1 // SPDX-License-Identifier: GPL-2.0
3 * Block rq-qos base io controller
5 * This works similar to wbt with a few exceptions
7 * - It's bio based, so the latency covers the whole block layer in addition to
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.
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.
23 * Consider the following
27 * fast (target=5ms) slow (target=10ms)
29 * a b normal(15ms) unloved
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.
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.
41 * There are 2 ways we throttle IO.
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.
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
55 * total_time += min_lat_nsec - actual_io_completion
57 * and then at throttle time will do
59 * throttle_time = min(total_time, NSEC_PER_SEC)
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.
65 * Copyright (C) 2018 Josef Bacik
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"
81 #define DEFAULT_SCALE_COOKIE 1000000U
83 static struct blkcg_policy blkcg_policy_iolatency;
86 struct blk_iolatency {
88 struct timer_list timer;
91 * ->enabled is the master enable switch gating the throttling logic and
92 * inflight tracking. The number of cgroups which have iolat enabled is
93 * tracked in ->enable_cnt, and ->enable is flipped on/off accordingly
94 * from ->enable_work with the request_queue frozen. For details, See
95 * blkiolatency_enable_work_fn().
99 struct work_struct enable_work;
102 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
104 return container_of(rqos, struct blk_iolatency, rqos);
107 struct child_latency_info {
110 /* Last time we adjusted the scale of everybody. */
111 u64 last_scale_event;
113 /* The latency that we missed. */
116 /* Total io's from all of our children for the last summation. */
119 /* The guy who actually changed the latency numbers. */
120 struct iolatency_grp *scale_grp;
122 /* Cookie to tell if we need to scale up or down. */
123 atomic_t scale_cookie;
126 struct percentile_stats {
131 struct latency_stat {
133 struct percentile_stats ps;
134 struct blk_rq_stat rqs;
138 struct iolatency_grp {
139 struct blkg_policy_data pd;
140 struct latency_stat __percpu *stats;
141 struct latency_stat cur_stat;
142 struct blk_iolatency *blkiolat;
143 struct rq_depth rq_depth;
144 struct rq_wait rq_wait;
145 atomic64_t window_start;
146 atomic_t scale_cookie;
150 /* total running average of our io latency. */
153 /* Our current number of IO's for the last summation. */
157 struct child_latency_info child_lat;
160 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
161 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
163 * These are the constants used to fake the fixed-point moving average
164 * calculation just like load average. The call to calc_load() folds
165 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
166 * window size is bucketed to try to approximately calculate average
167 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
168 * elapse immediately. Note, windows only elapse with IO activity. Idle
169 * periods extend the most recent window.
171 #define BLKIOLATENCY_NR_EXP_FACTORS 5
172 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
173 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
174 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
175 2045, // exp(1/600) - 600 samples
176 2039, // exp(1/240) - 240 samples
177 2031, // exp(1/120) - 120 samples
178 2023, // exp(1/80) - 80 samples
179 2014, // exp(1/60) - 60 samples
182 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
184 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
187 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
189 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
192 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
194 return pd_to_blkg(&iolat->pd);
197 static inline void latency_stat_init(struct iolatency_grp *iolat,
198 struct latency_stat *stat)
204 blk_rq_stat_init(&stat->rqs);
207 static inline void latency_stat_sum(struct iolatency_grp *iolat,
208 struct latency_stat *sum,
209 struct latency_stat *stat)
212 sum->ps.total += stat->ps.total;
213 sum->ps.missed += stat->ps.missed;
215 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
218 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
221 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
223 if (req_time >= iolat->min_lat_nsec)
227 blk_rq_stat_add(&stat->rqs, req_time);
231 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
232 struct latency_stat *stat)
235 u64 thresh = div64_u64(stat->ps.total, 10);
236 thresh = max(thresh, 1ULL);
237 return stat->ps.missed < thresh;
239 return stat->rqs.mean <= iolat->min_lat_nsec;
242 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
243 struct latency_stat *stat)
246 return stat->ps.total;
247 return stat->rqs.nr_samples;
250 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
251 struct latency_stat *stat)
259 * calc_load() takes in a number stored in fixed point representation.
260 * Because we are using this for IO time in ns, the values stored
261 * are significantly larger than the FIXED_1 denominator (2048).
