Line data Source code
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 1 : static const int cfq_quantum = 4;
23 1 : static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 : /* maximum backwards seek, in KiB */
25 1 : static const int cfq_back_max = 16 * 1024;
26 : /* penalty of a backwards seek */
27 1 : static const int cfq_back_penalty = 2;
28 1 : static const int cfq_slice_sync = HZ / 10;
29 1 : static int cfq_slice_async = HZ / 25;
30 1 : static const int cfq_slice_async_rq = 2;
31 1 : 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 : #define CFQ_SLICE_SCALE (5)
46 : #define CFQ_HW_QUEUE_MIN (5)
47 : #define CFQ_SERVICE_SHIFT 12
48 :
49 : #define CFQQ_SEEK_THR 8 * 1024
50 : #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
51 :
52 : #define RQ_CIC(rq) \
53 : ((struct cfq_io_context *) (rq)->elevator_private)
54 : #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
55 :
56 1 : static struct kmem_cache *cfq_pool;
57 1 : static struct kmem_cache *cfq_ioc_pool;
58 :
59 1 : static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
60 1 : static struct completion *ioc_gone;
61 1 : static DEFINE_SPINLOCK(ioc_gone_lock);
62 1 :
63 : #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 : #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 : #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66 :
67 : #define sample_valid(samples) ((samples) > 80)
68 : #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
69 :
70 : /*
71 : * Most of our rbtree usage is for sorting with min extraction, so
72 : * if we cache the leftmost node we don't have to walk down the tree
73 : * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74 : * move this into the elevator for the rq sorting as well.
75 : */
76 : struct cfq_rb_root {
77 : struct rb_root rb;
78 : struct rb_node *left;
79 : unsigned count;
80 : u64 min_vdisktime;
81 : struct rb_node *active;
82 : unsigned total_weight;
83 : };
84 3 : #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, 0, }
85 :
86 : /*
87 : * Per process-grouping structure
88 : */
89 : struct cfq_queue {
90 : /* reference count */
91 : atomic_t ref;
92 : /* various state flags, see below */
93 : unsigned int flags;
94 : /* parent cfq_data */
95 : struct cfq_data *cfqd;
96 : /* service_tree member */
97 : struct rb_node rb_node;
98 : /* service_tree key */
99 : unsigned long rb_key;
100 : /* prio tree member */
101 : struct rb_node p_node;
102 : /* prio tree root we belong to, if any */
103 : struct rb_root *p_root;
104 : /* sorted list of pending requests */
105 : struct rb_root sort_list;
106 : /* if fifo isn't expired, next request to serve */
107 : struct request *next_rq;
108 : /* requests queued in sort_list */
109 : int queued[2];
110 : /* currently allocated requests */
111 : int allocated[2];
112 : /* fifo list of requests in sort_list */
113 : struct list_head fifo;
114 :
115 : /* time when queue got scheduled in to dispatch first request. */
116 : unsigned long dispatch_start;
117 : unsigned int allocated_slice;
118 : /* time when first request from queue completed and slice started. */
119 : unsigned long slice_start;
120 : unsigned long slice_end;
121 : long slice_resid;
122 : unsigned int slice_dispatch;
123 :
124 : /* pending metadata requests */
125 : int meta_pending;
126 : /* number of requests that are on the dispatch list or inside driver */
127 : int dispatched;
128 :
129 : /* io prio of this group */
130 : unsigned short ioprio, org_ioprio;
131 : unsigned short ioprio_class, org_ioprio_class;
132 :
133 : unsigned int seek_samples;
134 : u64 seek_total;
135 : sector_t seek_mean;
136 : sector_t last_request_pos;
137 :
138 : pid_t pid;
139 :
140 : struct cfq_rb_root *service_tree;
141 : struct cfq_queue *new_cfqq;
142 : struct cfq_group *cfqg;
143 : struct cfq_group *orig_cfqg;
144 : /* Sectors dispatched in current dispatch round */
145 : unsigned long nr_sectors;
146 : };
147 1 :
148 : /*
149 : * First index in the service_trees.
150 : * IDLE is handled separately, so it has negative index
151 : */
152 : enum wl_prio_t {
153 1 : BE_WORKLOAD = 0,
154 : RT_WORKLOAD = 1,
155 : IDLE_WORKLOAD = 2,
156 : };
157 :
158 : /*
159 1 : * Second index in the service_trees.
160 : */
161 : enum wl_type_t {
162 : ASYNC_WORKLOAD = 0,
163 : SYNC_NOIDLE_WORKLOAD = 1,
164 : SYNC_WORKLOAD = 2
165 : };
166 :
167 : /* This is per cgroup per device grouping structure */
168 : struct cfq_group {
169 : /* group service_tree member */
170 : struct rb_node rb_node;
171 :
172 : /* group service_tree key */
173 : u64 vdisktime;
174 : unsigned int weight;
175 : bool on_st;
176 :
177 : /* number of cfqq currently on this group */
178 : int nr_cfqq;
179 :
180 : /* Per group busy queus average. Useful for workload slice calc. */
181 : unsigned int busy_queues_avg[2];
182 : /*
183 : * rr lists of queues with requests, onle rr for each priority class.
184 : * Counts are embedded in the cfq_rb_root
185 : */
186 : struct cfq_rb_root service_trees[2][3];
187 : struct cfq_rb_root service_tree_idle;
188 :
189 : unsigned long saved_workload_slice;
190 : enum wl_type_t saved_workload;
191 : enum wl_prio_t saved_serving_prio;
192 : struct blkio_group blkg;
193 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
194 1 : struct hlist_node cfqd_node;
195 : atomic_t ref;
196 : #endif
197 : };
198 :
199 : /*
200 : * Per block device queue structure
201 : */
202 : struct cfq_data {
203 : struct request_queue *queue;
204 : /* Root service tree for cfq_groups */
205 : struct cfq_rb_root grp_service_tree;
206 : struct cfq_group root_group;
207 :
208 : /*
209 : * The priority currently being served
210 : */
211 : enum wl_prio_t serving_prio;
212 : enum wl_type_t serving_type;
213 : unsigned long workload_expires;
214 : struct cfq_group *serving_group;
215 : bool noidle_tree_requires_idle;
216 :
217 : /*
218 : * Each priority tree is sorted by next_request position. These
219 : * trees are used when determining if two or more queues are
220 : * interleaving requests (see cfq_close_cooperator).
221 : */
222 : struct rb_root prio_trees[CFQ_PRIO_LISTS];
223 :
224 : unsigned int busy_queues;
225 :
226 : int rq_in_driver[2];
227 : int sync_flight;
228 :
229 : /*
230 : * queue-depth detection
231 : */
232 : int rq_queued;
233 : int hw_tag;
234 : /*
235 : * hw_tag can be
236 : * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
237 : * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
238 : * 0 => no NCQ
239 : */
240 : int hw_tag_est_depth;
241 : unsigned int hw_tag_samples;
242 :
243 : /*
244 : * idle window management
245 : */
246 : struct timer_list idle_slice_timer;
247 : struct work_struct unplug_work;
248 :
249 : struct cfq_queue *active_queue;
250 : struct cfq_io_context *active_cic;
251 :
252 : /*
253 : * async queue for each priority case
254 : */
255 : struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
256 : struct cfq_queue *async_idle_cfqq;
257 :
258 : sector_t last_position;
259 :
260 : /*
261 : * tunables, see top of file
262 : */
263 : unsigned int cfq_quantum;
264 : unsigned int cfq_fifo_expire[2];
265 : unsigned int cfq_back_penalty;
266 : unsigned int cfq_back_max;
267 : unsigned int cfq_slice[2];
268 : unsigned int cfq_slice_async_rq;
269 : unsigned int cfq_slice_idle;
270 : unsigned int cfq_latency;
271 : unsigned int cfq_group_isolation;
272 :
273 : struct list_head cic_list;
274 :
275 : /*
276 : * Fallback dummy cfqq for extreme OOM conditions
277 : */
278 : struct cfq_queue oom_cfqq;
279 :
280 : unsigned long last_delayed_sync;
281 :
282 : /* List of cfq groups being managed on this device*/
283 : struct hlist_head cfqg_list;
284 : struct rcu_head rcu;
285 : };
286 :
287 : static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
288 :
289 : static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
290 : enum wl_prio_t prio,
291 : enum wl_type_t type)
292 : {
293 0 : if (!cfqg)
294 0 : return NULL;
295 :
296 0 : if (prio == IDLE_WORKLOAD)
297 0 : return &cfqg->service_tree_idle;
298 :
299 0 : return &cfqg->service_trees[prio][type];
300 : }
301 :
302 : enum cfqq_state_flags {
303 : CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
304 : CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
305 : CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
306 : CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
307 : CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
308 : CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
309 : CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
310 : CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
311 : CFQ_CFQQ_FLAG_sync, /* synchronous queue */
312 : CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
313 : CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
314 : CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
315 : CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
316 : };
317 :
318 : #define CFQ_CFQQ_FNS(name) \
319 : static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
320 : { \
321 : (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
322 : } \
323 : static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
324 : { \
325 : (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
326 : } \
327 : static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
328 : { \
329 : return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
330 : }
331 :
332 0 : CFQ_CFQQ_FNS(on_rr);
333 0 : CFQ_CFQQ_FNS(wait_request);
334 0 : CFQ_CFQQ_FNS(must_dispatch);
335 0 : CFQ_CFQQ_FNS(must_alloc_slice);
336 0 : CFQ_CFQQ_FNS(fifo_expire);
337 0 : CFQ_CFQQ_FNS(idle_window);
338 0 : CFQ_CFQQ_FNS(prio_changed);
339 0 : CFQ_CFQQ_FNS(slice_new);
340 0 : CFQ_CFQQ_FNS(sync);
341 0 : CFQ_CFQQ_FNS(coop);
342 0 : CFQ_CFQQ_FNS(split_coop);
343 0 : CFQ_CFQQ_FNS(deep);
344 0 : CFQ_CFQQ_FNS(wait_busy);
345 0 : #undef CFQ_CFQQ_FNS
346 :
347 : #ifdef CONFIG_DEBUG_CFQ_IOSCHED
348 : #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
349 : blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
350 : cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
351 : blkg_path(&(cfqq)->cfqg->blkg), ##args);
352 :
353 : #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
354 : blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
355 : blkg_path(&(cfqg)->blkg), ##args); \
356 :
357 : #else
358 : #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
359 : blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
360 : #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
361 : #endif
362 : #define cfq_log(cfqd, fmt, args...) \
363 : blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
364 :
365 : /* Traverses through cfq group service trees */
366 : #define for_each_cfqg_st(cfqg, i, j, st) \
367 : for (i = 0; i <= IDLE_WORKLOAD; i++) \
368 : for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
369 : : &cfqg->service_tree_idle; \
370 : (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
371 : (i == IDLE_WORKLOAD && j == 0); \
372 : j++, st = i < IDLE_WORKLOAD ? \
373 : &cfqg->service_trees[i][j]: NULL) \
374 :
375 :
376 : static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
377 : {
378 0 : if (cfq_class_idle(cfqq))
379 0 : return IDLE_WORKLOAD;
380 0 : if (cfq_class_rt(cfqq))
381 0 : return RT_WORKLOAD;
382 0 : return BE_WORKLOAD;
383 : }
384 :
385 :
386 : static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
387 : {
388 0 : if (!cfq_cfqq_sync(cfqq))
389 0 : return ASYNC_WORKLOAD;
390 0 : if (!cfq_cfqq_idle_window(cfqq))
391 0 : return SYNC_NOIDLE_WORKLOAD;
392 0 : return SYNC_WORKLOAD;
393 : }
394 :
395 : static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
396 : struct cfq_data *cfqd,
397 : struct cfq_group *cfqg)
398 : {
399 0 : if (wl == IDLE_WORKLOAD)
400 0 : return cfqg->service_tree_idle.count;
401 :
402 0 : return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
403 : + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
404 : + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
405 : }
406 :
407 : static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
408 : struct cfq_group *cfqg)
409 : {
410 0 : return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
411 : + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
412 : }
413 :
414 : static void cfq_dispatch_insert(struct request_queue *, struct request *);
415 : static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
416 : struct io_context *, gfp_t);
417 : static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
418 : struct io_context *);
419 :
420 : static inline int rq_in_driver(struct cfq_data *cfqd)
421 : {
422 0 : return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
423 : }
424 :
425 : static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
426 : bool is_sync)
427 : {
428 0 : return cic->cfqq[is_sync];
429 : }
430 :
431 : static inline void cic_set_cfqq(struct cfq_io_context *cic,
432 : struct cfq_queue *cfqq, bool is_sync)
433 : {
434 0 : cic->cfqq[is_sync] = cfqq;
435 0 : }
436 :
437 : /*
438 : * We regard a request as SYNC, if it's either a read or has the SYNC bit
439 : * set (in which case it could also be direct WRITE).
440 : */
441 : static inline bool cfq_bio_sync(struct bio *bio)
442 : {
443 0 : return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
444 0 : }
445 :
446 : /*
447 : * scheduler run of queue, if there are requests pending and no one in the
448 : * driver that will restart queueing
449 : */
450 : static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
451 : {
452 0 : if (cfqd->busy_queues) {
453 : cfq_log(cfqd, "schedule dispatch");
454 0 : kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
455 0 : }
456 : }
457 :
458 : static int cfq_queue_empty(struct request_queue *q)
459 : {
460 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
461 :
462 0 : return !cfqd->rq_queued;
463 : }
464 :
465 : /*
466 : * Scale schedule slice based on io priority. Use the sync time slice only
467 : * if a queue is marked sync and has sync io queued. A sync queue with async
468 : * io only, should not get full sync slice length.
469 : */
470 : static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
471 : unsigned short prio)
472 0 : {
473 0 : const int base_slice = cfqd->cfq_slice[sync];
474 0 :
475 0 : WARN_ON(prio >= IOPRIO_BE_NR);
476 :
477 0 : return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
478 : }
479 :
480 : static inline int
481 : cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
482 : {
483 0 : return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
484 0 : }
485 :
486 : static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
487 : {
488 0 : u64 d = delta << CFQ_SERVICE_SHIFT;
489 0 :
490 0 : d = d * BLKIO_WEIGHT_DEFAULT;
491 0 : do_div(d, cfqg->weight);
492 0 : return d;
493 : }
494 :
495 : static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
496 : {
497 0 : s64 delta = (s64)(vdisktime - min_vdisktime);
498 0 : if (delta > 0)
499 0 : min_vdisktime = vdisktime;
500 :
501 0 : return min_vdisktime;
502 : }
503 :
504 : static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
505 : {
506 0 : s64 delta = (s64)(vdisktime - min_vdisktime);
507 0 : if (delta < 0)
508 0 : min_vdisktime = vdisktime;
509 :
510 0 : return min_vdisktime;
511 : }
512 :
513 : static void update_min_vdisktime(struct cfq_rb_root *st)
514 : {
515 0 : u64 vdisktime = st->min_vdisktime;
516 0 : struct cfq_group *cfqg;
517 0 :
518 0 : if (st->active) {
519 0 : cfqg = rb_entry_cfqg(st->active);
520 0 : vdisktime = cfqg->vdisktime;
521 : }
522 :
523 0 : if (st->left) {
524 0 : cfqg = rb_entry_cfqg(st->left);
525 0 : vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
526 : }
527 :
528 0 : st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
529 0 : }
530 :
531 : /*
532 : * get averaged number of queues of RT/BE priority.
533 : * average is updated, with a formula that gives more weight to higher numbers,
534 : * to quickly follows sudden increases and decrease slowly
535 : */
536 :
537 : static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
538 : struct cfq_group *cfqg, bool rt)
539 : {
540 0 : unsigned min_q, max_q;
541 0 : unsigned mult = cfq_hist_divisor - 1;
542 0 : unsigned round = cfq_hist_divisor / 2;
543 0 : unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
544 0 :
545 0 : min_q = min(cfqg->busy_queues_avg[rt], busy);
546 0 : max_q = max(cfqg->busy_queues_avg[rt], busy);
547 0 : cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
548 0 : cfq_hist_divisor;
549 0 : return cfqg->busy_queues_avg[rt];
550 : }
551 :
552 : static inline unsigned
553 : cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
554 : {
555 0 : struct cfq_rb_root *st = &cfqd->grp_service_tree;
556 :
557 0 : return cfq_target_latency * cfqg->weight / st->total_weight;
558 : }
559 :
560 : static inline void
561 : cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
562 : {
563 0 : unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
564 0 : if (cfqd->cfq_latency) {
565 0 : /*
566 0 : * interested queues (we consider only the ones with the same
567 0 : * priority class in the cfq group)
568 0 : */
569 0 : unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
570 0 : cfq_class_rt(cfqq));
571 0 : unsigned sync_slice = cfqd->cfq_slice[1];
572 0 : unsigned expect_latency = sync_slice * iq;
573 0 : unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
574 0 :
575 0 : if (expect_latency > group_slice) {
576 0 : unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
577 : /* scale low_slice according to IO priority
578 : * and sync vs async */
579 : unsigned low_slice =
580 0 : min(slice, base_low_slice * slice / sync_slice);
581 : /* the adapted slice value is scaled to fit all iqs
582 : * into the target latency */
583 0 : slice = max(slice * group_slice / expect_latency,
584 : low_slice);
585 : }
586 : }
587 0 : cfqq->slice_start = jiffies;
588 0 : cfqq->slice_end = jiffies + slice;
589 0 : cfqq->allocated_slice = slice;
590 0 : cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
591 : }
592 :
593 : /*
594 : * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
595 : * isn't valid until the first request from the dispatch is activated
596 : * and the slice time set.
