Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Block rq-qos base io controller
4 *
5 * This works similar to wbt with a few exceptions
6 *
7 * - It's bio based, so the latency covers the whole block layer in addition to
8 * the actual io.
9 * - We will throttle all IO that comes in here if we need to.
10 * - We use the mean latency over the 100ms window. This is because writes can
11 * be particularly fast, which could give us a false sense of the impact of
12 * other workloads on our protected workload.
13 * - By default there's no throttling, we set the queue_depth to UINT_MAX so
14 * that we can have as many outstanding bio's as we're allowed to. Only at
15 * throttle time do we pay attention to the actual queue depth.
16 *
17 * The hierarchy works like the cpu controller does, we track the latency at
18 * every configured node, and each configured node has it's own independent
19 * queue depth. This means that we only care about our latency targets at the
20 * peer level. Some group at the bottom of the hierarchy isn't going to affect
21 * a group at the end of some other path if we're only configred at leaf level.
22 *
23 * Consider the following
24 *
25 * root blkg
26 * / \
27 * fast (target=5ms) slow (target=10ms)
28 * / \ / \
29 * a b normal(15ms) unloved
30 *
31 * "a" and "b" have no target, but their combined io under "fast" cannot exceed
32 * an average latency of 5ms. If it does then we will throttle the "slow"
33 * group. In the case of "normal", if it exceeds its 15ms target, we will
34 * throttle "unloved", but nobody else.
35 *
36 * In this example "fast", "slow", and "normal" will be the only groups actually
37 * accounting their io latencies. We have to walk up the heirarchy to the root
38 * on every submit and complete so we can do the appropriate stat recording and
39 * adjust the queue depth of ourselves if needed.
40 *
41 * There are 2 ways we throttle IO.
42 *
43 * 1) Queue depth throttling. As we throttle down we will adjust the maximum
44 * number of IO's we're allowed to have in flight. This starts at (u64)-1 down
45 * to 1. If the group is only ever submitting IO for itself then this is the
46 * only way we throttle.
47 *
48 * 2) Induced delay throttling. This is for the case that a group is generating
49 * IO that has to be issued by the root cg to avoid priority inversion. So think
50 * REQ_META or REQ_SWAP. If we are already at qd == 1 and we're getting a lot
51 * of work done for us on behalf of the root cg and are being asked to scale
52 * down more then we induce a latency at userspace return. We accumulate the
53 * total amount of time we need to be punished by doing
54 *
55 * total_time += min_lat_nsec - actual_io_completion
56 *
57 * and then at throttle time will do
58 *
59 * throttle_time = min(total_time, NSEC_PER_SEC)
60 *
61 * This induced delay will throttle back the activity that is generating the
62 * root cg issued io's, wethere that's some metadata intensive operation or the
63 * group is using so much memory that it is pushing us into swap.
64 *
65 * Copyright (C) 2018 Josef Bacik
66 */
67#include <linux/kernel.h>
68#include <linux/blk_types.h>
69#include <linux/backing-dev.h>
70#include <linux/module.h>
71#include <linux/timer.h>
72#include <linux/memcontrol.h>
73#include <linux/sched/loadavg.h>
74#include <linux/sched/signal.h>
75#include <trace/events/block.h>
76#include <linux/blk-mq.h>
77#include "blk-rq-qos.h"
78#include "blk-stat.h"
79#include "blk-cgroup.h"
80#include "blk.h"
81
82#define DEFAULT_SCALE_COOKIE 1000000U
83
84static struct blkcg_policy blkcg_policy_iolatency;
85struct iolatency_grp;
86
87struct blk_iolatency {
88 struct rq_qos rqos;
89 struct timer_list timer;
90
91 /*
92 * ->enabled is the master enable switch gating the throttling logic and
93 * inflight tracking. The number of cgroups which have iolat enabled is
94 * tracked in ->enable_cnt, and ->enable is flipped on/off accordingly
95 * from ->enable_work with the request_queue frozen. For details, See
96 * blkiolatency_enable_work_fn().
97 */
98 bool enabled;
99 atomic_t enable_cnt;
100 struct work_struct enable_work;
101};
102
103static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
104{
105 return container_of(rqos, struct blk_iolatency, rqos);
106}
107
108struct child_latency_info {
109 spinlock_t lock;
110
111 /* Last time we adjusted the scale of everybody. */
112 u64 last_scale_event;
113
114 /* The latency that we missed. */
115 u64 scale_lat;
116
117 /* Total io's from all of our children for the last summation. */
118 u64 nr_samples;
119
120 /* The guy who actually changed the latency numbers. */
121 struct iolatency_grp *scale_grp;
122
123 /* Cookie to tell if we need to scale up or down. */
124 atomic_t scale_cookie;
125};
126
127struct percentile_stats {
128 u64 total;
129 u64 missed;
130};
131
132struct latency_stat {
133 union {
134 struct percentile_stats ps;
135 struct blk_rq_stat rqs;
136 };
137};
138
139struct iolatency_grp {
140 struct blkg_policy_data pd;
141 struct latency_stat __percpu *stats;
142 struct latency_stat cur_stat;
143 struct blk_iolatency *blkiolat;
144 unsigned int max_depth;
145 struct rq_wait rq_wait;
146 atomic64_t window_start;
147 atomic_t scale_cookie;
148 u64 min_lat_nsec;
149 u64 cur_win_nsec;
150
151 /* total running average of our io latency. */
152 u64 lat_avg;
153
154 /* Our current number of IO's for the last summation. */
155 u64 nr_samples;
156
157 bool ssd;
158 struct child_latency_info child_lat;
159};
160
161#define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
162#define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
163/*
164 * These are the constants used to fake the fixed-point moving average
165 * calculation just like load average. The call to calc_load() folds
166 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
167 * window size is bucketed to try to approximately calculate average
168 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
169 * elapse immediately. Note, windows only elapse with IO activity. Idle
170 * periods extend the most recent window.
171 */
172#define BLKIOLATENCY_NR_EXP_FACTORS 5
173#define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
174 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
175static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
176 2045, // exp(1/600) - 600 samples
177 2039, // exp(1/240) - 240 samples
178 2031, // exp(1/120) - 120 samples
179 2023, // exp(1/80) - 80 samples
180 2014, // exp(1/60) - 60 samples
181};
182
183static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
184{
185 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
186}
187
188static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
189{
190 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
191}
192
193static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
194{
195 return pd_to_blkg(&iolat->pd);
196}
197
198static inline void latency_stat_init(struct iolatency_grp *iolat,
199 struct latency_stat *stat)
200{
201 if (iolat->ssd) {
202 stat->ps.total = 0;
203 stat->ps.missed = 0;
204 } else
205 blk_rq_stat_init(&stat->rqs);
206}
207
208static inline void latency_stat_sum(struct iolatency_grp *iolat,
209 struct latency_stat *sum,
210 struct latency_stat *stat)
211{
212 if (iolat->ssd) {
213 sum->ps.total += stat->ps.total;
214 sum->ps.missed += stat->ps.missed;
215 } else
216 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
217}
218
219static inline void latency_stat_record_time(struct iolatency_grp *iolat,
220 u64 req_time)
221{
222 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
223 if (iolat->ssd) {
224 if (req_time >= iolat->min_lat_nsec)
225 stat->ps.missed++;
226 stat->ps.total++;
227 } else
228 blk_rq_stat_add(&stat->rqs, req_time);
229 put_cpu_ptr(stat);
230}
231
232static inline bool latency_sum_ok(struct iolatency_grp *iolat,
233 struct latency_stat *stat)
234{
235 if (iolat->ssd) {
236 u64 thresh = div64_u64(stat->ps.total, 10);
237 thresh = max(thresh, 1ULL);
238 return stat->ps.missed < thresh;
239 }
240 return stat->rqs.mean <= iolat->min_lat_nsec;
241}
242
243static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
244 struct latency_stat *stat)
245{
246 if (iolat->ssd)
247 return stat->ps.total;
248 return stat->rqs.nr_samples;
249}
250
251static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
252 struct latency_stat *stat)
253{
254 int exp_idx;
255
256 if (iolat->ssd)
257 return;
258
259 /*
260 * calc_load() takes in a number stored in fixed point representation.