262 * Therefore, rounding errors in the calculation are negligible and
265 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
266 div64_u64(iolat->cur_win_nsec,
267 BLKIOLATENCY_EXP_BUCKET_SIZE));
268 iolat->lat_avg = calc_load(iolat->lat_avg,
269 iolatency_exp_factors[exp_idx],
273 static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
275 atomic_dec(&rqw->inflight);
279 static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
281 struct iolatency_grp *iolat = private_data;
282 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
285 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
286 struct iolatency_grp *iolat,
290 struct rq_wait *rqw = &iolat->rq_wait;
291 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
294 blkcg_schedule_throttle(rqos->q, use_memdelay);
297 * To avoid priority inversions we want to just take a slot if we are
298 * issuing as root. If we're being killed off there's no point in
299 * delaying things, we may have been killed by OOM so throttling may
300 * make recovery take even longer, so just let the IO's through so the
303 if (issue_as_root || fatal_signal_pending(current)) {
304 atomic_inc(&rqw->inflight);
308 rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
311 #define SCALE_DOWN_FACTOR 2
312 #define SCALE_UP_FACTOR 4
314 static inline unsigned long scale_amount(unsigned long qd, bool up)
316 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
320 * We scale the qd down faster than we scale up, so we need to use this helper
321 * to adjust the scale_cookie accordingly so we don't prematurely get
322 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
324 * Each group has their own local copy of the last scale cookie they saw, so if
325 * the global scale cookie goes up or down they know which way they need to go
326 * based on their last knowledge of it.
328 static void scale_cookie_change(struct blk_iolatency *blkiolat,
329 struct child_latency_info *lat_info,
332 unsigned long qd = blkiolat->rqos.q->nr_requests;
333 unsigned long scale = scale_amount(qd, up);
334 unsigned long old = atomic_read(&lat_info->scale_cookie);
335 unsigned long max_scale = qd << 1;
336 unsigned long diff = 0;
338 if (old < DEFAULT_SCALE_COOKIE)
339 diff = DEFAULT_SCALE_COOKIE - old;
342 if (scale + old > DEFAULT_SCALE_COOKIE)
343 atomic_set(&lat_info->scale_cookie,
344 DEFAULT_SCALE_COOKIE);
346 atomic_inc(&lat_info->scale_cookie);
348 atomic_add(scale, &lat_info->scale_cookie);
351 * We don't want to dig a hole so deep that it takes us hours to
352 * dig out of it. Just enough that we don't throttle/unthrottle
353 * with jagged workloads but can still unthrottle once pressure
354 * has sufficiently dissipated.
357 if (diff < max_scale)
358 atomic_dec(&lat_info->scale_cookie);
360 atomic_sub(scale, &lat_info->scale_cookie);
366 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
367 * queue depth at a time so we don't get wild swings and hopefully dial in to
368 * fairer distribution of the overall queue depth.
370 static void scale_change(struct iolatency_grp *iolat, bool up)
372 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
373 unsigned long scale = scale_amount(qd, up);
374 unsigned long old = iolat->rq_depth.max_depth;
380 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
386 iolat->rq_depth.max_depth = old;
387 wake_up_all(&iolat->rq_wait.wait);
391 iolat->rq_depth.max_depth = max(old, 1UL);
395 /* Check our parent and see if the scale cookie has changed. */
396 static void check_scale_change(struct iolatency_grp *iolat)
398 struct iolatency_grp *parent;
399 struct child_latency_info *lat_info;
400 unsigned int cur_cookie;
401 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
406 if (lat_to_blkg(iolat)->parent == NULL)
409 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
413 lat_info = &parent->child_lat;
414 cur_cookie = atomic_read(&lat_info->scale_cookie);
415 scale_lat = READ_ONCE(lat_info->scale_lat);
417 if (cur_cookie < our_cookie)
419 else if (cur_cookie > our_cookie)
424 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
426 /* Somebody beat us to the punch, just bail. */
427 if (old != our_cookie)
430 if (direction < 0 && iolat->min_lat_nsec) {
433 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
437 * Sometimes high priority groups are their own worst enemy, so
438 * instead of taking it out on some poor other group that did 5%
439 * or less of the IO's for the last summation just skip this
442 samples_thresh = lat_info->nr_samples * 5;
443 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
444 if (iolat->nr_samples <= samples_thresh)
448 /* We're as low as we can go. */
449 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
450 blkcg_use_delay(lat_to_blkg(iolat));
454 /* We're back to the default cookie, unthrottle all the things. */
455 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
456 blkcg_clear_delay(lat_to_blkg(iolat));
457 iolat->rq_depth.max_depth = UINT_MAX;
458 wake_up_all(&iolat->rq_wait.wait);
462 scale_change(iolat, direction > 0);
465 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
467 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
468 struct blkcg_gq *blkg = bio->bi_blkg;
469 bool issue_as_root = bio_issue_as_root_blkg(bio);
471 if (!blkiolat->enabled)
474 while (blkg && blkg->parent) {
475 struct iolatency_grp *iolat = blkg_to_lat(blkg);
481 check_scale_change(iolat);
482 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
483 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
486 if (!timer_pending(&blkiolat->timer))
487 mod_timer(&blkiolat->timer, jiffies + HZ);
490 static void iolatency_record_time(struct iolatency_grp *iolat,
491 struct bio_issue *issue, u64 now,
494 u64 start = bio_issue_time(issue);
498 * Have to do this so we are truncated to the correct time that our
499 * issue is truncated to.
