)); | |
770 | } | |
771 | ||
772 | static inline s64 | |
773 | cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
774 | { | |
775 | return cfqg->vdisktime - st->min_vdisktime; | |
776 | } | |
777 | ||
778 | static void | |
779 | __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg) | |
780 | { | |
781 | struct rb_node **node = &st->rb.rb_node; | |
782 | struct rb_node *parent = NULL; | |
783 | struct cfq_group *__cfqg; | |
784 | s64 key = cfqg_key(st, cfqg); | |
785 | int left = 1; | |
786 | ||
787 | while (*node != NULL) { | |
788 | parent = *node; | |
789 | __cfqg = rb_entry_cfqg(parent); | |
790 | ||
791 | if (key < cfqg_key(st, __cfqg)) | |
792 | node = &parent->rb_left; | |
793 | else { | |
794 | node = &parent->rb_right; | |
795 | left = 0; | |
796 | } | |
797 | } | |
798 | ||
799 | if (left) | |
800 | st->left = &cfqg->rb_node; | |
801 | ||
802 | rb_link_node(&cfqg->rb_node, parent, node); | |
803 | rb_insert_color(&cfqg->rb_node, &st->rb); | |
804 | } | |
805 | ||
806 | static void | |
807 | cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
808 | { | |
809 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
810 | struct cfq_group *__cfqg; | |
811 | struct rb_node *n; | |
812 | ||
813 | cfqg->nr_cfqq++; | |
814 | if (cfqg->on_st) | |
815 | return; | |
816 | ||
817 | /* | |
818 | * Currently put the group at the end. Later implement something | |
819 | * so that groups get lesser vtime based on their weights, so that | |
820 | * if group does not loose all if it was not continously backlogged. | |
821 | */ | |
822 | n = rb_last(&st->rb); | |
823 | if (n) { | |
824 | __cfqg = rb_entry_cfqg(n); | |
825 | cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY; | |
826 | } else | |
827 | cfqg->vdisktime = st->min_vdisktime; | |
828 | ||
829 | __cfq_group_service_tree_add(st, cfqg); | |
830 | cfqg->on_st = true; | |
831 | cfqd->nr_groups++; | |
832 | st->total_weight += cfqg->weight; | |
833 | } | |
834 | ||
835 | static void | |
836 | cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
837 | { | |
838 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
839 | ||
840 | if (st->active == &cfqg->rb_node) | |
841 | st->active = NULL; | |
842 | ||
843 | BUG_ON(cfqg->nr_cfqq < 1); | |
844 | cfqg->nr_cfqq--; | |
845 | ||
846 | /* If there are other cfq queues under this group, don't delete it */ | |
847 | if (cfqg->nr_cfqq) | |
848 | return; | |
849 | ||
850 | cfq_log_cfqg(cfqd, cfqg, "del_from_rr group"); | |
851 | cfqg->on_st = false; | |
852 | cfqd->nr_groups--; | |
853 | st->total_weight -= cfqg->weight; | |
854 | if (!RB_EMPTY_NODE(&cfqg->rb_node)) | |
855 | cfq_rb_erase(&cfqg->rb_node, st); | |
856 | cfqg->saved_workload_slice = 0; | |
857 | blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1); | |
858 | } | |
859 | ||
860 | static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq) | |
861 | { | |
862 | unsigned int slice_used, allocated_slice; | |
863 | ||
864 | /* | |
865 | * Queue got expired before even a single request completed or | |
866 | * got expired immediately after first request completion. | |
867 | */ | |
868 | if (!cfqq->slice_start || cfqq->slice_start == jiffies) { | |
869 | /* | |
870 | * Also charge the seek time incurred to the group, otherwise | |
871 | * if there are mutiple queues in the group, each can dispatch | |
872 | * a single request on seeky media and cause lots of seek time | |
873 | * and group will never know it. | |
874 | */ | |
875 | slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start), | |
876 | 1); | |
877 | } else { | |
878 | slice_used = jiffies - cfqq->slice_start; | |
879 | allocated_slice = cfqq->slice_end - cfqq->slice_start; | |
880 | if (slice_used > allocated_slice) | |
881 | slice_used = allocated_slice; | |
882 | } | |
883 | ||
884 | cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used, | |
885 | cfqq->nr_sectors); | |
886 | return slice_used; | |
887 | } | |
888 | ||
889 | static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg, | |
890 | struct cfq_queue *cfqq) | |
891 | { | |
892 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
893 | unsigned int used_sl; | |
894 | ||
895 | used_sl = cfq_cfqq_slice_usage(cfqq); | |
896 | ||
897 | /* Can't update vdisktime while group is on service tree */ | |
898 | cfq_rb_erase(&cfqg->rb_node, st); | |
899 | cfqg->vdisktime += cfq_scale_slice(used_sl, cfqg); | |
900 | __cfq_group_service_tree_add(st, cfqg); | |
901 | ||
902 | /* This group is being expired. Save the context */ | |
903 | if (time_after(cfqd->workload_expires, jiffies)) { | |
904 | cfqg->saved_workload_slice = cfqd->workload_expires | |
905 | - jiffies; | |
906 | cfqg->saved_workload = cfqd->serving_type; | |
907 | cfqg->saved_serving_prio = cfqd->serving_prio; | |
908 | } else | |
909 | cfqg->saved_workload_slice = 0; | |
910 | ||
911 | cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime, | |
912 | st->min_vdisktime); | |
913 | blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl, | |
914 | cfqq->nr_sectors); | |
915 | } | |
916 | ||
917 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
918 | static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg) | |
919 | { | |
920 | if (blkg) | |
921 | return container_of(blkg, struct cfq_group, blkg); | |
922 | return NULL; | |
923 | } | |
924 | ||
925 | static struct cfq_group * | |
926 | cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create) | |
927 | { | |
928 | struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup); | |
929 | struct cfq_group *cfqg = NULL; | |
930 | void *key = cfqd; | |
931 | int i, j; | |
932 | struct cfq_rb_root *st; | |
933 | struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info; | |
934 | unsigned int major, minor; | |
935 | ||
936 | /* Do we need to take this reference */ | |
937 | if (!css_tryget(&blkcg->css)) | |
938 | return NULL;; | |
939 | ||
940 | cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key)); | |
941 | if (cfqg || !create) | |
942 | goto done; | |
943 | ||
944 | cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node); | |
945 | if (!cfqg) | |
946 | goto done; | |
947 | ||
948 | cfqg->weight = blkcg->weight; | |
949 | for_each_cfqg_st(cfqg, i, j, st) | |
950 | *st = CFQ_RB_ROOT; | |
951 | RB_CLEAR_NODE(&cfqg->rb_node); | |
952 | ||
953 | /* | |
954 | * Take the initial reference that will be released on destroy | |
955 | * This can be thought of a joint reference by cgroup and | |
956 | * elevator which will be dropped by either elevator exit | |
957 | * or cgroup deletion path depending on who is exiting first. | |
958 | */ | |
959 | atomic_set(&cfqg->ref, 1); | |
960 | ||
961 | /* Add group onto cgroup list */ | |
962 | sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor); | |
963 | blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd, | |
964 | MKDEV(major, minor)); | |
965 | ||
966 | /* Add group on cfqd list */ | |
967 | hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list); | |
968 | ||
969 | done: | |
970 | css_put(&blkcg->css); | |
971 | return cfqg; | |
972 | } | |
973 | ||
974 | /* | |
975 | * Search for the cfq group current task belongs to. If create = 1, then also | |
976 | * create the cfq group if it does not exist. request_queue lock must be held. | |
977 | */ | |
978 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
979 | { | |
980 | struct cgroup *cgroup; | |
981 | struct cfq_group *cfqg = NULL; | |
982 | ||
983 | rcu_read_lock(); | |
984 | cgroup = task_cgroup(current, blkio_subsys_id); | |
985 | cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create); | |
986 | if (!cfqg && create) | |
987 | cfqg = &cfqd->root_group; | |
988 | rcu_read_unlock(); | |
989 | return cfqg; | |
990 | } | |
991 | ||
992 | static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) | |
993 | { | |
994 | /* Currently, all async queues are mapped to root group */ | |
995 | if (!cfq_cfqq_sync(cfqq)) | |
996 | cfqg = &cfqq->cfqd->root_group; | |
997 | ||
998 | cfqq->cfqg = cfqg; | |
999 | /* cfqq reference on cfqg */ | |
1000 | atomic_inc(&cfqq->cfqg->ref); | |
1001 | } | |
1002 | ||
1003 | static void cfq_put_cfqg(struct cfq_group *cfqg) | |
1004 | { | |
1005 | struct cfq_rb_root *st; | |
1006 | int i, j; | |
1007 | ||
1008 | BUG_ON(atomic_read(&cfqg->ref) <= 0); | |
1009 | if (!atomic_dec_and_test(&cfqg->ref)) | |
1010 | return; | |
1011 | for_each_cfqg_st(cfqg, i, j, st) | |
1012 | BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL); | |
1013 | kfree(cfqg); | |
1014 | } | |
1015 | ||
1016 | static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1017 | { | |
1018 | /* Something wrong if we are trying to remove same group twice */ | |
1019 | BUG_ON(hlist_unhashed(&cfqg->cfqd_node)); | |
1020 | ||
1021 | hlist_del_init(&cfqg->cfqd_node); | |
1022 | ||
1023 | /* | |
1024 | * Put the reference taken at the time of creation so that when all | |
1025 | * queues are gone, group can be destroyed. | |
1026 | */ | |
1027 | cfq_put_cfqg(cfqg); | |
1028 | } | |
1029 | ||
1030 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) | |
1031 | { | |
1032 | struct hlist_node *pos, *n; | |
1033 | struct cfq_group *cfqg; | |
1034 | ||
1035 | hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) { | |
1036 | /* | |
1037 | * If cgroup removal path got to blk_group first and removed | |
1038 | * it from cgroup list, then it will take care of destroying | |
1039 | * cfqg also. | |
1040 | */ | |
1041 | if (!blkiocg_del_blkio_group(&cfqg->blkg)) | |
1042 | cfq_destroy_cfqg(cfqd, cfqg); | |
1043 | } | |
1044 | } | |
1045 | ||
1046 | /* | |
1047 | * Blk cgroup controller notification saying that blkio_group object is being | |
1048 | * delinked as associated cgroup object is going away. That also means that | |
1049 | * no new IO will come in this group. So get rid of this group as soon as | |
1050 | * any pending IO in the group is finished. | |
1051 | * | |
1052 | * This function is called under rcu_read_lock(). key is the rcu protected | |
1053 | * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu | |
1054 | * read lock. | |
1055 | * | |
1056 | * "key" was fetched from blkio_group under blkio_cgroup->lock. That means | |
1057 | * it should not be NULL as even if elevator was exiting, cgroup deltion | |
1058 | * path got to it first. | |
1059 | */ | |
1060 | void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg) | |
1061 | { | |
1062 | unsigned long flags; | |
1063 | struct cfq_data *cfqd = key; | |
1064 | ||
1065 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
1066 | cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg)); | |
1067 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
1068 | } | |
1069 | ||
1070 | #else /* GROUP_IOSCHED */ | |
1071 | static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create) | |
1072 | { | |
1073 | return &cfqd->root_group; | |
1074 | } | |
1075 | static inline void | |
1076 | cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) { | |
1077 | cfqq->cfqg = cfqg; | |
1078 | } | |
1079 | ||
1080 | static void cfq_release_cfq_groups(struct cfq_data *cfqd) {} | |
1081 | static inline void cfq_put_cfqg(struct cfq_group *cfqg) {} | |
1082 | ||
1083 | #endif /* GROUP_IOSCHED */ | |
1084 | ||
1085 | /* | |
1086 | * The cfqd->service_trees holds all pending cfq_queue's that have | |
1087 | * requests waiting to be processed. It is sorted in the order that | |
1088 | * we will service the queues. | |
1089 | */ | |
1090 | static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1091 | bool add_front) | |
1092 | { | |
1093 | struct rb_node **p, *parent; | |
1094 | struct cfq_queue *__cfqq; | |
1095 | unsigned long rb_key; | |
1096 | struct cfq_rb_root *service_tree; | |
1097 | int left; | |
1098 | int new_cfqq = 1; | |
1099 | ||
1100 | service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq), | |
1101 | cfqq_type(cfqq), cfqd); | |
1102 | if (cfq_class_idle(cfqq)) { | |
1103 | rb_key = CFQ_IDLE_DELAY; | |
1104 | parent = rb_last(&service_tree->rb); | |
1105 | if (parent && parent != &cfqq->rb_node) { | |
1106 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1107 | rb_key += __cfqq->rb_key; | |
1108 | } else | |
1109 | rb_key += jiffies; | |
1110 | } else if (!add_front) { | |
1111 | /* | |
1112 | * Get our rb key offset. Subtract any residual slice | |
1113 | * value carried from last service. A negative resid | |
1114 | * count indicates slice overrun, and this should position | |
1115 | * the next service time further away in the tree. | |
1116 | */ | |
1117 | rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies; | |
1118 | rb_key -= cfqq->slice_resid; | |
1119 | cfqq->slice_resid = 0; | |
1120 | } else { | |
1121 | rb_key = -HZ; | |
1122 | __cfqq = cfq_rb_first(service_tree); | |