597 : */
598 : static inline bool cfq_slice_used(struct cfq_queue *cfqq)
599 : {
600 0 : if (cfq_cfqq_slice_new(cfqq))
601 0 : return 0;
602 0 : if (time_before(jiffies, cfqq->slice_end))
603 0 : return 0;
604 :
605 0 : return 1;
606 : }
607 :
608 : /*
609 : * Lifted from AS - choose which of rq1 and rq2 that is best served now.
610 : * We choose the request that is closest to the head right now. Distance
611 : * behind the head is penalized and only allowed to a certain extent.
612 : */
613 : static struct request *
614 : cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
615 : {
616 0 : sector_t s1, s2, d1 = 0, d2 = 0;
617 0 : unsigned long back_max;
618 0 : #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
619 0 : #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
620 0 : unsigned wrap = 0; /* bit mask: requests behind the disk head? */
621 0 :
622 0 : if (rq1 == NULL || rq1 == rq2)
623 0 : return rq2;
624 0 : if (rq2 == NULL)
625 0 : return rq1;
626 0 :
627 0 : if (rq_is_sync(rq1) && !rq_is_sync(rq2))
628 0 : return rq1;
629 0 : else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
630 0 : return rq2;
631 0 : if (rq_is_meta(rq1) && !rq_is_meta(rq2))
632 0 : return rq1;
633 0 : else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
634 0 : return rq2;
635 :
636 0 : s1 = blk_rq_pos(rq1);
637 0 : s2 = blk_rq_pos(rq2);
638 :
639 : /*
640 : * by definition, 1KiB is 2 sectors
641 : */
642 0 : back_max = cfqd->cfq_back_max * 2;
643 :
644 : /*
645 : * Strict one way elevator _except_ in the case where we allow
646 : * short backward seeks which are biased as twice the cost of a
647 : * similar forward seek.
648 : */
649 0 : if (s1 >= last)
650 0 : d1 = s1 - last;
651 0 : else if (s1 + back_max >= last)
652 0 : d1 = (last - s1) * cfqd->cfq_back_penalty;
653 : else
654 0 : wrap |= CFQ_RQ1_WRAP;
655 :
656 0 : if (s2 >= last)
657 0 : d2 = s2 - last;
658 0 : else if (s2 + back_max >= last)
659 0 : d2 = (last - s2) * cfqd->cfq_back_penalty;
660 : else
661 0 : wrap |= CFQ_RQ2_WRAP;
662 :
663 : /* Found required data */
664 :
665 : /*
666 : * By doing switch() on the bit mask "wrap" we avoid having to
667 : * check two variables for all permutations: --> faster!
668 : */
669 : switch (wrap) {
670 0 : case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
671 0 : if (d1 < d2)
672 0 : return rq1;
673 0 : else if (d2 < d1)
674 0 : return rq2;
675 : else {
676 0 : if (s1 >= s2)
677 0 : return rq1;
678 : else
679 0 : return rq2;
680 : }
681 0 :
682 0 : case CFQ_RQ2_WRAP:
683 0 : return rq1;
684 0 : case CFQ_RQ1_WRAP:
685 0 : return rq2;
686 0 : case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
687 : default:
688 0 : /*
689 : * Since both rqs are wrapped,
690 : * start with the one that's further behind head
691 : * (--> only *one* back seek required),
692 : * since back seek takes more time than forward.
693 : */
694 0 : if (s1 <= s2)
695 0 : return rq1;
696 : else
697 0 : return rq2;
698 : }
699 : }
700 :
701 : /*
702 : * The below is leftmost cache rbtree addon
703 : */
704 : static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
705 : {
706 0 : /* Service tree is empty */
707 0 : if (!root->count)
708 0 : return NULL;
709 :
710 0 : if (!root->left)
711 0 : root->left = rb_first(&root->rb);
712 :
713 0 : if (root->left)
714 0 : return rb_entry(root->left, struct cfq_queue, rb_node);
715 :
716 0 : return NULL;
717 : }
718 :
719 : static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
720 : {
721 0 : if (!root->left)
722 0 : root->left = rb_first(&root->rb);
723 :
724 0 : if (root->left)
725 0 : return rb_entry_cfqg(root->left);
726 :
727 0 : return NULL;
728 : }
729 :
730 : static void rb_erase_init(struct rb_node *n, struct rb_root *root)
731 : {
732 0 : rb_erase(n, root);
733 0 : RB_CLEAR_NODE(n);
734 0 : }
735 :
736 : static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
737 : {
738 0 : if (root->left == n)
739 0 : root->left = NULL;
740 0 : rb_erase_init(n, &root->rb);
741 0 : --root->count;
742 0 : }
743 :
744 : /*
745 : * would be nice to take fifo expire time into account as well
746 : */
747 : static struct request *
748 : cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
749 : struct request *last)
750 : {
751 0 : struct rb_node *rbnext = rb_next(&last->rb_node);
752 0 : struct rb_node *rbprev = rb_prev(&last->rb_node);
753 0 : struct request *next = NULL, *prev = NULL;
754 0 :
755 0 : BUG_ON(RB_EMPTY_NODE(&last->rb_node));
756 0 :
757 0 : if (rbprev)
758 0 : prev = rb_entry_rq(rbprev);
759 0 :
760 0 : if (rbnext)
761 0 : next = rb_entry_rq(rbnext);
762 0 : else {
763 0 : rbnext = rb_first(&cfqq->sort_list);
764 0 : if (rbnext && rbnext != &last->rb_node)
765 0 : next = rb_entry_rq(rbnext);
766 : }
767 :
768 0 : return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
769 : }
770 :
771 : static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
772 : struct cfq_queue *cfqq)
773 0 : {
774 0 : /*
775 0 : * just an approximation, should be ok.
776 : */
777 0 : return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
778 : cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
779 : }
780 :
781 : static inline s64
782 : cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
783 : {
784 0 : return cfqg->vdisktime - st->min_vdisktime;
785 : }
786 :
787 : static void
788 : __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
789 : {
790 0 : struct rb_node **node = &st->rb.rb_node;
791 0 : struct rb_node *parent = NULL;
792 0 : struct cfq_group *__cfqg;
793 0 : s64 key = cfqg_key(st, cfqg);
794 0 : int left = 1;
795 0 :
796 0 : while (*node != NULL) {
797 0 : parent = *node;
798 0 : __cfqg = rb_entry_cfqg(parent);
799 :
800 0 : if (key < cfqg_key(st, __cfqg))
801 0 : node = &parent->rb_left;
802 : else {
803 0 : node = &parent->rb_right;
804 0 : left = 0;
805 : }
806 0 : }
807 :
808 0 : if (left)
809 0 : st->left = &cfqg->rb_node;
810 :
811 0 : rb_link_node(&cfqg->rb_node, parent, node);
812 0 : rb_insert_color(&cfqg->rb_node, &st->rb);
813 0 : }
814 :
815 : static void
816 : cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
817 : {
818 0 : struct cfq_rb_root *st = &cfqd->grp_service_tree;
819 0 : struct cfq_group *__cfqg;
820 0 : struct rb_node *n;
821 0 :
822 0 : cfqg->nr_cfqq++;
823 0 : if (cfqg->on_st)
824 0 : return;
825 :
826 : /*
827 : * Currently put the group at the end. Later implement something
828 : * so that groups get lesser vtime based on their weights, so that
829 : * if group does not loose all if it was not continously backlogged.
830 : */
831 0 : n = rb_last(&st->rb);
832 0 : if (n) {
833 0 : __cfqg = rb_entry_cfqg(n);
834 0 : cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
835 : } else
836 0 : cfqg->vdisktime = st->min_vdisktime;
837 :
838 0 : __cfq_group_service_tree_add(st, cfqg);
839 0 : cfqg->on_st = true;
840 0 : st->total_weight += cfqg->weight;
841 0 : }
842 :
843 : static void
844 : cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
845 : {
846 0 : struct cfq_rb_root *st = &cfqd->grp_service_tree;
847 0 :
848 0 : if (st->active == &cfqg->rb_node)
849 0 : st->active = NULL;
850 :
851 0 : BUG_ON(cfqg->nr_cfqq < 1);
852 0 : cfqg->nr_cfqq--;
853 :
854 : /* If there are other cfq queues under this group, don't delete it */
855 0 : if (cfqg->nr_cfqq)
856 0 : return;
857 :
858 : cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
859 0 : cfqg->on_st = false;
860 0 : st->total_weight -= cfqg->weight;
861 0 : if (!RB_EMPTY_NODE(&cfqg->rb_node))
862 0 : cfq_rb_erase(&cfqg->rb_node, st);
863 0 : cfqg->saved_workload_slice = 0;
864 0 : blkiocg_update_blkio_group_dequeue_stats(&cfqg->blkg, 1);
865 0 : }
866 :
867 : static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
868 : {
869 0 : unsigned int slice_used;
870 0 :
871 0 : /*
872 : * Queue got expired before even a single request completed or
873 : * got expired immediately after first request completion.
874 : */
875 0 : if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
876 : /*
877 : * Also charge the seek time incurred to the group, otherwise
878 : * if there are mutiple queues in the group, each can dispatch
879 : * a single request on seeky media and cause lots of seek time
880 : * and group will never know it.
881 : */
882 0 : slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
883 : 1);
884 : } else {
885 0 : slice_used = jiffies - cfqq->slice_start;
886 0 : if (slice_used > cfqq->allocated_slice)
887 0 : slice_used = cfqq->allocated_slice;
888 : }
889 :
890 : cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u sect=%lu", slice_used,
891 : cfqq->nr_sectors);
892 0 : return slice_used;
893 : }
894 :
895 : static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
896 : struct cfq_queue *cfqq)
897 0 : {
898 0 : struct cfq_rb_root *st = &cfqd->grp_service_tree;
899 0 : unsigned int used_sl, charge_sl;
900 0 : int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
901 0 : - cfqg->service_tree_idle.count;
902 0 :
903 0 : BUG_ON(nr_sync < 0);
904 0 : used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
905 :
906 0 : if (!cfq_cfqq_sync(cfqq) && !nr_sync)
907 0 : charge_sl = cfqq->allocated_slice;
908 :
909 : /* Can't update vdisktime while group is on service tree */
910 0 : cfq_rb_erase(&cfqg->rb_node, st);
911 0 : cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
912 0 : __cfq_group_service_tree_add(st, cfqg);
913 :
914 : /* This group is being expired. Save the context */
915 0 : if (time_after(cfqd->workload_expires, jiffies)) {
916 0 : cfqg->saved_workload_slice = cfqd->workload_expires
917 : - jiffies;
918 0 : cfqg->saved_workload = cfqd->serving_type;
919 0 : cfqg->saved_serving_prio = cfqd->serving_prio;
920 : } else
921 0 : cfqg->saved_workload_slice = 0;
922 :
923 : cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
924 : st->min_vdisktime);
925 0 : blkiocg_update_blkio_group_stats(&cfqg->blkg, used_sl,
926 : cfqq->nr_sectors);
927 0 : }
928 :
929 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
930 : static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
931 : {
932 : if (blkg)
933 : return container_of(blkg, struct cfq_group, blkg);
934 : return NULL;
935 : }
936 :
937 : void
938 : cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
939 : {
940 : cfqg_of_blkg(blkg)->weight = weight;
941 : }
942 :
943 : static struct cfq_group *
944 : cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
945 : {
946 : struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
947 : struct cfq_group *cfqg = NULL;
948 : void *key = cfqd;
949 : int i, j;
950 : struct cfq_rb_root *st;
951 : struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
952 : unsigned int major, minor;
953 :
954 : /* Do we need to take this reference */
955 : if (!blkiocg_css_tryget(blkcg))
956 : return NULL;;
957 :
958 : cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
959 : if (cfqg || !create)
960 : goto done;
961 :
962 : cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
963 : if (!cfqg)
964 : goto done;
965 :
966 : cfqg->weight = blkcg->weight;
967 : for_each_cfqg_st(cfqg, i, j, st)
968 : *st = CFQ_RB_ROOT;
969 : RB_CLEAR_NODE(&cfqg->rb_node);
970 :
971 : /*
972 : * Take the initial reference that will be released on destroy
973 : * This can be thought of a joint reference by cgroup and
974 : * elevator which will be dropped by either elevator exit
975 : * or cgroup deletion path depending on who is exiting first.
976 : */
977 : atomic_set(&cfqg->ref, 1);
978 :
979 : /* Add group onto cgroup list */
980 : sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
981 : blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
982 : MKDEV(major, minor));
983 :
984 : /* Add group on cfqd list */
985 : hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
986 :
987 : done:
988 : blkiocg_css_put(blkcg);
989 : return cfqg;
990 : }
991 :
992 : /*
993 : * Search for the cfq group current task belongs to. If create = 1, then also
994 : * create the cfq group if it does not exist. request_queue lock must be held.
995 : */
996 : static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
997 : {
998 : struct cgroup *cgroup;
999 : struct cfq_group *cfqg = NULL;
1000 :
1001 : rcu_read_lock();
1002 : cgroup = task_cgroup(current, blkio_subsys_id);
1003 : cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1004 : if (!cfqg && create)
1005 : cfqg = &cfqd->root_group;
1006 : rcu_read_unlock();
1007 : return cfqg;
1008 : }
1009 :
1010 : static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1011 : {
1012 : /* Currently, all async queues are mapped to root group */
1013 : if (!cfq_cfqq_sync(cfqq))
1014 : cfqg = &cfqq->cfqd->root_group;
1015 :
1016 : cfqq->cfqg = cfqg;
1017 : /* cfqq reference on cfqg */
1018 : atomic_inc(&cfqq->cfqg->ref);
1019 : }
1020 :
1021 : static void cfq_put_cfqg(struct cfq_group *cfqg)
1022 : {
1023 : struct cfq_rb_root *st;
1024 : int i, j;
1025 :
1026 : BUG_ON(atomic_read(&cfqg->ref) <= 0);
1027 : if (!atomic_dec_and_test(&cfqg->ref))
1028 : return;
1029 : for_each_cfqg_st(cfqg, i, j, st)
1030 : BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1031 : kfree(cfqg);
1032 : }
1033 :
1034 : static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1035 : {
1036 : /* Something wrong if we are trying to remove same group twice */
1037 : BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1038 :
1039 : hlist_del_init(&cfqg->cfqd_node);
1040 :
1041 : /*
1042 : * Put the reference taken at the time of creation so that when all
1043 : * queues are gone, group can be destroyed.
1044 : */
1045 : cfq_put_cfqg(cfqg);
1046 : }
1047 :
1048 : static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1049 : {
1050 : struct hlist_node *pos, *n;
1051 : struct cfq_group *cfqg;
1052 :
1053 : hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1054 : /*
1055 : * If cgroup removal path got to blk_group first and removed
1056 : * it from cgroup list, then it will take care of destroying
1057 : * cfqg also.
1058 : */
1059 : if (!blkiocg_del_blkio_group(&cfqg->blkg))
1060 : cfq_destroy_cfqg(cfqd, cfqg);
1061 : }
1062 : }
1063 :
1064 : /*
1065 : * Blk cgroup controller notification saying that blkio_group object is being
1066 : * delinked as associated cgroup object is going away. That also means that
1067 : * no new IO will come in this group. So get rid of this group as soon as
1068 : * any pending IO in the group is finished.
1069 : *
1070 : * This function is called under rcu_read_lock(). key is the rcu protected
1071 : * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1072 : * read lock.
1073 : *
1074 : * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1075 : * it should not be NULL as even if elevator was exiting, cgroup deltion
1076 : * path got to it first.
1077 : */
1078 : void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1079 : {
1080 : unsigned long flags;
1081 : struct cfq_data *cfqd = key;
1082 :
1083 : spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1084 : cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1085 : spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1086 : }
1087 :
1088 : #else /* GROUP_IOSCHED */
1089 : static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1090 : {
1091 0 : return &cfqd->root_group;
1092 : }
1093 : static inline void
1094 : cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1095 0 : cfqq->cfqg = cfqg;
1096 0 : }
1097 :
1098 : static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1099 : static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1100 0 :
1101 0 : #endif /* GROUP_IOSCHED */
1102 :
1103 : /*
1104 : * The cfqd->service_trees holds all pending cfq_queue's that have
1105 : * requests waiting to be processed. It is sorted in the order that
1106 : * we will service the queues.