261 * Because we are using this for IO time in ns, the values stored
262 * are significantly larger than the FIXED_1 denominator (2048).
263 * Therefore, rounding errors in the calculation are negligible and
264 * can be ignored.
265 */
266 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
267 div64_u64(iolat->cur_win_nsec,
268 BLKIOLATENCY_EXP_BUCKET_SIZE));
269 iolat->lat_avg = calc_load(iolat->lat_avg,
270 iolatency_exp_factors[exp_idx],
271 stat->rqs.mean);
272}
273
274static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
275{
276 atomic_dec(&rqw->inflight);
277 wake_up(&rqw->wait);
278}
279
280static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
281{
282 struct iolatency_grp *iolat = private_data;
283 return rq_wait_inc_below(rqw, iolat->max_depth);
284}
285
286static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
287 struct iolatency_grp *iolat,
288 bool issue_as_root,
289 bool use_memdelay)
290{
291 struct rq_wait *rqw = &iolat->rq_wait;
292 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
293
294 if (use_delay)
295 blkcg_schedule_throttle(rqos->q->disk, use_memdelay);
296
297 /*
298 * To avoid priority inversions we want to just take a slot if we are
299 * issuing as root. If we're being killed off there's no point in
300 * delaying things, we may have been killed by OOM so throttling may
301 * make recovery take even longer, so just let the IO's through so the
302 * task can go away.
303 */
304 if (issue_as_root || fatal_signal_pending(current)) {
305 atomic_inc(&rqw->inflight);
306 return;
307 }
308
309 rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
310}
311
312#define SCALE_DOWN_FACTOR 2
313#define SCALE_UP_FACTOR 4
314
315static inline unsigned long scale_amount(unsigned long qd, bool up)
316{
317 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
318}
319
320/*
321 * We scale the qd down faster than we scale up, so we need to use this helper
322 * to adjust the scale_cookie accordingly so we don't prematurely get
323 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
324 *
325 * Each group has their own local copy of the last scale cookie they saw, so if
326 * the global scale cookie goes up or down they know which way they need to go
327 * based on their last knowledge of it.
328 */
329static void scale_cookie_change(struct blk_iolatency *blkiolat,
330 struct child_latency_info *lat_info,
331 bool up)
332{
333 unsigned long qd = blkiolat->rqos.q->nr_requests;
334 unsigned long scale = scale_amount(qd, up);
335 unsigned long old = atomic_read(&lat_info->scale_cookie);
336 unsigned long max_scale = qd << 1;
337 unsigned long diff = 0;
338
339 if (old < DEFAULT_SCALE_COOKIE)
340 diff = DEFAULT_SCALE_COOKIE - old;
341
342 if (up) {
343 if (scale + old > DEFAULT_SCALE_COOKIE)
344 atomic_set(&lat_info->scale_cookie,
345 DEFAULT_SCALE_COOKIE);
346 else if (diff > qd)
347 atomic_inc(&lat_info->scale_cookie);
348 else
349 atomic_add(scale, &lat_info->scale_cookie);
350 } else {
351 /*
352 * We don't want to dig a hole so deep that it takes us hours to
353 * dig out of it. Just enough that we don't throttle/unthrottle
354 * with jagged workloads but can still unthrottle once pressure
355 * has sufficiently dissipated.
356 */
357 if (diff > qd) {
358 if (diff < max_scale)
359 atomic_dec(&lat_info->scale_cookie);
360 } else {
361 atomic_sub(scale, &lat_info->scale_cookie);
362 }
363 }
364}
365
366/*
367 * Change the queue depth of the iolatency_grp. We add 1/16th of the
368 * queue depth at a time so we don't get wild swings and hopefully dial in to
369 * fairer distribution of the overall queue depth. We halve the queue depth
370 * at a time so we can scale down queue depth quickly from default unlimited
371 * to target.
372 */
373static void scale_change(struct iolatency_grp *iolat, bool up)
374{
375 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
376 unsigned long scale = scale_amount(qd, up);
377 unsigned long old = iolat->max_depth;
378
379 if (old > qd)
380 old = qd;
381
382 if (up) {
383 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
384 return;
385
386 if (old < qd) {
387 old += scale;
388 old = min(old, qd);
389 iolat->max_depth = old;
390 wake_up_all(&iolat->rq_wait.wait);
391 }
392 } else {
393 old >>= 1;
394 iolat->max_depth = max(old, 1UL);
395 }
396}
397
398/* Check our parent and see if the scale cookie has changed. */
399static void check_scale_change(struct iolatency_grp *iolat)
400{
401 struct iolatency_grp *parent;
402 struct child_latency_info *lat_info;
403 unsigned int cur_cookie;
404 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
405 u64 scale_lat;
406 int direction = 0;
407
408 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
409 if (!parent)
410 return;
411
412 lat_info = &parent->child_lat;
413 cur_cookie = atomic_read(&lat_info->scale_cookie);
414 scale_lat = READ_ONCE(lat_info->scale_lat);
415
416 if (cur_cookie < our_cookie)
417 direction = -1;
418 else if (cur_cookie > our_cookie)
419 direction = 1;
420 else
421 return;
422
423 if (!atomic_try_cmpxchg(&iolat->scale_cookie, &our_cookie, cur_cookie)) {
424 /* Somebody beat us to the punch, just bail. */
425 return;
426 }
427
428 if (direction < 0 && iolat->min_lat_nsec) {
429 u64 samples_thresh;
430
431 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
432 return;
433
434 /*
435 * Sometimes high priority groups are their own worst enemy, so
436 * instead of taking it out on some poor other group that did 5%
437 * or less of the IO's for the last summation just skip this
438 * scale down event.
439 */
440 samples_thresh = lat_info->nr_samples * 5;
441 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
442 if (iolat->nr_samples <= samples_thresh)
443 return;
444 }
445
446 /* We're as low as we can go. */
447 if (iolat->max_depth == 1 && direction < 0) {
448 blkcg_use_delay(lat_to_blkg(iolat));
449 return;
450 }
451
452 /* We're back to the default cookie, unthrottle all the things. */
453 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
454 blkcg_clear_delay(lat_to_blkg(iolat));
455 iolat->max_depth = UINT_MAX;
456 wake_up_all(&iolat->rq_wait.wait);
457 return;
458 }
459
460 scale_change(iolat, direction > 0);
461}
462
463static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
464{
465 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
466 struct blkcg_gq *blkg = bio->bi_blkg;
467 bool issue_as_root = bio_issue_as_root_blkg(bio);
468
469 if (!blkiolat->enabled)
470 return;
471
472 while (blkg && blkg->parent) {
473 struct iolatency_grp *iolat = blkg_to_lat(blkg);
474 if (!iolat) {
475 blkg = blkg->parent;
476 continue;
477 }
478
479 check_scale_change(iolat);
480 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
481 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
482 blkg = blkg->parent;
483 }
484 if (!timer_pending(&blkiolat->timer))
485 mod_timer(&blkiolat->timer, jiffies + HZ);
486}
487
488static void iolatency_record_time(struct iolatency_grp *iolat,
489 struct bio_issue *issue, u64 now,
490 bool issue_as_root)
491{
492 u64 start = bio_issue_time(issue);
493 u64 req_time;
494
495 /*
496 * Have to do this so we are truncated to the correct time that our
497 * issue is truncated to.
498 */
499 now = __bio_issue_time(now);
500
501 if (now <= start)
502 return;
503
504 req_time = now - start;
505
506 /*
507 * We don't want to count issue_as_root bio's in the cgroups latency
508 * statistics as it could skew the numbers downwards.
