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;
92 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
94 return container_of(rqos, struct blk_iolatency, rqos);
97 static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
99 return atomic_read(&blkiolat->enabled) > 0;
102 struct child_latency_info {
105 /* Last time we adjusted the scale of everybody. */
106 u64 last_scale_event;
108 /* The latency that we missed. */
111 /* Total io's from all of our children for the last summation. */
114 /* The guy who actually changed the latency numbers. */
115 struct iolatency_grp *scale_grp;
117 /* Cookie to tell if we need to scale up or down. */
118 atomic_t scale_cookie;
121 struct percentile_stats {
126 struct latency_stat {
128 struct percentile_stats ps;
129 struct blk_rq_stat rqs;
133 struct 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;
145 /* total running average of our io latency. */
148 /* Our current number of IO's for the last summation. */
152 struct child_latency_info child_lat;
155 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
156 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
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.
166 #define BLKIOLATENCY_NR_EXP_FACTORS 5
167 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
168 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
169 static 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
177 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
179 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
182 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
184 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
187 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
189 return pd_to_blkg(&iolat->pd);
192 static inline void latency_stat_init(struct iolatency_grp *iolat,
193 struct latency_stat *stat)
199 blk_rq_stat_init(&stat->rqs);
202 static inline void latency_stat_sum(struct iolatency_grp *iolat,
203 struct latency_stat *sum,
204 struct latency_stat *stat)
207 sum->ps.total += stat->ps.total;
208 sum->ps.missed += stat->ps.missed;
210 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
213 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
216 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
218 if (req_time >= iolat->min_lat_nsec)
222 blk_rq_stat_add(&stat->rqs, req_time);
226 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
227 struct latency_stat *stat)
230 u64 thresh = div64_u64(stat->ps.total, 10);
231 thresh = max(thresh, 1ULL);
232 return stat->ps.missed < thresh;
234 return stat->rqs.mean <= iolat->min_lat_nsec;
237 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
238 struct latency_stat *stat)
241 return stat->ps.total;
242 return stat->rqs.nr_samples;
245 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
246 struct latency_stat *stat)
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
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],
268 static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
270 atomic_dec(&rqw->inflight);
274 static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
276 struct iolatency_grp *iolat = private_data;
277 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
280 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
281 struct iolatency_grp *iolat,
285 struct rq_wait *rqw = &iolat->rq_wait;
286 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
289 blkcg_schedule_throttle(rqos->q, use_memdelay);
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
298 if (issue_as_root || fatal_signal_pending(current)) {
299 atomic_inc(&rqw->inflight);
303 rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
306 #define SCALE_DOWN_FACTOR 2
307 #define SCALE_UP_FACTOR 4
309 static inline unsigned long scale_amount(unsigned long qd, bool up)
311 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
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.
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.
323 static void scale_cookie_change(struct blk_iolatency *blkiolat,
324 struct child_latency_info *lat_info,
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;
333 if (old < DEFAULT_SCALE_COOKIE)
334 diff = DEFAULT_SCALE_COOKIE - old;
337 if (scale + old > DEFAULT_SCALE_COOKIE)
338 atomic_set(&lat_info->scale_cookie,
339 DEFAULT_SCALE_COOKIE);
341 atomic_inc(&lat_info->scale_cookie);
343 atomic_add(scale, &lat_info->scale_cookie);
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.
352 if (diff < max_scale)
353 atomic_dec(&lat_info->scale_cookie);
355 atomic_sub(scale, &lat_info->scale_cookie);
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.