501 now = __bio_issue_time(now);
506 req_time = now - start;
509 * We don't want to count issue_as_root bio's in the cgroups latency
510 * statistics as it could skew the numbers downwards.
512 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
513 u64 sub = iolat->min_lat_nsec;
515 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
519 latency_stat_record_time(iolat, req_time);
522 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
523 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
525 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
527 struct blkcg_gq *blkg = lat_to_blkg(iolat);
528 struct iolatency_grp *parent;
529 struct child_latency_info *lat_info;
530 struct latency_stat stat;
534 latency_stat_init(iolat, &stat);
536 for_each_online_cpu(cpu) {
537 struct latency_stat *s;
538 s = per_cpu_ptr(iolat->stats, cpu);
539 latency_stat_sum(iolat, &stat, s);
540 latency_stat_init(iolat, s);
544 parent = blkg_to_lat(blkg->parent);
548 lat_info = &parent->child_lat;
550 iolat_update_total_lat_avg(iolat, &stat);
552 /* Everything is ok and we don't need to adjust the scale. */
553 if (latency_sum_ok(iolat, &stat) &&
554 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
557 /* Somebody beat us to the punch, just bail. */
558 spin_lock_irqsave(&lat_info->lock, flags);
560 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
561 lat_info->nr_samples -= iolat->nr_samples;
562 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
563 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
565 if ((lat_info->last_scale_event >= now ||
566 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
569 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
570 latency_sum_ok(iolat, &stat)) {
571 if (latency_stat_samples(iolat, &iolat->cur_stat) <
572 BLKIOLATENCY_MIN_GOOD_SAMPLES)
574 if (lat_info->scale_grp == iolat) {
575 lat_info->last_scale_event = now;
576 scale_cookie_change(iolat->blkiolat, lat_info, true);
578 } else if (lat_info->scale_lat == 0 ||
579 lat_info->scale_lat >= iolat->min_lat_nsec) {
580 lat_info->last_scale_event = now;
581 if (!lat_info->scale_grp ||
582 lat_info->scale_lat > iolat->min_lat_nsec) {
583 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
584 lat_info->scale_grp = iolat;
586 scale_cookie_change(iolat->blkiolat, lat_info, false);
588 latency_stat_init(iolat, &iolat->cur_stat);
590 spin_unlock_irqrestore(&lat_info->lock, flags);
593 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
595 struct blkcg_gq *blkg;
597 struct iolatency_grp *iolat;
600 bool issue_as_root = bio_issue_as_root_blkg(bio);
604 if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
607 iolat = blkg_to_lat(bio->bi_blkg);
611 if (!iolat->blkiolat->enabled)
614 now = ktime_to_ns(ktime_get());
615 while (blkg && blkg->parent) {
616 iolat = blkg_to_lat(blkg);
621 rqw = &iolat->rq_wait;
623 inflight = atomic_dec_return(&rqw->inflight);
624 WARN_ON_ONCE(inflight < 0);
626 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
627 * submitted, so do not account for it.
629 if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
630 iolatency_record_time(iolat, &bio->bi_issue, now,
632 window_start = atomic64_read(&iolat->window_start);
633 if (now > window_start &&
634 (now - window_start) >= iolat->cur_win_nsec) {
635 if (atomic64_cmpxchg(&iolat->window_start,
636 window_start, now) == window_start)
637 iolatency_check_latencies(iolat, now);
645 static void blkcg_iolatency_exit(struct rq_qos *rqos)
647 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
649 del_timer_sync(&blkiolat->timer);
650 flush_work(&blkiolat->enable_work);
651 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
655 static struct rq_qos_ops blkcg_iolatency_ops = {
656 .throttle = blkcg_iolatency_throttle,
657 .done_bio = blkcg_iolatency_done_bio,
658 .exit = blkcg_iolatency_exit,
661 static void blkiolatency_timer_fn(struct timer_list *t)
663 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
664 struct blkcg_gq *blkg;
665 struct cgroup_subsys_state *pos_css;
666 u64 now = ktime_to_ns(ktime_get());
669 blkg_for_each_descendant_pre(blkg, pos_css,
670 blkiolat->rqos.q->root_blkg) {
671 struct iolatency_grp *iolat;
672 struct child_latency_info *lat_info;
677 * We could be exiting, don't access the pd unless we have a
680 if (!blkg_tryget(blkg))
683 iolat = blkg_to_lat(blkg);
687 lat_info = &iolat->child_lat;
688 cookie = atomic_read(&lat_info->scale_cookie);
690 if (cookie >= DEFAULT_SCALE_COOKIE)
693 spin_lock_irqsave(&lat_info->lock, flags);
694 if (lat_info->last_scale_event >= now)
698 * We scaled down but don't have a scale_grp, scale up and carry
701 if (lat_info->scale_grp == NULL) {
702 scale_cookie_change(iolat->blkiolat, lat_info, true);
707 * It's been 5 seconds since our last scale event, clear the
708 * scale grp in case the group that needed the scale down isn't
709 * doing any IO currently.