1123 | rb_key += __cfqq ? __cfqq->rb_key : jiffies; | |
1124 | } | |
1125 | ||
1126 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
1127 | new_cfqq = 0; | |
1128 | /* | |
1129 | * same position, nothing more to do | |
1130 | */ | |
1131 | if (rb_key == cfqq->rb_key && | |
1132 | cfqq->service_tree == service_tree) | |
1133 | return; | |
1134 | ||
1135 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1136 | cfqq->service_tree = NULL; | |
1137 | } | |
1138 | ||
1139 | left = 1; | |
1140 | parent = NULL; | |
1141 | cfqq->service_tree = service_tree; | |
1142 | p = &service_tree->rb.rb_node; | |
1143 | while (*p) { | |
1144 | struct rb_node **n; | |
1145 | ||
1146 | parent = *p; | |
1147 | __cfqq = rb_entry(parent, struct cfq_queue, rb_node); | |
1148 | ||
1149 | /* | |
1150 | * sort by key, that represents service time. | |
1151 | */ | |
1152 | if (time_before(rb_key, __cfqq->rb_key)) | |
1153 | n = &(*p)->rb_left; | |
1154 | else { | |
1155 | n = &(*p)->rb_right; | |
1156 | left = 0; | |
1157 | } | |
1158 | ||
1159 | p = n; | |
1160 | } | |
1161 | ||
1162 | if (left) | |
1163 | service_tree->left = &cfqq->rb_node; | |
1164 | ||
1165 | cfqq->rb_key = rb_key; | |
1166 | rb_link_node(&cfqq->rb_node, parent, p); | |
1167 | rb_insert_color(&cfqq->rb_node, &service_tree->rb); | |
1168 | service_tree->count++; | |
1169 | if (add_front || !new_cfqq) | |
1170 | return; | |
1171 | cfq_group_service_tree_add(cfqd, cfqq->cfqg); | |
1172 | } | |
1173 | ||
1174 | static struct cfq_queue * | |
1175 | cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root, | |
1176 | sector_t sector, struct rb_node **ret_parent, | |
1177 | struct rb_node ***rb_link) | |
1178 | { | |
1179 | struct rb_node **p, *parent; | |
1180 | struct cfq_queue *cfqq = NULL; | |
1181 | ||
1182 | parent = NULL; | |
1183 | p = &root->rb_node; | |
1184 | while (*p) { | |
1185 | struct rb_node **n; | |
1186 | ||
1187 | parent = *p; | |
1188 | cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1189 | ||
1190 | /* | |
1191 | * Sort strictly based on sector. Smallest to the left, | |
1192 | * largest to the right. | |
1193 | */ | |
1194 | if (sector > blk_rq_pos(cfqq->next_rq)) | |
1195 | n = &(*p)->rb_right; | |
1196 | else if (sector < blk_rq_pos(cfqq->next_rq)) | |
1197 | n = &(*p)->rb_left; | |
1198 | else | |
1199 | break; | |
1200 | p = n; | |
1201 | cfqq = NULL; | |
1202 | } | |
1203 | ||
1204 | *ret_parent = parent; | |
1205 | if (rb_link) | |
1206 | *rb_link = p; | |
1207 | return cfqq; | |
1208 | } | |
1209 | ||
1210 | static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1211 | { | |
1212 | struct rb_node **p, *parent; | |
1213 | struct cfq_queue *__cfqq; | |
1214 | ||
1215 | if (cfqq->p_root) { | |
1216 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1217 | cfqq->p_root = NULL; | |
1218 | } | |
1219 | ||
1220 | if (cfq_class_idle(cfqq)) | |
1221 | return; | |
1222 | if (!cfqq->next_rq) | |
1223 | return; | |
1224 | ||
1225 | cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio]; | |
1226 | __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root, | |
1227 | blk_rq_pos(cfqq->next_rq), &parent, &p); | |
1228 | if (!__cfqq) { | |
1229 | rb_link_node(&cfqq->p_node, parent, p); | |
1230 | rb_insert_color(&cfqq->p_node, cfqq->p_root); | |
1231 | } else | |
1232 | cfqq->p_root = NULL; | |
1233 | } | |
1234 | ||
1235 | /* | |
1236 | * Update cfqq's position in the service tree. | |
1237 | */ | |
1238 | static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1239 | { | |
1240 | /* | |
1241 | * Resorting requires the cfqq to be on the RR list already. | |
1242 | */ | |
1243 | if (cfq_cfqq_on_rr(cfqq)) { | |
1244 | cfq_service_tree_add(cfqd, cfqq, 0); | |
1245 | cfq_prio_tree_add(cfqd, cfqq); | |
1246 | } | |
1247 | } | |
1248 | ||
1249 | /* | |
1250 | * add to busy list of queues for service, trying to be fair in ordering | |
1251 | * the pending list according to last request service | |
1252 | */ | |
1253 | static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1254 | { | |
1255 | cfq_log_cfqq(cfqd, cfqq, "add_to_rr"); | |
1256 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
1257 | cfq_mark_cfqq_on_rr(cfqq); | |
1258 | cfqd->busy_queues++; | |
1259 | ||
1260 | cfq_resort_rr_list(cfqd, cfqq); | |
1261 | } | |
1262 | ||
1263 | /* | |
1264 | * Called when the cfqq no longer has requests pending, remove it from | |
1265 | * the service tree. | |
1266 | */ | |
1267 | static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1268 | { | |
1269 | cfq_log_cfqq(cfqd, cfqq, "del_from_rr"); | |
1270 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
1271 | cfq_clear_cfqq_on_rr(cfqq); | |
1272 | ||
1273 | if (!RB_EMPTY_NODE(&cfqq->rb_node)) { | |
1274 | cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree); | |
1275 | cfqq->service_tree = NULL; | |
1276 | } | |
1277 | if (cfqq->p_root) { | |
1278 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1279 | cfqq->p_root = NULL; | |
1280 | } | |
1281 | ||
1282 | cfq_group_service_tree_del(cfqd, cfqq->cfqg); | |
1283 | BUG_ON(!cfqd->busy_queues); | |
1284 | cfqd->busy_queues--; | |
1285 | } | |
1286 | ||
1287 | /* | |
1288 | * rb tree support functions | |
1289 | */ | |
1290 | static void cfq_del_rq_rb(struct request *rq) | |
1291 | { | |
1292 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1293 | const int sync = rq_is_sync(rq); | |
1294 | ||
1295 | BUG_ON(!cfqq->queued[sync]); | |
1296 | cfqq->queued[sync]--; | |
1297 | ||
1298 | elv_rb_del(&cfqq->sort_list, rq); | |
1299 | ||
1300 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) { | |
1301 | /* | |
1302 | * Queue will be deleted from service tree when we actually | |
1303 | * expire it later. Right now just remove it from prio tree | |
1304 | * as it is empty. | |
1305 | */ | |
1306 | if (cfqq->p_root) { | |
1307 | rb_erase(&cfqq->p_node, cfqq->p_root); | |
1308 | cfqq->p_root = NULL; | |
1309 | } | |
1310 | } | |
1311 | } | |
1312 | ||
1313 | static void cfq_add_rq_rb(struct request *rq) | |
1314 | { | |
1315 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1316 | struct cfq_data *cfqd = cfqq->cfqd; | |
1317 | struct request *__alias, *prev; | |
1318 | ||
1319 | cfqq->queued[rq_is_sync(rq)]++; | |
1320 | ||
1321 | /* | |
1322 | * looks a little odd, but the first insert might return an alias. | |
1323 | * if that happens, put the alias on the dispatch list | |
1324 | */ | |
1325 | while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL) | |
1326 | cfq_dispatch_insert(cfqd->queue, __alias); | |
1327 | ||
1328 | if (!cfq_cfqq_on_rr(cfqq)) | |
1329 | cfq_add_cfqq_rr(cfqd, cfqq); | |
1330 | ||
1331 | /* | |
1332 | * check if this request is a better next-serve candidate | |
1333 | */ | |
1334 | prev = cfqq->next_rq; | |
1335 | cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position); | |
1336 | ||
1337 | /* | |
1338 | * adjust priority tree position, if ->next_rq changes | |
1339 | */ | |
1340 | if (prev != cfqq->next_rq) | |
1341 | cfq_prio_tree_add(cfqd, cfqq); | |
1342 | ||
1343 | BUG_ON(!cfqq->next_rq); | |
1344 | } | |
1345 | ||
1346 | static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq) | |
1347 | { | |
1348 | elv_rb_del(&cfqq->sort_list, rq); | |
1349 | cfqq->queued[rq_is_sync(rq)]--; | |
1350 | cfq_add_rq_rb(rq); | |
1351 | } | |
1352 | ||
1353 | static struct request * | |
1354 | cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio) | |
1355 | { | |
1356 | struct task_struct *tsk = current; | |
1357 | struct cfq_io_context *cic; | |
1358 | struct cfq_queue *cfqq; | |
1359 | ||
1360 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
1361 | if (!cic) | |
1362 | return NULL; | |
1363 | ||
1364 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
1365 | if (cfqq) { | |
1366 | sector_t sector = bio->bi_sector + bio_sectors(bio); | |
1367 | ||
1368 | return elv_rb_find(&cfqq->sort_list, sector); | |
1369 | } | |
1370 | ||
1371 | return NULL; | |
1372 | } | |
1373 | ||
1374 | static void cfq_activate_request(struct request_queue *q, struct request *rq) | |
1375 | { | |
1376 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1377 | ||
1378 | cfqd->rq_in_driver[rq_is_sync(rq)]++; | |
1379 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d", | |
1380 | rq_in_driver(cfqd)); | |
1381 | ||
1382 | cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
1383 | } | |
1384 | ||
1385 | static void cfq_deactivate_request(struct request_queue *q, struct request *rq) | |
1386 | { | |
1387 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1388 | const int sync = rq_is_sync(rq); | |
1389 | ||
1390 | WARN_ON(!cfqd->rq_in_driver[sync]); | |
1391 | cfqd->rq_in_driver[sync]--; | |
1392 | cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d", | |
1393 | rq_in_driver(cfqd)); | |
1394 | } | |
1395 | ||
1396 | static void cfq_remove_request(struct request *rq) | |
1397 | { | |
1398 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1399 | ||
1400 | if (cfqq->next_rq == rq) | |
1401 | cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq); | |
1402 | ||
1403 | list_del_init(&rq->queuelist); | |
1404 | cfq_del_rq_rb(rq); | |
1405 | ||
1406 | cfqq->cfqd->rq_queued--; | |
1407 | if (rq_is_meta(rq)) { | |
1408 | WARN_ON(!cfqq->meta_pending); | |
1409 | cfqq->meta_pending--; | |
1410 | } | |
1411 | } | |
1412 | ||
1413 | static int cfq_merge(struct request_queue *q, struct request **req, | |
1414 | struct bio *bio) | |
1415 | { | |
1416 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1417 | struct request *__rq; | |
1418 | ||
1419 | __rq = cfq_find_rq_fmerge(cfqd, bio); | |
1420 | if (__rq && elv_rq_merge_ok(__rq, bio)) { | |
1421 | *req = __rq; | |
1422 | return ELEVATOR_FRONT_MERGE; | |
1423 | } | |
1424 | ||
1425 | return ELEVATOR_NO_MERGE; | |
1426 | } | |
1427 | ||
1428 | static void cfq_merged_request(struct request_queue *q, struct request *req, | |
1429 | int type) | |
1430 | { | |
1431 | if (type == ELEVATOR_FRONT_MERGE) { | |
1432 | struct cfq_queue *cfqq = RQ_CFQQ(req); | |
1433 | ||
1434 | cfq_reposition_rq_rb(cfqq, req); | |
1435 | } | |
1436 | } | |
1437 | ||
1438 | static void | |
1439 | cfq_merged_requests(struct request_queue *q, struct request *rq, | |
1440 | struct request *next) | |
1441 | { | |
1442 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1443 | /* | |
1444 | * reposition in fifo if next is older than rq | |
1445 | */ | |
1446 | if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && | |
1447 | time_before(rq_fifo_time(next), rq_fifo_time(rq))) { | |
1448 | list_move(&rq->queuelist, &next->queuelist); | |
1449 | rq_set_fifo_time(rq, rq_fifo_time(next)); | |
1450 | } | |
1451 | ||
1452 | if (cfqq->next_rq == next) | |
1453 | cfqq->next_rq = rq; | |
1454 | cfq_remove_request(next); | |
1455 | } | |
1456 | ||
1457 | static int cfq_allow_merge(struct request_queue *q, struct request *rq, | |
1458 | struct bio *bio) | |
1459 | { | |
1460 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1461 | struct cfq_io_context *cic; | |
1462 | struct cfq_queue *cfqq; | |
1463 | ||
1464 | /* | |
1465 | * Disallow merge of a sync bio into an async request. | |
1466 | */ | |
1467 | if (cfq_bio_sync(bio) && !rq_is_sync(rq)) | |
1468 | return false; | |
1469 | ||
1470 | /* | |
1471 | * Lookup the cfqq that this bio will be queued with. Allow | |
1472 | * merge only if rq is queued there. | |
1473 | */ | |
1474 | cic = cfq_cic_lookup(cfqd, current->io_context); | |
1475 | if (!cic) | |
1476 | return false; | |
1477 | ||
1478 | cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio)); | |
1479 | return cfqq == RQ_CFQQ(rq); | |
1480 | } | |
1481 | ||
1482 | static void __cfq_set_active_queue(struct cfq_data *cfqd, | |
1483 | struct cfq_queue *cfqq) | |
1484 | { | |
1485 | if (cfqq) { | |
1486 | cfq_log_cfqq(cfqd, cfqq, "set_active"); | |
1487 | cfqq->slice_start = 0; | |
1488 | cfqq->dispatch_start = jiffies; | |
1489 | cfqq->slice_end = 0; | |
1490 | cfqq->slice_dispatch = 0; | |
1491 | cfqq->nr_sectors = 0; | |
1492 | ||
1493 | cfq_clear_cfqq_wait_request(cfqq); | |
1494 | cfq_clear_cfqq_must_dispatch(cfqq); | |
1495 | cfq_clear_cfqq_must_alloc_slice(cfqq); | |
1496 | cfq_clear_cfqq_fifo_expire(cfqq); | |
1497 | cfq_mark_cfqq_slice_new(cfqq); | |
1498 | ||
1499 | del_timer(&cfqd->idle_slice_timer); | |
1500 | } | |
1501 | ||
1502 | cfqd->active_queue = cfqq; | |
1503 | } | |
1504 | ||
1505 | /* | |
1506 | * current cfqq expired its slice (or was too idle), select new one | |
1507 | */ | |
1508 | static void | |
1509 | __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1510 | bool timed_out) | |
1511 | { | |
1512 | cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out); | |
1513 | ||
1514 | if (cfq_cfqq_wait_request(cfqq)) | |
1515 | del_timer(&cfqd->idle_slice_timer); | |
1516 | ||