509 */
510 if (unlikely(issue_as_root && iolat->max_depth != UINT_MAX)) {
511 u64 sub = iolat->min_lat_nsec;
512 if (req_time < sub)
513 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
514 return;
515 }
516
517 latency_stat_record_time(iolat, req_time);
518}
519
520#define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
521#define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
522
523static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
524{
525 struct blkcg_gq *blkg = lat_to_blkg(iolat);
526 struct iolatency_grp *parent;
527 struct child_latency_info *lat_info;
528 struct latency_stat stat;
529 unsigned long flags;
530 int cpu;
531
532 latency_stat_init(iolat, &stat);
533 preempt_disable();
534 for_each_online_cpu(cpu) {
535 struct latency_stat *s;
536 s = per_cpu_ptr(iolat->stats, cpu);
537 latency_stat_sum(iolat, &stat, s);
538 latency_stat_init(iolat, s);
539 }
540 preempt_enable();
541
542 parent = blkg_to_lat(blkg->parent);
543 if (!parent)
544 return;
545
546 lat_info = &parent->child_lat;
547
548 iolat_update_total_lat_avg(iolat, &stat);
549
550 /* Everything is ok and we don't need to adjust the scale. */
551 if (latency_sum_ok(iolat, &stat) &&
552 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
553 return;
554
555 /* Somebody beat us to the punch, just bail. */
556 spin_lock_irqsave(&lat_info->lock, flags);
557
558 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
559 lat_info->nr_samples -= iolat->nr_samples;
560 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
561 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
562
563 if ((lat_info->last_scale_event >= now ||
564 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
565 goto out;
566
567 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
568 latency_sum_ok(iolat, &stat)) {
569 if (latency_stat_samples(iolat, &iolat->cur_stat) <
570 BLKIOLATENCY_MIN_GOOD_SAMPLES)
571 goto out;
572 if (lat_info->scale_grp == iolat) {
573 lat_info->last_scale_event = now;
574 scale_cookie_change(iolat->blkiolat, lat_info, true);
575 }
576 } else if (lat_info->scale_lat == 0 ||
577 lat_info->scale_lat >= iolat->min_lat_nsec) {
578 lat_info->last_scale_event = now;
579 if (!lat_info->scale_grp ||
580 lat_info->scale_lat > iolat->min_lat_nsec) {
581 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
582 lat_info->scale_grp = iolat;
583 }
584 scale_cookie_change(iolat->blkiolat, lat_info, false);
585 }
586 latency_stat_init(iolat, &iolat->cur_stat);
587out:
588 spin_unlock_irqrestore(&lat_info->lock, flags);
589}
590
591static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
592{
593 struct blkcg_gq *blkg;
594 struct rq_wait *rqw;
595 struct iolatency_grp *iolat;
596 u64 window_start;
597 u64 now;
598 bool issue_as_root = bio_issue_as_root_blkg(bio);
599 int inflight = 0;
600
601 blkg = bio->bi_blkg;
602 if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
603 return;
604
605 iolat = blkg_to_lat(bio->bi_blkg);
606 if (!iolat)
607 return;
608
609 if (!iolat->blkiolat->enabled)
610 return;
611
612 now = ktime_to_ns(ktime_get());
613 while (blkg && blkg->parent) {
614 iolat = blkg_to_lat(blkg);
615 if (!iolat) {
616 blkg = blkg->parent;
617 continue;
618 }
619 rqw = &iolat->rq_wait;
620
621 inflight = atomic_dec_return(&rqw->inflight);
622 WARN_ON_ONCE(inflight < 0);
623 /*
624 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
625 * submitted, so do not account for it.
626 */
627 if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
628 iolatency_record_time(iolat, &bio->bi_issue, now,
629 issue_as_root);
630 window_start = atomic64_read(&iolat->window_start);
631 if (now > window_start &&
632 (now - window_start) >= iolat->cur_win_nsec) {
633 if (atomic64_try_cmpxchg(&iolat->window_start,
634 &window_start, now))
635 iolatency_check_latencies(iolat, now);
636 }
637 }
638 wake_up(&rqw->wait);
639 blkg = blkg->parent;
640 }
641}
642
643static void blkcg_iolatency_exit(struct rq_qos *rqos)
644{
645 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
646
647 timer_shutdown_sync(&blkiolat->timer);
648 flush_work(&blkiolat->enable_work);
649 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
650 kfree(blkiolat);
651}
652
653static struct rq_qos_ops blkcg_iolatency_ops = {
654 .throttle = blkcg_iolatency_throttle,
655 .done_bio = blkcg_iolatency_done_bio,
656 .exit = blkcg_iolatency_exit,
657};
658
659static void blkiolatency_timer_fn(struct timer_list *t)
660{
661 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
662 struct blkcg_gq *blkg;
663 struct cgroup_subsys_state *pos_css;
664 u64 now = ktime_to_ns(ktime_get());
665
666 rcu_read_lock();
667 blkg_for_each_descendant_pre(blkg, pos_css,
668 blkiolat->rqos.q->root_blkg) {
669 struct iolatency_grp *iolat;
670 struct child_latency_info *lat_info;
671 unsigned long flags;
672 u64 cookie;
673
674 /*
675 * We could be exiting, don't access the pd unless we have a
676 * ref on the blkg.
677 */
678 if (!blkg_tryget(blkg))
679 continue;
680
681 iolat = blkg_to_lat(blkg);
682 if (!iolat)
683 goto next;
684
685 lat_info = &iolat->child_lat;
686 cookie = atomic_read(&lat_info->scale_cookie);
687
688 if (cookie >= DEFAULT_SCALE_COOKIE)
689 goto next;
690
691 spin_lock_irqsave(&lat_info->lock, flags);
692 if (lat_info->last_scale_event >= now)
693 goto next_lock;
694
695 /*
696 * We scaled down but don't have a scale_grp, scale up and carry
697 * on.
698 */
699 if (lat_info->scale_grp == NULL) {
700 scale_cookie_change(iolat->blkiolat, lat_info, true);
701 goto next_lock;
702 }
703
704 /*
705 * It's been 5 seconds since our last scale event, clear the
706 * scale grp in case the group that needed the scale down isn't
707 * doing any IO currently.
708 */
709 if (now - lat_info->last_scale_event >=
710 ((u64)NSEC_PER_SEC * 5))
711 lat_info->scale_grp = NULL;
712next_lock:
713 spin_unlock_irqrestore(&lat_info->lock, flags);
714next:
715 blkg_put(blkg);
716 }
717 rcu_read_unlock();
718}
719
720/**
721 * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
722 * @work: enable_work of the blk_iolatency of interest
723 *
724 * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
725 * is relatively expensive as it involves walking up the hierarchy twice for
726 * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
727 * want to disable the in-flight tracking.
728 *
729 * We have to make sure that the counting is balanced - we don't want to leak
730 * the in-flight counts by disabling accounting in the completion path while IOs
731 * are in flight. This is achieved by ensuring that no IO is in flight by
732 * freezing the queue while flipping ->enabled. As this requires a sleepable
733 * context, ->enabled flipping is punted to this work function.
734 */
735static void blkiolatency_enable_work_fn(struct work_struct *work)
736{
737 struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
738 enable_work);
739 bool enabled;
740
741 /*
742 * There can only be one instance of this function running for @blkiolat
743 * and it's guaranteed to be executed at least once after the latest
744 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
745 * sufficient.
746 *
747 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
748 * in blkcg_iolatency_exit().