711 if (now - lat_info->last_scale_event >=
712 ((u64)NSEC_PER_SEC * 5))
713 lat_info->scale_grp = NULL;
715 spin_unlock_irqrestore(&lat_info->lock, flags);
723 * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
724 * @work: enable_work of the blk_iolatency of interest
726 * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
727 * is relatively expensive as it involves walking up the hierarchy twice for
728 * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
729 * want to disable the in-flight tracking.
731 * We have to make sure that the counting is balanced - we don't want to leak
732 * the in-flight counts by disabling accounting in the completion path while IOs
733 * are in flight. This is achieved by ensuring that no IO is in flight by
734 * freezing the queue while flipping ->enabled. As this requires a sleepable
735 * context, ->enabled flipping is punted to this work function.
737 static void blkiolatency_enable_work_fn(struct work_struct *work)
739 struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
744 * There can only be one instance of this function running for @blkiolat
745 * and it's guaranteed to be executed at least once after the latest
746 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
749 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
750 * in blkcg_iolatency_exit().
752 enabled = atomic_read(&blkiolat->enable_cnt);
753 if (enabled != blkiolat->enabled) {
754 blk_mq_freeze_queue(blkiolat->rqos.q);
755 blkiolat->enabled = enabled;
756 blk_mq_unfreeze_queue(blkiolat->rqos.q);
760 int blk_iolatency_init(struct request_queue *q)
762 struct blk_iolatency *blkiolat;
766 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
770 rqos = &blkiolat->rqos;
771 rqos->id = RQ_QOS_LATENCY;
772 rqos->ops = &blkcg_iolatency_ops;
775 ret = rq_qos_add(q, rqos);
778 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
782 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
783 INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
794 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
796 struct iolatency_grp *iolat = blkg_to_lat(blkg);
797 struct blk_iolatency *blkiolat = iolat->blkiolat;
798 u64 oldval = iolat->min_lat_nsec;
800 iolat->min_lat_nsec = val;
801 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
802 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
803 BLKIOLATENCY_MAX_WIN_SIZE);
805 if (!oldval && val) {
806 if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
807 schedule_work(&blkiolat->enable_work);
809 if (oldval && !val) {
810 blkcg_clear_delay(blkg);
811 if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
812 schedule_work(&blkiolat->enable_work);
816 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
819 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
820 struct child_latency_info *lat_info;
824 lat_info = &iolat->child_lat;
825 spin_lock(&lat_info->lock);
826 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
827 lat_info->last_scale_event = 0;
828 lat_info->scale_grp = NULL;
829 lat_info->scale_lat = 0;
830 spin_unlock(&lat_info->lock);
834 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
835 size_t nbytes, loff_t off)
837 struct blkcg *blkcg = css_to_blkcg(of_css(of));
838 struct blkcg_gq *blkg;
839 struct blkg_conf_ctx ctx;
840 struct iolatency_grp *iolat;
846 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
850 iolat = blkg_to_lat(ctx.blkg);
854 while ((tok = strsep(&p, " "))) {
856 char val[21]; /* 18446744073709551616 */
858 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
861 if (!strcmp(key, "target")) {
864 if (!strcmp(val, "max"))
866 else if (sscanf(val, "%llu", &v) == 1)
867 lat_val = v * NSEC_PER_USEC;
875 /* Walk up the tree to see if our new val is lower than it should be. */
877 oldval = iolat->min_lat_nsec;
879 iolatency_set_min_lat_nsec(blkg, lat_val);
880 if (oldval != iolat->min_lat_nsec)
881 iolatency_clear_scaling(blkg);
884 blkg_conf_finish(&ctx);
885 return ret ?: nbytes;
888 static u64 iolatency_prfill_limit(struct seq_file *sf,
889 struct blkg_policy_data *pd, int off)
891 struct iolatency_grp *iolat = pd_to_lat(pd);
892 const char *dname = blkg_dev_name(pd->blkg);
894 if (!dname || !iolat->min_lat_nsec)
896 seq_printf(sf, "%s target=%llu\n",