1107 : */
1108 : static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1109 : bool add_front)
1110 : {
1111 0 : struct rb_node **p, *parent;
1112 0 : struct cfq_queue *__cfqq;
1113 0 : unsigned long rb_key;
1114 0 : struct cfq_rb_root *service_tree;
1115 0 : int left;
1116 0 : int new_cfqq = 1;
1117 0 : int group_changed = 0;
1118 0 :
1119 0 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
1120 0 : if (!cfqd->cfq_group_isolation
1121 0 : && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1122 0 : && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1123 0 : /* Move this cfq to root group */
1124 0 : cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1125 : if (!RB_EMPTY_NODE(&cfqq->rb_node))
1126 : cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1127 : cfqq->orig_cfqg = cfqq->cfqg;
1128 : cfqq->cfqg = &cfqd->root_group;
1129 : atomic_inc(&cfqd->root_group.ref);
1130 : group_changed = 1;
1131 : } else if (!cfqd->cfq_group_isolation
1132 : && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1133 : /* cfqq is sequential now needs to go to its original group */
1134 : BUG_ON(cfqq->cfqg != &cfqd->root_group);
1135 : if (!RB_EMPTY_NODE(&cfqq->rb_node))
1136 : cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1137 : cfq_put_cfqg(cfqq->cfqg);
1138 : cfqq->cfqg = cfqq->orig_cfqg;
1139 : cfqq->orig_cfqg = NULL;
1140 : group_changed = 1;
1141 : cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1142 : }
1143 : #endif
1144 :
1145 0 : service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1146 : cfqq_type(cfqq));
1147 0 : if (cfq_class_idle(cfqq)) {
1148 0 : rb_key = CFQ_IDLE_DELAY;
1149 0 : parent = rb_last(&service_tree->rb);
1150 0 : if (parent && parent != &cfqq->rb_node) {
1151 0 : __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1152 0 : rb_key += __cfqq->rb_key;
1153 : } else
1154 0 : rb_key += jiffies;
1155 0 : } else if (!add_front) {
1156 : /*
1157 : * Get our rb key offset. Subtract any residual slice
1158 : * value carried from last service. A negative resid
1159 : * count indicates slice overrun, and this should position
1160 : * the next service time further away in the tree.
1161 : */
1162 0 : rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1163 0 : rb_key -= cfqq->slice_resid;
1164 0 : cfqq->slice_resid = 0;
1165 : } else {
1166 0 : rb_key = -HZ;
1167 0 : __cfqq = cfq_rb_first(service_tree);
1168 0 : rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1169 : }
1170 :
1171 0 : if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1172 0 : new_cfqq = 0;
1173 : /*
1174 : * same position, nothing more to do
1175 : */
1176 0 : if (rb_key == cfqq->rb_key &&
1177 : cfqq->service_tree == service_tree)
1178 0 : return;
1179 :
1180 0 : cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1181 0 : cfqq->service_tree = NULL;
1182 : }
1183 :
1184 0 : left = 1;
1185 0 : parent = NULL;
1186 0 : cfqq->service_tree = service_tree;
1187 0 : p = &service_tree->rb.rb_node;
1188 0 : while (*p) {
1189 0 : struct rb_node **n;
1190 0 :
1191 0 : parent = *p;
1192 0 : __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1193 :
1194 : /*
1195 : * sort by key, that represents service time.
1196 : */
1197 0 : if (time_before(rb_key, __cfqq->rb_key))
1198 0 : n = &(*p)->rb_left;
1199 : else {
1200 0 : n = &(*p)->rb_right;
1201 0 : left = 0;
1202 : }
1203 :
1204 0 : p = n;
1205 0 : }
1206 :
1207 0 : if (left)
1208 0 : service_tree->left = &cfqq->rb_node;
1209 :
1210 0 : cfqq->rb_key = rb_key;
1211 0 : rb_link_node(&cfqq->rb_node, parent, p);
1212 0 : rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1213 0 : service_tree->count++;
1214 0 : if ((add_front || !new_cfqq) && !group_changed)
1215 0 : return;
1216 0 : cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1217 0 : }
1218 :
1219 : static struct cfq_queue *
1220 : cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1221 : sector_t sector, struct rb_node **ret_parent,
1222 : struct rb_node ***rb_link)
1223 : {
1224 0 : struct rb_node **p, *parent;
1225 0 : struct cfq_queue *cfqq = NULL;
1226 0 :
1227 0 : parent = NULL;
1228 0 : p = &root->rb_node;
1229 0 : while (*p) {
1230 0 : struct rb_node **n;
1231 0 :
1232 0 : parent = *p;
1233 0 : cfqq = rb_entry(parent, struct cfq_queue, p_node);
1234 :
1235 0 : /*
1236 : * Sort strictly based on sector. Smallest to the left,
1237 : * largest to the right.
1238 : */
1239 0 : if (sector > blk_rq_pos(cfqq->next_rq))
1240 0 : n = &(*p)->rb_right;
1241 0 : else if (sector < blk_rq_pos(cfqq->next_rq))
1242 0 : n = &(*p)->rb_left;
1243 : else
1244 0 : break;
1245 0 : p = n;
1246 0 : cfqq = NULL;
1247 0 : }
1248 :
1249 0 : *ret_parent = parent;
1250 0 : if (rb_link)
1251 0 : *rb_link = p;
1252 0 : return cfqq;
1253 : }
1254 :
1255 : static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1256 : {
1257 0 : struct rb_node **p, *parent;
1258 0 : struct cfq_queue *__cfqq;
1259 0 :
1260 0 : if (cfqq->p_root) {
1261 0 : rb_erase(&cfqq->p_node, cfqq->p_root);
1262 0 : cfqq->p_root = NULL;
1263 : }
1264 :
1265 0 : if (cfq_class_idle(cfqq))
1266 0 : return;
1267 0 : if (!cfqq->next_rq)
1268 0 : return;
1269 :
1270 0 : cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1271 0 : __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1272 : blk_rq_pos(cfqq->next_rq), &parent, &p);
1273 0 : if (!__cfqq) {
1274 0 : rb_link_node(&cfqq->p_node, parent, p);
1275 0 : rb_insert_color(&cfqq->p_node, cfqq->p_root);
1276 : } else
1277 0 : cfqq->p_root = NULL;
1278 0 : }
1279 :
1280 : /*
1281 : * Update cfqq's position in the service tree.
1282 : */
1283 : static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1284 : {
1285 0 : /*
1286 : * Resorting requires the cfqq to be on the RR list already.
1287 : */
1288 0 : if (cfq_cfqq_on_rr(cfqq)) {
1289 0 : cfq_service_tree_add(cfqd, cfqq, 0);
1290 0 : cfq_prio_tree_add(cfqd, cfqq);
1291 : }
1292 0 : }
1293 :
1294 : /*
1295 : * add to busy list of queues for service, trying to be fair in ordering
1296 : * the pending list according to last request service
1297 : */
1298 : static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1299 : {
1300 0 : cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1301 0 : BUG_ON(cfq_cfqq_on_rr(cfqq));
1302 0 : cfq_mark_cfqq_on_rr(cfqq);
1303 0 : cfqd->busy_queues++;
1304 :
1305 0 : cfq_resort_rr_list(cfqd, cfqq);
1306 0 : }
1307 :
1308 : /*
1309 : * Called when the cfqq no longer has requests pending, remove it from
1310 : * the service tree.
1311 : */
1312 : static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1313 : {
1314 0 : cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1315 0 : BUG_ON(!cfq_cfqq_on_rr(cfqq));
1316 0 : cfq_clear_cfqq_on_rr(cfqq);
1317 :
1318 0 : if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1319 0 : cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1320 0 : cfqq->service_tree = NULL;
1321 : }
1322 0 : if (cfqq->p_root) {
1323 0 : rb_erase(&cfqq->p_node, cfqq->p_root);
1324 0 : cfqq->p_root = NULL;
1325 : }
1326 :
1327 0 : cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1328 0 : BUG_ON(!cfqd->busy_queues);
1329 0 : cfqd->busy_queues--;
1330 0 : }
1331 :
1332 : /*
1333 : * rb tree support functions
1334 : */
1335 : static void cfq_del_rq_rb(struct request *rq)
1336 : {
1337 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
1338 0 : const int sync = rq_is_sync(rq);
1339 0 :
1340 0 : BUG_ON(!cfqq->queued[sync]);
1341 0 : cfqq->queued[sync]--;
1342 :
1343 0 : elv_rb_del(&cfqq->sort_list, rq);
1344 :
1345 0 : if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1346 : /*
1347 : * Queue will be deleted from service tree when we actually
1348 : * expire it later. Right now just remove it from prio tree
1349 : * as it is empty.
1350 : */
1351 0 : if (cfqq->p_root) {
1352 0 : rb_erase(&cfqq->p_node, cfqq->p_root);
1353 0 : cfqq->p_root = NULL;
1354 : }
1355 0 : }
1356 : }
1357 :
1358 : static void cfq_add_rq_rb(struct request *rq)
1359 : {
1360 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
1361 0 : struct cfq_data *cfqd = cfqq->cfqd;
1362 0 : struct request *__alias, *prev;
1363 0 :
1364 0 : cfqq->queued[rq_is_sync(rq)]++;
1365 0 :
1366 0 : /*
1367 : * looks a little odd, but the first insert might return an alias.
1368 : * if that happens, put the alias on the dispatch list
1369 : */
1370 0 : while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1371 0 : cfq_dispatch_insert(cfqd->queue, __alias);
1372 0 :
1373 0 : if (!cfq_cfqq_on_rr(cfqq))
1374 0 : cfq_add_cfqq_rr(cfqd, cfqq);
1375 :
1376 : /*
1377 : * check if this request is a better next-serve candidate
1378 : */
1379 0 : prev = cfqq->next_rq;
1380 0 : cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1381 :
1382 : /*
1383 : * adjust priority tree position, if ->next_rq changes
1384 : */
1385 0 : if (prev != cfqq->next_rq)
1386 0 : cfq_prio_tree_add(cfqd, cfqq);
1387 :
1388 0 : BUG_ON(!cfqq->next_rq);
1389 : }
1390 0 :
1391 : static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1392 : {
1393 0 : elv_rb_del(&cfqq->sort_list, rq);
1394 0 : cfqq->queued[rq_is_sync(rq)]--;
1395 0 : cfq_add_rq_rb(rq);
1396 0 : }
1397 :
1398 : static struct request *
1399 : cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1400 : {
1401 0 : struct task_struct *tsk = current;
1402 0 : struct cfq_io_context *cic;
1403 0 : struct cfq_queue *cfqq;
1404 0 :
1405 0 : cic = cfq_cic_lookup(cfqd, tsk->io_context);
1406 0 : if (!cic)
1407 0 : return NULL;
1408 :
1409 0 : cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1410 0 : if (cfqq) {
1411 0 : sector_t sector = bio->bi_sector + bio_sectors(bio);
1412 :
1413 0 : return elv_rb_find(&cfqq->sort_list, sector);
1414 : }
1415 :
1416 0 : return NULL;
1417 : }
1418 :
1419 : static void cfq_activate_request(struct request_queue *q, struct request *rq)
1420 : {
1421 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
1422 0 :
1423 0 : cfqd->rq_in_driver[rq_is_sync(rq)]++;
1424 0 : cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1425 : rq_in_driver(cfqd));
1426 :
1427 0 : cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1428 0 : }
1429 :
1430 : static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1431 : {
1432 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
1433 0 : const int sync = rq_is_sync(rq);
1434 0 :
1435 0 : WARN_ON(!cfqd->rq_in_driver[sync]);
1436 0 : cfqd->rq_in_driver[sync]--;
1437 0 : cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1438 : rq_in_driver(cfqd));
1439 : }
1440 :
1441 : static void cfq_remove_request(struct request *rq)
1442 : {
1443 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
1444 0 :
1445 0 : if (cfqq->next_rq == rq)
1446 0 : cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1447 :
1448 0 : list_del_init(&rq->queuelist);
1449 0 : cfq_del_rq_rb(rq);
1450 :
1451 0 : cfqq->cfqd->rq_queued--;
1452 0 : if (rq_is_meta(rq)) {
1453 0 : WARN_ON(!cfqq->meta_pending);
1454 0 : cfqq->meta_pending--;
1455 : }
1456 0 : }
1457 :
1458 : static int cfq_merge(struct request_queue *q, struct request **req,
1459 : struct bio *bio)
1460 0 : {
1461 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
1462 0 : struct request *__rq;
1463 :
1464 0 : __rq = cfq_find_rq_fmerge(cfqd, bio);
1465 0 : if (__rq && elv_rq_merge_ok(__rq, bio)) {
1466 0 : *req = __rq;
1467 0 : return ELEVATOR_FRONT_MERGE;
1468 : }
1469 :
1470 0 : return ELEVATOR_NO_MERGE;
1471 : }
1472 :
1473 : static void cfq_merged_request(struct request_queue *q, struct request *req,
1474 : int type)
1475 0 : {
1476 0 : if (type == ELEVATOR_FRONT_MERGE) {
1477 0 : struct cfq_queue *cfqq = RQ_CFQQ(req);
1478 :
1479 0 : cfq_reposition_rq_rb(cfqq, req);
1480 : }
1481 0 : }
1482 :
1483 : static void
1484 : cfq_merged_requests(struct request_queue *q, struct request *rq,
1485 : struct request *next)
1486 0 : {
1487 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
1488 0 : /*
1489 : * reposition in fifo if next is older than rq
1490 : */
1491 0 : if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1492 0 : time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1493 0 : list_move(&rq->queuelist, &next->queuelist);
1494 0 : rq_set_fifo_time(rq, rq_fifo_time(next));
1495 : }
1496 :
1497 0 : if (cfqq->next_rq == next)
1498 0 : cfqq->next_rq = rq;
1499 0 : cfq_remove_request(next);
1500 0 : }
1501 :
1502 : static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1503 : struct bio *bio)
1504 0 : {
1505 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
1506 0 : struct cfq_io_context *cic;
1507 0 : struct cfq_queue *cfqq;
1508 0 :
1509 0 : /*
1510 0 : * Disallow merge of a sync bio into an async request.
1511 0 : */
1512 0 : if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1513 0 : return false;
1514 :
1515 : /*
1516 : * Lookup the cfqq that this bio will be queued with. Allow
1517 : * merge only if rq is queued there.
1518 : */
1519 0 : cic = cfq_cic_lookup(cfqd, current->io_context);
1520 0 : if (!cic)
1521 0 : return false;
1522 :
1523 0 : cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1524 0 : return cfqq == RQ_CFQQ(rq);
1525 : }
1526 :
1527 : static void __cfq_set_active_queue(struct cfq_data *cfqd,
1528 : struct cfq_queue *cfqq)
1529 : {
1530 0 : if (cfqq) {
1531 : cfq_log_cfqq(cfqd, cfqq, "set_active");
1532 0 : cfqq->slice_start = 0;
1533 0 : cfqq->dispatch_start = jiffies;
1534 0 : cfqq->allocated_slice = 0;
1535 0 : cfqq->slice_end = 0;
1536 0 : cfqq->slice_dispatch = 0;
1537 0 : cfqq->nr_sectors = 0;
1538 :
1539 0 : cfq_clear_cfqq_wait_request(cfqq);
1540 0 : cfq_clear_cfqq_must_dispatch(cfqq);
1541 0 : cfq_clear_cfqq_must_alloc_slice(cfqq);
1542 0 : cfq_clear_cfqq_fifo_expire(cfqq);
1543 0 : cfq_mark_cfqq_slice_new(cfqq);
1544 :
1545 0 : del_timer(&cfqd->idle_slice_timer);
1546 : }
1547 :
1548 0 : cfqd->active_queue = cfqq;
1549 0 : }
1550 :
1551 : /*
1552 : * current cfqq expired its slice (or was too idle), select new one
1553 : */
1554 : static void
1555 : __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1556 : bool timed_out)
1557 0 : {
1558 0 : cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1559 0 :
1560 0 : if (cfq_cfqq_wait_request(cfqq))
1561 0 : del_timer(&cfqd->idle_slice_timer);
1562 :
1563 0 : cfq_clear_cfqq_wait_request(cfqq);
1564 0 : cfq_clear_cfqq_wait_busy(cfqq);
1565 :
1566 : /*
1567 : * If this cfqq is shared between multiple processes, check to
1568 : * make sure that those processes are still issuing I/Os within
1569 : * the mean seek distance. If not, it may be time to break the
1570 : * queues apart again.