1517 | cfq_clear_cfqq_wait_request(cfqq); | |
1518 | ||
1519 | /* | |
1520 | * store what was left of this slice, if the queue idled/timed out | |
1521 | */ | |
1522 | if (timed_out && !cfq_cfqq_slice_new(cfqq)) { | |
1523 | cfqq->slice_resid = cfqq->slice_end - jiffies; | |
1524 | cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid); | |
1525 | } | |
1526 | ||
1527 | cfq_group_served(cfqd, cfqq->cfqg, cfqq); | |
1528 | ||
1529 | if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) | |
1530 | cfq_del_cfqq_rr(cfqd, cfqq); | |
1531 | ||
1532 | cfq_resort_rr_list(cfqd, cfqq); | |
1533 | ||
1534 | if (cfqq == cfqd->active_queue) | |
1535 | cfqd->active_queue = NULL; | |
1536 | ||
1537 | if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active) | |
1538 | cfqd->grp_service_tree.active = NULL; | |
1539 | ||
1540 | if (cfqd->active_cic) { | |
1541 | put_io_context(cfqd->active_cic->ioc); | |
1542 | cfqd->active_cic = NULL; | |
1543 | } | |
1544 | } | |
1545 | ||
1546 | static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out) | |
1547 | { | |
1548 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1549 | ||
1550 | if (cfqq) | |
1551 | __cfq_slice_expired(cfqd, cfqq, timed_out); | |
1552 | } | |
1553 | ||
1554 | /* | |
1555 | * Get next queue for service. Unless we have a queue preemption, | |
1556 | * we'll simply select the first cfqq in the service tree. | |
1557 | */ | |
1558 | static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd) | |
1559 | { | |
1560 | struct cfq_rb_root *service_tree = | |
1561 | service_tree_for(cfqd->serving_group, cfqd->serving_prio, | |
1562 | cfqd->serving_type, cfqd); | |
1563 | ||
1564 | if (!cfqd->rq_queued) | |
1565 | return NULL; | |
1566 | ||
1567 | /* There is nothing to dispatch */ | |
1568 | if (!service_tree) | |
1569 | return NULL; | |
1570 | if (RB_EMPTY_ROOT(&service_tree->rb)) | |
1571 | return NULL; | |
1572 | return cfq_rb_first(service_tree); | |
1573 | } | |
1574 | ||
1575 | static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd) | |
1576 | { | |
1577 | struct cfq_group *cfqg; | |
1578 | struct cfq_queue *cfqq; | |
1579 | int i, j; | |
1580 | struct cfq_rb_root *st; | |
1581 | ||
1582 | if (!cfqd->rq_queued) | |
1583 | return NULL; | |
1584 | ||
1585 | cfqg = cfq_get_next_cfqg(cfqd); | |
1586 | if (!cfqg) | |
1587 | return NULL; | |
1588 | ||
1589 | for_each_cfqg_st(cfqg, i, j, st) | |
1590 | if ((cfqq = cfq_rb_first(st)) != NULL) | |
1591 | return cfqq; | |
1592 | return NULL; | |
1593 | } | |
1594 | ||
1595 | /* | |
1596 | * Get and set a new active queue for service. | |
1597 | */ | |
1598 | static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd, | |
1599 | struct cfq_queue *cfqq) | |
1600 | { | |
1601 | if (!cfqq) | |
1602 | cfqq = cfq_get_next_queue(cfqd); | |
1603 | ||
1604 | __cfq_set_active_queue(cfqd, cfqq); | |
1605 | return cfqq; | |
1606 | } | |
1607 | ||
1608 | static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd, | |
1609 | struct request *rq) | |
1610 | { | |
1611 | if (blk_rq_pos(rq) >= cfqd->last_position) | |
1612 | return blk_rq_pos(rq) - cfqd->last_position; | |
1613 | else | |
1614 | return cfqd->last_position - blk_rq_pos(rq); | |
1615 | } | |
1616 | ||
1617 | #define CFQQ_SEEK_THR 8 * 1024 | |
1618 | #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR) | |
1619 | ||
1620 | static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
1621 | struct request *rq) | |
1622 | { | |
1623 | sector_t sdist = cfqq->seek_mean; | |
1624 | ||
1625 | if (!sample_valid(cfqq->seek_samples)) | |
1626 | sdist = CFQQ_SEEK_THR; | |
1627 | ||
1628 | return cfq_dist_from_last(cfqd, rq) <= sdist; | |
1629 | } | |
1630 | ||
1631 | static struct cfq_queue *cfqq_close(struct cfq_data *cfqd, | |
1632 | struct cfq_queue *cur_cfqq) | |
1633 | { | |
1634 | struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio]; | |
1635 | struct rb_node *parent, *node; | |
1636 | struct cfq_queue *__cfqq; | |
1637 | sector_t sector = cfqd->last_position; | |
1638 | ||
1639 | if (RB_EMPTY_ROOT(root)) | |
1640 | return NULL; | |
1641 | ||
1642 | /* | |
1643 | * First, if we find a request starting at the end of the last | |
1644 | * request, choose it. | |
1645 | */ | |
1646 | __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL); | |
1647 | if (__cfqq) | |
1648 | return __cfqq; | |
1649 | ||
1650 | /* | |
1651 | * If the exact sector wasn't found, the parent of the NULL leaf | |
1652 | * will contain the closest sector. | |
1653 | */ | |
1654 | __cfqq = rb_entry(parent, struct cfq_queue, p_node); | |
1655 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
1656 | return __cfqq; | |
1657 | ||
1658 | if (blk_rq_pos(__cfqq->next_rq) < sector) | |
1659 | node = rb_next(&__cfqq->p_node); | |
1660 | else | |
1661 | node = rb_prev(&__cfqq->p_node); | |
1662 | if (!node) | |
1663 | return NULL; | |
1664 | ||
1665 | __cfqq = rb_entry(node, struct cfq_queue, p_node); | |
1666 | if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq)) | |
1667 | return __cfqq; | |
1668 | ||
1669 | return NULL; | |
1670 | } | |
1671 | ||
1672 | /* | |
1673 | * cfqd - obvious | |
1674 | * cur_cfqq - passed in so that we don't decide that the current queue is | |
1675 | * closely cooperating with itself. | |
1676 | * | |
1677 | * So, basically we're assuming that that cur_cfqq has dispatched at least | |
1678 | * one request, and that cfqd->last_position reflects a position on the disk | |
1679 | * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid | |
1680 | * assumption. | |
1681 | */ | |
1682 | static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd, | |
1683 | struct cfq_queue *cur_cfqq) | |
1684 | { | |
1685 | struct cfq_queue *cfqq; | |
1686 | ||
1687 | if (!cfq_cfqq_sync(cur_cfqq)) | |
1688 | return NULL; | |
1689 | if (CFQQ_SEEKY(cur_cfqq)) | |
1690 | return NULL; | |
1691 | ||
1692 | /* | |
1693 | * We should notice if some of the queues are cooperating, eg | |
1694 | * working closely on the same area of the disk. In that case, | |
1695 | * we can group them together and don't waste time idling. | |
1696 | */ | |
1697 | cfqq = cfqq_close(cfqd, cur_cfqq); | |
1698 | if (!cfqq) | |
1699 | return NULL; | |
1700 | ||
1701 | /* | |
1702 | * It only makes sense to merge sync queues. | |
1703 | */ | |
1704 | if (!cfq_cfqq_sync(cfqq)) | |
1705 | return NULL; | |
1706 | if (CFQQ_SEEKY(cfqq)) | |
1707 | return NULL; | |
1708 | ||
1709 | /* | |
1710 | * Do not merge queues of different priority classes | |
1711 | */ | |
1712 | if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq)) | |
1713 | return NULL; | |
1714 | ||
1715 | return cfqq; | |
1716 | } | |
1717 | ||
1718 | /* | |
1719 | * Determine whether we should enforce idle window for this queue. | |
1720 | */ | |
1721 | ||
1722 | static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1723 | { | |
1724 | enum wl_prio_t prio = cfqq_prio(cfqq); | |
1725 | struct cfq_rb_root *service_tree = cfqq->service_tree; | |
1726 | ||
1727 | BUG_ON(!service_tree); | |
1728 | BUG_ON(!service_tree->count); | |
1729 | ||
1730 | /* We never do for idle class queues. */ | |
1731 | if (prio == IDLE_WORKLOAD) | |
1732 | return false; | |
1733 | ||
1734 | /* We do for queues that were marked with idle window flag. */ | |
1735 | if (cfq_cfqq_idle_window(cfqq)) | |
1736 | return true; | |
1737 | ||
1738 | /* | |
1739 | * Otherwise, we do only if they are the last ones | |
1740 | * in their service tree. | |
1741 | */ | |
1742 | return service_tree->count == 1; | |
1743 | } | |
1744 | ||
1745 | static void cfq_arm_slice_timer(struct cfq_data *cfqd) | |
1746 | { | |
1747 | struct cfq_queue *cfqq = cfqd->active_queue; | |
1748 | struct cfq_io_context *cic; | |
1749 | unsigned long sl; | |
1750 | ||
1751 | /* | |
1752 | * SSD device without seek penalty, disable idling. But only do so | |
1753 | * for devices that support queuing, otherwise we still have a problem | |
1754 | * with sync vs async workloads. | |
1755 | */ | |
1756 | if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag) | |
1757 | return; | |
1758 | ||
1759 | WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list)); | |
1760 | WARN_ON(cfq_cfqq_slice_new(cfqq)); | |
1761 | ||
1762 | /* | |
1763 | * idle is disabled, either manually or by past process history | |
1764 | */ | |
1765 | if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq)) | |
1766 | return; | |
1767 | ||
1768 | /* | |
1769 | * still active requests from this queue, don't idle | |
1770 | */ | |
1771 | if (cfqq->dispatched) | |
1772 | return; | |
1773 | ||
1774 | /* | |
1775 | * task has exited, don't wait | |
1776 | */ | |
1777 | cic = cfqd->active_cic; | |
1778 | if (!cic || !atomic_read(&cic->ioc->nr_tasks)) | |
1779 | return; | |
1780 | ||
1781 | /* | |
1782 | * If our average think time is larger than the remaining time | |
1783 | * slice, then don't idle. This avoids overrunning the allotted | |
1784 | * time slice. | |
1785 | */ | |
1786 | if (sample_valid(cic->ttime_samples) && | |
1787 | (cfqq->slice_end - jiffies < cic->ttime_mean)) | |
1788 | return; | |
1789 | ||
1790 | cfq_mark_cfqq_wait_request(cfqq); | |
1791 | ||
1792 | sl = cfqd->cfq_slice_idle; | |
1793 | ||
1794 | mod_timer(&cfqd->idle_slice_timer, jiffies + sl); | |
1795 | cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl); | |
1796 | } | |
1797 | ||
1798 | /* | |
1799 | * Move request from internal lists to the request queue dispatch list. | |
1800 | */ | |
1801 | static void cfq_dispatch_insert(struct request_queue *q, struct request *rq) | |
1802 | { | |
1803 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
1804 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
1805 | ||
1806 | cfq_log_cfqq(cfqd, cfqq, "dispatch_insert"); | |
1807 | ||
1808 | cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq); | |
1809 | cfq_remove_request(rq); | |
1810 | cfqq->dispatched++; | |
1811 | elv_dispatch_sort(q, rq); | |
1812 | ||
1813 | if (cfq_cfqq_sync(cfqq)) | |
1814 | cfqd->sync_flight++; | |
1815 | cfqq->nr_sectors += blk_rq_sectors(rq); | |
1816 | } | |
1817 | ||
1818 | /* | |
1819 | * return expired entry, or NULL to just start from scratch in rbtree | |
1820 | */ | |
1821 | static struct request *cfq_check_fifo(struct cfq_queue *cfqq) | |
1822 | { | |
1823 | struct request *rq = NULL; | |
1824 | ||
1825 | if (cfq_cfqq_fifo_expire(cfqq)) | |
1826 | return NULL; | |
1827 | ||
1828 | cfq_mark_cfqq_fifo_expire(cfqq); | |
1829 | ||
1830 | if (list_empty(&cfqq->fifo)) | |
1831 | return NULL; | |
1832 | ||
1833 | rq = rq_entry_fifo(cfqq->fifo.next); | |
1834 | if (time_before(jiffies, rq_fifo_time(rq))) | |
1835 | rq = NULL; | |
1836 | ||
1837 | cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq); | |
1838 | return rq; | |
1839 | } | |
1840 | ||
1841 | static inline int | |
1842 | cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
1843 | { | |
1844 | const int base_rq = cfqd->cfq_slice_async_rq; | |
1845 | ||
1846 | WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR); | |
1847 | ||
1848 | return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio)); | |
1849 | } | |
1850 | ||
1851 | /* | |
1852 | * Must be called with the queue_lock held. | |
1853 | */ | |
1854 | static int cfqq_process_refs(struct cfq_queue *cfqq) | |
1855 | { | |
1856 | int process_refs, io_refs; | |
1857 | ||
1858 | io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE]; | |
1859 | process_refs = atomic_read(&cfqq->ref) - io_refs; | |
1860 | BUG_ON(process_refs < 0); | |
1861 | return process_refs; | |
1862 | } | |
1863 | ||
1864 | static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq) | |
1865 | { | |
1866 | int process_refs, new_process_refs; | |
1867 | struct cfq_queue *__cfqq; | |
1868 | ||
1869 | /* Avoid a circular list and skip interim queue merges */ | |
1870 | while ((__cfqq = new_cfqq->new_cfqq)) { | |
1871 | if (__cfqq == cfqq) | |
1872 | return; | |
1873 | new_cfqq = __cfqq; | |
1874 | } | |
1875 | ||
1876 | process_refs = cfqq_process_refs(cfqq); | |
1877 | /* | |
1878 | * If the process for the cfqq has gone away, there is no | |
1879 | * sense in merging the queues. | |
1880 | */ | |
1881 | if (process_refs == 0) | |
1882 | return; | |
1883 | ||
1884 | /* | |
1885 | * Merge in the direction of the lesser amount of work. | |
1886 | */ | |
1887 | new_process_refs = cfqq_process_refs(new_cfqq); | |
1888 | if (new_process_refs >= process_refs) { | |
1889 | cfqq->new_cfqq = new_cfqq; | |
1890 | atomic_add(process_refs, &new_cfqq->ref); | |
1891 | } else { | |
1892 | new_cfqq->new_cfqq = cfqq; | |
1893 | atomic_add(new_process_refs, &cfqq->ref); | |
1894 | } | |
1895 | } | |
1896 | ||
1897 | static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, | |
1898 | struct cfq_group *cfqg, enum wl_prio_t prio, | |