749 */
750 enabled = atomic_read(&blkiolat->enable_cnt);
751 if (enabled != blkiolat->enabled) {
752 blk_mq_freeze_queue(blkiolat->rqos.q);
753 blkiolat->enabled = enabled;
754 blk_mq_unfreeze_queue(blkiolat->rqos.q);
755 }
756}
757
758int blk_iolatency_init(struct gendisk *disk)
759{
760 struct request_queue *q = disk->queue;
761 struct blk_iolatency *blkiolat;
762 struct rq_qos *rqos;
763 int ret;
764
765 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
766 if (!blkiolat)
767 return -ENOMEM;
768
769 rqos = &blkiolat->rqos;
770 rqos->id = RQ_QOS_LATENCY;
771 rqos->ops = &blkcg_iolatency_ops;
772 rqos->q = q;
773
774 ret = rq_qos_add(q, rqos);
775 if (ret)
776 goto err_free;
777 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
778 if (ret)
779 goto err_qos_del;
780
781 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
782 INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
783
784 return 0;
785
786err_qos_del:
787 rq_qos_del(q, rqos);
788err_free:
789 kfree(blkiolat);
790 return ret;
791}
792
793static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
794{
795 struct iolatency_grp *iolat = blkg_to_lat(blkg);
796 struct blk_iolatency *blkiolat = iolat->blkiolat;
797 u64 oldval = iolat->min_lat_nsec;
798
799 iolat->min_lat_nsec = val;
800 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
801 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
802 BLKIOLATENCY_MAX_WIN_SIZE);
803
804 if (!oldval && val) {
805 if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
806 schedule_work(&blkiolat->enable_work);
807 }
808 if (oldval && !val) {
809 blkcg_clear_delay(blkg);
810 if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
811 schedule_work(&blkiolat->enable_work);
812 }
813}
814
815static void iolatency_clear_scaling(struct blkcg_gq *blkg)
816{
817 if (blkg->parent) {
818 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
819 struct child_latency_info *lat_info;
820 if (!iolat)
821 return;
822
823 lat_info = &iolat->child_lat;
824 spin_lock(&lat_info->lock);
825 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
826 lat_info->last_scale_event = 0;
827 lat_info->scale_grp = NULL;
828 lat_info->scale_lat = 0;
829 spin_unlock(&lat_info->lock);
830 }
831}
832
833static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
834 size_t nbytes, loff_t off)
835{
836 struct blkcg *blkcg = css_to_blkcg(of_css(of));
837 struct blkcg_gq *blkg;
838 struct blkg_conf_ctx ctx;
839 struct iolatency_grp *iolat;
840 char *p, *tok;
841 u64 lat_val = 0;
842 u64 oldval;
843 int ret;
844
845 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
846 if (ret)
847 return ret;
848
849 iolat = blkg_to_lat(ctx.blkg);
850 p = ctx.body;
851
852 ret = -EINVAL;
853 while ((tok = strsep(&p, " "))) {
854 char key[16];
855 char val[21]; /* 18446744073709551616 */
856
857 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
858 goto out;
859
860 if (!strcmp(key, "target")) {
861 u64 v;
862
863 if (!strcmp(val, "max"))
864 lat_val = 0;
865 else if (sscanf(val, "%llu", &v) == 1)
866 lat_val = v * NSEC_PER_USEC;
867 else
868 goto out;
869 } else {
870 goto out;
871 }
872 }
873
874 /* Walk up the tree to see if our new val is lower than it should be. */
875 blkg = ctx.blkg;
876 oldval = iolat->min_lat_nsec;
877
878 iolatency_set_min_lat_nsec(blkg, lat_val);
879 if (oldval != iolat->min_lat_nsec)
880 iolatency_clear_scaling(blkg);
881 ret = 0;
882out:
883 blkg_conf_finish(&ctx);
884 return ret ?: nbytes;
885}
886
887static u64 iolatency_prfill_limit(struct seq_file *sf,
888 struct blkg_policy_data *pd, int off)
889{
890 struct iolatency_grp *iolat = pd_to_lat(pd);
891 const char *dname = blkg_dev_name(pd->blkg);
892
893 if (!dname || !iolat->min_lat_nsec)
894 return 0;
895 seq_printf(sf, "%s target=%llu\n",
896 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
897 return 0;
898}
899
900static int iolatency_print_limit(struct seq_file *sf, void *v)
901{
902 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
903 iolatency_prfill_limit,
904 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
905 return 0;
906}
907
908static void iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
909{
910 struct latency_stat stat;
911 int cpu;
912
913 latency_stat_init(iolat, &stat);
914 preempt_disable();
915 for_each_online_cpu(cpu) {
916 struct latency_stat *s;
917 s = per_cpu_ptr(iolat->stats, cpu);
918 latency_stat_sum(iolat, &stat, s);
919 }
920 preempt_enable();
921
922 if (iolat->max_depth == UINT_MAX)
923 seq_printf(s, " missed=%llu total=%llu depth=max",
924 (unsigned long long)stat.ps.missed,
925 (unsigned long long)stat.ps.total);
926 else
927 seq_printf(s, " missed=%llu total=%llu depth=%u",
928 (unsigned long long)stat.ps.missed,
929 (unsigned long long)stat.ps.total,
930 iolat->max_depth);
931}
932
933static void iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
934{
935 struct iolatency_grp *iolat = pd_to_lat(pd);
936 unsigned long long avg_lat;
937 unsigned long long cur_win;
938
939 if (!blkcg_debug_stats)
940 return;
941
942 if (iolat->ssd)
943 return iolatency_ssd_stat(iolat, s);
944
945 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
946 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
947 if (iolat->max_depth == UINT_MAX)
948 seq_printf(s, " depth=max avg_lat=%llu win=%llu",
949 avg_lat, cur_win);
950 else
951 seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
952 iolat->max_depth, avg_lat, cur_win);
953}
954
955static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
956 struct request_queue *q,
957 struct blkcg *blkcg)
958{
959 struct iolatency_grp *iolat;
960
961 iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
962 if (!iolat)
963 return NULL;
964 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
965 __alignof__(struct latency_stat), gfp);
966 if (!iolat->stats) {
967 kfree(iolat);
968 return NULL;
969 }
970 return &iolat->pd;
971}
972
973static void iolatency_pd_init(struct blkg_policy_data *pd)
974{
975 struct iolatency_grp *iolat = pd_to_lat(pd);
976 struct blkcg_gq *blkg = lat_to_blkg(iolat);
977 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
978 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
979 u64 now = ktime_to_ns(ktime_get());
980 int cpu;
981
982 if (blk_queue_nonrot(blkg->q))
983 iolat->ssd = true;
984 else
985 iolat->ssd = false;
986
987 for_each_possible_cpu(cpu) {
988 struct latency_stat *stat;
989 stat = per_cpu_ptr(iolat->stats, cpu);
990 latency_stat_init(iolat, stat);
991 }
992
993 latency_stat_init(iolat, &iolat->cur_stat);
994 rq_wait_init(&iolat->rq_wait);
995 spin_lock_init(&iolat->child_lat.lock);
996 iolat->max_depth = UINT_MAX;
997 iolat->blkiolat = blkiolat;
998 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
999 atomic64_set(&iolat->window_start, now);
1000
1001 /*
1002 * We init things in list order, so the pd for the parent may not be
1003 * init'ed yet for whatever reason.
1004 */
1005 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1006 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1007 atomic_set(&iolat->scale_cookie,
1008 atomic_read(&parent->child_lat.scale_cookie));
1009 } else {
1010 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1011 }
1012
1013 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1014}
1015
1016static void iolatency_pd_offline(struct blkg_policy_data *pd)
1017{
1018 struct iolatency_grp *iolat = pd_to_lat(pd);
1019 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1020
1021 iolatency_set_min_lat_nsec(blkg, 0);
1022 iolatency_clear_scaling(blkg);
1023}
1024
1025static void iolatency_pd_free(struct blkg_policy_data *pd)
1026{
1027 struct iolatency_grp *iolat = pd_to_lat(pd);
1028 free_percpu(iolat->stats);
1029 kfree(iolat);
1030}
1031
1032static struct cftype iolatency_files[] = {
1033 {
1034 .name = "latency",
1035 .flags = CFTYPE_NOT_ON_ROOT,
1036 .seq_show = iolatency_print_limit,
1037 .write = iolatency_set_limit,
1038 },
1039 {}
1040};
1041
1042static struct blkcg_policy blkcg_policy_iolatency = {
1043 .dfl_cftypes = iolatency_files,
1044 .pd_alloc_fn = iolatency_pd_alloc,
1045 .pd_init_fn = iolatency_pd_init,
1046 .pd_offline_fn = iolatency_pd_offline,
1047 .pd_free_fn = iolatency_pd_free,
1048 .pd_stat_fn = iolatency_pd_stat,
1049};
1050
1051static int __init iolatency_init(void)
1052{
1053 return blkcg_policy_register(&blkcg_policy_iolatency);
1054}
1055
1056static void __exit iolatency_exit(void)
1057{
1058 blkcg_policy_unregister(&blkcg_policy_iolatency);
1059}
1060
1061module_init(iolatency_init);
1062module_exit(iolatency_exit);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Block rq-qos base io controller
4 *
5 * This works similar to wbt with a few exceptions
6 *
7 * - It's bio based, so the latency covers the whole block layer in addition to
8 * the actual io.