1571 : */
1572 0 : if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1573 0 : cfq_mark_cfqq_split_coop(cfqq);
1574 :
1575 : /*
1576 : * store what was left of this slice, if the queue idled/timed out
1577 : */
1578 0 : if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1579 0 : cfqq->slice_resid = cfqq->slice_end - jiffies;
1580 : cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1581 : }
1582 :
1583 0 : cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1584 :
1585 0 : if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1586 0 : cfq_del_cfqq_rr(cfqd, cfqq);
1587 :
1588 0 : cfq_resort_rr_list(cfqd, cfqq);
1589 :
1590 0 : if (cfqq == cfqd->active_queue)
1591 0 : cfqd->active_queue = NULL;
1592 :
1593 0 : if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1594 0 : cfqd->grp_service_tree.active = NULL;
1595 :
1596 0 : if (cfqd->active_cic) {
1597 0 : put_io_context(cfqd->active_cic->ioc);
1598 0 : cfqd->active_cic = NULL;
1599 : }
1600 0 : }
1601 :
1602 : static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1603 : {
1604 0 : struct cfq_queue *cfqq = cfqd->active_queue;
1605 :
1606 0 : if (cfqq)
1607 0 : __cfq_slice_expired(cfqd, cfqq, timed_out);
1608 0 : }
1609 :
1610 : /*
1611 : * Get next queue for service. Unless we have a queue preemption,
1612 : * we'll simply select the first cfqq in the service tree.
1613 : */
1614 : static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1615 : {
1616 0 : struct cfq_rb_root *service_tree =
1617 0 : service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1618 0 : cfqd->serving_type);
1619 :
1620 0 : if (!cfqd->rq_queued)
1621 0 : return NULL;
1622 :
1623 : /* There is nothing to dispatch */
1624 0 : if (!service_tree)
1625 0 : return NULL;
1626 0 : if (RB_EMPTY_ROOT(&service_tree->rb))
1627 0 : return NULL;
1628 0 : return cfq_rb_first(service_tree);
1629 : }
1630 :
1631 : static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1632 : {
1633 0 : struct cfq_group *cfqg;
1634 0 : struct cfq_queue *cfqq;
1635 0 : int i, j;
1636 0 : struct cfq_rb_root *st;
1637 0 :
1638 0 : if (!cfqd->rq_queued)
1639 0 : return NULL;
1640 :
1641 0 : cfqg = cfq_get_next_cfqg(cfqd);
1642 0 : if (!cfqg)
1643 0 : return NULL;
1644 :
1645 0 : for_each_cfqg_st(cfqg, i, j, st)
1646 0 : if ((cfqq = cfq_rb_first(st)) != NULL)
1647 0 : return cfqq;
1648 0 : return NULL;
1649 : }
1650 :
1651 : /*
1652 : * Get and set a new active queue for service.
1653 : */
1654 : static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1655 : struct cfq_queue *cfqq)
1656 : {
1657 0 : if (!cfqq)
1658 0 : cfqq = cfq_get_next_queue(cfqd);
1659 :
1660 0 : __cfq_set_active_queue(cfqd, cfqq);
1661 0 : return cfqq;
1662 : }
1663 :
1664 : static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1665 : struct request *rq)
1666 0 : {
1667 0 : if (blk_rq_pos(rq) >= cfqd->last_position)
1668 0 : return blk_rq_pos(rq) - cfqd->last_position;
1669 : else
1670 0 : return cfqd->last_position - blk_rq_pos(rq);
1671 : }
1672 :
1673 : static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1674 : struct request *rq, bool for_preempt)
1675 : {
1676 0 : sector_t sdist = cfqq->seek_mean;
1677 0 :
1678 0 : if (!sample_valid(cfqq->seek_samples))
1679 0 : sdist = CFQQ_SEEK_THR;
1680 :
1681 : /* if seek_mean is big, using it as close criteria is meaningless */
1682 0 : if (sdist > CFQQ_SEEK_THR && !for_preempt)
1683 0 : sdist = CFQQ_SEEK_THR;
1684 :
1685 0 : return cfq_dist_from_last(cfqd, rq) <= sdist;
1686 : }
1687 :
1688 : static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1689 : struct cfq_queue *cur_cfqq)
1690 0 : {
1691 0 : struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1692 0 : struct rb_node *parent, *node;
1693 0 : struct cfq_queue *__cfqq;
1694 0 : sector_t sector = cfqd->last_position;
1695 0 :
1696 0 : if (RB_EMPTY_ROOT(root))
1697 0 : return NULL;
1698 0 :
1699 0 : /*
1700 : * First, if we find a request starting at the end of the last
1701 : * request, choose it.
1702 : */
1703 0 : __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1704 0 : if (__cfqq)
1705 0 : return __cfqq;
1706 :
1707 : /*
1708 : * If the exact sector wasn't found, the parent of the NULL leaf
1709 : * will contain the closest sector.
1710 : */
1711 0 : __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1712 0 : if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1713 0 : return __cfqq;
1714 :
1715 0 : if (blk_rq_pos(__cfqq->next_rq) < sector)
1716 0 : node = rb_next(&__cfqq->p_node);
1717 : else
1718 0 : node = rb_prev(&__cfqq->p_node);
1719 0 : if (!node)
1720 0 : return NULL;
1721 :
1722 0 : __cfqq = rb_entry(node, struct cfq_queue, p_node);
1723 0 : if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq, false))
1724 0 : return __cfqq;
1725 :
1726 0 : return NULL;
1727 : }
1728 :
1729 : /*
1730 : * cfqd - obvious
1731 : * cur_cfqq - passed in so that we don't decide that the current queue is
1732 : * closely cooperating with itself.
1733 : *
1734 : * So, basically we're assuming that that cur_cfqq has dispatched at least
1735 : * one request, and that cfqd->last_position reflects a position on the disk
1736 : * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1737 : * assumption.
1738 : */
1739 : static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1740 : struct cfq_queue *cur_cfqq)
1741 0 : {
1742 0 : struct cfq_queue *cfqq;
1743 0 :
1744 0 : if (!cfq_cfqq_sync(cur_cfqq))
1745 0 : return NULL;
1746 0 : if (CFQQ_SEEKY(cur_cfqq))
1747 0 : return NULL;
1748 :
1749 : /*
1750 : * Don't search priority tree if it's the only queue in the group.
1751 : */
1752 0 : if (cur_cfqq->cfqg->nr_cfqq == 1)
1753 0 : return NULL;
1754 :
1755 : /*
1756 : * We should notice if some of the queues are cooperating, eg
1757 : * working closely on the same area of the disk. In that case,
1758 : * we can group them together and don't waste time idling.
1759 : */
1760 0 : cfqq = cfqq_close(cfqd, cur_cfqq);
1761 0 : if (!cfqq)
1762 0 : return NULL;
1763 :
1764 : /* If new queue belongs to different cfq_group, don't choose it */
1765 0 : if (cur_cfqq->cfqg != cfqq->cfqg)
1766 0 : return NULL;
1767 :
1768 : /*
1769 : * It only makes sense to merge sync queues.
1770 : */
1771 0 : if (!cfq_cfqq_sync(cfqq))
1772 0 : return NULL;
1773 0 : if (CFQQ_SEEKY(cfqq))
1774 0 : return NULL;
1775 :
1776 : /*
1777 : * Do not merge queues of different priority classes
1778 : */
1779 0 : if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1780 0 : return NULL;
1781 :
1782 0 : return cfqq;
1783 : }
1784 :
1785 : /*
1786 : * Determine whether we should enforce idle window for this queue.
1787 : */
1788 :
1789 : static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1790 : {
1791 0 : enum wl_prio_t prio = cfqq_prio(cfqq);
1792 0 : struct cfq_rb_root *service_tree = cfqq->service_tree;
1793 0 :
1794 0 : BUG_ON(!service_tree);
1795 0 : BUG_ON(!service_tree->count);
1796 0 :
1797 0 : /* We never do for idle class queues. */
1798 0 : if (prio == IDLE_WORKLOAD)
1799 0 : return false;
1800 :
1801 : /* We do for queues that were marked with idle window flag. */
1802 0 : if (cfq_cfqq_idle_window(cfqq) &&
1803 : !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1804 0 : return true;
1805 :
1806 : /*
1807 : * Otherwise, we do only if they are the last ones
1808 : * in their service tree.
1809 : */
1810 0 : return service_tree->count == 1 && cfq_cfqq_sync(cfqq);
1811 : }
1812 :
1813 : static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1814 : {
1815 0 : struct cfq_queue *cfqq = cfqd->active_queue;
1816 0 : struct cfq_io_context *cic;
1817 0 : unsigned long sl;
1818 0 :
1819 0 : /*
1820 0 : * SSD device without seek penalty, disable idling. But only do so
1821 0 : * for devices that support queuing, otherwise we still have a problem
1822 0 : * with sync vs async workloads.
1823 0 : */
1824 0 : if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1825 0 : return;
1826 0 :
1827 0 : WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1828 0 : WARN_ON(cfq_cfqq_slice_new(cfqq));
1829 :
1830 : /*
1831 : * idle is disabled, either manually or by past process history
1832 : */
1833 0 : if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1834 0 : return;
1835 :
1836 : /*
1837 : * still active requests from this queue, don't idle
1838 : */
1839 0 : if (cfqq->dispatched)
1840 0 : return;
1841 :
1842 : /*
1843 : * task has exited, don't wait
1844 : */
1845 0 : cic = cfqd->active_cic;
1846 0 : if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1847 0 : return;
1848 :
1849 : /*
1850 : * If our average think time is larger than the remaining time
1851 : * slice, then don't idle. This avoids overrunning the allotted
1852 : * time slice.
1853 : */
1854 0 : if (sample_valid(cic->ttime_samples) &&
1855 : (cfqq->slice_end - jiffies < cic->ttime_mean))
1856 0 : return;
1857 :
1858 0 : cfq_mark_cfqq_wait_request(cfqq);
1859 :
1860 0 : sl = cfqd->cfq_slice_idle;
1861 :
1862 0 : mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1863 0 : cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1864 : }
1865 :
1866 : /*
1867 : * Move request from internal lists to the request queue dispatch list.
1868 : */
1869 : static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1870 : {
1871 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
1872 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
1873 0 :
1874 0 : cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1875 :
1876 0 : cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1877 0 : cfq_remove_request(rq);
1878 0 : cfqq->dispatched++;
1879 0 : elv_dispatch_sort(q, rq);
1880 :
1881 0 : if (cfq_cfqq_sync(cfqq))
1882 0 : cfqd->sync_flight++;
1883 0 : cfqq->nr_sectors += blk_rq_sectors(rq);
1884 0 : }
1885 :
1886 : /*
1887 : * return expired entry, or NULL to just start from scratch in rbtree
1888 : */
1889 : static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1890 : {
1891 0 : struct request *rq = NULL;
1892 0 :
1893 0 : if (cfq_cfqq_fifo_expire(cfqq))
1894 0 : return NULL;
1895 :
1896 0 : cfq_mark_cfqq_fifo_expire(cfqq);
1897 :
1898 0 : if (list_empty(&cfqq->fifo))
1899 0 : return NULL;
1900 :
1901 0 : rq = rq_entry_fifo(cfqq->fifo.next);
1902 0 : if (time_before(jiffies, rq_fifo_time(rq)))
1903 0 : rq = NULL;
1904 :
1905 : cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1906 0 : return rq;
1907 : }
1908 :
1909 : static inline int
1910 : cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1911 : {
1912 0 : const int base_rq = cfqd->cfq_slice_async_rq;
1913 0 :
1914 0 : WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1915 :
1916 0 : return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1917 : }
1918 :
1919 : /*
1920 : * Must be called with the queue_lock held.
1921 : */
1922 : static int cfqq_process_refs(struct cfq_queue *cfqq)
1923 : {
1924 0 : int process_refs, io_refs;
1925 0 :
1926 0 : io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1927 0 : process_refs = atomic_read(&cfqq->ref) - io_refs;
1928 0 : BUG_ON(process_refs < 0);
1929 0 : return process_refs;
1930 : }
1931 :
1932 : static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1933 : {
1934 0 : int process_refs, new_process_refs;
1935 0 : struct cfq_queue *__cfqq;
1936 0 :
1937 0 : /*
1938 : * If there are no process references on the new_cfqq, then it is
1939 : * unsafe to follow the ->new_cfqq chain as other cfqq's in the
1940 : * chain may have dropped their last reference (not just their
1941 : * last process reference).
1942 : */
1943 0 : if (!cfqq_process_refs(new_cfqq))
1944 0 : return;
1945 :
1946 : /* Avoid a circular list and skip interim queue merges */
1947 0 : while ((__cfqq = new_cfqq->new_cfqq)) {
1948 0 : if (__cfqq == cfqq)
1949 0 : return;
1950 0 : new_cfqq = __cfqq;
1951 0 : }
1952 :
1953 0 : process_refs = cfqq_process_refs(cfqq);
1954 0 : new_process_refs = cfqq_process_refs(new_cfqq);
1955 : /*
1956 : * If the process for the cfqq has gone away, there is no
1957 : * sense in merging the queues.
1958 : */
1959 0 : if (process_refs == 0 || new_process_refs == 0)
1960 0 : return;
1961 :
1962 : /*
1963 : * Merge in the direction of the lesser amount of work.
1964 : */
1965 0 : if (new_process_refs >= process_refs) {
1966 0 : cfqq->new_cfqq = new_cfqq;
1967 0 : atomic_add(process_refs, &new_cfqq->ref);
1968 : } else {
1969 0 : new_cfqq->new_cfqq = cfqq;
1970 0 : atomic_add(new_process_refs, &cfqq->ref);
1971 : }
1972 0 : }
1973 :
1974 : static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1975 : struct cfq_group *cfqg, enum wl_prio_t prio)
1976 : {
1977 0 : struct cfq_queue *queue;
1978 0 : int i;
1979 0 : bool key_valid = false;
1980 0 : unsigned long lowest_key = 0;
1981 0 : enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1982 0 :
1983 0 : for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1984 0 : /* select the one with lowest rb_key */
1985 0 : queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1986 : if (queue &&
1987 0 : (!key_valid || time_before(queue->rb_key, lowest_key))) {
1988 0 : lowest_key = queue->rb_key;
1989 0 : cur_best = i;
1990 0 : key_valid = true;
1991 : }
1992 : }
1993 :
1994 0 : return cur_best;
1995 : }
1996 :
1997 : static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1998 : {
1999 0 : unsigned slice;
2000 0 : unsigned count;
2001 0 : struct cfq_rb_root *st;
2002 0 : unsigned group_slice;
2003 0 :
2004 0 : if (!cfqg) {
2005 0 : cfqd->serving_prio = IDLE_WORKLOAD;
2006 0 : cfqd->workload_expires = jiffies + 1;
2007 0 : return;
2008 0 : }
2009 0 :
2010 0 : /* Choose next priority. RT > BE > IDLE */
2011 0 : if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2012 0 : cfqd->serving_prio = RT_WORKLOAD;
2013 0 : else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2014 0 : cfqd->serving_prio = BE_WORKLOAD;
2015 0 : else {
2016 0 : cfqd->serving_prio = IDLE_WORKLOAD;
2017 0 : cfqd->workload_expires = jiffies + 1;
2018 0 : return;
2019 : }
2020 :
2021 : /*
2022 : * For RT and BE, we have to choose also the type
2023 : * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2024 : * expiration time
2025 : */
2026 0 : st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2027 0 : count = st->count;
2028 :
2029 : /*
2030 : * check workload expiration, and that we still have other queues ready
2031 : */
2032 0 : if (count && !time_after(jiffies, cfqd->workload_expires))
2033 0 : return;
2034 :
2035 : /* otherwise select new workload type */
2036 0 : cfqd->serving_type =
2037 : cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2038 0 : st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2039 0 : count = st->count;
2040 :
2041 : /*
2042 : * the workload slice is computed as a fraction of target latency
2043 : * proportional to the number of queues in that workload, over
2044 : * all the queues in the same priority class
2045 : */
2046 0 : group_slice = cfq_group_slice(cfqd, cfqg);
2047 :
2048 : slice = group_slice * count /
2049 0 : max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2050 : cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2051 :
2052 0 : if (cfqd->serving_type == ASYNC_WORKLOAD) {
2053 : unsigned int tmp;
2054 :
2055 : /*
2056 : * Async queues are currently system wide. Just taking
2057 : * proportion of queues with-in same group will lead to higher
2058 : * async ratio system wide as generally root group is going
2059 : * to have higher weight. A more accurate thing would be to
2060 : * calculate system wide asnc/sync ratio.