365 static void scale_change(struct iolatency_grp *iolat, bool up)
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;
375 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
381 iolat->rq_depth.max_depth = old;
382 wake_up_all(&iolat->rq_wait.wait);
386 iolat->rq_depth.max_depth = max(old, 1UL);
390 /* Check our parent and see if the scale cookie has changed. */
391 static void check_scale_change(struct iolatency_grp *iolat)
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);
401 if (lat_to_blkg(iolat)->parent == NULL)
404 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
408 lat_info = &parent->child_lat;
409 cur_cookie = atomic_read(&lat_info->scale_cookie);
410 scale_lat = READ_ONCE(lat_info->scale_lat);
412 if (cur_cookie < our_cookie)
414 else if (cur_cookie > our_cookie)
419 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
421 /* Somebody beat us to the punch, just bail. */
422 if (old != our_cookie)
425 if (direction < 0 && iolat->min_lat_nsec) {
428 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
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
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)
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));
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);
457 scale_change(iolat, direction > 0);
460 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
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);
466 if (!blk_iolatency_enabled(blkiolat))
469 while (blkg && blkg->parent) {
470 struct iolatency_grp *iolat = blkg_to_lat(blkg);
476 check_scale_change(iolat);
477 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
478 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
481 if (!timer_pending(&blkiolat->timer))
482 mod_timer(&blkiolat->timer, jiffies + HZ);
485 static void iolatency_record_time(struct iolatency_grp *iolat,
486 struct bio_issue *issue, u64 now,
489 u64 start = bio_issue_time(issue);
493 * Have to do this so we are truncated to the correct time that our
494 * issue is truncated to.
496 now = __bio_issue_time(now);
501 req_time = now - start;
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.
507 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
508 u64 sub = iolat->min_lat_nsec;
510 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
514 latency_stat_record_time(iolat, req_time);
517 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
518 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
520 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
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;
529 latency_stat_init(iolat, &stat);
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);
539 parent = blkg_to_lat(blkg->parent);
543 lat_info = &parent->child_lat;
545 iolat_update_total_lat_avg(iolat, &stat);
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)
552 /* Somebody beat us to the punch, just bail. */
553 spin_lock_irqsave(&lat_info->lock, flags);
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);
560 if ((lat_info->last_scale_event >= now ||
561 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
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)
569 if (lat_info->scale_grp == iolat) {
570 lat_info->last_scale_event = now;
571 scale_cookie_change(iolat->blkiolat, lat_info, true);
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;
581 scale_cookie_change(iolat->blkiolat, lat_info, false);
583 latency_stat_init(iolat, &iolat->cur_stat);
585 spin_unlock_irqrestore(&lat_info->lock, flags);
588 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
590 struct blkcg_gq *blkg;
592 struct iolatency_grp *iolat;
594 u64 now = ktime_to_ns(ktime_get());
595 bool issue_as_root = bio_issue_as_root_blkg(bio);
596 bool enabled = false;
600 if (!blkg || !bio_flagged(bio, BIO_TRACKED))
603 iolat = blkg_to_lat(bio->bi_blkg);
607 enabled = blk_iolatency_enabled(iolat->blkiolat);
611 while (blkg && blkg->parent) {
612 iolat = blkg_to_lat(blkg);
617 rqw = &iolat->rq_wait;
619 inflight = atomic_dec_return(&rqw->inflight);
620 WARN_ON_ONCE(inflight < 0);
622 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
623 * submitted, so do not account for it.
625 if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
626 iolatency_record_time(iolat, &bio->bi_issue, now,
628 window_start = atomic64_read(&iolat->window_start);
629 if (now > window_start &&
630 (now - window_start) >= iolat->cur_win_nsec) {
631 if (atomic64_cmpxchg(&iolat->window_start,
632 window_start, now) == window_start)
633 iolatency_check_latencies(iolat, now);
641 static void blkcg_iolatency_exit(struct rq_qos *rqos)
643 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
645 del_timer_sync(&blkiolat->timer);
646 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
650 static struct rq_qos_ops blkcg_iolatency_ops = {
651 .throttle = blkcg_iolatency_throttle,
652 .done_bio = blkcg_iolatency_done_bio,
653 .exit = blkcg_iolatency_exit,
656 static void blkiolatency_timer_fn(struct timer_list *t)
658 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
659 struct blkcg_gq *blkg;
660 struct cgroup_subsys_state *pos_css;
661 u64 now = ktime_to_ns(ktime_get());
664 blkg_for_each_descendant_pre(blkg, pos_css,
665 blkiolat->rqos.q->root_blkg) {
666 struct iolatency_grp *iolat;
667 struct child_latency_info *lat_info;
672 * We could be exiting, don't access the pd unless we have a
675 if (!blkg_tryget(blkg))
678 iolat = blkg_to_lat(blkg);
682 lat_info = &iolat->child_lat;
683 cookie = atomic_read(&lat_info->scale_cookie);
685 if (cookie >= DEFAULT_SCALE_COOKIE)
688 spin_lock_irqsave(&lat_info->lock, flags);
689 if (lat_info->last_scale_event >= now)
693 * We scaled down but don't have a scale_grp, scale up and carry
696 if (lat_info->scale_grp == NULL) {
697 scale_cookie_change(iolat->blkiolat, lat_info, true);
702 * It's been 5 seconds since our last scale event, clear the
703 * scale grp in case the group that needed the scale down isn't
704 * doing any IO currently.