9 * - We will throttle all IO that comes in here if we need to.
10 * - We use the mean latency over the 100ms window. This is because writes can
11 * be particularly fast, which could give us a false sense of the impact of
12 * other workloads on our protected workload.
13 * - By default there's no throttling, we set the queue_depth to UINT_MAX so
14 * that we can have as many outstanding bio's as we're allowed to. Only at
15 * throttle time do we pay attention to the actual queue depth.
16 *
17 * The hierarchy works like the cpu controller does, we track the latency at
18 * every configured node, and each configured node has it's own independent
19 * queue depth. This means that we only care about our latency targets at the
20 * peer level. Some group at the bottom of the hierarchy isn't going to affect
21 * a group at the end of some other path if we're only configred at leaf level.
22 *
23 * Consider the following
24 *
25 * root blkg
26 * / \
27 * fast (target=5ms) slow (target=10ms)
28 * / \ / \
29 * a b normal(15ms) unloved
30 *
31 * "a" and "b" have no target, but their combined io under "fast" cannot exceed
32 * an average latency of 5ms. If it does then we will throttle the "slow"
33 * group. In the case of "normal", if it exceeds its 15ms target, we will
34 * throttle "unloved", but nobody else.
35 *
36 * In this example "fast", "slow", and "normal" will be the only groups actually
37 * accounting their io latencies. We have to walk up the heirarchy to the root
38 * on every submit and complete so we can do the appropriate stat recording and
39 * adjust the queue depth of ourselves if needed.
40 *
41 * There are 2 ways we throttle IO.
42 *
43 * 1) Queue depth throttling. As we throttle down we will adjust the maximum
44 * number of IO's we're allowed to have in flight. This starts at (u64)-1 down
45 * to 1. If the group is only ever submitting IO for itself then this is the
46 * only way we throttle.
47 *
48 * 2) Induced delay throttling. This is for the case that a group is generating
49 * IO that has to be issued by the root cg to avoid priority inversion. So think
50 * REQ_META or REQ_SWAP. If we are already at qd == 1 and we're getting a lot
51 * of work done for us on behalf of the root cg and are being asked to scale
52 * down more then we induce a latency at userspace return. We accumulate the
53 * total amount of time we need to be punished by doing
54 *
55 * total_time += min_lat_nsec - actual_io_completion
56 *
57 * and then at throttle time will do
58 *
59 * throttle_time = min(total_time, NSEC_PER_SEC)
60 *
61 * This induced delay will throttle back the activity that is generating the
62 * root cg issued io's, wethere that's some metadata intensive operation or the
63 * group is using so much memory that it is pushing us into swap.
64 *
65 * Copyright (C) 2018 Josef Bacik
66 */
67#include <linux/kernel.h>
68#include <linux/blk_types.h>
69#include <linux/backing-dev.h>
70#include <linux/module.h>
71#include <linux/timer.h>
72#include <linux/memcontrol.h>
73#include <linux/sched/loadavg.h>
74#include <linux/sched/signal.h>
75#include <trace/events/block.h>
76#include <linux/blk-mq.h>
77#include "blk-rq-qos.h"
78#include "blk-stat.h"
79#include "blk.h"
80
81#define DEFAULT_SCALE_COOKIE 1000000U
82
83static struct blkcg_policy blkcg_policy_iolatency;
84struct iolatency_grp;
85
86struct blk_iolatency {
87 struct rq_qos rqos;
88 struct timer_list timer;
89 atomic_t enabled;
90};
91
92static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
93{
94 return container_of(rqos, struct blk_iolatency, rqos);
95}
96
97static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
98{
99 return atomic_read(&blkiolat->enabled) > 0;
100}
101
102struct child_latency_info {
103 spinlock_t lock;
104
105 /* Last time we adjusted the scale of everybody. */
106 u64 last_scale_event;
107
108 /* The latency that we missed. */
109 u64 scale_lat;
110
111 /* Total io's from all of our children for the last summation. */
112 u64 nr_samples;
113
114 /* The guy who actually changed the latency numbers. */
115 struct iolatency_grp *scale_grp;
116
117 /* Cookie to tell if we need to scale up or down. */
118 atomic_t scale_cookie;
119};
120
121struct percentile_stats {
122 u64 total;
123 u64 missed;
124};
125
126struct latency_stat {
127 union {
128 struct percentile_stats ps;
129 struct blk_rq_stat rqs;
130 };
131};
132
133struct iolatency_grp {
134 struct blkg_policy_data pd;
135 struct latency_stat __percpu *stats;
136 struct latency_stat cur_stat;
137 struct blk_iolatency *blkiolat;
138 struct rq_depth rq_depth;
139 struct rq_wait rq_wait;
140 atomic64_t window_start;
141 atomic_t scale_cookie;
142 u64 min_lat_nsec;
143 u64 cur_win_nsec;
144
145 /* total running average of our io latency. */
146 u64 lat_avg;
147
148 /* Our current number of IO's for the last summation. */
149 u64 nr_samples;
150
151 bool ssd;
152 struct child_latency_info child_lat;
153};
154
155#define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
156#define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
157/*
158 * These are the constants used to fake the fixed-point moving average
159 * calculation just like load average. The call to calc_load() folds
160 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
161 * window size is bucketed to try to approximately calculate average
162 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
163 * elapse immediately. Note, windows only elapse with IO activity. Idle
164 * periods extend the most recent window.
165 */
166#define BLKIOLATENCY_NR_EXP_FACTORS 5
167#define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
168 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
169static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
170 2045, // exp(1/600) - 600 samples
171 2039, // exp(1/240) - 240 samples
172 2031, // exp(1/120) - 120 samples
173 2023, // exp(1/80) - 80 samples
174 2014, // exp(1/60) - 60 samples
175};
176
177static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
178{
179 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
180}
181
182static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
183{
184 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
185}
186
187static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
188{
189 return pd_to_blkg(&iolat->pd);
190}
191
192static inline void latency_stat_init(struct iolatency_grp *iolat,
193 struct latency_stat *stat)
194{
195 if (iolat->ssd) {
196 stat->ps.total = 0;
197 stat->ps.missed = 0;
198 } else
199 blk_rq_stat_init(&stat->rqs);
200}
201
202static inline void latency_stat_sum(struct iolatency_grp *iolat,
203 struct latency_stat *sum,
204 struct latency_stat *stat)
205{
206 if (iolat->ssd) {
207 sum->ps.total += stat->ps.total;
208 sum->ps.missed += stat->ps.missed;
209 } else
210 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
211}
212
213static inline void latency_stat_record_time(struct iolatency_grp *iolat,
214 u64 req_time)
215{
216 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
217 if (iolat->ssd) {
218 if (req_time >= iolat->min_lat_nsec)
219 stat->ps.missed++;
220 stat->ps.total++;
221 } else
222 blk_rq_stat_add(&stat->rqs, req_time);
223 put_cpu_ptr(stat);
224}
225
226static inline bool latency_sum_ok(struct iolatency_grp *iolat,
227 struct latency_stat *stat)
228{
229 if (iolat->ssd) {
230 u64 thresh = div64_u64(stat->ps.total, 10);
231 thresh = max(thresh, 1ULL);
232 return stat->ps.missed < thresh;
233 }
234 return stat->rqs.mean <= iolat->min_lat_nsec;
235}
236
237static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
238 struct latency_stat *stat)
239{
240 if (iolat->ssd)
241 return stat->ps.total;
242 return stat->rqs.nr_samples;
243}
244
245static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
246 struct latency_stat *stat)
247{
248 int exp_idx;
249
250 if (iolat->ssd)
251 return;
252
253 /*
254 * calc_load() takes in a number stored in fixed point representation.