1899 | bool prio_changed) | |
1900 | { | |
1901 | struct cfq_queue *queue; | |
1902 | int i; | |
1903 | bool key_valid = false; | |
1904 | unsigned long lowest_key = 0; | |
1905 | enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD; | |
1906 | ||
1907 | if (prio_changed) { | |
1908 | /* | |
1909 | * When priorities switched, we prefer starting | |
1910 | * from SYNC_NOIDLE (first choice), or just SYNC | |
1911 | * over ASYNC | |
1912 | */ | |
1913 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) | |
1914 | return cur_best; | |
1915 | cur_best = SYNC_WORKLOAD; | |
1916 | if (service_tree_for(cfqg, prio, cur_best, cfqd)->count) | |
1917 | return cur_best; | |
1918 | ||
1919 | return ASYNC_WORKLOAD; | |
1920 | } | |
1921 | ||
1922 | for (i = 0; i < 3; ++i) { | |
1923 | /* otherwise, select the one with lowest rb_key */ | |
1924 | queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd)); | |
1925 | if (queue && | |
1926 | (!key_valid || time_before(queue->rb_key, lowest_key))) { | |
1927 | lowest_key = queue->rb_key; | |
1928 | cur_best = i; | |
1929 | key_valid = true; | |
1930 | } | |
1931 | } | |
1932 | ||
1933 | return cur_best; | |
1934 | } | |
1935 | ||
1936 | static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg) | |
1937 | { | |
1938 | enum wl_prio_t previous_prio = cfqd->serving_prio; | |
1939 | bool prio_changed; | |
1940 | unsigned slice; | |
1941 | unsigned count; | |
1942 | struct cfq_rb_root *st; | |
1943 | unsigned group_slice; | |
1944 | ||
1945 | if (!cfqg) { | |
1946 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1947 | cfqd->workload_expires = jiffies + 1; | |
1948 | return; | |
1949 | } | |
1950 | ||
1951 | /* Choose next priority. RT > BE > IDLE */ | |
1952 | if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg)) | |
1953 | cfqd->serving_prio = RT_WORKLOAD; | |
1954 | else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg)) | |
1955 | cfqd->serving_prio = BE_WORKLOAD; | |
1956 | else { | |
1957 | cfqd->serving_prio = IDLE_WORKLOAD; | |
1958 | cfqd->workload_expires = jiffies + 1; | |
1959 | return; | |
1960 | } | |
1961 | ||
1962 | /* | |
1963 | * For RT and BE, we have to choose also the type | |
1964 | * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload | |
1965 | * expiration time | |
1966 | */ | |
1967 | prio_changed = (cfqd->serving_prio != previous_prio); | |
1968 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
1969 | cfqd); | |
1970 | count = st->count; | |
1971 | ||
1972 | /* | |
1973 | * If priority didn't change, check workload expiration, | |
1974 | * and that we still have other queues ready | |
1975 | */ | |
1976 | if (!prio_changed && count && | |
1977 | !time_after(jiffies, cfqd->workload_expires)) | |
1978 | return; | |
1979 | ||
1980 | /* otherwise select new workload type */ | |
1981 | cfqd->serving_type = | |
1982 | cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed); | |
1983 | st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type, | |
1984 | cfqd); | |
1985 | count = st->count; | |
1986 | ||
1987 | /* | |
1988 | * the workload slice is computed as a fraction of target latency | |
1989 | * proportional to the number of queues in that workload, over | |
1990 | * all the queues in the same priority class | |
1991 | */ | |
1992 | group_slice = cfq_group_slice(cfqd, cfqg); | |
1993 | ||
1994 | slice = group_slice * count / | |
1995 | max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio], | |
1996 | cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg)); | |
1997 | ||
1998 | if (cfqd->serving_type == ASYNC_WORKLOAD) | |
1999 | /* async workload slice is scaled down according to | |
2000 | * the sync/async slice ratio. */ | |
2001 | slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1]; | |
2002 | else | |
2003 | /* sync workload slice is at least 2 * cfq_slice_idle */ | |
2004 | slice = max(slice, 2 * cfqd->cfq_slice_idle); | |
2005 | ||
2006 | slice = max_t(unsigned, slice, CFQ_MIN_TT); | |
2007 | cfqd->workload_expires = jiffies + slice; | |
2008 | cfqd->noidle_tree_requires_idle = false; | |
2009 | } | |
2010 | ||
2011 | static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd) | |
2012 | { | |
2013 | struct cfq_rb_root *st = &cfqd->grp_service_tree; | |
2014 | struct cfq_group *cfqg; | |
2015 | ||
2016 | if (RB_EMPTY_ROOT(&st->rb)) | |
2017 | return NULL; | |
2018 | cfqg = cfq_rb_first_group(st); | |
2019 | st->active = &cfqg->rb_node; | |
2020 | update_min_vdisktime(st); | |
2021 | return cfqg; | |
2022 | } | |
2023 | ||
2024 | static void cfq_choose_cfqg(struct cfq_data *cfqd) | |
2025 | { | |
2026 | struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd); | |
2027 | ||
2028 | cfqd->serving_group = cfqg; | |
2029 | ||
2030 | /* Restore the workload type data */ | |
2031 | if (cfqg->saved_workload_slice) { | |
2032 | cfqd->workload_expires = jiffies + cfqg->saved_workload_slice; | |
2033 | cfqd->serving_type = cfqg->saved_workload; | |
2034 | cfqd->serving_prio = cfqg->saved_serving_prio; | |
2035 | } | |
2036 | choose_service_tree(cfqd, cfqg); | |
2037 | } | |
2038 | ||
2039 | /* | |
2040 | * Select a queue for service. If we have a current active queue, | |
2041 | * check whether to continue servicing it, or retrieve and set a new one. | |
2042 | */ | |
2043 | static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd) | |
2044 | { | |
2045 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
2046 | ||
2047 | cfqq = cfqd->active_queue; | |
2048 | if (!cfqq) | |
2049 | goto new_queue; | |
2050 | ||
2051 | if (!cfqd->rq_queued) | |
2052 | return NULL; | |
2053 | /* | |
2054 | * The active queue has run out of time, expire it and select new. | |
2055 | */ | |
2056 | if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) | |
2057 | goto expire; | |
2058 | ||
2059 | /* | |
2060 | * The active queue has requests and isn't expired, allow it to | |
2061 | * dispatch. | |
2062 | */ | |
2063 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2064 | goto keep_queue; | |
2065 | ||
2066 | /* | |
2067 | * If another queue has a request waiting within our mean seek | |
2068 | * distance, let it run. The expire code will check for close | |
2069 | * cooperators and put the close queue at the front of the service | |
2070 | * tree. If possible, merge the expiring queue with the new cfqq. | |
2071 | */ | |
2072 | new_cfqq = cfq_close_cooperator(cfqd, cfqq); | |
2073 | if (new_cfqq) { | |
2074 | if (!cfqq->new_cfqq) | |
2075 | cfq_setup_merge(cfqq, new_cfqq); | |
2076 | goto expire; | |
2077 | } | |
2078 | ||
2079 | /* | |
2080 | * No requests pending. If the active queue still has requests in | |
2081 | * flight or is idling for a new request, allow either of these | |
2082 | * conditions to happen (or time out) before selecting a new queue. | |
2083 | */ | |
2084 | if (timer_pending(&cfqd->idle_slice_timer) || | |
2085 | (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) { | |
2086 | cfqq = NULL; | |
2087 | goto keep_queue; | |
2088 | } | |
2089 | ||
2090 | expire: | |
2091 | cfq_slice_expired(cfqd, 0); | |
2092 | new_queue: | |
2093 | /* | |
2094 | * Current queue expired. Check if we have to switch to a new | |
2095 | * service tree | |
2096 | */ | |
2097 | if (!new_cfqq) | |
2098 | cfq_choose_cfqg(cfqd); | |
2099 | ||
2100 | cfqq = cfq_set_active_queue(cfqd, new_cfqq); | |
2101 | keep_queue: | |
2102 | return cfqq; | |
2103 | } | |
2104 | ||
2105 | static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq) | |
2106 | { | |
2107 | int dispatched = 0; | |
2108 | ||
2109 | while (cfqq->next_rq) { | |
2110 | cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq); | |
2111 | dispatched++; | |
2112 | } | |
2113 | ||
2114 | BUG_ON(!list_empty(&cfqq->fifo)); | |
2115 | ||
2116 | /* By default cfqq is not expired if it is empty. Do it explicitly */ | |
2117 | __cfq_slice_expired(cfqq->cfqd, cfqq, 0); | |
2118 | return dispatched; | |
2119 | } | |
2120 | ||
2121 | /* | |
2122 | * Drain our current requests. Used for barriers and when switching | |
2123 | * io schedulers on-the-fly. | |
2124 | */ | |
2125 | static int cfq_forced_dispatch(struct cfq_data *cfqd) | |
2126 | { | |
2127 | struct cfq_queue *cfqq; | |
2128 | int dispatched = 0; | |
2129 | ||
2130 | while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) | |
2131 | dispatched += __cfq_forced_dispatch_cfqq(cfqq); | |
2132 | ||
2133 | cfq_slice_expired(cfqd, 0); | |
2134 | BUG_ON(cfqd->busy_queues); | |
2135 | ||
2136 | cfq_log(cfqd, "forced_dispatch=%d", dispatched); | |
2137 | return dispatched; | |
2138 | } | |
2139 | ||
2140 | static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2141 | { | |
2142 | unsigned int max_dispatch; | |
2143 | ||
2144 | /* | |
2145 | * Drain async requests before we start sync IO | |
2146 | */ | |
2147 | if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC]) | |
2148 | return false; | |
2149 | ||
2150 | /* | |
2151 | * If this is an async queue and we have sync IO in flight, let it wait | |
2152 | */ | |
2153 | if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq)) | |
2154 | return false; | |
2155 | ||
2156 | max_dispatch = cfqd->cfq_quantum; | |
2157 | if (cfq_class_idle(cfqq)) | |
2158 | max_dispatch = 1; | |
2159 | ||
2160 | /* | |
2161 | * Does this cfqq already have too much IO in flight? | |
2162 | */ | |
2163 | if (cfqq->dispatched >= max_dispatch) { | |
2164 | /* | |
2165 | * idle queue must always only have a single IO in flight | |
2166 | */ | |
2167 | if (cfq_class_idle(cfqq)) | |
2168 | return false; | |
2169 | ||
2170 | /* | |
2171 | * We have other queues, don't allow more IO from this one | |
2172 | */ | |
2173 | if (cfqd->busy_queues > 1) | |
2174 | return false; | |
2175 | ||
2176 | /* | |
2177 | * Sole queue user, no limit | |
2178 | */ | |
2179 | max_dispatch = -1; | |
2180 | } | |
2181 | ||
2182 | /* | |
2183 | * Async queues must wait a bit before being allowed dispatch. | |
2184 | * We also ramp up the dispatch depth gradually for async IO, | |
2185 | * based on the last sync IO we serviced | |
2186 | */ | |
2187 | if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) { | |
2188 | unsigned long last_sync = jiffies - cfqd->last_end_sync_rq; | |
2189 | unsigned int depth; | |
2190 | ||
2191 | depth = last_sync / cfqd->cfq_slice[1]; | |
2192 | if (!depth && !cfqq->dispatched) | |
2193 | depth = 1; | |
2194 | if (depth < max_dispatch) | |
2195 | max_dispatch = depth; | |
2196 | } | |
2197 | ||
2198 | /* | |
2199 | * If we're below the current max, allow a dispatch | |
2200 | */ | |
2201 | return cfqq->dispatched < max_dispatch; | |
2202 | } | |
2203 | ||
2204 | /* | |
2205 | * Dispatch a request from cfqq, moving them to the request queue | |
2206 | * dispatch list. | |
2207 | */ | |
2208 | static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2209 | { | |
2210 | struct request *rq; | |
2211 | ||
2212 | BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list)); | |
2213 | ||
2214 | if (!cfq_may_dispatch(cfqd, cfqq)) | |
2215 | return false; | |
2216 | ||
2217 | /* | |
2218 | * follow expired path, else get first next available | |
2219 | */ | |
2220 | rq = cfq_check_fifo(cfqq); | |
2221 | if (!rq) | |
2222 | rq = cfqq->next_rq; | |
2223 | ||
2224 | /* | |
2225 | * insert request into driver dispatch list | |
2226 | */ | |
2227 | cfq_dispatch_insert(cfqd->queue, rq); | |
2228 | ||
2229 | if (!cfqd->active_cic) { | |
2230 | struct cfq_io_context *cic = RQ_CIC(rq); | |
2231 | ||
2232 | atomic_long_inc(&cic->ioc->refcount); | |
2233 | cfqd->active_cic = cic; | |
2234 | } | |
2235 | ||
2236 | return true; | |
2237 | } | |
2238 | ||
2239 | /* | |
2240 | * Find the cfqq that we need to service and move a request from that to the | |
2241 | * dispatch list | |
2242 | */ | |
2243 | static int cfq_dispatch_requests(struct request_queue *q, int force) | |
2244 | { | |
2245 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
2246 | struct cfq_queue *cfqq; | |
2247 | ||
2248 | if (!cfqd->busy_queues) | |
2249 | return 0; | |
2250 | ||
2251 | if (unlikely(force)) | |
2252 | return cfq_forced_dispatch(cfqd); | |
2253 | ||
2254 | cfqq = cfq_select_queue(cfqd); | |
2255 | if (!cfqq) | |
2256 | return 0; | |
2257 | ||
2258 | /* | |
2259 | * Dispatch a request from this cfqq, if it is allowed | |
2260 | */ | |
2261 | if (!cfq_dispatch_request(cfqd, cfqq)) | |
2262 | return 0; | |
2263 | ||
2264 | cfqq->slice_dispatch++; | |
2265 | cfq_clear_cfqq_must_dispatch(cfqq); | |
2266 | ||
2267 | /* | |
2268 | * expire an async queue immediately if it has used up its slice. idle | |
2269 | * queue always expire after 1 dispatch round. | |
2270 | */ | |
2271 | if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) && | |
2272 | cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) || | |