897 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
901 static int iolatency_print_limit(struct seq_file *sf, void *v)
903 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
904 iolatency_prfill_limit,
905 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
909 static bool iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
911 struct latency_stat stat;
914 latency_stat_init(iolat, &stat);
916 for_each_online_cpu(cpu) {
917 struct latency_stat *s;
918 s = per_cpu_ptr(iolat->stats, cpu);
919 latency_stat_sum(iolat, &stat, s);
923 if (iolat->rq_depth.max_depth == UINT_MAX)
924 seq_printf(s, " missed=%llu total=%llu depth=max",
925 (unsigned long long)stat.ps.missed,
926 (unsigned long long)stat.ps.total);
928 seq_printf(s, " missed=%llu total=%llu depth=%u",
929 (unsigned long long)stat.ps.missed,
930 (unsigned long long)stat.ps.total,
931 iolat->rq_depth.max_depth);
935 static bool iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
937 struct iolatency_grp *iolat = pd_to_lat(pd);
938 unsigned long long avg_lat;
939 unsigned long long cur_win;
941 if (!blkcg_debug_stats)
945 return iolatency_ssd_stat(iolat, s);
947 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
948 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
949 if (iolat->rq_depth.max_depth == UINT_MAX)
950 seq_printf(s, " depth=max avg_lat=%llu win=%llu",
953 seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
954 iolat->rq_depth.max_depth, avg_lat, cur_win);
958 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
959 struct request_queue *q,
962 struct iolatency_grp *iolat;
964 iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
967 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
968 __alignof__(struct latency_stat), gfp);
976 static void iolatency_pd_init(struct blkg_policy_data *pd)
978 struct iolatency_grp *iolat = pd_to_lat(pd);
979 struct blkcg_gq *blkg = lat_to_blkg(iolat);
980 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
981 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
982 u64 now = ktime_to_ns(ktime_get());
985 if (blk_queue_nonrot(blkg->q))
990 for_each_possible_cpu(cpu) {
991 struct latency_stat *stat;
992 stat = per_cpu_ptr(iolat->stats, cpu);
993 latency_stat_init(iolat, stat);
996 latency_stat_init(iolat, &iolat->cur_stat);
997 rq_wait_init(&iolat->rq_wait);
998 spin_lock_init(&iolat->child_lat.lock);
999 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
1000 iolat->rq_depth.max_depth = UINT_MAX;
1001 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
1002 iolat->blkiolat = blkiolat;
1003 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
1004 atomic64_set(&iolat->window_start, now);
1007 * We init things in list order, so the pd for the parent may not be
1008 * init'ed yet for whatever reason.
1010 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1011 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1012 atomic_set(&iolat->scale_cookie,
1013 atomic_read(&parent->child_lat.scale_cookie));
1015 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1018 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1021 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1023 struct iolatency_grp *iolat = pd_to_lat(pd);
1024 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1026 iolatency_set_min_lat_nsec(blkg, 0);
1027 iolatency_clear_scaling(blkg);
1030 static void iolatency_pd_free(struct blkg_policy_data *pd)
1032 struct iolatency_grp *iolat = pd_to_lat(pd);
1033 free_percpu(iolat->stats);
1037 static struct cftype iolatency_files[] = {
1040 .flags = CFTYPE_NOT_ON_ROOT,
1041 .seq_show = iolatency_print_limit,
1042 .write = iolatency_set_limit,
1047 static struct blkcg_policy blkcg_policy_iolatency = {
1048 .dfl_cftypes = iolatency_files,
1049 .pd_alloc_fn = iolatency_pd_alloc,
1050 .pd_init_fn = iolatency_pd_init,
1051 .pd_offline_fn = iolatency_pd_offline,
1052 .pd_free_fn = iolatency_pd_free,
1053 .pd_stat_fn = iolatency_pd_stat,
1056 static int __init iolatency_init(void)
1058 return blkcg_policy_register(&blkcg_policy_iolatency);
1061 static void __exit iolatency_exit(void)
1063 blkcg_policy_unregister(&blkcg_policy_iolatency);
1066 module_init(iolatency_init);
1067 module_exit(iolatency_exit);