255 * Because we are using this for IO time in ns, the values stored
256 * are significantly larger than the FIXED_1 denominator (2048).
257 * Therefore, rounding errors in the calculation are negligible and
258 * can be ignored.
259 */
260 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
261 div64_u64(iolat->cur_win_nsec,
262 BLKIOLATENCY_EXP_BUCKET_SIZE));
263 iolat->lat_avg = calc_load(iolat->lat_avg,
264 iolatency_exp_factors[exp_idx],
265 stat->rqs.mean);
266}
267
268static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
269{
270 atomic_dec(&rqw->inflight);
271 wake_up(&rqw->wait);
272}
273
274static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
275{
276 struct iolatency_grp *iolat = private_data;
277 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
278}
279
280static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
281 struct iolatency_grp *iolat,
282 bool issue_as_root,
283 bool use_memdelay)
284{
285 struct rq_wait *rqw = &iolat->rq_wait;
286 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
287
288 if (use_delay)
289 blkcg_schedule_throttle(rqos->q, use_memdelay);
290
291 /*
292 * To avoid priority inversions we want to just take a slot if we are
293 * issuing as root. If we're being killed off there's no point in
294 * delaying things, we may have been killed by OOM so throttling may
295 * make recovery take even longer, so just let the IO's through so the
296 * task can go away.
297 */
298 if (issue_as_root || fatal_signal_pending(current)) {
299 atomic_inc(&rqw->inflight);
300 return;
301 }
302
303 rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
304}
305
306#define SCALE_DOWN_FACTOR 2
307#define SCALE_UP_FACTOR 4
308
309static inline unsigned long scale_amount(unsigned long qd, bool up)
310{
311 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
312}
313
314/*
315 * We scale the qd down faster than we scale up, so we need to use this helper
316 * to adjust the scale_cookie accordingly so we don't prematurely get
317 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
318 *
319 * Each group has their own local copy of the last scale cookie they saw, so if
320 * the global scale cookie goes up or down they know which way they need to go
321 * based on their last knowledge of it.
322 */
323static void scale_cookie_change(struct blk_iolatency *blkiolat,
324 struct child_latency_info *lat_info,
325 bool up)
326{
327 unsigned long qd = blkiolat->rqos.q->nr_requests;
328 unsigned long scale = scale_amount(qd, up);
329 unsigned long old = atomic_read(&lat_info->scale_cookie);
330 unsigned long max_scale = qd << 1;
331 unsigned long diff = 0;
332
333 if (old < DEFAULT_SCALE_COOKIE)
334 diff = DEFAULT_SCALE_COOKIE - old;
335
336 if (up) {
337 if (scale + old > DEFAULT_SCALE_COOKIE)
338 atomic_set(&lat_info->scale_cookie,
339 DEFAULT_SCALE_COOKIE);
340 else if (diff > qd)
341 atomic_inc(&lat_info->scale_cookie);
342 else
343 atomic_add(scale, &lat_info->scale_cookie);
344 } else {
345 /*
346 * We don't want to dig a hole so deep that it takes us hours to
347 * dig out of it. Just enough that we don't throttle/unthrottle
348 * with jagged workloads but can still unthrottle once pressure
349 * has sufficiently dissipated.
350 */
351 if (diff > qd) {
352 if (diff < max_scale)
353 atomic_dec(&lat_info->scale_cookie);
354 } else {
355 atomic_sub(scale, &lat_info->scale_cookie);
356 }
357 }
358}
359
360/*
361 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
362 * queue depth at a time so we don't get wild swings and hopefully dial in to
363 * fairer distribution of the overall queue depth.
364 */
365static void scale_change(struct iolatency_grp *iolat, bool up)
366{
367 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
368 unsigned long scale = scale_amount(qd, up);
369 unsigned long old = iolat->rq_depth.max_depth;
370
371 if (old > qd)
372 old = qd;
373
374 if (up) {
375 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
376 return;
377
378 if (old < qd) {
379 old += scale;
380 old = min(old, qd);
381 iolat->rq_depth.max_depth = old;
382 wake_up_all(&iolat->rq_wait.wait);
383 }
384 } else {
385 old >>= 1;
386 iolat->rq_depth.max_depth = max(old, 1UL);
387 }
388}
389
390/* Check our parent and see if the scale cookie has changed. */
391static void check_scale_change(struct iolatency_grp *iolat)
392{
393 struct iolatency_grp *parent;
394 struct child_latency_info *lat_info;
395 unsigned int cur_cookie;
396 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
397 u64 scale_lat;
398 unsigned int old;
399 int direction = 0;
400
401 if (lat_to_blkg(iolat)->parent == NULL)
402 return;
403
404 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
405 if (!parent)
406 return;
407
408 lat_info = &parent->child_lat;
409 cur_cookie = atomic_read(&lat_info->scale_cookie);
410 scale_lat = READ_ONCE(lat_info->scale_lat);
411
412 if (cur_cookie < our_cookie)
413 direction = -1;
414 else if (cur_cookie > our_cookie)
415 direction = 1;
416 else
417 return;
418
419 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
420
421 /* Somebody beat us to the punch, just bail. */
422 if (old != our_cookie)
423 return;
424
425 if (direction < 0 && iolat->min_lat_nsec) {
426 u64 samples_thresh;
427
428 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
429 return;
430
431 /*
432 * Sometimes high priority groups are their own worst enemy, so
433 * instead of taking it out on some poor other group that did 5%
434 * or less of the IO's for the last summation just skip this
435 * scale down event.
436 */
437 samples_thresh = lat_info->nr_samples * 5;
438 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
439 if (iolat->nr_samples <= samples_thresh)
440 return;
441 }
442
443 /* We're as low as we can go. */
444 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
445 blkcg_use_delay(lat_to_blkg(iolat));
446 return;
447 }
448
449 /* We're back to the default cookie, unthrottle all the things. */
450 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
451 blkcg_clear_delay(lat_to_blkg(iolat));
452 iolat->rq_depth.max_depth = UINT_MAX;
453 wake_up_all(&iolat->rq_wait.wait);
454 return;
455 }
456
457 scale_change(iolat, direction > 0);
458}
459
460static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
461{
462 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
463 struct blkcg_gq *blkg = bio->bi_blkg;
464 bool issue_as_root = bio_issue_as_root_blkg(bio);
465
466 if (!blk_iolatency_enabled(blkiolat))
467 return;
468
469 while (blkg && blkg->parent) {
470 struct iolatency_grp *iolat = blkg_to_lat(blkg);
471 if (!iolat) {
472 blkg = blkg->parent;
473 continue;
474 }
475
476 check_scale_change(iolat);
477 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
478 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
479 blkg = blkg->parent;
480 }
481 if (!timer_pending(&blkiolat->timer))
482 mod_timer(&blkiolat->timer, jiffies + HZ);
483}
484
485static void iolatency_record_time(struct iolatency_grp *iolat,
486 struct bio_issue *issue, u64 now,
487 bool issue_as_root)
488{
489 u64 start = bio_issue_time(issue);
490 u64 req_time;
491
492 /*
493 * Have to do this so we are truncated to the correct time that our
494 * issue is truncated to.
495 */
496 now = __bio_issue_time(now);
497
498 if (now <= start)
499 return;
500
501 req_time = now - start;
502
503 /*
504 * We don't want to count issue_as_root bio's in the cgroups latency
505 * statistics as it could skew the numbers downwards.