2061 : */
2062 0 : tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2063 0 : tmp = tmp/cfqd->busy_queues;
2064 0 : slice = min_t(unsigned, slice, tmp);
2065 :
2066 : /* async workload slice is scaled down according to
2067 : * the sync/async slice ratio. */
2068 0 : slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2069 : } else
2070 : /* sync workload slice is at least 2 * cfq_slice_idle */
2071 0 : slice = max(slice, 2 * cfqd->cfq_slice_idle);
2072 :
2073 0 : slice = max_t(unsigned, slice, CFQ_MIN_TT);
2074 0 : cfqd->workload_expires = jiffies + slice;
2075 0 : cfqd->noidle_tree_requires_idle = false;
2076 0 : }
2077 :
2078 : static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2079 : {
2080 0 : struct cfq_rb_root *st = &cfqd->grp_service_tree;
2081 0 : struct cfq_group *cfqg;
2082 :
2083 0 : if (RB_EMPTY_ROOT(&st->rb))
2084 0 : return NULL;
2085 0 : cfqg = cfq_rb_first_group(st);
2086 0 : st->active = &cfqg->rb_node;
2087 0 : update_min_vdisktime(st);
2088 0 : return cfqg;
2089 : }
2090 :
2091 : static void cfq_choose_cfqg(struct cfq_data *cfqd)
2092 : {
2093 0 : struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2094 0 :
2095 0 : cfqd->serving_group = cfqg;
2096 :
2097 : /* Restore the workload type data */
2098 0 : if (cfqg->saved_workload_slice) {
2099 0 : cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2100 0 : cfqd->serving_type = cfqg->saved_workload;
2101 0 : cfqd->serving_prio = cfqg->saved_serving_prio;
2102 : } else
2103 0 : cfqd->workload_expires = jiffies - 1;
2104 :
2105 0 : choose_service_tree(cfqd, cfqg);
2106 0 : }
2107 :
2108 : /*
2109 : * Select a queue for service. If we have a current active queue,
2110 : * check whether to continue servicing it, or retrieve and set a new one.
2111 : */
2112 : static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2113 : {
2114 0 : struct cfq_queue *cfqq, *new_cfqq = NULL;
2115 0 :
2116 0 : cfqq = cfqd->active_queue;
2117 0 : if (!cfqq)
2118 0 : goto new_queue;
2119 0 :
2120 0 : if (!cfqd->rq_queued)
2121 0 : return NULL;
2122 :
2123 : /*
2124 : * We were waiting for group to get backlogged. Expire the queue
2125 : */
2126 0 : if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2127 0 : goto expire;
2128 :
2129 : /*
2130 : * The active queue has run out of time, expire it and select new.
2131 : */
2132 0 : if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2133 : /*
2134 : * If slice had not expired at the completion of last request
2135 : * we might not have turned on wait_busy flag. Don't expire
2136 : * the queue yet. Allow the group to get backlogged.
2137 : *
2138 : * The very fact that we have used the slice, that means we
2139 : * have been idling all along on this queue and it should be
2140 : * ok to wait for this request to complete.
2141 : */
2142 0 : if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2143 : && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2144 0 : cfqq = NULL;
2145 0 : goto keep_queue;
2146 : } else
2147 0 : goto expire;
2148 : }
2149 :
2150 : /*
2151 : * The active queue has requests and isn't expired, allow it to
2152 : * dispatch.
2153 : */
2154 0 : if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2155 0 : goto keep_queue;
2156 :
2157 : /*
2158 : * If another queue has a request waiting within our mean seek
2159 : * distance, let it run. The expire code will check for close
2160 : * cooperators and put the close queue at the front of the service
2161 : * tree. If possible, merge the expiring queue with the new cfqq.
2162 : */
2163 0 : new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2164 0 : if (new_cfqq) {
2165 0 : if (!cfqq->new_cfqq)
2166 0 : cfq_setup_merge(cfqq, new_cfqq);
2167 0 : goto expire;
2168 : }
2169 :
2170 : /*
2171 : * No requests pending. If the active queue still has requests in
2172 : * flight or is idling for a new request, allow either of these
2173 : * conditions to happen (or time out) before selecting a new queue.
2174 : */
2175 0 : if (timer_pending(&cfqd->idle_slice_timer) ||
2176 : (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2177 0 : cfqq = NULL;
2178 0 : goto keep_queue;
2179 : }
2180 :
2181 : expire:
2182 0 : cfq_slice_expired(cfqd, 0);
2183 : new_queue:
2184 0 : /*
2185 0 : * Current queue expired. Check if we have to switch to a new
2186 : * service tree
2187 : */
2188 0 : if (!new_cfqq)
2189 0 : cfq_choose_cfqg(cfqd);
2190 :
2191 0 : cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2192 : keep_queue:
2193 0 : return cfqq;
2194 : }
2195 :
2196 : static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2197 : {
2198 0 : int dispatched = 0;
2199 0 :
2200 0 : while (cfqq->next_rq) {
2201 0 : cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2202 0 : dispatched++;
2203 : }
2204 0 :
2205 0 : BUG_ON(!list_empty(&cfqq->fifo));
2206 :
2207 : /* By default cfqq is not expired if it is empty. Do it explicitly */
2208 0 : __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2209 0 : return dispatched;
2210 : }
2211 :
2212 : /*
2213 : * Drain our current requests. Used for barriers and when switching
2214 : * io schedulers on-the-fly.
2215 : */
2216 : static int cfq_forced_dispatch(struct cfq_data *cfqd)
2217 : {
2218 0 : struct cfq_queue *cfqq;
2219 0 : int dispatched = 0;
2220 0 :
2221 0 : while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2222 0 : dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2223 0 :
2224 0 : cfq_slice_expired(cfqd, 0);
2225 0 : BUG_ON(cfqd->busy_queues);
2226 :
2227 : cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2228 0 : return dispatched;
2229 : }
2230 :
2231 : static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2232 : {
2233 0 : unsigned int max_dispatch;
2234 0 :
2235 0 : /*
2236 0 : * Drain async requests before we start sync IO
2237 0 : */
2238 0 : if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
2239 0 : return false;
2240 :
2241 : /*
2242 : * If this is an async queue and we have sync IO in flight, let it wait
2243 : */
2244 0 : if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
2245 0 : return false;
2246 :
2247 0 : max_dispatch = cfqd->cfq_quantum;
2248 0 : if (cfq_class_idle(cfqq))
2249 0 : max_dispatch = 1;
2250 :
2251 : /*
2252 : * Does this cfqq already have too much IO in flight?
2253 : */
2254 0 : if (cfqq->dispatched >= max_dispatch) {
2255 : /*
2256 : * idle queue must always only have a single IO in flight
2257 : */
2258 0 : if (cfq_class_idle(cfqq))
2259 0 : return false;
2260 :
2261 : /*
2262 : * We have other queues, don't allow more IO from this one
2263 : */
2264 0 : if (cfqd->busy_queues > 1)
2265 0 : return false;
2266 :
2267 : /*
2268 : * Sole queue user, no limit
2269 : */
2270 0 : max_dispatch = -1;
2271 : }
2272 :
2273 : /*
2274 : * Async queues must wait a bit before being allowed dispatch.
2275 : * We also ramp up the dispatch depth gradually for async IO,
2276 : * based on the last sync IO we serviced
2277 : */
2278 0 : if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2279 0 : unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2280 : unsigned int depth;
2281 :
2282 0 : depth = last_sync / cfqd->cfq_slice[1];
2283 0 : if (!depth && !cfqq->dispatched)
2284 0 : depth = 1;
2285 0 : if (depth < max_dispatch)
2286 0 : max_dispatch = depth;
2287 : }
2288 :
2289 : /*
2290 : * If we're below the current max, allow a dispatch
2291 : */
2292 0 : return cfqq->dispatched < max_dispatch;
2293 : }
2294 :
2295 : /*
2296 : * Dispatch a request from cfqq, moving them to the request queue
2297 : * dispatch list.
2298 : */
2299 : static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2300 : {
2301 0 : struct request *rq;
2302 0 :
2303 0 : BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2304 0 :
2305 0 : if (!cfq_may_dispatch(cfqd, cfqq))
2306 0 : return false;
2307 :
2308 : /*
2309 : * follow expired path, else get first next available
2310 : */
2311 0 : rq = cfq_check_fifo(cfqq);
2312 0 : if (!rq)
2313 0 : rq = cfqq->next_rq;
2314 :
2315 : /*
2316 : * insert request into driver dispatch list
2317 : */
2318 0 : cfq_dispatch_insert(cfqd->queue, rq);
2319 :
2320 0 : if (!cfqd->active_cic) {
2321 0 : struct cfq_io_context *cic = RQ_CIC(rq);
2322 :
2323 0 : atomic_long_inc(&cic->ioc->refcount);
2324 0 : cfqd->active_cic = cic;
2325 : }
2326 :
2327 0 : return true;
2328 : }
2329 :
2330 : /*
2331 : * Find the cfqq that we need to service and move a request from that to the
2332 : * dispatch list
2333 : */
2334 : static int cfq_dispatch_requests(struct request_queue *q, int force)
2335 : {
2336 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
2337 0 : struct cfq_queue *cfqq;
2338 0 :
2339 0 : if (!cfqd->busy_queues)
2340 0 : return 0;
2341 0 :
2342 0 : if (unlikely(force))
2343 0 : return cfq_forced_dispatch(cfqd);
2344 :
2345 0 : cfqq = cfq_select_queue(cfqd);
2346 0 : if (!cfqq)
2347 0 : return 0;
2348 :
2349 : /*
2350 : * Dispatch a request from this cfqq, if it is allowed
2351 : */
2352 0 : if (!cfq_dispatch_request(cfqd, cfqq))
2353 0 : return 0;
2354 :
2355 0 : cfqq->slice_dispatch++;
2356 0 : cfq_clear_cfqq_must_dispatch(cfqq);
2357 :
2358 : /*
2359 : * expire an async queue immediately if it has used up its slice. idle
2360 : * queue always expire after 1 dispatch round.
2361 : */
2362 0 : if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2363 : cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2364 : cfq_class_idle(cfqq))) {
2365 0 : cfqq->slice_end = jiffies + 1;
2366 0 : cfq_slice_expired(cfqd, 0);
2367 : }
2368 :
2369 : cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2370 0 : return 1;
2371 : }
2372 :
2373 : /*
2374 : * task holds one reference to the queue, dropped when task exits. each rq
2375 : * in-flight on this queue also holds a reference, dropped when rq is freed.
2376 : *
2377 : * Each cfq queue took a reference on the parent group. Drop it now.
2378 : * queue lock must be held here.
2379 : */
2380 : static void cfq_put_queue(struct cfq_queue *cfqq)
2381 : {
2382 0 : struct cfq_data *cfqd = cfqq->cfqd;
2383 0 : struct cfq_group *cfqg, *orig_cfqg;
2384 0 :
2385 0 : BUG_ON(atomic_read(&cfqq->ref) <= 0);
2386 0 :
2387 0 : if (!atomic_dec_and_test(&cfqq->ref))
2388 0 : return;
2389 0 :
2390 0 : cfq_log_cfqq(cfqd, cfqq, "put_queue");
2391 0 : BUG_ON(rb_first(&cfqq->sort_list));
2392 0 : BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2393 0 : cfqg = cfqq->cfqg;
2394 0 : orig_cfqg = cfqq->orig_cfqg;
2395 :
2396 0 : if (unlikely(cfqd->active_queue == cfqq)) {
2397 0 : __cfq_slice_expired(cfqd, cfqq, 0);
2398 0 : cfq_schedule_dispatch(cfqd);
2399 : }
2400 :
2401 0 : BUG_ON(cfq_cfqq_on_rr(cfqq));
2402 0 : kmem_cache_free(cfq_pool, cfqq);
2403 0 : cfq_put_cfqg(cfqg);
2404 0 : if (orig_cfqg)
2405 0 : cfq_put_cfqg(orig_cfqg);
2406 0 : }
2407 :
2408 : /*
2409 : * Must always be called with the rcu_read_lock() held
2410 : */
2411 : static void
2412 : __call_for_each_cic(struct io_context *ioc,
2413 : void (*func)(struct io_context *, struct cfq_io_context *))
2414 : {
2415 11 : struct cfq_io_context *cic;
2416 11 : struct hlist_node *n;
2417 11 :
2418 132 : hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2419 66 : func(ioc, cic);
2420 22 : }
2421 :
2422 : /*
2423 : * Call func for each cic attached to this ioc.
2424 : */
2425 11 : static void
2426 : call_for_each_cic(struct io_context *ioc,
2427 : void (*func)(struct io_context *, struct cfq_io_context *))
2428 : {
2429 0 : rcu_read_lock();
2430 0 : __call_for_each_cic(ioc, func);
2431 0 : rcu_read_unlock();
2432 0 : }
2433 :
2434 : static void cfq_cic_free_rcu(struct rcu_head *head)
2435 : {
2436 0 : struct cfq_io_context *cic;
2437 0 :
2438 0 : cic = container_of(head, struct cfq_io_context, rcu_head);
2439 0 :
2440 0 : kmem_cache_free(cfq_ioc_pool, cic);
2441 0 : elv_ioc_count_dec(cfq_ioc_count);
2442 0 :
2443 0 : if (ioc_gone) {
2444 : /*
2445 : * CFQ scheduler is exiting, grab exit lock and check
2446 : * the pending io context count. If it hits zero,
2447 : * complete ioc_gone and set it back to NULL
2448 : */
2449 0 : spin_lock(&ioc_gone_lock);
2450 0 : if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2451 0 : complete(ioc_gone);
2452 0 : ioc_gone = NULL;
2453 : }
2454 0 : spin_unlock(&ioc_gone_lock);
2455 : }
2456 0 : }
2457 :
2458 : static void cfq_cic_free(struct cfq_io_context *cic)
2459 : {
2460 11 : call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2461 11 : }
2462 :
2463 : static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2464 : {
2465 11 : unsigned long flags;
2466 11 :
2467 77 : BUG_ON(!cic->dead_key);
2468 :
2469 33 : spin_lock_irqsave(&ioc->lock, flags);
2470 11 : radix_tree_delete(&ioc->radix_root, cic->dead_key);
2471 22 : hlist_del_rcu(&cic->cic_list);
2472 22 : spin_unlock_irqrestore(&ioc->lock, flags);
2473 :
2474 22 : cfq_cic_free(cic);
2475 11 : }
2476 :
2477 : /*
2478 : * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2479 : * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2480 : * and ->trim() which is called with the task lock held
2481 : */
2482 : static void cfq_free_io_context(struct io_context *ioc)
2483 : {
2484 : /*
2485 : * ioc->refcount is zero here, or we are called from elv_unregister(),
2486 : * so no more cic's are allowed to be linked into this ioc. So it
2487 : * should be ok to iterate over the known list, we will see all cic's
2488 : * since no new ones are added.
2489 : */
2490 22 : __call_for_each_cic(ioc, cic_free_func);
2491 11 : }
2492 :
2493 : static void cfq_put_cooperator(struct cfq_queue *cfqq)
2494 : {
2495 0 : struct cfq_queue *__cfqq, *next;
2496 0 :
2497 0 : /*
2498 0 : * If this queue was scheduled to merge with another queue, be
2499 : * sure to drop the reference taken on that queue (and others in
2500 : * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2501 : */
2502 0 : __cfqq = cfqq->new_cfqq;
2503 0 : while (__cfqq) {
2504 0 : if (__cfqq == cfqq) {
2505 0 : WARN(1, "cfqq->new_cfqq loop detected\n");
2506 0 : break;
2507 : }
2508 0 : next = __cfqq->new_cfqq;
2509 0 : cfq_put_queue(__cfqq);
2510 0 : __cfqq = next;
2511 : }
2512 0 : }
2513 :
2514 : static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2515 : {
2516 0 : if (unlikely(cfqq == cfqd->active_queue)) {
2517 0 : __cfq_slice_expired(cfqd, cfqq, 0);
2518 0 : cfq_schedule_dispatch(cfqd);
2519 : }
2520 :
2521 0 : cfq_put_cooperator(cfqq);
2522 :
2523 0 : cfq_put_queue(cfqq);
2524 0 : }
2525 :
2526 : static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2527 : struct cfq_io_context *cic)
2528 0 : {
2529 0 : struct io_context *ioc = cic->ioc;
2530 :
2531 0 : list_del_init(&cic->queue_list);
2532 :
2533 : /*
2534 : * Make sure key == NULL is seen for dead queues
2535 : */
2536 0 : smp_wmb();
2537 0 : cic->dead_key = (unsigned long) cic->key;
2538 0 : cic->key = NULL;
2539 :
2540 0 : rcu_read_lock();
2541 0 : if (rcu_dereference(ioc->ioc_data) == cic) {
2542 0 : rcu_read_unlock();
2543 0 : spin_lock(&ioc->lock);
2544 0 : rcu_assign_pointer(ioc->ioc_data, NULL);
2545 0 : spin_unlock(&ioc->lock);
2546 : } else
2547 0 : rcu_read_unlock();
2548 :
2549 0 : if (cic->cfqq[BLK_RW_ASYNC]) {
2550 0 : cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2551 0 : cic->cfqq[BLK_RW_ASYNC] = NULL;
2552 : }
2553 :
2554 0 : if (cic->cfqq[BLK_RW_SYNC]) {
2555 0 : cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2556 0 : cic->cfqq[BLK_RW_SYNC] = NULL;
2557 : }
2558 0 : }
2559 :
2560 : static void cfq_exit_single_io_context(struct io_context *ioc,
2561 : struct cfq_io_context *cic)
2562 0 : {
2563 0 : struct cfq_data *cfqd = cic->key;
2564 0 :
2565 0 : if (cfqd) {
2566 0 : struct request_queue *q = cfqd->queue;
2567 : unsigned long flags;
2568 :
2569 0 : spin_lock_irqsave(q->queue_lock, flags);
2570 :
2571 : /*
2572 : * Ensure we get a fresh copy of the ->key to prevent
2573 : * race between exiting task and queue
2574 : */
2575 : smp_read_barrier_depends();
2576 0 : if (cic->key)
2577 0 : __cfq_exit_single_io_context(cfqd, cic);
2578 :
2579 0 : spin_unlock_irqrestore(q->queue_lock, flags);
2580 : }
2581 0 : }
2582 :
2583 : /*
2584 : * The process that ioc belongs to has exited, we need to clean up
2585 : * and put the internal structures we have that belongs to that process.