2273 | cfq_class_idle(cfqq))) { | |
2274 | cfqq->slice_end = jiffies + 1; | |
2275 | cfq_slice_expired(cfqd, 0); | |
2276 | } | |
2277 | ||
2278 | cfq_log_cfqq(cfqd, cfqq, "dispatched a request"); | |
2279 | return 1; | |
2280 | } | |
2281 | ||
2282 | /* | |
2283 | * task holds one reference to the queue, dropped when task exits. each rq | |
2284 | * in-flight on this queue also holds a reference, dropped when rq is freed. | |
2285 | * | |
2286 | * Each cfq queue took a reference on the parent group. Drop it now. | |
2287 | * queue lock must be held here. | |
2288 | */ | |
2289 | static void cfq_put_queue(struct cfq_queue *cfqq) | |
2290 | { | |
2291 | struct cfq_data *cfqd = cfqq->cfqd; | |
2292 | struct cfq_group *cfqg; | |
2293 | ||
2294 | BUG_ON(atomic_read(&cfqq->ref) <= 0); | |
2295 | ||
2296 | if (!atomic_dec_and_test(&cfqq->ref)) | |
2297 | return; | |
2298 | ||
2299 | cfq_log_cfqq(cfqd, cfqq, "put_queue"); | |
2300 | BUG_ON(rb_first(&cfqq->sort_list)); | |
2301 | BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]); | |
2302 | cfqg = cfqq->cfqg; | |
2303 | ||
2304 | if (unlikely(cfqd->active_queue == cfqq)) { | |
2305 | __cfq_slice_expired(cfqd, cfqq, 0); | |
2306 | cfq_schedule_dispatch(cfqd); | |
2307 | } | |
2308 | ||
2309 | BUG_ON(cfq_cfqq_on_rr(cfqq)); | |
2310 | kmem_cache_free(cfq_pool, cfqq); | |
2311 | cfq_put_cfqg(cfqg); | |
2312 | } | |
2313 | ||
2314 | /* | |
2315 | * Must always be called with the rcu_read_lock() held | |
2316 | */ | |
2317 | static void | |
2318 | __call_for_each_cic(struct io_context *ioc, | |
2319 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2320 | { | |
2321 | struct cfq_io_context *cic; | |
2322 | struct hlist_node *n; | |
2323 | ||
2324 | hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list) | |
2325 | func(ioc, cic); | |
2326 | } | |
2327 | ||
2328 | /* | |
2329 | * Call func for each cic attached to this ioc. | |
2330 | */ | |
2331 | static void | |
2332 | call_for_each_cic(struct io_context *ioc, | |
2333 | void (*func)(struct io_context *, struct cfq_io_context *)) | |
2334 | { | |
2335 | rcu_read_lock(); | |
2336 | __call_for_each_cic(ioc, func); | |
2337 | rcu_read_unlock(); | |
2338 | } | |
2339 | ||
2340 | static void cfq_cic_free_rcu(struct rcu_head *head) | |
2341 | { | |
2342 | struct cfq_io_context *cic; | |
2343 | ||
2344 | cic = container_of(head, struct cfq_io_context, rcu_head); | |
2345 | ||
2346 | kmem_cache_free(cfq_ioc_pool, cic); | |
2347 | elv_ioc_count_dec(cfq_ioc_count); | |
2348 | ||
2349 | if (ioc_gone) { | |
2350 | /* | |
2351 | * CFQ scheduler is exiting, grab exit lock and check | |
2352 | * the pending io context count. If it hits zero, | |
2353 | * complete ioc_gone and set it back to NULL | |
2354 | */ | |
2355 | spin_lock(&ioc_gone_lock); | |
2356 | if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) { | |
2357 | complete(ioc_gone); | |
2358 | ioc_gone = NULL; | |
2359 | } | |
2360 | spin_unlock(&ioc_gone_lock); | |
2361 | } | |
2362 | } | |
2363 | ||
2364 | static void cfq_cic_free(struct cfq_io_context *cic) | |
2365 | { | |
2366 | call_rcu(&cic->rcu_head, cfq_cic_free_rcu); | |
2367 | } | |
2368 | ||
2369 | static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic) | |
2370 | { | |
2371 | unsigned long flags; | |
2372 | ||
2373 | BUG_ON(!cic->dead_key); | |
2374 | ||
2375 | spin_lock_irqsave(&ioc->lock, flags); | |
2376 | radix_tree_delete(&ioc->radix_root, cic->dead_key); | |
2377 | hlist_del_rcu(&cic->cic_list); | |
2378 | spin_unlock_irqrestore(&ioc->lock, flags); | |
2379 | ||
2380 | cfq_cic_free(cic); | |
2381 | } | |
2382 | ||
2383 | /* | |
2384 | * Must be called with rcu_read_lock() held or preemption otherwise disabled. | |
2385 | * Only two callers of this - ->dtor() which is called with the rcu_read_lock(), | |
2386 | * and ->trim() which is called with the task lock held | |
2387 | */ | |
2388 | static void cfq_free_io_context(struct io_context *ioc) | |
2389 | { | |
2390 | /* | |
2391 | * ioc->refcount is zero here, or we are called from elv_unregister(), | |
2392 | * so no more cic's are allowed to be linked into this ioc. So it | |
2393 | * should be ok to iterate over the known list, we will see all cic's | |
2394 | * since no new ones are added. | |
2395 | */ | |
2396 | __call_for_each_cic(ioc, cic_free_func); | |
2397 | } | |
2398 | ||
2399 | static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
2400 | { | |
2401 | struct cfq_queue *__cfqq, *next; | |
2402 | ||
2403 | if (unlikely(cfqq == cfqd->active_queue)) { | |
2404 | __cfq_slice_expired(cfqd, cfqq, 0); | |
2405 | cfq_schedule_dispatch(cfqd); | |
2406 | } | |
2407 | ||
2408 | /* | |
2409 | * If this queue was scheduled to merge with another queue, be | |
2410 | * sure to drop the reference taken on that queue (and others in | |
2411 | * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs. | |
2412 | */ | |
2413 | __cfqq = cfqq->new_cfqq; | |
2414 | while (__cfqq) { | |
2415 | if (__cfqq == cfqq) { | |
2416 | WARN(1, "cfqq->new_cfqq loop detected\n"); | |
2417 | break; | |
2418 | } | |
2419 | next = __cfqq->new_cfqq; | |
2420 | cfq_put_queue(__cfqq); | |
2421 | __cfqq = next; | |
2422 | } | |
2423 | ||
2424 | cfq_put_queue(cfqq); | |
2425 | } | |
2426 | ||
2427 | static void __cfq_exit_single_io_context(struct cfq_data *cfqd, | |
2428 | struct cfq_io_context *cic) | |
2429 | { | |
2430 | struct io_context *ioc = cic->ioc; | |
2431 | ||
2432 | list_del_init(&cic->queue_list); | |
2433 | ||
2434 | /* | |
2435 | * Make sure key == NULL is seen for dead queues | |
2436 | */ | |
2437 | smp_wmb(); | |
2438 | cic->dead_key = (unsigned long) cic->key; | |
2439 | cic->key = NULL; | |
2440 | ||
2441 | if (ioc->ioc_data == cic) | |
2442 | rcu_assign_pointer(ioc->ioc_data, NULL); | |
2443 | ||
2444 | if (cic->cfqq[BLK_RW_ASYNC]) { | |
2445 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]); | |
2446 | cic->cfqq[BLK_RW_ASYNC] = NULL; | |
2447 | } | |
2448 | ||
2449 | if (cic->cfqq[BLK_RW_SYNC]) { | |
2450 | cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]); | |
2451 | cic->cfqq[BLK_RW_SYNC] = NULL; | |
2452 | } | |
2453 | } | |
2454 | ||
2455 | static void cfq_exit_single_io_context(struct io_context *ioc, | |
2456 | struct cfq_io_context *cic) | |
2457 | { | |
2458 | struct cfq_data *cfqd = cic->key; | |
2459 | ||
2460 | if (cfqd) { | |
2461 | struct request_queue *q = cfqd->queue; | |
2462 | unsigned long flags; | |
2463 | ||
2464 | spin_lock_irqsave(q->queue_lock, flags); | |
2465 | ||
2466 | /* | |
2467 | * Ensure we get a fresh copy of the ->key to prevent | |
2468 | * race between exiting task and queue | |
2469 | */ | |
2470 | smp_read_barrier_depends(); | |
2471 | if (cic->key) | |
2472 | __cfq_exit_single_io_context(cfqd, cic); | |
2473 | ||
2474 | spin_unlock_irqrestore(q->queue_lock, flags); | |
2475 | } | |
2476 | } | |
2477 | ||
2478 | /* | |
2479 | * The process that ioc belongs to has exited, we need to clean up | |
2480 | * and put the internal structures we have that belongs to that process. | |
2481 | */ | |
2482 | static void cfq_exit_io_context(struct io_context *ioc) | |
2483 | { | |
2484 | call_for_each_cic(ioc, cfq_exit_single_io_context); | |
2485 | } | |
2486 | ||
2487 | static struct cfq_io_context * | |
2488 | cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) | |
2489 | { | |
2490 | struct cfq_io_context *cic; | |
2491 | ||
2492 | cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO, | |
2493 | cfqd->queue->node); | |
2494 | if (cic) { | |
2495 | cic->last_end_request = jiffies; | |
2496 | INIT_LIST_HEAD(&cic->queue_list); | |
2497 | INIT_HLIST_NODE(&cic->cic_list); | |
2498 | cic->dtor = cfq_free_io_context; | |
2499 | cic->exit = cfq_exit_io_context; | |
2500 | elv_ioc_count_inc(cfq_ioc_count); | |
2501 | } | |
2502 | ||
2503 | return cic; | |
2504 | } | |
2505 | ||
2506 | static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc) | |
2507 | { | |
2508 | struct task_struct *tsk = current; | |
2509 | int ioprio_class; | |
2510 | ||
2511 | if (!cfq_cfqq_prio_changed(cfqq)) | |
2512 | return; | |
2513 | ||
2514 | ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio); | |
2515 | switch (ioprio_class) { | |
2516 | default: | |
2517 | printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class); | |
2518 | case IOPRIO_CLASS_NONE: | |
2519 | /* | |
2520 | * no prio set, inherit CPU scheduling settings | |
2521 | */ | |
2522 | cfqq->ioprio = task_nice_ioprio(tsk); | |
2523 | cfqq->ioprio_class = task_nice_ioclass(tsk); | |
2524 | break; | |
2525 | case IOPRIO_CLASS_RT: | |
2526 | cfqq->ioprio = task_ioprio(ioc); | |
2527 | cfqq->ioprio_class = IOPRIO_CLASS_RT; | |
2528 | break; | |
2529 | case IOPRIO_CLASS_BE: | |
2530 | cfqq->ioprio = task_ioprio(ioc); | |
2531 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
2532 | break; | |
2533 | case IOPRIO_CLASS_IDLE: | |
2534 | cfqq->ioprio_class = IOPRIO_CLASS_IDLE; | |
2535 | cfqq->ioprio = 7; | |
2536 | cfq_clear_cfqq_idle_window(cfqq); | |
2537 | break; | |
2538 | } | |
2539 | ||
2540 | /* | |
2541 | * keep track of original prio settings in case we have to temporarily | |
2542 | * elevate the priority of this queue | |
2543 | */ | |
2544 | cfqq->org_ioprio = cfqq->ioprio; | |
2545 | cfqq->org_ioprio_class = cfqq->ioprio_class; | |
2546 | cfq_clear_cfqq_prio_changed(cfqq); | |
2547 | } | |
2548 | ||
2549 | static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic) | |
2550 | { | |
2551 | struct cfq_data *cfqd = cic->key; | |
2552 | struct cfq_queue *cfqq; | |
2553 | unsigned long flags; | |
2554 | ||
2555 | if (unlikely(!cfqd)) | |
2556 | return; | |
2557 | ||
2558 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2559 | ||
2560 | cfqq = cic->cfqq[BLK_RW_ASYNC]; | |
2561 | if (cfqq) { | |
2562 | struct cfq_queue *new_cfqq; | |
2563 | new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc, | |
2564 | GFP_ATOMIC); | |
2565 | if (new_cfqq) { | |
2566 | cic->cfqq[BLK_RW_ASYNC] = new_cfqq; | |
2567 | cfq_put_queue(cfqq); | |
2568 | } | |
2569 | } | |
2570 | ||
2571 | cfqq = cic->cfqq[BLK_RW_SYNC]; | |
2572 | if (cfqq) | |
2573 | cfq_mark_cfqq_prio_changed(cfqq); | |
2574 | ||
2575 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2576 | } | |
2577 | ||
2578 | static void cfq_ioc_set_ioprio(struct io_context *ioc) | |
2579 | { | |
2580 | call_for_each_cic(ioc, changed_ioprio); | |
2581 | ioc->ioprio_changed = 0; | |
2582 | } | |
2583 | ||
2584 | static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2585 | pid_t pid, bool is_sync) | |
2586 | { | |
2587 | RB_CLEAR_NODE(&cfqq->rb_node); | |
2588 | RB_CLEAR_NODE(&cfqq->p_node); | |
2589 | INIT_LIST_HEAD(&cfqq->fifo); | |
2590 | ||
2591 | atomic_set(&cfqq->ref, 0); | |
2592 | cfqq->cfqd = cfqd; | |
2593 | ||
2594 | cfq_mark_cfqq_prio_changed(cfqq); | |
2595 | ||
2596 | if (is_sync) { | |
2597 | if (!cfq_class_idle(cfqq)) | |
2598 | cfq_mark_cfqq_idle_window(cfqq); | |
2599 | cfq_mark_cfqq_sync(cfqq); | |
2600 | } | |
2601 | cfqq->pid = pid; | |
2602 | } | |
2603 | ||
2604 | static struct cfq_queue * | |
2605 | cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, | |
2606 | struct io_context *ioc, gfp_t gfp_mask) | |
2607 | { | |
2608 | struct cfq_queue *cfqq, *new_cfqq = NULL; | |
2609 | struct cfq_io_context *cic; | |
2610 | struct cfq_group *cfqg; | |
2611 | ||
2612 | retry: | |
2613 | cfqg = cfq_get_cfqg(cfqd, 1); | |
2614 | cic = cfq_cic_lookup(cfqd, ioc); | |
2615 | /* cic always exists here */ | |
2616 | cfqq = cic_to_cfqq(cic, is_sync); | |
2617 | ||
2618 | /* | |
2619 | * Always try a new alloc if we fell back to the OOM cfqq | |
2620 | * originally, since it should just be a temporary situation. | |
2621 | */ | |
2622 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
2623 | cfqq = NULL; | |
2624 | if (new_cfqq) { | |
2625 | cfqq = new_cfqq; | |
2626 | new_cfqq = NULL; | |
2627 | } else if (gfp_mask & __GFP_WAIT) { | |
2628 | spin_unlock_irq(cfqd->queue->queue_lock); | |
2629 | new_cfqq = kmem_cache_alloc_node(cfq_pool, | |
2630 | gfp_mask | __GFP_ZERO, | |
2631 | cfqd->queue->node); | |
2632 | spin_lock_irq(cfqd->queue->queue_lock); | |
2633 | if (new_cfqq) | |
2634 | goto retry; | |
2635 | } else { | |
2636 | cfqq = kmem_cache_alloc_node(cfq_pool, | |
2637 | gfp_mask | __GFP_ZERO, | |
2638 | cfqd->queue->node); | |
2639 | } | |
2640 | ||
2641 | if (cfqq) { | |
2642 | cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync); | |
2643 | cfq_init_prio_data(cfqq, ioc); | |
2644 | cfq_link_cfqq_cfqg(cfqq, cfqg); | |
2645 | cfq_log_cfqq(cfqd, cfqq, "alloced"); | |
2646 | } else | |
2647 | cfqq = &cfqd->oom_cfqq; | |
2648 | } | |
2649 | ||
2650 | if (new_cfqq) | |
2651 | kmem_cache_free(cfq_pool, new_cfqq); | |
2652 | ||
2653 | return cfqq; | |
2654 | } | |
2655 | ||