506 */
507 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
508 u64 sub = iolat->min_lat_nsec;
509 if (req_time < sub)
510 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
511 return;
512 }
513
514 latency_stat_record_time(iolat, req_time);
515}
516
517#define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
518#define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
519
520static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
521{
522 struct blkcg_gq *blkg = lat_to_blkg(iolat);
523 struct iolatency_grp *parent;
524 struct child_latency_info *lat_info;
525 struct latency_stat stat;
526 unsigned long flags;
527 int cpu;
528
529 latency_stat_init(iolat, &stat);
530 preempt_disable();
531 for_each_online_cpu(cpu) {
532 struct latency_stat *s;
533 s = per_cpu_ptr(iolat->stats, cpu);
534 latency_stat_sum(iolat, &stat, s);
535 latency_stat_init(iolat, s);
536 }
537 preempt_enable();
538
539 parent = blkg_to_lat(blkg->parent);
540 if (!parent)
541 return;
542
543 lat_info = &parent->child_lat;
544
545 iolat_update_total_lat_avg(iolat, &stat);
546
547 /* Everything is ok and we don't need to adjust the scale. */
548 if (latency_sum_ok(iolat, &stat) &&
549 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
550 return;
551
552 /* Somebody beat us to the punch, just bail. */
553 spin_lock_irqsave(&lat_info->lock, flags);
554
555 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
556 lat_info->nr_samples -= iolat->nr_samples;
557 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
558 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
559
560 if ((lat_info->last_scale_event >= now ||
561 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
562 goto out;
563
564 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
565 latency_sum_ok(iolat, &stat)) {
566 if (latency_stat_samples(iolat, &iolat->cur_stat) <
567 BLKIOLATENCY_MIN_GOOD_SAMPLES)
568 goto out;
569 if (lat_info->scale_grp == iolat) {
570 lat_info->last_scale_event = now;
571 scale_cookie_change(iolat->blkiolat, lat_info, true);
572 }
573 } else if (lat_info->scale_lat == 0 ||
574 lat_info->scale_lat >= iolat->min_lat_nsec) {
575 lat_info->last_scale_event = now;
576 if (!lat_info->scale_grp ||
577 lat_info->scale_lat > iolat->min_lat_nsec) {
578 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
579 lat_info->scale_grp = iolat;
580 }
581 scale_cookie_change(iolat->blkiolat, lat_info, false);
582 }
583 latency_stat_init(iolat, &iolat->cur_stat);
584out:
585 spin_unlock_irqrestore(&lat_info->lock, flags);
586}
587
588static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
589{
590 struct blkcg_gq *blkg;
591 struct rq_wait *rqw;
592 struct iolatency_grp *iolat;
593 u64 window_start;
594 u64 now;
595 bool issue_as_root = bio_issue_as_root_blkg(bio);
596 bool enabled = false;
597 int inflight = 0;
598
599 blkg = bio->bi_blkg;
600 if (!blkg || !bio_flagged(bio, BIO_TRACKED))
601 return;
602
603 iolat = blkg_to_lat(bio->bi_blkg);
604 if (!iolat)
605 return;
606
607 enabled = blk_iolatency_enabled(iolat->blkiolat);
608 if (!enabled)
609 return;
610
611 now = ktime_to_ns(ktime_get());
612 while (blkg && blkg->parent) {
613 iolat = blkg_to_lat(blkg);
614 if (!iolat) {
615 blkg = blkg->parent;
616 continue;
617 }
618 rqw = &iolat->rq_wait;
619
620 inflight = atomic_dec_return(&rqw->inflight);
621 WARN_ON_ONCE(inflight < 0);
622 /*
623 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
624 * submitted, so do not account for it.
625 */
626 if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
627 iolatency_record_time(iolat, &bio->bi_issue, now,
628 issue_as_root);
629 window_start = atomic64_read(&iolat->window_start);
630 if (now > window_start &&
631 (now - window_start) >= iolat->cur_win_nsec) {
632 if (atomic64_cmpxchg(&iolat->window_start,
633 window_start, now) == window_start)
634 iolatency_check_latencies(iolat, now);
635 }
636 }
637 wake_up(&rqw->wait);
638 blkg = blkg->parent;
639 }
640}
641
642static void blkcg_iolatency_exit(struct rq_qos *rqos)
643{
644 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
645
646 del_timer_sync(&blkiolat->timer);
647 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
648 kfree(blkiolat);
649}
650
651static struct rq_qos_ops blkcg_iolatency_ops = {
652 .throttle = blkcg_iolatency_throttle,
653 .done_bio = blkcg_iolatency_done_bio,
654 .exit = blkcg_iolatency_exit,
655};
656
657static void blkiolatency_timer_fn(struct timer_list *t)
658{
659 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
660 struct blkcg_gq *blkg;
661 struct cgroup_subsys_state *pos_css;
662 u64 now = ktime_to_ns(ktime_get());
663
664 rcu_read_lock();
665 blkg_for_each_descendant_pre(blkg, pos_css,
666 blkiolat->rqos.q->root_blkg) {
667 struct iolatency_grp *iolat;
668 struct child_latency_info *lat_info;
669 unsigned long flags;
670 u64 cookie;
671
672 /*
673 * We could be exiting, don't access the pd unless we have a
674 * ref on the blkg.
675 */
676 if (!blkg_tryget(blkg))
677 continue;
678
679 iolat = blkg_to_lat(blkg);
680 if (!iolat)
681 goto next;
682
683 lat_info = &iolat->child_lat;
684 cookie = atomic_read(&lat_info->scale_cookie);
685
686 if (cookie >= DEFAULT_SCALE_COOKIE)
687 goto next;
688
689 spin_lock_irqsave(&lat_info->lock, flags);
690 if (lat_info->last_scale_event >= now)
691 goto next_lock;
692
693 /*
694 * We scaled down but don't have a scale_grp, scale up and carry
695 * on.
696 */
697 if (lat_info->scale_grp == NULL) {
698 scale_cookie_change(iolat->blkiolat, lat_info, true);
699 goto next_lock;
700 }
701
702 /*
703 * It's been 5 seconds since our last scale event, clear the
704 * scale grp in case the group that needed the scale down isn't
705 * doing any IO currently.
706 */
707 if (now - lat_info->last_scale_event >=
708 ((u64)NSEC_PER_SEC * 5))
709 lat_info->scale_grp = NULL;
710next_lock:
711 spin_unlock_irqrestore(&lat_info->lock, flags);
712next:
713 blkg_put(blkg);
714 }
715 rcu_read_unlock();
716}
717
718int blk_iolatency_init(struct request_queue *q)
719{
720 struct blk_iolatency *blkiolat;
721 struct rq_qos *rqos;
722 int ret;
723
724 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
725 if (!blkiolat)
726 return -ENOMEM;
727
728 rqos = &blkiolat->rqos;
729 rqos->id = RQ_QOS_LATENCY;
730 rqos->ops = &blkcg_iolatency_ops;
731 rqos->q = q;
732
733 rq_qos_add(q, rqos);
734
735 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
736 if (ret) {
737 rq_qos_del(q, rqos);
738 kfree(blkiolat);
739 return ret;
740 }
741
742 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
743
744 return 0;
745}
746
747/*
748 * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
749 * return 0.