706 if (now - lat_info->last_scale_event >=
707 ((u64)NSEC_PER_SEC * 5))
708 lat_info->scale_grp = NULL;
710 spin_unlock_irqrestore(&lat_info->lock, flags);
717 int blk_iolatency_init(struct request_queue *q)
719 struct blk_iolatency *blkiolat;
723 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
727 rqos = &blkiolat->rqos;
728 rqos->id = RQ_QOS_LATENCY;
729 rqos->ops = &blkcg_iolatency_ops;
734 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
741 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
747 * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
750 static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
752 struct iolatency_grp *iolat = blkg_to_lat(blkg);
753 u64 oldval = iolat->min_lat_nsec;
755 iolat->min_lat_nsec = val;
756 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
757 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
758 BLKIOLATENCY_MAX_WIN_SIZE);
762 if (oldval && !val) {
763 blkcg_clear_delay(blkg);
769 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
772 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
773 struct child_latency_info *lat_info;
777 lat_info = &iolat->child_lat;
778 spin_lock(&lat_info->lock);
779 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
780 lat_info->last_scale_event = 0;
781 lat_info->scale_grp = NULL;
782 lat_info->scale_lat = 0;
783 spin_unlock(&lat_info->lock);
787 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
788 size_t nbytes, loff_t off)
790 struct blkcg *blkcg = css_to_blkcg(of_css(of));
791 struct blkcg_gq *blkg;
792 struct blkg_conf_ctx ctx;
793 struct iolatency_grp *iolat;
800 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
804 iolat = blkg_to_lat(ctx.blkg);
808 while ((tok = strsep(&p, " "))) {
810 char val[21]; /* 18446744073709551616 */
812 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
815 if (!strcmp(key, "target")) {
818 if (!strcmp(val, "max"))
820 else if (sscanf(val, "%llu", &v) == 1)
821 lat_val = v * NSEC_PER_USEC;
829 /* Walk up the tree to see if our new val is lower than it should be. */
831 oldval = iolat->min_lat_nsec;
833 enable = iolatency_set_min_lat_nsec(blkg, lat_val);
835 WARN_ON_ONCE(!blk_get_queue(blkg->q));
839 if (oldval != iolat->min_lat_nsec) {
840 iolatency_clear_scaling(blkg);
845 blkg_conf_finish(&ctx);
846 if (ret == 0 && enable) {
847 struct iolatency_grp *tmp = blkg_to_lat(blkg);
848 struct blk_iolatency *blkiolat = tmp->blkiolat;
850 blk_mq_freeze_queue(blkg->q);
853 atomic_inc(&blkiolat->enabled);
854 else if (enable == -1)
855 atomic_dec(&blkiolat->enabled);
859 blk_mq_unfreeze_queue(blkg->q);
862 blk_put_queue(blkg->q);
864 return ret ?: nbytes;
867 static u64 iolatency_prfill_limit(struct seq_file *sf,
868 struct blkg_policy_data *pd, int off)
870 struct iolatency_grp *iolat = pd_to_lat(pd);
871 const char *dname = blkg_dev_name(pd->blkg);
873 if (!dname || !iolat->min_lat_nsec)
875 seq_printf(sf, "%s target=%llu\n",
876 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
880 static int iolatency_print_limit(struct seq_file *sf, void *v)
882 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
883 iolatency_prfill_limit,
884 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
888 static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
891 struct latency_stat stat;
894 latency_stat_init(iolat, &stat);
896 for_each_online_cpu(cpu) {
897 struct latency_stat *s;
898 s = per_cpu_ptr(iolat->stats, cpu);
899 latency_stat_sum(iolat, &stat, s);
903 if (iolat->rq_depth.max_depth == UINT_MAX)
904 return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
905 (unsigned long long)stat.ps.missed,
906 (unsigned long long)stat.ps.total);