2656 | static struct cfq_queue ** | |
2657 | cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio) | |
2658 | { | |
2659 | switch (ioprio_class) { | |
2660 | case IOPRIO_CLASS_RT: | |
2661 | return &cfqd->async_cfqq[0][ioprio]; | |
2662 | case IOPRIO_CLASS_BE: | |
2663 | return &cfqd->async_cfqq[1][ioprio]; | |
2664 | case IOPRIO_CLASS_IDLE: | |
2665 | return &cfqd->async_idle_cfqq; | |
2666 | default: | |
2667 | BUG(); | |
2668 | } | |
2669 | } | |
2670 | ||
2671 | static struct cfq_queue * | |
2672 | cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc, | |
2673 | gfp_t gfp_mask) | |
2674 | { | |
2675 | const int ioprio = task_ioprio(ioc); | |
2676 | const int ioprio_class = task_ioprio_class(ioc); | |
2677 | struct cfq_queue **async_cfqq = NULL; | |
2678 | struct cfq_queue *cfqq = NULL; | |
2679 | ||
2680 | if (!is_sync) { | |
2681 | async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio); | |
2682 | cfqq = *async_cfqq; | |
2683 | } | |
2684 | ||
2685 | if (!cfqq) | |
2686 | cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask); | |
2687 | ||
2688 | /* | |
2689 | * pin the queue now that it's allocated, scheduler exit will prune it | |
2690 | */ | |
2691 | if (!is_sync && !(*async_cfqq)) { | |
2692 | atomic_inc(&cfqq->ref); | |
2693 | *async_cfqq = cfqq; | |
2694 | } | |
2695 | ||
2696 | atomic_inc(&cfqq->ref); | |
2697 | return cfqq; | |
2698 | } | |
2699 | ||
2700 | /* | |
2701 | * We drop cfq io contexts lazily, so we may find a dead one. | |
2702 | */ | |
2703 | static void | |
2704 | cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc, | |
2705 | struct cfq_io_context *cic) | |
2706 | { | |
2707 | unsigned long flags; | |
2708 | ||
2709 | WARN_ON(!list_empty(&cic->queue_list)); | |
2710 | ||
2711 | spin_lock_irqsave(&ioc->lock, flags); | |
2712 | ||
2713 | BUG_ON(ioc->ioc_data == cic); | |
2714 | ||
2715 | radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd); | |
2716 | hlist_del_rcu(&cic->cic_list); | |
2717 | spin_unlock_irqrestore(&ioc->lock, flags); | |
2718 | ||
2719 | cfq_cic_free(cic); | |
2720 | } | |
2721 | ||
2722 | static struct cfq_io_context * | |
2723 | cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc) | |
2724 | { | |
2725 | struct cfq_io_context *cic; | |
2726 | unsigned long flags; | |
2727 | void *k; | |
2728 | ||
2729 | if (unlikely(!ioc)) | |
2730 | return NULL; | |
2731 | ||
2732 | rcu_read_lock(); | |
2733 | ||
2734 | /* | |
2735 | * we maintain a last-hit cache, to avoid browsing over the tree | |
2736 | */ | |
2737 | cic = rcu_dereference(ioc->ioc_data); | |
2738 | if (cic && cic->key == cfqd) { | |
2739 | rcu_read_unlock(); | |
2740 | return cic; | |
2741 | } | |
2742 | ||
2743 | do { | |
2744 | cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd); | |
2745 | rcu_read_unlock(); | |
2746 | if (!cic) | |
2747 | break; | |
2748 | /* ->key must be copied to avoid race with cfq_exit_queue() */ | |
2749 | k = cic->key; | |
2750 | if (unlikely(!k)) { | |
2751 | cfq_drop_dead_cic(cfqd, ioc, cic); | |
2752 | rcu_read_lock(); | |
2753 | continue; | |
2754 | } | |
2755 | ||
2756 | spin_lock_irqsave(&ioc->lock, flags); | |
2757 | rcu_assign_pointer(ioc->ioc_data, cic); | |
2758 | spin_unlock_irqrestore(&ioc->lock, flags); | |
2759 | break; | |
2760 | } while (1); | |
2761 | ||
2762 | return cic; | |
2763 | } | |
2764 | ||
2765 | /* | |
2766 | * Add cic into ioc, using cfqd as the search key. This enables us to lookup | |
2767 | * the process specific cfq io context when entered from the block layer. | |
2768 | * Also adds the cic to a per-cfqd list, used when this queue is removed. | |
2769 | */ | |
2770 | static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc, | |
2771 | struct cfq_io_context *cic, gfp_t gfp_mask) | |
2772 | { | |
2773 | unsigned long flags; | |
2774 | int ret; | |
2775 | ||
2776 | ret = radix_tree_preload(gfp_mask); | |
2777 | if (!ret) { | |
2778 | cic->ioc = ioc; | |
2779 | cic->key = cfqd; | |
2780 | ||
2781 | spin_lock_irqsave(&ioc->lock, flags); | |
2782 | ret = radix_tree_insert(&ioc->radix_root, | |
2783 | (unsigned long) cfqd, cic); | |
2784 | if (!ret) | |
2785 | hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list); | |
2786 | spin_unlock_irqrestore(&ioc->lock, flags); | |
2787 | ||
2788 | radix_tree_preload_end(); | |
2789 | ||
2790 | if (!ret) { | |
2791 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
2792 | list_add(&cic->queue_list, &cfqd->cic_list); | |
2793 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
2794 | } | |
2795 | } | |
2796 | ||
2797 | if (ret) | |
2798 | printk(KERN_ERR "cfq: cic link failed!\n"); | |
2799 | ||
2800 | return ret; | |
2801 | } | |
2802 | ||
2803 | /* | |
2804 | * Setup general io context and cfq io context. There can be several cfq | |
2805 | * io contexts per general io context, if this process is doing io to more | |
2806 | * than one device managed by cfq. | |
2807 | */ | |
2808 | static struct cfq_io_context * | |
2809 | cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask) | |
2810 | { | |
2811 | struct io_context *ioc = NULL; | |
2812 | struct cfq_io_context *cic; | |
2813 | ||
2814 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
2815 | ||
2816 | ioc = get_io_context(gfp_mask, cfqd->queue->node); | |
2817 | if (!ioc) | |
2818 | return NULL; | |
2819 | ||
2820 | cic = cfq_cic_lookup(cfqd, ioc); | |
2821 | if (cic) | |
2822 | goto out; | |
2823 | ||
2824 | cic = cfq_alloc_io_context(cfqd, gfp_mask); | |
2825 | if (cic == NULL) | |
2826 | goto err; | |
2827 | ||
2828 | if (cfq_cic_link(cfqd, ioc, cic, gfp_mask)) | |
2829 | goto err_free; | |
2830 | ||
2831 | out: | |
2832 | smp_read_barrier_depends(); | |
2833 | if (unlikely(ioc->ioprio_changed)) | |
2834 | cfq_ioc_set_ioprio(ioc); | |
2835 | ||
2836 | return cic; | |
2837 | err_free: | |
2838 | cfq_cic_free(cic); | |
2839 | err: | |
2840 | put_io_context(ioc); | |
2841 | return NULL; | |
2842 | } | |
2843 | ||
2844 | static void | |
2845 | cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic) | |
2846 | { | |
2847 | unsigned long elapsed = jiffies - cic->last_end_request; | |
2848 | unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle); | |
2849 | ||
2850 | cic->ttime_samples = (7*cic->ttime_samples + 256) / 8; | |
2851 | cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8; | |
2852 | cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples; | |
2853 | } | |
2854 | ||
2855 | static void | |
2856 | cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2857 | struct request *rq) | |
2858 | { | |
2859 | sector_t sdist; | |
2860 | u64 total; | |
2861 | ||
2862 | if (!cfqq->last_request_pos) | |
2863 | sdist = 0; | |
2864 | else if (cfqq->last_request_pos < blk_rq_pos(rq)) | |
2865 | sdist = blk_rq_pos(rq) - cfqq->last_request_pos; | |
2866 | else | |
2867 | sdist = cfqq->last_request_pos - blk_rq_pos(rq); | |
2868 | ||
2869 | /* | |
2870 | * Don't allow the seek distance to get too large from the | |
2871 | * odd fragment, pagein, etc | |
2872 | */ | |
2873 | if (cfqq->seek_samples <= 60) /* second&third seek */ | |
2874 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024); | |
2875 | else | |
2876 | sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64); | |
2877 | ||
2878 | cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8; | |
2879 | cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8; | |
2880 | total = cfqq->seek_total + (cfqq->seek_samples/2); | |
2881 | do_div(total, cfqq->seek_samples); | |
2882 | cfqq->seek_mean = (sector_t)total; | |
2883 | ||
2884 | /* | |
2885 | * If this cfqq is shared between multiple processes, check to | |
2886 | * make sure that those processes are still issuing I/Os within | |
2887 | * the mean seek distance. If not, it may be time to break the | |
2888 | * queues apart again. | |
2889 | */ | |
2890 | if (cfq_cfqq_coop(cfqq)) { | |
2891 | if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start) | |
2892 | cfqq->seeky_start = jiffies; | |
2893 | else if (!CFQQ_SEEKY(cfqq)) | |
2894 | cfqq->seeky_start = 0; | |
2895 | } | |
2896 | } | |
2897 | ||
2898 | /* | |
2899 | * Disable idle window if the process thinks too long or seeks so much that | |
2900 | * it doesn't matter | |
2901 | */ | |
2902 | static void | |
2903 | cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
2904 | struct cfq_io_context *cic) | |
2905 | { | |
2906 | int old_idle, enable_idle; | |
2907 | ||
2908 | /* | |
2909 | * Don't idle for async or idle io prio class | |
2910 | */ | |
2911 | if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq)) | |
2912 | return; | |
2913 | ||
2914 | enable_idle = old_idle = cfq_cfqq_idle_window(cfqq); | |
2915 | ||
2916 | if (cfqq->queued[0] + cfqq->queued[1] >= 4) | |
2917 | cfq_mark_cfqq_deep(cfqq); | |
2918 | ||
2919 | if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle || | |
2920 | (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples) | |
2921 | && CFQQ_SEEKY(cfqq))) | |
2922 | enable_idle = 0; | |
2923 | else if (sample_valid(cic->ttime_samples)) { | |
2924 | if (cic->ttime_mean > cfqd->cfq_slice_idle) | |
2925 | enable_idle = 0; | |
2926 | else | |
2927 | enable_idle = 1; | |
2928 | } | |
2929 | ||
2930 | if (old_idle != enable_idle) { | |
2931 | cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle); | |
2932 | if (enable_idle) | |
2933 | cfq_mark_cfqq_idle_window(cfqq); | |
2934 | else | |
2935 | cfq_clear_cfqq_idle_window(cfqq); | |
2936 | } | |
2937 | } | |
2938 | ||
2939 | /* | |
2940 | * Check if new_cfqq should preempt the currently active queue. Return 0 for | |
2941 | * no or if we aren't sure, a 1 will cause a preempt. | |
2942 | */ | |
2943 | static bool | |
2944 | cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq, | |
2945 | struct request *rq) | |
2946 | { | |
2947 | struct cfq_queue *cfqq; | |
2948 | ||
2949 | cfqq = cfqd->active_queue; | |
2950 | if (!cfqq) | |
2951 | return false; | |
2952 | ||
2953 | if (cfq_slice_used(cfqq)) | |
2954 | return true; | |
2955 | ||
2956 | if (cfq_class_idle(new_cfqq)) | |
2957 | return false; | |
2958 | ||
2959 | if (cfq_class_idle(cfqq)) | |
2960 | return true; | |
2961 | ||
2962 | /* Allow preemption only if we are idling on sync-noidle tree */ | |
2963 | if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD && | |
2964 | cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD && | |
2965 | new_cfqq->service_tree->count == 2 && | |
2966 | RB_EMPTY_ROOT(&cfqq->sort_list)) | |
2967 | return true; | |
2968 | ||
2969 | /* | |
2970 | * if the new request is sync, but the currently running queue is | |
2971 | * not, let the sync request have priority. | |
2972 | */ | |
2973 | if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq)) | |
2974 | return true; | |
2975 | ||
2976 | /* | |
2977 | * So both queues are sync. Let the new request get disk time if | |
2978 | * it's a metadata request and the current queue is doing regular IO. | |
2979 | */ | |
2980 | if (rq_is_meta(rq) && !cfqq->meta_pending) | |
2981 | return true; | |
2982 | ||
2983 | /* | |
2984 | * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice. | |
2985 | */ | |
2986 | if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq)) | |
2987 | return true; | |
2988 | ||
2989 | if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq)) | |
2990 | return false; | |
2991 | ||
2992 | /* | |
2993 | * if this request is as-good as one we would expect from the | |
2994 | * current cfqq, let it preempt | |
2995 | */ | |
2996 | if (cfq_rq_close(cfqd, cfqq, rq)) | |
2997 | return true; | |
2998 | ||
2999 | return false; | |
3000 | } | |
3001 | ||
3002 | /* | |
3003 | * cfqq preempts the active queue. if we allowed preempt with no slice left, | |
3004 | * let it have half of its nominal slice. | |
3005 | */ | |
3006 | static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq) | |
3007 | { | |
3008 | cfq_log_cfqq(cfqd, cfqq, "preempt"); | |
3009 | cfq_slice_expired(cfqd, 1); | |
3010 | ||
3011 | /* | |
3012 | * Put the new queue at the front of the of the current list, | |
3013 | * so we know that it will be selected next. | |
3014 | */ | |
3015 | BUG_ON(!cfq_cfqq_on_rr(cfqq)); | |
3016 | ||
3017 | cfq_service_tree_add(cfqd, cfqq, 1); | |
3018 | ||
3019 | cfqq->slice_end = 0; | |
3020 | cfq_mark_cfqq_slice_new(cfqq); | |
3021 | } | |
3022 | ||
3023 | /* | |
3024 | * Called when a new fs request (rq) is added (to cfqq). Check if there's | |
3025 | * something we should do about it | |
3026 | */ | |
3027 | static void | |
3028 | cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq, | |
3029 | struct request *rq) | |
3030 | { | |
3031 | struct cfq_io_context *cic = RQ_CIC(rq); | |
3032 | ||
3033 | cfqd->rq_queued++; | |
3034 | if (rq_is_meta(rq)) | |
3035 | cfqq->meta_pending++; | |
3036 | ||