2586 : */
2587 : static void cfq_exit_io_context(struct io_context *ioc)
2588 : {
2589 0 : call_for_each_cic(ioc, cfq_exit_single_io_context);
2590 0 : }
2591 :
2592 : static struct cfq_io_context *
2593 : cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2594 : {
2595 0 : struct cfq_io_context *cic;
2596 0 :
2597 0 : cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2598 0 : cfqd->queue->node);
2599 0 : if (cic) {
2600 0 : cic->last_end_request = jiffies;
2601 0 : INIT_LIST_HEAD(&cic->queue_list);
2602 0 : INIT_HLIST_NODE(&cic->cic_list);
2603 0 : cic->dtor = cfq_free_io_context;
2604 0 : cic->exit = cfq_exit_io_context;
2605 0 : elv_ioc_count_inc(cfq_ioc_count);
2606 0 : }
2607 :
2608 0 : return cic;
2609 : }
2610 :
2611 : static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2612 : {
2613 0 : struct task_struct *tsk = current;
2614 0 : int ioprio_class;
2615 0 :
2616 0 : if (!cfq_cfqq_prio_changed(cfqq))
2617 0 : return;
2618 0 :
2619 0 : ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2620 0 : switch (ioprio_class) {
2621 0 : default:
2622 0 : printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2623 0 : case IOPRIO_CLASS_NONE:
2624 0 : /*
2625 : * no prio set, inherit CPU scheduling settings
2626 : */
2627 0 : cfqq->ioprio = task_nice_ioprio(tsk);
2628 0 : cfqq->ioprio_class = task_nice_ioclass(tsk);
2629 0 : break;
2630 0 : case IOPRIO_CLASS_RT:
2631 0 : cfqq->ioprio = task_ioprio(ioc);
2632 0 : cfqq->ioprio_class = IOPRIO_CLASS_RT;
2633 0 : break;
2634 0 : case IOPRIO_CLASS_BE:
2635 0 : cfqq->ioprio = task_ioprio(ioc);
2636 0 : cfqq->ioprio_class = IOPRIO_CLASS_BE;
2637 0 : break;
2638 0 : case IOPRIO_CLASS_IDLE:
2639 0 : cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2640 0 : cfqq->ioprio = 7;
2641 0 : cfq_clear_cfqq_idle_window(cfqq);
2642 0 : break;
2643 : }
2644 :
2645 : /*
2646 : * keep track of original prio settings in case we have to temporarily
2647 : * elevate the priority of this queue
2648 : */
2649 0 : cfqq->org_ioprio = cfqq->ioprio;
2650 0 : cfqq->org_ioprio_class = cfqq->ioprio_class;
2651 0 : cfq_clear_cfqq_prio_changed(cfqq);
2652 0 : }
2653 :
2654 : static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2655 : {
2656 0 : struct cfq_data *cfqd = cic->key;
2657 0 : struct cfq_queue *cfqq;
2658 0 : unsigned long flags;
2659 0 :
2660 0 : if (unlikely(!cfqd))
2661 0 : return;
2662 :
2663 0 : spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2664 :
2665 0 : cfqq = cic->cfqq[BLK_RW_ASYNC];
2666 0 : if (cfqq) {
2667 : struct cfq_queue *new_cfqq;
2668 0 : new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2669 : GFP_ATOMIC);
2670 0 : if (new_cfqq) {
2671 0 : cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2672 0 : cfq_put_queue(cfqq);
2673 : }
2674 : }
2675 :
2676 0 : cfqq = cic->cfqq[BLK_RW_SYNC];
2677 0 : if (cfqq)
2678 0 : cfq_mark_cfqq_prio_changed(cfqq);
2679 :
2680 0 : spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2681 0 : }
2682 :
2683 : static void cfq_ioc_set_ioprio(struct io_context *ioc)
2684 : {
2685 0 : call_for_each_cic(ioc, changed_ioprio);
2686 0 : ioc->ioprio_changed = 0;
2687 0 : }
2688 :
2689 : static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2690 : pid_t pid, bool is_sync)
2691 : {
2692 0 : RB_CLEAR_NODE(&cfqq->rb_node);
2693 0 : RB_CLEAR_NODE(&cfqq->p_node);
2694 0 : INIT_LIST_HEAD(&cfqq->fifo);
2695 :
2696 0 : atomic_set(&cfqq->ref, 0);
2697 0 : cfqq->cfqd = cfqd;
2698 :
2699 0 : cfq_mark_cfqq_prio_changed(cfqq);
2700 :
2701 0 : if (is_sync) {
2702 0 : if (!cfq_class_idle(cfqq))
2703 0 : cfq_mark_cfqq_idle_window(cfqq);
2704 0 : cfq_mark_cfqq_sync(cfqq);
2705 : }
2706 0 : cfqq->pid = pid;
2707 0 : }
2708 :
2709 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
2710 : static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2711 : {
2712 : struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2713 : struct cfq_data *cfqd = cic->key;
2714 : unsigned long flags;
2715 : struct request_queue *q;
2716 :
2717 : if (unlikely(!cfqd))
2718 : return;
2719 :
2720 : q = cfqd->queue;
2721 :
2722 : spin_lock_irqsave(q->queue_lock, flags);
2723 :
2724 : if (sync_cfqq) {
2725 : /*
2726 : * Drop reference to sync queue. A new sync queue will be
2727 : * assigned in new group upon arrival of a fresh request.
2728 : */
2729 : cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2730 : cic_set_cfqq(cic, NULL, 1);
2731 : cfq_put_queue(sync_cfqq);
2732 : }
2733 :
2734 : spin_unlock_irqrestore(q->queue_lock, flags);
2735 : }
2736 :
2737 : static void cfq_ioc_set_cgroup(struct io_context *ioc)
2738 : {
2739 : call_for_each_cic(ioc, changed_cgroup);
2740 : ioc->cgroup_changed = 0;
2741 : }
2742 : #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2743 :
2744 : static struct cfq_queue *
2745 : cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2746 : struct io_context *ioc, gfp_t gfp_mask)
2747 : {
2748 0 : struct cfq_queue *cfqq, *new_cfqq = NULL;
2749 0 : struct cfq_io_context *cic;
2750 0 : struct cfq_group *cfqg;
2751 0 :
2752 0 : retry:
2753 0 : cfqg = cfq_get_cfqg(cfqd, 1);
2754 0 : cic = cfq_cic_lookup(cfqd, ioc);
2755 : /* cic always exists here */
2756 0 : cfqq = cic_to_cfqq(cic, is_sync);
2757 :
2758 : /*
2759 : * Always try a new alloc if we fell back to the OOM cfqq
2760 : * originally, since it should just be a temporary situation.
2761 : */
2762 0 : if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2763 0 : cfqq = NULL;
2764 0 : if (new_cfqq) {
2765 0 : cfqq = new_cfqq;
2766 0 : new_cfqq = NULL;
2767 0 : } else if (gfp_mask & __GFP_WAIT) {
2768 0 : spin_unlock_irq(cfqd->queue->queue_lock);
2769 0 : new_cfqq = kmem_cache_alloc_node(cfq_pool,
2770 : gfp_mask | __GFP_ZERO,
2771 : cfqd->queue->node);
2772 0 : spin_lock_irq(cfqd->queue->queue_lock);
2773 0 : if (new_cfqq)
2774 0 : goto retry;
2775 : } else {
2776 0 : cfqq = kmem_cache_alloc_node(cfq_pool,
2777 : gfp_mask | __GFP_ZERO,
2778 : cfqd->queue->node);
2779 : }
2780 :
2781 0 : if (cfqq) {
2782 0 : cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2783 0 : cfq_init_prio_data(cfqq, ioc);
2784 0 : cfq_link_cfqq_cfqg(cfqq, cfqg);
2785 : cfq_log_cfqq(cfqd, cfqq, "alloced");
2786 : } else
2787 0 : cfqq = &cfqd->oom_cfqq;
2788 : }
2789 :
2790 0 : if (new_cfqq)
2791 0 : kmem_cache_free(cfq_pool, new_cfqq);
2792 :
2793 0 : return cfqq;
2794 : }
2795 :
2796 : static struct cfq_queue **
2797 : cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2798 : {
2799 : switch (ioprio_class) {
2800 0 : case IOPRIO_CLASS_RT:
2801 0 : return &cfqd->async_cfqq[0][ioprio];
2802 0 : case IOPRIO_CLASS_BE:
2803 0 : return &cfqd->async_cfqq[1][ioprio];
2804 0 : case IOPRIO_CLASS_IDLE:
2805 0 : return &cfqd->async_idle_cfqq;
2806 0 : default:
2807 0 : BUG();
2808 : }
2809 0 : }
2810 :
2811 : static struct cfq_queue *
2812 : cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2813 : gfp_t gfp_mask)
2814 : {
2815 0 : const int ioprio = task_ioprio(ioc);
2816 0 : const int ioprio_class = task_ioprio_class(ioc);
2817 0 : struct cfq_queue **async_cfqq = NULL;
2818 0 : struct cfq_queue *cfqq = NULL;
2819 0 :
2820 0 : if (!is_sync) {
2821 0 : async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2822 0 : cfqq = *async_cfqq;
2823 : }
2824 :
2825 0 : if (!cfqq)
2826 0 : cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2827 :
2828 : /*
2829 : * pin the queue now that it's allocated, scheduler exit will prune it
2830 : */
2831 0 : if (!is_sync && !(*async_cfqq)) {
2832 0 : atomic_inc(&cfqq->ref);
2833 0 : *async_cfqq = cfqq;
2834 : }
2835 :
2836 0 : atomic_inc(&cfqq->ref);
2837 0 : return cfqq;
2838 : }
2839 :
2840 : /*
2841 : * We drop cfq io contexts lazily, so we may find a dead one.
2842 : */
2843 : static void
2844 : cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2845 : struct cfq_io_context *cic)
2846 0 : {
2847 0 : unsigned long flags;
2848 0 :
2849 0 : WARN_ON(!list_empty(&cic->queue_list));
2850 0 :
2851 0 : spin_lock_irqsave(&ioc->lock, flags);
2852 :
2853 0 : BUG_ON(ioc->ioc_data == cic);
2854 :
2855 0 : radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2856 0 : hlist_del_rcu(&cic->cic_list);
2857 0 : spin_unlock_irqrestore(&ioc->lock, flags);
2858 :
2859 0 : cfq_cic_free(cic);
2860 0 : }
2861 :
2862 : static struct cfq_io_context *
2863 : cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2864 : {
2865 0 : struct cfq_io_context *cic;
2866 0 : unsigned long flags;
2867 0 : void *k;
2868 0 :
2869 0 : if (unlikely(!ioc))
2870 0 : return NULL;
2871 0 :
2872 0 : rcu_read_lock();
2873 :
2874 : /*
2875 : * we maintain a last-hit cache, to avoid browsing over the tree
2876 : */
2877 0 : cic = rcu_dereference(ioc->ioc_data);
2878 0 : if (cic && cic->key == cfqd) {
2879 0 : rcu_read_unlock();
2880 0 : return cic;
2881 : }
2882 :
2883 : do {
2884 0 : cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2885 0 : rcu_read_unlock();
2886 0 : if (!cic)
2887 0 : break;
2888 : /* ->key must be copied to avoid race with cfq_exit_queue() */
2889 0 : k = cic->key;
2890 0 : if (unlikely(!k)) {
2891 0 : cfq_drop_dead_cic(cfqd, ioc, cic);
2892 0 : rcu_read_lock();
2893 0 : continue;
2894 : }
2895 :
2896 0 : spin_lock_irqsave(&ioc->lock, flags);
2897 0 : rcu_assign_pointer(ioc->ioc_data, cic);
2898 0 : spin_unlock_irqrestore(&ioc->lock, flags);
2899 0 : break;
2900 0 : } while (1);
2901 0 :
2902 0 : return cic;
2903 : }
2904 :
2905 : /*
2906 : * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2907 : * the process specific cfq io context when entered from the block layer.
2908 : * Also adds the cic to a per-cfqd list, used when this queue is removed.
2909 : */
2910 : static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2911 : struct cfq_io_context *cic, gfp_t gfp_mask)
2912 : {
2913 0 : unsigned long flags;
2914 0 : int ret;
2915 0 :
2916 0 : ret = radix_tree_preload(gfp_mask);
2917 0 : if (!ret) {
2918 0 : cic->ioc = ioc;
2919 0 : cic->key = cfqd;
2920 :
2921 0 : spin_lock_irqsave(&ioc->lock, flags);
2922 0 : ret = radix_tree_insert(&ioc->radix_root,
2923 : (unsigned long) cfqd, cic);
2924 0 : if (!ret)
2925 0 : hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2926 0 : spin_unlock_irqrestore(&ioc->lock, flags);
2927 :
2928 0 : radix_tree_preload_end();
2929 :
2930 0 : if (!ret) {
2931 0 : spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2932 0 : list_add(&cic->queue_list, &cfqd->cic_list);
2933 0 : spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2934 : }
2935 : }
2936 :
2937 0 : if (ret)
2938 0 : printk(KERN_ERR "cfq: cic link failed!\n");
2939 :
2940 0 : return ret;
2941 : }
2942 :
2943 : /*
2944 : * Setup general io context and cfq io context. There can be several cfq
2945 : * io contexts per general io context, if this process is doing io to more
2946 : * than one device managed by cfq.
2947 : */
2948 : static struct cfq_io_context *
2949 : cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2950 : {
2951 0 : struct io_context *ioc = NULL;
2952 0 : struct cfq_io_context *cic;
2953 0 :
2954 0 : might_sleep_if(gfp_mask & __GFP_WAIT);
2955 :
2956 0 : ioc = get_io_context(gfp_mask, cfqd->queue->node);
2957 0 : if (!ioc)
2958 0 : return NULL;
2959 :
2960 0 : cic = cfq_cic_lookup(cfqd, ioc);
2961 0 : if (cic)
2962 0 : goto out;
2963 :
2964 0 : cic = cfq_alloc_io_context(cfqd, gfp_mask);
2965 0 : if (cic == NULL)
2966 0 : goto err;
2967 :
2968 0 : if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2969 0 : goto err_free;
2970 :
2971 : out:
2972 : smp_read_barrier_depends();
2973 0 : if (unlikely(ioc->ioprio_changed))
2974 0 : cfq_ioc_set_ioprio(ioc);
2975 :
2976 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
2977 : if (unlikely(ioc->cgroup_changed))
2978 : cfq_ioc_set_cgroup(ioc);
2979 : #endif
2980 0 : return cic;
2981 0 : err_free:
2982 0 : cfq_cic_free(cic);
2983 : err:
2984 0 : put_io_context(ioc);
2985 0 : return NULL;
2986 : }
2987 :
2988 : static void
2989 : cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2990 : {
2991 0 : unsigned long elapsed = jiffies - cic->last_end_request;
2992 0 : unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2993 0 :
2994 0 : cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2995 0 : cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2996 0 : cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2997 0 : }
2998 :
2999 : static void
3000 : cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3001 : struct request *rq)
3002 : {
3003 0 : sector_t sdist;
3004 0 : u64 total;
3005 0 :
3006 0 : if (!cfqq->last_request_pos)
3007 0 : sdist = 0;
3008 0 : else if (cfqq->last_request_pos < blk_rq_pos(rq))
3009 0 : sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3010 0 : else
3011 0 : sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3012 0 :
3013 0 : /*
3014 : * Don't allow the seek distance to get too large from the
3015 : * odd fragment, pagein, etc
3016 : */
3017 0 : if (cfqq->seek_samples <= 60) /* second&third seek */
3018 0 : sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
3019 : else
3020 0 : sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
3021 :
3022 0 : cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
3023 0 : cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
3024 0 : total = cfqq->seek_total + (cfqq->seek_samples/2);
3025 0 : do_div(total, cfqq->seek_samples);
3026 0 : cfqq->seek_mean = (sector_t)total;
3027 0 : }
3028 :
3029 : /*
3030 : * Disable idle window if the process thinks too long or seeks so much that
3031 : * it doesn't matter
3032 : */
3033 : static void
3034 : cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3035 : struct cfq_io_context *cic)
3036 : {
3037 0 : int old_idle, enable_idle;
3038 0 :
3039 0 : /*
3040 0 : * Don't idle for async or idle io prio class
3041 0 : */
3042 0 : if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3043 0 : return;
3044 :
3045 0 : enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3046 :
3047 0 : if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3048 0 : cfq_mark_cfqq_deep(cfqq);
3049 :
3050 0 : if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3051 : (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
3052 : && CFQQ_SEEKY(cfqq)))
3053 0 : enable_idle = 0;
3054 0 : else if (sample_valid(cic->ttime_samples)) {
3055 0 : if (cic->ttime_mean > cfqd->cfq_slice_idle)
3056 0 : enable_idle = 0;
3057 : else
3058 0 : enable_idle = 1;
3059 : }
3060 :
3061 0 : if (old_idle != enable_idle) {
3062 : cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3063 0 : if (enable_idle)
3064 0 : cfq_mark_cfqq_idle_window(cfqq);
3065 : else
3066 0 : cfq_clear_cfqq_idle_window(cfqq);
3067 0 : }
3068 : }
3069 :
3070 : /*
3071 : * Check if new_cfqq should preempt the currently active queue. Return 0 for
3072 : * no or if we aren't sure, a 1 will cause a preempt.