907 return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
908 (unsigned long long)stat.ps.missed,
909 (unsigned long long)stat.ps.total,
910 iolat->rq_depth.max_depth);
913 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
916 struct iolatency_grp *iolat = pd_to_lat(pd);
917 unsigned long long avg_lat;
918 unsigned long long cur_win;
920 if (!blkcg_debug_stats)
924 return iolatency_ssd_stat(iolat, buf, size);
926 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
927 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
928 if (iolat->rq_depth.max_depth == UINT_MAX)
929 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
932 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
933 iolat->rq_depth.max_depth, avg_lat, cur_win);
937 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
938 struct request_queue *q,
941 struct iolatency_grp *iolat;
943 iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
946 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
947 __alignof__(struct latency_stat), gfp);
955 static void iolatency_pd_init(struct blkg_policy_data *pd)
957 struct iolatency_grp *iolat = pd_to_lat(pd);
958 struct blkcg_gq *blkg = lat_to_blkg(iolat);
959 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
960 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
961 u64 now = ktime_to_ns(ktime_get());
964 if (blk_queue_nonrot(blkg->q))
969 for_each_possible_cpu(cpu) {
970 struct latency_stat *stat;
971 stat = per_cpu_ptr(iolat->stats, cpu);
972 latency_stat_init(iolat, stat);
975 latency_stat_init(iolat, &iolat->cur_stat);
976 rq_wait_init(&iolat->rq_wait);
977 spin_lock_init(&iolat->child_lat.lock);
978 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
979 iolat->rq_depth.max_depth = UINT_MAX;
980 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
981 iolat->blkiolat = blkiolat;
982 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
983 atomic64_set(&iolat->window_start, now);
986 * We init things in list order, so the pd for the parent may not be
987 * init'ed yet for whatever reason.
989 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
990 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
991 atomic_set(&iolat->scale_cookie,
992 atomic_read(&parent->child_lat.scale_cookie));
994 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
997 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1000 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1002 struct iolatency_grp *iolat = pd_to_lat(pd);
1003 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1004 struct blk_iolatency *blkiolat = iolat->blkiolat;
1007 ret = iolatency_set_min_lat_nsec(blkg, 0);
1009 atomic_inc(&blkiolat->enabled);
1011 atomic_dec(&blkiolat->enabled);
1012 iolatency_clear_scaling(blkg);
1015 static void iolatency_pd_free(struct blkg_policy_data *pd)
1017 struct iolatency_grp *iolat = pd_to_lat(pd);
1018 free_percpu(iolat->stats);
1022 static struct cftype iolatency_files[] = {
1025 .flags = CFTYPE_NOT_ON_ROOT,
1026 .seq_show = iolatency_print_limit,
1027 .write = iolatency_set_limit,
1032 static struct blkcg_policy blkcg_policy_iolatency = {
1033 .dfl_cftypes = iolatency_files,
1034 .pd_alloc_fn = iolatency_pd_alloc,
1035 .pd_init_fn = iolatency_pd_init,
1036 .pd_offline_fn = iolatency_pd_offline,
1037 .pd_free_fn = iolatency_pd_free,
1038 .pd_stat_fn = iolatency_pd_stat,
1041 static int __init iolatency_init(void)
1043 return blkcg_policy_register(&blkcg_policy_iolatency);
1046 static void __exit iolatency_exit(void)
1048 return blkcg_policy_unregister(&blkcg_policy_iolatency);
1051 module_init(iolatency_init);
1052 module_exit(iolatency_exit);