3037 | cfq_update_io_thinktime(cfqd, cic); | |
3038 | cfq_update_io_seektime(cfqd, cfqq, rq); | |
3039 | cfq_update_idle_window(cfqd, cfqq, cic); | |
3040 | ||
3041 | cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); | |
3042 | ||
3043 | if (cfqq == cfqd->active_queue) { | |
3044 | /* | |
3045 | * Remember that we saw a request from this process, but | |
3046 | * don't start queuing just yet. Otherwise we risk seeing lots | |
3047 | * of tiny requests, because we disrupt the normal plugging | |
3048 | * and merging. If the request is already larger than a single | |
3049 | * page, let it rip immediately. For that case we assume that | |
3050 | * merging is already done. Ditto for a busy system that | |
3051 | * has other work pending, don't risk delaying until the | |
3052 | * idle timer unplug to continue working. | |
3053 | */ | |
3054 | if (cfq_cfqq_wait_request(cfqq)) { | |
3055 | if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE || | |
3056 | cfqd->busy_queues > 1) { | |
3057 | del_timer(&cfqd->idle_slice_timer); | |
3058 | __blk_run_queue(cfqd->queue); | |
3059 | } else | |
3060 | cfq_mark_cfqq_must_dispatch(cfqq); | |
3061 | } | |
3062 | } else if (cfq_should_preempt(cfqd, cfqq, rq)) { | |
3063 | /* | |
3064 | * not the active queue - expire current slice if it is | |
3065 | * idle and has expired it's mean thinktime or this new queue | |
3066 | * has some old slice time left and is of higher priority or | |
3067 | * this new queue is RT and the current one is BE | |
3068 | */ | |
3069 | cfq_preempt_queue(cfqd, cfqq); | |
3070 | __blk_run_queue(cfqd->queue); | |
3071 | } | |
3072 | } | |
3073 | ||
3074 | static void cfq_insert_request(struct request_queue *q, struct request *rq) | |
3075 | { | |
3076 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3077 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3078 | ||
3079 | cfq_log_cfqq(cfqd, cfqq, "insert_request"); | |
3080 | cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc); | |
3081 | ||
3082 | rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]); | |
3083 | list_add_tail(&rq->queuelist, &cfqq->fifo); | |
3084 | cfq_add_rq_rb(rq); | |
3085 | ||
3086 | cfq_rq_enqueued(cfqd, cfqq, rq); | |
3087 | } | |
3088 | ||
3089 | /* | |
3090 | * Update hw_tag based on peak queue depth over 50 samples under | |
3091 | * sufficient load. | |
3092 | */ | |
3093 | static void cfq_update_hw_tag(struct cfq_data *cfqd) | |
3094 | { | |
3095 | struct cfq_queue *cfqq = cfqd->active_queue; | |
3096 | ||
3097 | if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth) | |
3098 | cfqd->hw_tag_est_depth = rq_in_driver(cfqd); | |
3099 | ||
3100 | if (cfqd->hw_tag == 1) | |
3101 | return; | |
3102 | ||
3103 | if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN && | |
3104 | rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN) | |
3105 | return; | |
3106 | ||
3107 | /* | |
3108 | * If active queue hasn't enough requests and can idle, cfq might not | |
3109 | * dispatch sufficient requests to hardware. Don't zero hw_tag in this | |
3110 | * case | |
3111 | */ | |
3112 | if (cfqq && cfq_cfqq_idle_window(cfqq) && | |
3113 | cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] < | |
3114 | CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN) | |
3115 | return; | |
3116 | ||
3117 | if (cfqd->hw_tag_samples++ < 50) | |
3118 | return; | |
3119 | ||
3120 | if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN) | |
3121 | cfqd->hw_tag = 1; | |
3122 | else | |
3123 | cfqd->hw_tag = 0; | |
3124 | } | |
3125 | ||
3126 | static void cfq_completed_request(struct request_queue *q, struct request *rq) | |
3127 | { | |
3128 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3129 | struct cfq_data *cfqd = cfqq->cfqd; | |
3130 | const int sync = rq_is_sync(rq); | |
3131 | unsigned long now; | |
3132 | ||
3133 | now = jiffies; | |
3134 | cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq)); | |
3135 | ||
3136 | cfq_update_hw_tag(cfqd); | |
3137 | ||
3138 | WARN_ON(!cfqd->rq_in_driver[sync]); | |
3139 | WARN_ON(!cfqq->dispatched); | |
3140 | cfqd->rq_in_driver[sync]--; | |
3141 | cfqq->dispatched--; | |
3142 | ||
3143 | if (cfq_cfqq_sync(cfqq)) | |
3144 | cfqd->sync_flight--; | |
3145 | ||
3146 | if (sync) { | |
3147 | RQ_CIC(rq)->last_end_request = now; | |
3148 | cfqd->last_end_sync_rq = now; | |
3149 | } | |
3150 | ||
3151 | /* | |
3152 | * If this is the active queue, check if it needs to be expired, | |
3153 | * or if we want to idle in case it has no pending requests. | |
3154 | */ | |
3155 | if (cfqd->active_queue == cfqq) { | |
3156 | const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list); | |
3157 | ||
3158 | if (cfq_cfqq_slice_new(cfqq)) { | |
3159 | cfq_set_prio_slice(cfqd, cfqq); | |
3160 | cfq_clear_cfqq_slice_new(cfqq); | |
3161 | } | |
3162 | /* | |
3163 | * Idling is not enabled on: | |
3164 | * - expired queues | |
3165 | * - idle-priority queues | |
3166 | * - async queues | |
3167 | * - queues with still some requests queued | |
3168 | * - when there is a close cooperator | |
3169 | */ | |
3170 | if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq)) | |
3171 | cfq_slice_expired(cfqd, 1); | |
3172 | else if (sync && cfqq_empty && | |
3173 | !cfq_close_cooperator(cfqd, cfqq)) { | |
3174 | cfqd->noidle_tree_requires_idle |= !rq_noidle(rq); | |
3175 | /* | |
3176 | * Idling is enabled for SYNC_WORKLOAD. | |
3177 | * SYNC_NOIDLE_WORKLOAD idles at the end of the tree | |
3178 | * only if we processed at least one !rq_noidle request | |
3179 | */ | |
3180 | if (cfqd->serving_type == SYNC_WORKLOAD | |
3181 | || cfqd->noidle_tree_requires_idle) | |
3182 | cfq_arm_slice_timer(cfqd); | |
3183 | } | |
3184 | } | |
3185 | ||
3186 | if (!rq_in_driver(cfqd)) | |
3187 | cfq_schedule_dispatch(cfqd); | |
3188 | } | |
3189 | ||
3190 | /* | |
3191 | * we temporarily boost lower priority queues if they are holding fs exclusive | |
3192 | * resources. they are boosted to normal prio (CLASS_BE/4) | |
3193 | */ | |
3194 | static void cfq_prio_boost(struct cfq_queue *cfqq) | |
3195 | { | |
3196 | if (has_fs_excl()) { | |
3197 | /* | |
3198 | * boost idle prio on transactions that would lock out other | |
3199 | * users of the filesystem | |
3200 | */ | |
3201 | if (cfq_class_idle(cfqq)) | |
3202 | cfqq->ioprio_class = IOPRIO_CLASS_BE; | |
3203 | if (cfqq->ioprio > IOPRIO_NORM) | |
3204 | cfqq->ioprio = IOPRIO_NORM; | |
3205 | } else { | |
3206 | /* | |
3207 | * unboost the queue (if needed) | |
3208 | */ | |
3209 | cfqq->ioprio_class = cfqq->org_ioprio_class; | |
3210 | cfqq->ioprio = cfqq->org_ioprio; | |
3211 | } | |
3212 | } | |
3213 | ||
3214 | static inline int __cfq_may_queue(struct cfq_queue *cfqq) | |
3215 | { | |
3216 | if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) { | |
3217 | cfq_mark_cfqq_must_alloc_slice(cfqq); | |
3218 | return ELV_MQUEUE_MUST; | |
3219 | } | |
3220 | ||
3221 | return ELV_MQUEUE_MAY; | |
3222 | } | |
3223 | ||
3224 | static int cfq_may_queue(struct request_queue *q, int rw) | |
3225 | { | |
3226 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3227 | struct task_struct *tsk = current; | |
3228 | struct cfq_io_context *cic; | |
3229 | struct cfq_queue *cfqq; | |
3230 | ||
3231 | /* | |
3232 | * don't force setup of a queue from here, as a call to may_queue | |
3233 | * does not necessarily imply that a request actually will be queued. | |
3234 | * so just lookup a possibly existing queue, or return 'may queue' | |
3235 | * if that fails | |
3236 | */ | |
3237 | cic = cfq_cic_lookup(cfqd, tsk->io_context); | |
3238 | if (!cic) | |
3239 | return ELV_MQUEUE_MAY; | |
3240 | ||
3241 | cfqq = cic_to_cfqq(cic, rw_is_sync(rw)); | |
3242 | if (cfqq) { | |
3243 | cfq_init_prio_data(cfqq, cic->ioc); | |
3244 | cfq_prio_boost(cfqq); | |
3245 | ||
3246 | return __cfq_may_queue(cfqq); | |
3247 | } | |
3248 | ||
3249 | return ELV_MQUEUE_MAY; | |
3250 | } | |
3251 | ||
3252 | /* | |
3253 | * queue lock held here | |
3254 | */ | |
3255 | static void cfq_put_request(struct request *rq) | |
3256 | { | |
3257 | struct cfq_queue *cfqq = RQ_CFQQ(rq); | |
3258 | ||
3259 | if (cfqq) { | |
3260 | const int rw = rq_data_dir(rq); | |
3261 | ||
3262 | BUG_ON(!cfqq->allocated[rw]); | |
3263 | cfqq->allocated[rw]--; | |
3264 | ||
3265 | put_io_context(RQ_CIC(rq)->ioc); | |
3266 | ||
3267 | rq->elevator_private = NULL; | |
3268 | rq->elevator_private2 = NULL; | |
3269 | ||
3270 | cfq_put_queue(cfqq); | |
3271 | } | |
3272 | } | |
3273 | ||
3274 | static struct cfq_queue * | |
3275 | cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic, | |
3276 | struct cfq_queue *cfqq) | |
3277 | { | |
3278 | cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq); | |
3279 | cic_set_cfqq(cic, cfqq->new_cfqq, 1); | |
3280 | cfq_mark_cfqq_coop(cfqq->new_cfqq); | |
3281 | cfq_put_queue(cfqq); | |
3282 | return cic_to_cfqq(cic, 1); | |
3283 | } | |
3284 | ||
3285 | static int should_split_cfqq(struct cfq_queue *cfqq) | |
3286 | { | |
3287 | if (cfqq->seeky_start && | |
3288 | time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT)) | |
3289 | return 1; | |
3290 | return 0; | |
3291 | } | |
3292 | ||
3293 | /* | |
3294 | * Returns NULL if a new cfqq should be allocated, or the old cfqq if this | |
3295 | * was the last process referring to said cfqq. | |
3296 | */ | |
3297 | static struct cfq_queue * | |
3298 | split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq) | |
3299 | { | |
3300 | if (cfqq_process_refs(cfqq) == 1) { | |
3301 | cfqq->seeky_start = 0; | |
3302 | cfqq->pid = current->pid; | |
3303 | cfq_clear_cfqq_coop(cfqq); | |
3304 | return cfqq; | |
3305 | } | |
3306 | ||
3307 | cic_set_cfqq(cic, NULL, 1); | |
3308 | cfq_put_queue(cfqq); | |
3309 | return NULL; | |
3310 | } | |
3311 | /* | |
3312 | * Allocate cfq data structures associated with this request. | |
3313 | */ | |
3314 | static int | |
3315 | cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask) | |
3316 | { | |
3317 | struct cfq_data *cfqd = q->elevator->elevator_data; | |
3318 | struct cfq_io_context *cic; | |
3319 | const int rw = rq_data_dir(rq); | |
3320 | const bool is_sync = rq_is_sync(rq); | |
3321 | struct cfq_queue *cfqq; | |
3322 | unsigned long flags; | |
3323 | ||
3324 | might_sleep_if(gfp_mask & __GFP_WAIT); | |
3325 | ||
3326 | cic = cfq_get_io_context(cfqd, gfp_mask); | |
3327 | ||
3328 | spin_lock_irqsave(q->queue_lock, flags); | |
3329 | ||
3330 | if (!cic) | |
3331 | goto queue_fail; | |
3332 | ||
3333 | new_queue: | |
3334 | cfqq = cic_to_cfqq(cic, is_sync); | |
3335 | if (!cfqq || cfqq == &cfqd->oom_cfqq) { | |
3336 | cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask); | |
3337 | cic_set_cfqq(cic, cfqq, is_sync); | |
3338 | } else { | |
3339 | /* | |
3340 | * If the queue was seeky for too long, break it apart. | |
3341 | */ | |
3342 | if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) { | |
3343 | cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq"); | |
3344 | cfqq = split_cfqq(cic, cfqq); | |
3345 | if (!cfqq) | |
3346 | goto new_queue; | |
3347 | } | |
3348 | ||
3349 | /* | |
3350 | * Check to see if this queue is scheduled to merge with | |
3351 | * another, closely cooperating queue. The merging of | |
3352 | * queues happens here as it must be done in process context. | |
3353 | * The reference on new_cfqq was taken in merge_cfqqs. | |
3354 | */ | |
3355 | if (cfqq->new_cfqq) | |
3356 | cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq); | |
3357 | } | |
3358 | ||
3359 | cfqq->allocated[rw]++; | |
3360 | atomic_inc(&cfqq->ref); | |
3361 | ||
3362 | spin_unlock_irqrestore(q->queue_lock, flags); | |
3363 | ||
3364 | rq->elevator_private = cic; | |
3365 | rq->elevator_private2 = cfqq; | |
3366 | return 0; | |
3367 | ||
3368 | queue_fail: | |
3369 | if (cic) | |
3370 | put_io_context(cic->ioc); | |
3371 | ||
3372 | cfq_schedule_dispatch(cfqd); | |
3373 | spin_unlock_irqrestore(q->queue_lock, flags); | |
3374 | cfq_log(cfqd, "set_request fail"); | |
3375 | return 1; | |
3376 | } | |
3377 | ||
3378 | static void cfq_kick_queue(struct work_struct *work) | |
3379 | { | |
3380 | struct cfq_data *cfqd = | |
3381 | container_of(work, struct cfq_data, unplug_work); | |
3382 | struct request_queue *q = cfqd->queue; | |
3383 | ||
3384 | spin_lock_irq(q->queue_lock); | |
3385 | __blk_run_queue(cfqd->queue); | |
3386 | spin_unlock_irq(q->queue_lock); | |
3387 | } | |
3388 | ||
3389 | /* | |
3390 | * Timer running if the active_queue is currently idling inside its time slice | |
3391 | */ | |
3392 | static void cfq_idle_slice_timer(unsigned long data) | |
3393 | { | |
3394 | struct cfq_data *cfqd = (struct cfq_data *) data; | |
3395 | struct cfq_queue *cfqq; | |
3396 | unsigned long flags; | |
3397 | int timed_out = 1; | |
3398 | ||