3073 : */
3074 : static bool
3075 : cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3076 : struct request *rq)
3077 : {
3078 0 : struct cfq_queue *cfqq;
3079 0 :
3080 0 : cfqq = cfqd->active_queue;
3081 0 : if (!cfqq)
3082 0 : return false;
3083 0 :
3084 0 : if (cfq_class_idle(new_cfqq))
3085 0 : return false;
3086 :
3087 0 : if (cfq_class_idle(cfqq))
3088 0 : return true;
3089 :
3090 : /*
3091 : * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3092 : */
3093 0 : if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3094 0 : return false;
3095 :
3096 : /*
3097 : * if the new request is sync, but the currently running queue is
3098 : * not, let the sync request have priority.
3099 : */
3100 0 : if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3101 0 : return true;
3102 :
3103 0 : if (new_cfqq->cfqg != cfqq->cfqg)
3104 0 : return false;
3105 :
3106 0 : if (cfq_slice_used(cfqq))
3107 0 : return true;
3108 :
3109 : /* Allow preemption only if we are idling on sync-noidle tree */
3110 0 : if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3111 : cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3112 : new_cfqq->service_tree->count == 2 &&
3113 : RB_EMPTY_ROOT(&cfqq->sort_list))
3114 0 : return true;
3115 :
3116 : /*
3117 : * So both queues are sync. Let the new request get disk time if
3118 : * it's a metadata request and the current queue is doing regular IO.
3119 : */
3120 0 : if (rq_is_meta(rq) && !cfqq->meta_pending)
3121 0 : return true;
3122 :
3123 : /*
3124 : * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3125 : */
3126 0 : if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3127 0 : return true;
3128 :
3129 0 : if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3130 0 : return false;
3131 :
3132 : /*
3133 : * if this request is as-good as one we would expect from the
3134 : * current cfqq, let it preempt
3135 : */
3136 0 : if (cfq_rq_close(cfqd, cfqq, rq, true))
3137 0 : return true;
3138 :
3139 0 : return false;
3140 : }
3141 :
3142 : /*
3143 : * cfqq preempts the active queue. if we allowed preempt with no slice left,
3144 : * let it have half of its nominal slice.
3145 : */
3146 : static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3147 : {
3148 0 : cfq_log_cfqq(cfqd, cfqq, "preempt");
3149 0 : cfq_slice_expired(cfqd, 1);
3150 :
3151 : /*
3152 : * Put the new queue at the front of the of the current list,
3153 : * so we know that it will be selected next.
3154 : */
3155 0 : BUG_ON(!cfq_cfqq_on_rr(cfqq));
3156 :
3157 0 : cfq_service_tree_add(cfqd, cfqq, 1);
3158 :
3159 0 : cfqq->slice_end = 0;
3160 0 : cfq_mark_cfqq_slice_new(cfqq);
3161 0 : }
3162 :
3163 : /*
3164 : * Called when a new fs request (rq) is added (to cfqq). Check if there's
3165 : * something we should do about it
3166 : */
3167 : static void
3168 : cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3169 : struct request *rq)
3170 0 : {
3171 0 : struct cfq_io_context *cic = RQ_CIC(rq);
3172 0 :
3173 0 : cfqd->rq_queued++;
3174 0 : if (rq_is_meta(rq))
3175 0 : cfqq->meta_pending++;
3176 :
3177 0 : cfq_update_io_thinktime(cfqd, cic);
3178 0 : cfq_update_io_seektime(cfqd, cfqq, rq);
3179 0 : cfq_update_idle_window(cfqd, cfqq, cic);
3180 :
3181 0 : cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3182 :
3183 0 : if (cfqq == cfqd->active_queue) {
3184 : /*
3185 : * Remember that we saw a request from this process, but
3186 : * don't start queuing just yet. Otherwise we risk seeing lots
3187 : * of tiny requests, because we disrupt the normal plugging
3188 : * and merging. If the request is already larger than a single
3189 : * page, let it rip immediately. For that case we assume that
3190 : * merging is already done. Ditto for a busy system that
3191 : * has other work pending, don't risk delaying until the
3192 : * idle timer unplug to continue working.
3193 : */
3194 0 : if (cfq_cfqq_wait_request(cfqq)) {
3195 0 : if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3196 : cfqd->busy_queues > 1) {
3197 0 : del_timer(&cfqd->idle_slice_timer);
3198 0 : cfq_clear_cfqq_wait_request(cfqq);
3199 0 : __blk_run_queue(cfqd->queue);
3200 : } else
3201 0 : cfq_mark_cfqq_must_dispatch(cfqq);
3202 : }
3203 0 : } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3204 : /*
3205 : * not the active queue - expire current slice if it is
3206 : * idle and has expired it's mean thinktime or this new queue
3207 : * has some old slice time left and is of higher priority or
3208 : * this new queue is RT and the current one is BE
3209 : */
3210 0 : cfq_preempt_queue(cfqd, cfqq);
3211 0 : __blk_run_queue(cfqd->queue);
3212 : }
3213 0 : }
3214 :
3215 : static void cfq_insert_request(struct request_queue *q, struct request *rq)
3216 : {
3217 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
3218 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
3219 0 :
3220 : cfq_log_cfqq(cfqd, cfqq, "insert_request");
3221 0 : cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3222 :
3223 0 : rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3224 0 : list_add_tail(&rq->queuelist, &cfqq->fifo);
3225 0 : cfq_add_rq_rb(rq);
3226 :
3227 0 : cfq_rq_enqueued(cfqd, cfqq, rq);
3228 0 : }
3229 :
3230 : /*
3231 : * Update hw_tag based on peak queue depth over 50 samples under
3232 : * sufficient load.
3233 : */
3234 : static void cfq_update_hw_tag(struct cfq_data *cfqd)
3235 : {
3236 0 : struct cfq_queue *cfqq = cfqd->active_queue;
3237 0 :
3238 0 : if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
3239 0 : cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
3240 0 :
3241 0 : if (cfqd->hw_tag == 1)
3242 0 : return;
3243 :
3244 0 : if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3245 : rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
3246 0 : return;
3247 :
3248 : /*
3249 : * If active queue hasn't enough requests and can idle, cfq might not
3250 : * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3251 : * case
3252 : */
3253 0 : if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3254 : cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3255 : CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
3256 0 : return;
3257 :
3258 0 : if (cfqd->hw_tag_samples++ < 50)
3259 0 : return;
3260 :
3261 0 : if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3262 0 : cfqd->hw_tag = 1;
3263 : else
3264 0 : cfqd->hw_tag = 0;
3265 0 : }
3266 :
3267 : static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3268 : {
3269 0 : struct cfq_io_context *cic = cfqd->active_cic;
3270 0 :
3271 : /* If the queue already has requests, don't wait */
3272 0 : if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3273 0 : return false;
3274 :
3275 : /* If there are other queues in the group, don't wait */
3276 0 : if (cfqq->cfqg->nr_cfqq > 1)
3277 0 : return false;
3278 :
3279 0 : if (cfq_slice_used(cfqq))
3280 0 : return true;
3281 :
3282 : /* if slice left is less than think time, wait busy */
3283 0 : if (cic && sample_valid(cic->ttime_samples)
3284 : && (cfqq->slice_end - jiffies < cic->ttime_mean))
3285 0 : return true;
3286 :
3287 : /*
3288 : * If think times is less than a jiffy than ttime_mean=0 and above
3289 : * will not be true. It might happen that slice has not expired yet
3290 : * but will expire soon (4-5 ns) during select_queue(). To cover the
3291 : * case where think time is less than a jiffy, mark the queue wait
3292 : * busy if only 1 jiffy is left in the slice.
3293 : */
3294 0 : if (cfqq->slice_end - jiffies == 1)
3295 0 : return true;
3296 :
3297 0 : return false;
3298 : }
3299 :
3300 : static void cfq_completed_request(struct request_queue *q, struct request *rq)
3301 : {
3302 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
3303 0 : struct cfq_data *cfqd = cfqq->cfqd;
3304 0 : const int sync = rq_is_sync(rq);
3305 0 : unsigned long now;
3306 0 :
3307 0 : now = jiffies;
3308 0 : cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3309 0 :
3310 0 : cfq_update_hw_tag(cfqd);
3311 0 :
3312 0 : WARN_ON(!cfqd->rq_in_driver[sync]);
3313 0 : WARN_ON(!cfqq->dispatched);
3314 0 : cfqd->rq_in_driver[sync]--;
3315 0 : cfqq->dispatched--;
3316 0 :
3317 0 : if (cfq_cfqq_sync(cfqq))
3318 0 : cfqd->sync_flight--;
3319 :
3320 0 : if (sync) {
3321 0 : RQ_CIC(rq)->last_end_request = now;
3322 0 : if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3323 0 : cfqd->last_delayed_sync = now;
3324 : }
3325 :
3326 : /*
3327 : * If this is the active queue, check if it needs to be expired,
3328 : * or if we want to idle in case it has no pending requests.
3329 : */
3330 0 : if (cfqd->active_queue == cfqq) {
3331 0 : const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3332 :
3333 0 : if (cfq_cfqq_slice_new(cfqq)) {
3334 0 : cfq_set_prio_slice(cfqd, cfqq);
3335 0 : cfq_clear_cfqq_slice_new(cfqq);
3336 : }
3337 :
3338 : /*
3339 : * Should we wait for next request to come in before we expire
3340 : * the queue.
3341 : */
3342 0 : if (cfq_should_wait_busy(cfqd, cfqq)) {
3343 0 : cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3344 0 : cfq_mark_cfqq_wait_busy(cfqq);
3345 : }
3346 :
3347 : /*
3348 : * Idling is not enabled on:
3349 : * - expired queues
3350 : * - idle-priority queues
3351 : * - async queues
3352 : * - queues with still some requests queued
3353 : * - when there is a close cooperator
3354 : */
3355 0 : if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3356 0 : cfq_slice_expired(cfqd, 1);
3357 0 : else if (sync && cfqq_empty &&
3358 : !cfq_close_cooperator(cfqd, cfqq)) {
3359 0 : cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3360 : /*
3361 : * Idling is enabled for SYNC_WORKLOAD.
3362 : * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3363 : * only if we processed at least one !rq_noidle request
3364 : */
3365 0 : if (cfqd->serving_type == SYNC_WORKLOAD
3366 : || cfqd->noidle_tree_requires_idle
3367 : || cfqq->cfqg->nr_cfqq == 1)
3368 0 : cfq_arm_slice_timer(cfqd);
3369 : }
3370 : }
3371 :
3372 0 : if (!rq_in_driver(cfqd))
3373 0 : cfq_schedule_dispatch(cfqd);
3374 0 : }
3375 :
3376 : /*
3377 : * we temporarily boost lower priority queues if they are holding fs exclusive
3378 : * resources. they are boosted to normal prio (CLASS_BE/4)
3379 : */
3380 : static void cfq_prio_boost(struct cfq_queue *cfqq)
3381 : {
3382 0 : if (has_fs_excl()) {
3383 0 : /*
3384 : * boost idle prio on transactions that would lock out other
3385 : * users of the filesystem
3386 : */
3387 0 : if (cfq_class_idle(cfqq))
3388 0 : cfqq->ioprio_class = IOPRIO_CLASS_BE;
3389 0 : if (cfqq->ioprio > IOPRIO_NORM)
3390 0 : cfqq->ioprio = IOPRIO_NORM;
3391 : } else {
3392 : /*
3393 : * unboost the queue (if needed)
3394 : */
3395 0 : cfqq->ioprio_class = cfqq->org_ioprio_class;
3396 0 : cfqq->ioprio = cfqq->org_ioprio;
3397 : }
3398 0 : }
3399 :
3400 : static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3401 : {
3402 0 : if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3403 0 : cfq_mark_cfqq_must_alloc_slice(cfqq);
3404 0 : return ELV_MQUEUE_MUST;
3405 : }
3406 :
3407 0 : return ELV_MQUEUE_MAY;
3408 : }
3409 :
3410 : static int cfq_may_queue(struct request_queue *q, int rw)
3411 : {
3412 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
3413 0 : struct task_struct *tsk = current;
3414 0 : struct cfq_io_context *cic;
3415 0 : struct cfq_queue *cfqq;
3416 0 :
3417 0 : /*
3418 0 : * don't force setup of a queue from here, as a call to may_queue
3419 : * does not necessarily imply that a request actually will be queued.
3420 : * so just lookup a possibly existing queue, or return 'may queue'
3421 : * if that fails
3422 : */
3423 0 : cic = cfq_cic_lookup(cfqd, tsk->io_context);
3424 0 : if (!cic)
3425 0 : return ELV_MQUEUE_MAY;
3426 :
3427 0 : cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3428 0 : if (cfqq) {
3429 0 : cfq_init_prio_data(cfqq, cic->ioc);
3430 0 : cfq_prio_boost(cfqq);
3431 :
3432 0 : return __cfq_may_queue(cfqq);
3433 : }
3434 :
3435 0 : return ELV_MQUEUE_MAY;
3436 : }
3437 :
3438 : /*
3439 : * queue lock held here
3440 : */
3441 : static void cfq_put_request(struct request *rq)
3442 : {
3443 0 : struct cfq_queue *cfqq = RQ_CFQQ(rq);
3444 0 :
3445 0 : if (cfqq) {
3446 0 : const int rw = rq_data_dir(rq);
3447 :
3448 0 : BUG_ON(!cfqq->allocated[rw]);
3449 0 : cfqq->allocated[rw]--;
3450 :
3451 0 : put_io_context(RQ_CIC(rq)->ioc);
3452 :
3453 0 : rq->elevator_private = NULL;
3454 0 : rq->elevator_private2 = NULL;
3455 :
3456 0 : cfq_put_queue(cfqq);
3457 : }
3458 0 : }
3459 :
3460 : static struct cfq_queue *
3461 : cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3462 : struct cfq_queue *cfqq)
3463 : {
3464 0 : cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3465 0 : cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3466 0 : cfq_mark_cfqq_coop(cfqq->new_cfqq);
3467 0 : cfq_put_queue(cfqq);
3468 0 : return cic_to_cfqq(cic, 1);
3469 : }
3470 :
3471 : /*
3472 : * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3473 : * was the last process referring to said cfqq.
3474 : */
3475 : static struct cfq_queue *
3476 : split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3477 : {
3478 0 : if (cfqq_process_refs(cfqq) == 1) {
3479 0 : cfqq->pid = current->pid;
3480 0 : cfq_clear_cfqq_coop(cfqq);
3481 0 : cfq_clear_cfqq_split_coop(cfqq);
3482 0 : return cfqq;
3483 : }
3484 :
3485 0 : cic_set_cfqq(cic, NULL, 1);
3486 :
3487 0 : cfq_put_cooperator(cfqq);
3488 :
3489 0 : cfq_put_queue(cfqq);
3490 0 : return NULL;
3491 : }
3492 : /*
3493 : * Allocate cfq data structures associated with this request.