750 */
751static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
752{
753 struct iolatency_grp *iolat = blkg_to_lat(blkg);
754 u64 oldval = iolat->min_lat_nsec;
755
756 iolat->min_lat_nsec = val;
757 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
758 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
759 BLKIOLATENCY_MAX_WIN_SIZE);
760
761 if (!oldval && val)
762 return 1;
763 if (oldval && !val) {
764 blkcg_clear_delay(blkg);
765 return -1;
766 }
767 return 0;
768}
769
770static void iolatency_clear_scaling(struct blkcg_gq *blkg)
771{
772 if (blkg->parent) {
773 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
774 struct child_latency_info *lat_info;
775 if (!iolat)
776 return;
777
778 lat_info = &iolat->child_lat;
779 spin_lock(&lat_info->lock);
780 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
781 lat_info->last_scale_event = 0;
782 lat_info->scale_grp = NULL;
783 lat_info->scale_lat = 0;
784 spin_unlock(&lat_info->lock);
785 }
786}
787
788static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
789 size_t nbytes, loff_t off)
790{
791 struct blkcg *blkcg = css_to_blkcg(of_css(of));
792 struct blkcg_gq *blkg;
793 struct blkg_conf_ctx ctx;
794 struct iolatency_grp *iolat;
795 char *p, *tok;
796 u64 lat_val = 0;
797 u64 oldval;
798 int ret;
799 int enable = 0;
800
801 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
802 if (ret)
803 return ret;
804
805 iolat = blkg_to_lat(ctx.blkg);
806 p = ctx.body;
807
808 ret = -EINVAL;
809 while ((tok = strsep(&p, " "))) {
810 char key[16];
811 char val[21]; /* 18446744073709551616 */
812
813 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
814 goto out;
815
816 if (!strcmp(key, "target")) {
817 u64 v;
818
819 if (!strcmp(val, "max"))
820 lat_val = 0;
821 else if (sscanf(val, "%llu", &v) == 1)
822 lat_val = v * NSEC_PER_USEC;
823 else
824 goto out;
825 } else {
826 goto out;
827 }
828 }
829
830 /* Walk up the tree to see if our new val is lower than it should be. */
831 blkg = ctx.blkg;
832 oldval = iolat->min_lat_nsec;
833
834 enable = iolatency_set_min_lat_nsec(blkg, lat_val);
835 if (enable) {
836 WARN_ON_ONCE(!blk_get_queue(blkg->q));
837 blkg_get(blkg);
838 }
839
840 if (oldval != iolat->min_lat_nsec) {
841 iolatency_clear_scaling(blkg);
842 }
843
844 ret = 0;
845out:
846 blkg_conf_finish(&ctx);
847 if (ret == 0 && enable) {
848 struct iolatency_grp *tmp = blkg_to_lat(blkg);
849 struct blk_iolatency *blkiolat = tmp->blkiolat;
850
851 blk_mq_freeze_queue(blkg->q);
852
853 if (enable == 1)
854 atomic_inc(&blkiolat->enabled);
855 else if (enable == -1)
856 atomic_dec(&blkiolat->enabled);
857 else
858 WARN_ON_ONCE(1);
859
860 blk_mq_unfreeze_queue(blkg->q);
861
862 blkg_put(blkg);
863 blk_put_queue(blkg->q);
864 }
865 return ret ?: nbytes;
866}
867
868static u64 iolatency_prfill_limit(struct seq_file *sf,
869 struct blkg_policy_data *pd, int off)
870{
871 struct iolatency_grp *iolat = pd_to_lat(pd);
872 const char *dname = blkg_dev_name(pd->blkg);
873
874 if (!dname || !iolat->min_lat_nsec)
875 return 0;
876 seq_printf(sf, "%s target=%llu\n",
877 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
878 return 0;
879}
880
881static int iolatency_print_limit(struct seq_file *sf, void *v)
882{
883 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
884 iolatency_prfill_limit,
885 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
886 return 0;
887}
888
889static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
890 size_t size)
891{
892 struct latency_stat stat;
893 int cpu;
894
895 latency_stat_init(iolat, &stat);
896 preempt_disable();
897 for_each_online_cpu(cpu) {
898 struct latency_stat *s;
899 s = per_cpu_ptr(iolat->stats, cpu);
900 latency_stat_sum(iolat, &stat, s);
901 }
902 preempt_enable();
903
904 if (iolat->rq_depth.max_depth == UINT_MAX)
905 return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
906 (unsigned long long)stat.ps.missed,
907 (unsigned long long)stat.ps.total);
908 return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
909 (unsigned long long)stat.ps.missed,
910 (unsigned long long)stat.ps.total,
911 iolat->rq_depth.max_depth);
912}
913
914static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
915 size_t size)
916{
917 struct iolatency_grp *iolat = pd_to_lat(pd);
918 unsigned long long avg_lat;
919 unsigned long long cur_win;
920
921 if (!blkcg_debug_stats)
922 return 0;
923
924 if (iolat->ssd)
925 return iolatency_ssd_stat(iolat, buf, size);
926
927 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
928 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
929 if (iolat->rq_depth.max_depth == UINT_MAX)
930 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
931 avg_lat, cur_win);
932
933 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
934 iolat->rq_depth.max_depth, avg_lat, cur_win);
935}
936
937
938static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
939 struct request_queue *q,
940 struct blkcg *blkcg)
941{
942 struct iolatency_grp *iolat;
943
944 iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
945 if (!iolat)
946 return NULL;
947 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
948 __alignof__(struct latency_stat), gfp);
949 if (!iolat->stats) {
950 kfree(iolat);
951 return NULL;
952 }
953 return &iolat->pd;
954}
955
956static void iolatency_pd_init(struct blkg_policy_data *pd)
957{
958 struct iolatency_grp *iolat = pd_to_lat(pd);
959 struct blkcg_gq *blkg = lat_to_blkg(iolat);
960 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
961 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
962 u64 now = ktime_to_ns(ktime_get());
963 int cpu;
964
965 if (blk_queue_nonrot(blkg->q))
966 iolat->ssd = true;
967 else
968 iolat->ssd = false;
969
970 for_each_possible_cpu(cpu) {
971 struct latency_stat *stat;
972 stat = per_cpu_ptr(iolat->stats, cpu);
973 latency_stat_init(iolat, stat);
974 }
975
976 latency_stat_init(iolat, &iolat->cur_stat);
977 rq_wait_init(&iolat->rq_wait);
978 spin_lock_init(&iolat->child_lat.lock);
979 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
980 iolat->rq_depth.max_depth = UINT_MAX;
981 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
982 iolat->blkiolat = blkiolat;
983 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
984 atomic64_set(&iolat->window_start, now);
985
986 /*
987 * We init things in list order, so the pd for the parent may not be
988 * init'ed yet for whatever reason.
989 */
990 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
991 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
992 atomic_set(&iolat->scale_cookie,
993 atomic_read(&parent->child_lat.scale_cookie));
994 } else {
995 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
996 }
997
998 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
999}
1000
1001static void iolatency_pd_offline(struct blkg_policy_data *pd)
1002{
1003 struct iolatency_grp *iolat = pd_to_lat(pd);
1004 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1005 struct blk_iolatency *blkiolat = iolat->blkiolat;
1006 int ret;
1007
1008 ret = iolatency_set_min_lat_nsec(blkg, 0);
1009 if (ret == 1)
1010 atomic_inc(&blkiolat->enabled);
1011 if (ret == -1)
1012 atomic_dec(&blkiolat->enabled);
1013 iolatency_clear_scaling(blkg);
1014}
1015
1016static void iolatency_pd_free(struct blkg_policy_data *pd)
1017{
1018 struct iolatency_grp *iolat = pd_to_lat(pd);
1019 free_percpu(iolat->stats);
1020 kfree(iolat);
1021}
1022
1023static struct cftype iolatency_files[] = {
1024 {
1025 .name = "latency",
1026 .flags = CFTYPE_NOT_ON_ROOT,
1027 .seq_show = iolatency_print_limit,
1028 .write = iolatency_set_limit,
1029 },
1030 {}
1031};
1032
1033static struct blkcg_policy blkcg_policy_iolatency = {
1034 .dfl_cftypes = iolatency_files,
1035 .pd_alloc_fn = iolatency_pd_alloc,
1036 .pd_init_fn = iolatency_pd_init,
1037 .pd_offline_fn = iolatency_pd_offline,
1038 .pd_free_fn = iolatency_pd_free,
1039 .pd_stat_fn = iolatency_pd_stat,
1040};
1041
1042static int __init iolatency_init(void)
1043{
1044 return blkcg_policy_register(&blkcg_policy_iolatency);
1045}
1046
1047static void __exit iolatency_exit(void)
1048{
1049 return blkcg_policy_unregister(&blkcg_policy_iolatency);
1050}
1051
1052module_init(iolatency_init);
1053module_exit(iolatency_exit);