3399 | cfq_log(cfqd, "idle timer fired"); | |
3400 | ||
3401 | spin_lock_irqsave(cfqd->queue->queue_lock, flags); | |
3402 | ||
3403 | cfqq = cfqd->active_queue; | |
3404 | if (cfqq) { | |
3405 | timed_out = 0; | |
3406 | ||
3407 | /* | |
3408 | * We saw a request before the queue expired, let it through | |
3409 | */ | |
3410 | if (cfq_cfqq_must_dispatch(cfqq)) | |
3411 | goto out_kick; | |
3412 | ||
3413 | /* | |
3414 | * expired | |
3415 | */ | |
3416 | if (cfq_slice_used(cfqq)) | |
3417 | goto expire; | |
3418 | ||
3419 | /* | |
3420 | * only expire and reinvoke request handler, if there are | |
3421 | * other queues with pending requests | |
3422 | */ | |
3423 | if (!cfqd->busy_queues) | |
3424 | goto out_cont; | |
3425 | ||
3426 | /* | |
3427 | * not expired and it has a request pending, let it dispatch | |
3428 | */ | |
3429 | if (!RB_EMPTY_ROOT(&cfqq->sort_list)) | |
3430 | goto out_kick; | |
3431 | ||
3432 | /* | |
3433 | * Queue depth flag is reset only when the idle didn't succeed | |
3434 | */ | |
3435 | cfq_clear_cfqq_deep(cfqq); | |
3436 | } | |
3437 | expire: | |
3438 | cfq_slice_expired(cfqd, timed_out); | |
3439 | out_kick: | |
3440 | cfq_schedule_dispatch(cfqd); | |
3441 | out_cont: | |
3442 | spin_unlock_irqrestore(cfqd->queue->queue_lock, flags); | |
3443 | } | |
3444 | ||
3445 | static void cfq_shutdown_timer_wq(struct cfq_data *cfqd) | |
3446 | { | |
3447 | del_timer_sync(&cfqd->idle_slice_timer); | |
3448 | cancel_work_sync(&cfqd->unplug_work); | |
3449 | } | |
3450 | ||
3451 | static void cfq_put_async_queues(struct cfq_data *cfqd) | |
3452 | { | |
3453 | int i; | |
3454 | ||
3455 | for (i = 0; i < IOPRIO_BE_NR; i++) { | |
3456 | if (cfqd->async_cfqq[0][i]) | |
3457 | cfq_put_queue(cfqd->async_cfqq[0][i]); | |
3458 | if (cfqd->async_cfqq[1][i]) | |
3459 | cfq_put_queue(cfqd->async_cfqq[1][i]); | |
3460 | } | |
3461 | ||
3462 | if (cfqd->async_idle_cfqq) | |
3463 | cfq_put_queue(cfqd->async_idle_cfqq); | |
3464 | } | |
3465 | ||
3466 | static void cfq_exit_queue(struct elevator_queue *e) | |
3467 | { | |
3468 | struct cfq_data *cfqd = e->elevator_data; | |
3469 | struct request_queue *q = cfqd->queue; | |
3470 | ||
3471 | cfq_shutdown_timer_wq(cfqd); | |
3472 | ||
3473 | spin_lock_irq(q->queue_lock); | |
3474 | ||
3475 | if (cfqd->active_queue) | |
3476 | __cfq_slice_expired(cfqd, cfqd->active_queue, 0); | |
3477 | ||
3478 | while (!list_empty(&cfqd->cic_list)) { | |
3479 | struct cfq_io_context *cic = list_entry(cfqd->cic_list.next, | |
3480 | struct cfq_io_context, | |
3481 | queue_list); | |
3482 | ||
3483 | __cfq_exit_single_io_context(cfqd, cic); | |
3484 | } | |
3485 | ||
3486 | cfq_put_async_queues(cfqd); | |
3487 | cfq_release_cfq_groups(cfqd); | |
3488 | blkiocg_del_blkio_group(&cfqd->root_group.blkg); | |
3489 | ||
3490 | spin_unlock_irq(q->queue_lock); | |
3491 | ||
3492 | cfq_shutdown_timer_wq(cfqd); | |
3493 | ||
3494 | /* Wait for cfqg->blkg->key accessors to exit their grace periods. */ | |
3495 | synchronize_rcu(); | |
3496 | kfree(cfqd); | |
3497 | } | |
3498 | ||
3499 | static void *cfq_init_queue(struct request_queue *q) | |
3500 | { | |
3501 | struct cfq_data *cfqd; | |
3502 | int i, j; | |
3503 | struct cfq_group *cfqg; | |
3504 | struct cfq_rb_root *st; | |
3505 | ||
3506 | cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node); | |
3507 | if (!cfqd) | |
3508 | return NULL; | |
3509 | ||
3510 | /* Init root service tree */ | |
3511 | cfqd->grp_service_tree = CFQ_RB_ROOT; | |
3512 | ||
3513 | /* Init root group */ | |
3514 | cfqg = &cfqd->root_group; | |
3515 | for_each_cfqg_st(cfqg, i, j, st) | |
3516 | *st = CFQ_RB_ROOT; | |
3517 | RB_CLEAR_NODE(&cfqg->rb_node); | |
3518 | ||
3519 | /* Give preference to root group over other groups */ | |
3520 | cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT; | |
3521 | ||
3522 | #ifdef CONFIG_CFQ_GROUP_IOSCHED | |
3523 | /* | |
3524 | * Take a reference to root group which we never drop. This is just | |
3525 | * to make sure that cfq_put_cfqg() does not try to kfree root group | |
3526 | */ | |
3527 | atomic_set(&cfqg->ref, 1); | |
3528 | blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd, | |
3529 | 0); | |
3530 | #endif | |
3531 | /* | |
3532 | * Not strictly needed (since RB_ROOT just clears the node and we | |
3533 | * zeroed cfqd on alloc), but better be safe in case someone decides | |
3534 | * to add magic to the rb code | |
3535 | */ | |
3536 | for (i = 0; i < CFQ_PRIO_LISTS; i++) | |
3537 | cfqd->prio_trees[i] = RB_ROOT; | |
3538 | ||
3539 | /* | |
3540 | * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues. | |
3541 | * Grab a permanent reference to it, so that the normal code flow | |
3542 | * will not attempt to free it. | |
3543 | */ | |
3544 | cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0); | |
3545 | atomic_inc(&cfqd->oom_cfqq.ref); | |
3546 | cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group); | |
3547 | ||
3548 | INIT_LIST_HEAD(&cfqd->cic_list); | |
3549 | ||
3550 | cfqd->queue = q; | |
3551 | ||
3552 | init_timer(&cfqd->idle_slice_timer); | |
3553 | cfqd->idle_slice_timer.function = cfq_idle_slice_timer; | |
3554 | cfqd->idle_slice_timer.data = (unsigned long) cfqd; | |
3555 | ||
3556 | INIT_WORK(&cfqd->unplug_work, cfq_kick_queue); | |
3557 | ||
3558 | cfqd->cfq_quantum = cfq_quantum; | |
3559 | cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0]; | |
3560 | cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1]; | |
3561 | cfqd->cfq_back_max = cfq_back_max; | |
3562 | cfqd->cfq_back_penalty = cfq_back_penalty; | |
3563 | cfqd->cfq_slice[0] = cfq_slice_async; | |
3564 | cfqd->cfq_slice[1] = cfq_slice_sync; | |
3565 | cfqd->cfq_slice_async_rq = cfq_slice_async_rq; | |
3566 | cfqd->cfq_slice_idle = cfq_slice_idle; | |
3567 | cfqd->cfq_latency = 1; | |
3568 | cfqd->hw_tag = -1; | |
3569 | cfqd->last_end_sync_rq = jiffies; | |
3570 | return cfqd; | |
3571 | } | |
3572 | ||
3573 | static void cfq_slab_kill(void) | |
3574 | { | |
3575 | /* | |
3576 | * Caller already ensured that pending RCU callbacks are completed, | |
3577 | * so we should have no busy allocations at this point. | |
3578 | */ | |
3579 | if (cfq_pool) | |
3580 | kmem_cache_destroy(cfq_pool); | |
3581 | if (cfq_ioc_pool) | |
3582 | kmem_cache_destroy(cfq_ioc_pool); | |
3583 | } | |
3584 | ||
3585 | static int __init cfq_slab_setup(void) | |
3586 | { | |
3587 | cfq_pool = KMEM_CACHE(cfq_queue, 0); | |
3588 | if (!cfq_pool) | |
3589 | goto fail; | |
3590 | ||
3591 | cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0); | |
3592 | if (!cfq_ioc_pool) | |
3593 | goto fail; | |
3594 | ||
3595 | return 0; | |
3596 | fail: | |
3597 | cfq_slab_kill(); | |
3598 | return -ENOMEM; | |
3599 | } | |
3600 | ||
3601 | /* | |
3602 | * sysfs parts below --> | |
3603 | */ | |
3604 | static ssize_t | |
3605 | cfq_var_show(unsigned int var, char *page) | |
3606 | { | |
3607 | return sprintf(page, "%d\n", var); | |
3608 | } | |
3609 | ||
3610 | static ssize_t | |
3611 | cfq_var_store(unsigned int *var, const char *page, size_t count) | |
3612 | { | |
3613 | char *p = (char *) page; | |
3614 | ||
3615 | *var = simple_strtoul(p, &p, 10); | |
3616 | return count; | |
3617 | } | |
3618 | ||
3619 | #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ | |
3620 | static ssize_t __FUNC(struct elevator_queue *e, char *page) \ | |
3621 | { \ | |
3622 | struct cfq_data *cfqd = e->elevator_data; \ | |
3623 | unsigned int __data = __VAR; \ | |
3624 | if (__CONV) \ | |
3625 | __data = jiffies_to_msecs(__data); \ | |
3626 | return cfq_var_show(__data, (page)); \ | |
3627 | } | |
3628 | SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0); | |
3629 | SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1); | |
3630 | SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1); | |
3631 | SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0); | |
3632 | SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0); | |
3633 | SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1); | |
3634 | SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1); | |
3635 | SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1); | |
3636 | SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0); | |
3637 | SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0); | |
3638 | #undef SHOW_FUNCTION | |
3639 | ||
3640 | #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ | |
3641 | static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \ | |
3642 | { \ | |
3643 | struct cfq_data *cfqd = e->elevator_data; \ | |
3644 | unsigned int __data; \ | |
3645 | int ret = cfq_var_store(&__data, (page), count); \ | |
3646 | if (__data < (MIN)) \ | |
3647 | __data = (MIN); \ | |
3648 | else if (__data > (MAX)) \ | |
3649 | __data = (MAX); \ | |
3650 | if (__CONV) \ | |
3651 | *(__PTR) = msecs_to_jiffies(__data); \ | |
3652 | else \ | |
3653 | *(__PTR) = __data; \ | |
3654 | return ret; \ | |
3655 | } | |
3656 | STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0); | |
3657 | STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1, | |
3658 | UINT_MAX, 1); | |
3659 | STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1, | |
3660 | UINT_MAX, 1); | |
3661 | STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0); | |
3662 | STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1, | |
3663 | UINT_MAX, 0); | |
3664 | STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1); | |
3665 | STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1); | |
3666 | STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1); | |
3667 | STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1, | |
3668 | UINT_MAX, 0); | |
3669 | STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0); | |
3670 | #undef STORE_FUNCTION | |
3671 | ||
3672 | #define CFQ_ATTR(name) \ | |
3673 | __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store) | |
3674 | ||
3675 | static struct elv_fs_entry cfq_attrs[] = { | |
3676 | CFQ_ATTR(quantum), | |
3677 | CFQ_ATTR(fifo_expire_sync), | |
3678 | CFQ_ATTR(fifo_expire_async), | |
3679 | CFQ_ATTR(back_seek_max), | |
3680 | CFQ_ATTR(back_seek_penalty), | |
3681 | CFQ_ATTR(slice_sync), | |
3682 | CFQ_ATTR(slice_async), | |
3683 | CFQ_ATTR(slice_async_rq), | |
3684 | CFQ_ATTR(slice_idle), | |
3685 | CFQ_ATTR(low_latency), | |
3686 | __ATTR_NULL | |
3687 | }; | |
3688 | ||
3689 | static struct elevator_type iosched_cfq = { | |
3690 | .ops = { | |
3691 | .elevator_merge_fn = cfq_merge, | |
3692 | .elevator_merged_fn = cfq_merged_request, | |
3693 | .elevator_merge_req_fn = cfq_merged_requests, | |
3694 | .elevator_allow_merge_fn = cfq_allow_merge, | |
3695 | .elevator_dispatch_fn = cfq_dispatch_requests, | |
3696 | .elevator_add_req_fn = cfq_insert_request, | |
3697 | .elevator_activate_req_fn = cfq_activate_request, | |
3698 | .elevator_deactivate_req_fn = cfq_deactivate_request, | |
3699 | .elevator_queue_empty_fn = cfq_queue_empty, | |
3700 | .elevator_completed_req_fn = cfq_completed_request, | |
3701 | .elevator_former_req_fn = elv_rb_former_request, | |
3702 | .elevator_latter_req_fn = elv_rb_latter_request, | |
3703 | .elevator_set_req_fn = cfq_set_request, | |
3704 | .elevator_put_req_fn = cfq_put_request, | |
3705 | .elevator_may_queue_fn = cfq_may_queue, | |
3706 | .elevator_init_fn = cfq_init_queue, | |
3707 | .elevator_exit_fn = cfq_exit_queue, | |
3708 | .trim = cfq_free_io_context, | |
3709 | }, | |
3710 | .elevator_attrs = cfq_attrs, | |
3711 | .elevator_name = "cfq", | |
3712 | .elevator_owner = THIS_MODULE, | |
3713 | }; | |
3714 | ||
3715 | static int __init cfq_init(void) | |
3716 | { | |
3717 | /* | |
3718 | * could be 0 on HZ < 1000 setups | |
3719 | */ | |
3720 | if (!cfq_slice_async) | |
3721 | cfq_slice_async = 1; | |
3722 | if (!cfq_slice_idle) | |
3723 | cfq_slice_idle = 1; | |
3724 | ||
3725 | if (cfq_slab_setup()) | |
3726 | return -ENOMEM; | |
3727 | ||
3728 | elv_register(&iosched_cfq); | |
3729 | ||
3730 | return 0; | |
3731 | } | |
3732 | ||
3733 | static void __exit cfq_exit(void) | |
3734 | { | |
3735 | DECLARE_COMPLETION_ONSTACK(all_gone); | |
3736 | elv_unregister(&iosched_cfq); | |
3737 | ioc_gone = &all_gone; | |
3738 | /* ioc_gone's update must be visible before reading ioc_count */ | |
3739 | smp_wmb(); | |
3740 | ||
3741 | /* | |
3742 | * this also protects us from entering cfq_slab_kill() with | |
3743 | * pending RCU callbacks | |
3744 | */ | |
3745 | if (elv_ioc_count_read(cfq_ioc_count)) | |
3746 | wait_for_completion(&all_gone); | |
3747 | cfq_slab_kill(); | |
3748 | } | |
3749 | ||
3750 | module_init(cfq_init); | |
3751 | module_exit(cfq_exit); | |
3752 | ||
3753 | MODULE_AUTHOR("Jens Axboe"); | |
3754 | MODULE_LICENSE("GPL"); | |
3755 | MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler"); |