3494 : */
3495 : static int
3496 : cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3497 : {
3498 0 : struct cfq_data *cfqd = q->elevator->elevator_data;
3499 0 : struct cfq_io_context *cic;
3500 0 : const int rw = rq_data_dir(rq);
3501 0 : const bool is_sync = rq_is_sync(rq);
3502 0 : struct cfq_queue *cfqq;
3503 0 : unsigned long flags;
3504 0 :
3505 0 : might_sleep_if(gfp_mask & __GFP_WAIT);
3506 0 :
3507 0 : cic = cfq_get_io_context(cfqd, gfp_mask);
3508 :
3509 0 : spin_lock_irqsave(q->queue_lock, flags);
3510 :
3511 0 : if (!cic)
3512 0 : goto queue_fail;
3513 :
3514 : new_queue:
3515 0 : cfqq = cic_to_cfqq(cic, is_sync);
3516 0 : if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3517 0 : cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3518 0 : cic_set_cfqq(cic, cfqq, is_sync);
3519 : } else {
3520 : /*
3521 : * If the queue was seeky for too long, break it apart.
3522 : */
3523 0 : if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3524 : cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3525 0 : cfqq = split_cfqq(cic, cfqq);
3526 0 : if (!cfqq)
3527 0 : goto new_queue;
3528 : }
3529 :
3530 : /*
3531 : * Check to see if this queue is scheduled to merge with
3532 : * another, closely cooperating queue. The merging of
3533 : * queues happens here as it must be done in process context.
3534 : * The reference on new_cfqq was taken in merge_cfqqs.
3535 : */
3536 0 : if (cfqq->new_cfqq)
3537 0 : cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3538 : }
3539 :
3540 0 : cfqq->allocated[rw]++;
3541 0 : atomic_inc(&cfqq->ref);
3542 :
3543 0 : spin_unlock_irqrestore(q->queue_lock, flags);
3544 :
3545 0 : rq->elevator_private = cic;
3546 0 : rq->elevator_private2 = cfqq;
3547 0 : return 0;
3548 0 :
3549 : queue_fail:
3550 0 : if (cic)
3551 0 : put_io_context(cic->ioc);
3552 :
3553 0 : cfq_schedule_dispatch(cfqd);
3554 0 : spin_unlock_irqrestore(q->queue_lock, flags);
3555 : cfq_log(cfqd, "set_request fail");
3556 0 : return 1;
3557 : }
3558 :
3559 : static void cfq_kick_queue(struct work_struct *work)
3560 : {
3561 0 : struct cfq_data *cfqd =
3562 0 : container_of(work, struct cfq_data, unplug_work);
3563 0 : struct request_queue *q = cfqd->queue;
3564 :
3565 0 : spin_lock_irq(q->queue_lock);
3566 0 : __blk_run_queue(cfqd->queue);
3567 0 : spin_unlock_irq(q->queue_lock);
3568 0 : }
3569 :
3570 : /*
3571 : * Timer running if the active_queue is currently idling inside its time slice
3572 : */
3573 : static void cfq_idle_slice_timer(unsigned long data)
3574 : {
3575 0 : struct cfq_data *cfqd = (struct cfq_data *) data;
3576 0 : struct cfq_queue *cfqq;
3577 0 : unsigned long flags;
3578 0 : int timed_out = 1;
3579 0 :
3580 0 : cfq_log(cfqd, "idle timer fired");
3581 0 :
3582 0 : spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3583 :
3584 0 : cfqq = cfqd->active_queue;
3585 0 : if (cfqq) {
3586 0 : timed_out = 0;
3587 :
3588 : /*
3589 : * We saw a request before the queue expired, let it through
3590 : */
3591 0 : if (cfq_cfqq_must_dispatch(cfqq))
3592 0 : goto out_kick;
3593 :
3594 : /*
3595 : * expired
3596 : */
3597 0 : if (cfq_slice_used(cfqq))
3598 0 : goto expire;
3599 :
3600 : /*
3601 : * only expire and reinvoke request handler, if there are
3602 : * other queues with pending requests
3603 : */
3604 0 : if (!cfqd->busy_queues)
3605 0 : goto out_cont;
3606 :
3607 : /*
3608 : * not expired and it has a request pending, let it dispatch
3609 : */
3610 0 : if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3611 0 : goto out_kick;
3612 :
3613 : /*
3614 : * Queue depth flag is reset only when the idle didn't succeed
3615 : */
3616 0 : cfq_clear_cfqq_deep(cfqq);
3617 : }
3618 : expire:
3619 0 : cfq_slice_expired(cfqd, timed_out);
3620 : out_kick:
3621 0 : cfq_schedule_dispatch(cfqd);
3622 : out_cont:
3623 0 : spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3624 0 : }
3625 :
3626 : static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3627 : {
3628 0 : del_timer_sync(&cfqd->idle_slice_timer);
3629 0 : cancel_work_sync(&cfqd->unplug_work);
3630 0 : }
3631 :
3632 : static void cfq_put_async_queues(struct cfq_data *cfqd)
3633 : {
3634 0 : int i;
3635 :
3636 0 : for (i = 0; i < IOPRIO_BE_NR; i++) {
3637 0 : if (cfqd->async_cfqq[0][i])
3638 0 : cfq_put_queue(cfqd->async_cfqq[0][i]);
3639 0 : if (cfqd->async_cfqq[1][i])
3640 0 : cfq_put_queue(cfqd->async_cfqq[1][i]);
3641 : }
3642 :
3643 0 : if (cfqd->async_idle_cfqq)
3644 0 : cfq_put_queue(cfqd->async_idle_cfqq);
3645 0 : }
3646 :
3647 : static void cfq_cfqd_free(struct rcu_head *head)
3648 : {
3649 0 : kfree(container_of(head, struct cfq_data, rcu));
3650 : }
3651 0 :
3652 : static void cfq_exit_queue(struct elevator_queue *e)
3653 : {
3654 0 : struct cfq_data *cfqd = e->elevator_data;
3655 0 : struct request_queue *q = cfqd->queue;
3656 0 :
3657 0 : cfq_shutdown_timer_wq(cfqd);
3658 0 :
3659 0 : spin_lock_irq(q->queue_lock);
3660 :
3661 0 : if (cfqd->active_queue)
3662 0 : __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3663 :
3664 0 : while (!list_empty(&cfqd->cic_list)) {
3665 0 : struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3666 0 : struct cfq_io_context,
3667 : queue_list);
3668 :
3669 0 : __cfq_exit_single_io_context(cfqd, cic);
3670 : }
3671 0 :
3672 0 : cfq_put_async_queues(cfqd);
3673 0 : cfq_release_cfq_groups(cfqd);
3674 0 : blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3675 :
3676 0 : spin_unlock_irq(q->queue_lock);
3677 :
3678 0 : cfq_shutdown_timer_wq(cfqd);
3679 :
3680 : /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3681 0 : call_rcu(&cfqd->rcu, cfq_cfqd_free);
3682 0 : }
3683 :
3684 : static void *cfq_init_queue(struct request_queue *q)
3685 : {
3686 0 : struct cfq_data *cfqd;
3687 0 : int i, j;
3688 0 : struct cfq_group *cfqg;
3689 0 : struct cfq_rb_root *st;
3690 0 :
3691 0 : cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3692 0 : if (!cfqd)
3693 0 : return NULL;
3694 0 :
3695 0 : /* Init root service tree */
3696 0 : cfqd->grp_service_tree = CFQ_RB_ROOT;
3697 :
3698 : /* Init root group */
3699 0 : cfqg = &cfqd->root_group;
3700 0 : for_each_cfqg_st(cfqg, i, j, st)
3701 0 : *st = CFQ_RB_ROOT;
3702 0 : RB_CLEAR_NODE(&cfqg->rb_node);
3703 :
3704 : /* Give preference to root group over other groups */
3705 0 : cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3706 :
3707 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
3708 : /*
3709 : * Take a reference to root group which we never drop. This is just
3710 : * to make sure that cfq_put_cfqg() does not try to kfree root group
3711 : */
3712 : atomic_set(&cfqg->ref, 1);
3713 : blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3714 : 0);
3715 : #endif
3716 : /*
3717 : * Not strictly needed (since RB_ROOT just clears the node and we
3718 : * zeroed cfqd on alloc), but better be safe in case someone decides
3719 : * to add magic to the rb code
3720 : */
3721 0 : for (i = 0; i < CFQ_PRIO_LISTS; i++)
3722 0 : cfqd->prio_trees[i] = RB_ROOT;
3723 0 :
3724 : /*
3725 : * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3726 : * Grab a permanent reference to it, so that the normal code flow
3727 : * will not attempt to free it.
3728 : */
3729 0 : cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3730 0 : atomic_inc(&cfqd->oom_cfqq.ref);
3731 0 : cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3732 :
3733 0 : INIT_LIST_HEAD(&cfqd->cic_list);
3734 :
3735 0 : cfqd->queue = q;
3736 :
3737 0 : init_timer(&cfqd->idle_slice_timer);
3738 0 : cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3739 0 : cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3740 :
3741 0 : INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3742 :
3743 0 : cfqd->cfq_quantum = cfq_quantum;
3744 0 : cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3745 0 : cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3746 0 : cfqd->cfq_back_max = cfq_back_max;
3747 0 : cfqd->cfq_back_penalty = cfq_back_penalty;
3748 0 : cfqd->cfq_slice[0] = cfq_slice_async;
3749 0 : cfqd->cfq_slice[1] = cfq_slice_sync;
3750 0 : cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3751 0 : cfqd->cfq_slice_idle = cfq_slice_idle;
3752 0 : cfqd->cfq_latency = 1;
3753 0 : cfqd->cfq_group_isolation = 0;
3754 0 : cfqd->hw_tag = -1;
3755 : /*
3756 : * we optimistically start assuming sync ops weren't delayed in last
3757 : * second, in order to have larger depth for async operations.
3758 : */
3759 0 : cfqd->last_delayed_sync = jiffies - HZ;
3760 0 : INIT_RCU_HEAD(&cfqd->rcu);
3761 0 : return cfqd;
3762 : }
3763 :
3764 : static void cfq_slab_kill(void)
3765 : {
3766 : /*
3767 : * Caller already ensured that pending RCU callbacks are completed,
3768 : * so we should have no busy allocations at this point.
3769 : */
3770 6 : if (cfq_pool)
3771 3 : kmem_cache_destroy(cfq_pool);
3772 6 : if (cfq_ioc_pool)
3773 3 : kmem_cache_destroy(cfq_ioc_pool);
3774 3 : }
3775 :
3776 : static int __init cfq_slab_setup(void)
3777 : {
3778 1 : cfq_pool = KMEM_CACHE(cfq_queue, 0);
3779 2 : if (!cfq_pool)
3780 1 : goto fail;
3781 :
3782 1 : cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3783 2 : if (!cfq_ioc_pool)
3784 1 : goto fail;
3785 :
3786 1 : return 0;
3787 1 : fail:
3788 2 : cfq_slab_kill();
3789 1 : return -ENOMEM;
3790 : }
3791 :
3792 : /*
3793 : * sysfs parts below -->
3794 : */
3795 : static ssize_t
3796 : cfq_var_show(unsigned int var, char *page)
3797 : {
3798 33 : return sprintf(page, "%d\n", var);
3799 : }
3800 :
3801 : static ssize_t
3802 : cfq_var_store(unsigned int *var, const char *page, size_t count)
3803 : {
3804 22 : char *p = (char *) page;
3805 11 :
3806 22 : *var = simple_strtoul(p, &p, 10);
3807 11 : return count;
3808 : }
3809 :
3810 : #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3811 : static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3812 : { \
3813 : struct cfq_data *cfqd = e->elevator_data; \
3814 : unsigned int __data = __VAR; \
3815 : if (__CONV) \
3816 : __data = jiffies_to_msecs(__data); \
3817 : return cfq_var_show(__data, (page)); \
3818 : }
3819 6 : SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3820 7 : SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3821 8 : SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3822 8 : SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3823 9 : SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3824 10 : SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3825 10 : SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3826 10 : SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3827 9 : SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3828 9 : SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3829 9 : SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3830 3 : #undef SHOW_FUNCTION
3831 3 :
3832 2 : #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3833 1 : static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3834 : { \
3835 : struct cfq_data *cfqd = e->elevator_data; \
3836 : unsigned int __data; \
3837 : int ret = cfq_var_store(&__data, (page), count); \
3838 : if (__data < (MIN)) \
3839 : __data = (MIN); \
3840 : else if (__data > (MAX)) \
3841 : __data = (MAX); \
3842 : if (__CONV) \
3843 : *(__PTR) = msecs_to_jiffies(__data); \
3844 : else \
3845 : *(__PTR) = __data; \
3846 : return ret; \
3847 : }
3848 10 : STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3849 11 : STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3850 1 : UINT_MAX, 1);
3851 12 : STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3852 2 : UINT_MAX, 1);
3853 9 : STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3854 12 : STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3855 2 : UINT_MAX, 0);
3856 11 : STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3857 14 : STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3858 14 : STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3859 13 : STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3860 3 : UINT_MAX, 0);
3861 13 : STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3862 14 : STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3863 4 : #undef STORE_FUNCTION
3864 4 :
3865 4 : #define CFQ_ATTR(name) \
3866 2 : __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3867 2 :
3868 2 : static struct elv_fs_entry cfq_attrs[] = {
3869 : CFQ_ATTR(quantum),
3870 : CFQ_ATTR(fifo_expire_sync),
3871 : CFQ_ATTR(fifo_expire_async),
3872 : CFQ_ATTR(back_seek_max),
3873 : CFQ_ATTR(back_seek_penalty),
3874 : CFQ_ATTR(slice_sync),
3875 : CFQ_ATTR(slice_async),
3876 : CFQ_ATTR(slice_async_rq),
3877 : CFQ_ATTR(slice_idle),
3878 : CFQ_ATTR(low_latency),
3879 : CFQ_ATTR(group_isolation),
3880 : __ATTR_NULL
3881 : };
3882 :
3883 1 : static struct elevator_type iosched_cfq = {
3884 : .ops = {
3885 : .elevator_merge_fn = cfq_merge,
3886 : .elevator_merged_fn = cfq_merged_request,
3887 : .elevator_merge_req_fn = cfq_merged_requests,
3888 : .elevator_allow_merge_fn = cfq_allow_merge,
3889 : .elevator_dispatch_fn = cfq_dispatch_requests,
3890 : .elevator_add_req_fn = cfq_insert_request,
3891 : .elevator_activate_req_fn = cfq_activate_request,
3892 : .elevator_deactivate_req_fn = cfq_deactivate_request,
3893 : .elevator_queue_empty_fn = cfq_queue_empty,
3894 : .elevator_completed_req_fn = cfq_completed_request,
3895 : .elevator_former_req_fn = elv_rb_former_request,
3896 : .elevator_latter_req_fn = elv_rb_latter_request,
3897 : .elevator_set_req_fn = cfq_set_request,
3898 : .elevator_put_req_fn = cfq_put_request,
3899 : .elevator_may_queue_fn = cfq_may_queue,
3900 : .elevator_init_fn = cfq_init_queue,
3901 : .elevator_exit_fn = cfq_exit_queue,
3902 : .trim = cfq_free_io_context,
3903 : },
3904 : .elevator_attrs = cfq_attrs,
3905 : .elevator_name = "cfq",
3906 : .elevator_owner = THIS_MODULE,
3907 : };
3908 :
3909 : #ifdef CONFIG_CFQ_GROUP_IOSCHED
3910 : static struct blkio_policy_type blkio_policy_cfq = {
3911 : .ops = {
3912 : .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3913 : .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3914 : },
3915 : };
3916 : #else
3917 1 : static struct blkio_policy_type blkio_policy_cfq;
3918 : #endif
3919 :
3920 : static int __init cfq_init(void)
3921 : {
3922 1 : /*
3923 : * could be 0 on HZ < 1000 setups
3924 : */
3925 2 : if (!cfq_slice_async)
3926 1 : cfq_slice_async = 1;
3927 2 : if (!cfq_slice_idle)
3928 1 : cfq_slice_idle = 1;
3929 :
3930 5 : if (cfq_slab_setup())
3931 1 : return -ENOMEM;
3932 :
3933 1 : elv_register(&iosched_cfq);
3934 2 : blkio_policy_register(&blkio_policy_cfq);
3935 :
3936 1 : return 0;
3937 : }
3938 :
3939 : static void __exit cfq_exit(void)
3940 : {
3941 10 : DECLARE_COMPLETION_ONSTACK(all_gone);
3942 6 : blkio_policy_unregister(&blkio_policy_cfq);
3943 4 : elv_unregister(&iosched_cfq);
3944 4 : ioc_gone = &all_gone;
3945 2 : /* ioc_gone's update must be visible before reading ioc_count */
3946 2 : smp_wmb();
3947 :
3948 : /*
3949 : * this also protects us from entering cfq_slab_kill() with
3950 : * pending RCU callbacks
3951 : */
3952 28 : if (elv_ioc_count_read(cfq_ioc_count))
3953 6 : wait_for_completion(&all_gone);
3954 6 : cfq_slab_kill();
3955 2 : }
3956 :
3957 : module_init(cfq_init);
3958 : module_exit(cfq_exit);
3959 1 :
3960 : MODULE_AUTHOR("Jens Axboe");
3961 : MODULE_LICENSE("GPL");
3962 : MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");
|