2 * Block rq-qos base io controller
4 * This works similar to wbt with a few exceptions
6 * - It's bio based, so the latency covers the whole block layer in addition to
8 * - We will throttle all IO that comes in here if we need to.
9 * - We use the mean latency over the 100ms window. This is because writes can
10 * be particularly fast, which could give us a false sense of the impact of
11 * other workloads on our protected workload.
12 * - By default there's no throttling, we set the queue_depth to UINT_MAX so
13 * that we can have as many outstanding bio's as we're allowed to. Only at
14 * throttle time do we pay attention to the actual queue depth.
16 * The hierarchy works like the cpu controller does, we track the latency at
17 * every configured node, and each configured node has it's own independent
18 * queue depth. This means that we only care about our latency targets at the
19 * peer level. Some group at the bottom of the hierarchy isn't going to affect
20 * a group at the end of some other path if we're only configred at leaf level.
22 * Consider the following
26 * fast (target=5ms) slow (target=10ms)
28 * a b normal(15ms) unloved
30 * "a" and "b" have no target, but their combined io under "fast" cannot exceed
31 * an average latency of 5ms. If it does then we will throttle the "slow"
32 * group. In the case of "normal", if it exceeds its 15ms target, we will
33 * throttle "unloved", but nobody else.
35 * In this example "fast", "slow", and "normal" will be the only groups actually
36 * accounting their io latencies. We have to walk up the heirarchy to the root
37 * on every submit and complete so we can do the appropriate stat recording and
38 * adjust the queue depth of ourselves if needed.
40 * There are 2 ways we throttle IO.
42 * 1) Queue depth throttling. As we throttle down we will adjust the maximum
43 * number of IO's we're allowed to have in flight. This starts at (u64)-1 down
44 * to 1. If the group is only ever submitting IO for itself then this is the
45 * only way we throttle.
47 * 2) Induced delay throttling. This is for the case that a group is generating
48 * IO that has to be issued by the root cg to avoid priority inversion. So think
49 * REQ_META or REQ_SWAP. If we are already at qd == 1 and we're getting a lot
50 * of work done for us on behalf of the root cg and are being asked to scale
51 * down more then we induce a latency at userspace return. We accumulate the
52 * total amount of time we need to be punished by doing
54 * total_time += min_lat_nsec - actual_io_completion
56 * and then at throttle time will do
58 * throttle_time = min(total_time, NSEC_PER_SEC)
60 * This induced delay will throttle back the activity that is generating the
61 * root cg issued io's, wethere that's some metadata intensive operation or the
62 * group is using so much memory that it is pushing us into swap.
64 * Copyright (C) 2018 Josef Bacik
66 #include <linux/kernel.h>
67 #include <linux/blk_types.h>
68 #include <linux/backing-dev.h>
69 #include <linux/module.h>
70 #include <linux/timer.h>
71 #include <linux/memcontrol.h>
72 #include <linux/sched/loadavg.h>
73 #include <linux/sched/signal.h>
74 #include <trace/events/block.h>
75 #include "blk-rq-qos.h"
78 #define DEFAULT_SCALE_COOKIE 1000000U
80 static struct blkcg_policy blkcg_policy_iolatency;
83 struct blk_iolatency {
85 struct timer_list timer;
89 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
91 return container_of(rqos, struct blk_iolatency, rqos);
94 static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
96 return atomic_read(&blkiolat->enabled) > 0;
99 struct child_latency_info {
102 /* Last time we adjusted the scale of everybody. */
103 u64 last_scale_event;
105 /* The latency that we missed. */
108 /* Total io's from all of our children for the last summation. */
111 /* The guy who actually changed the latency numbers. */
112 struct iolatency_grp *scale_grp;
114 /* Cookie to tell if we need to scale up or down. */
115 atomic_t scale_cookie;
118 struct percentile_stats {
123 struct latency_stat {
125 struct percentile_stats ps;
126 struct blk_rq_stat rqs;
130 struct iolatency_grp {
131 struct blkg_policy_data pd;
132 struct latency_stat __percpu *stats;
133 struct latency_stat cur_stat;
134 struct blk_iolatency *blkiolat;
135 struct rq_depth rq_depth;
136 struct rq_wait rq_wait;
137 atomic64_t window_start;
138 atomic_t scale_cookie;
142 /* total running average of our io latency. */
145 /* Our current number of IO's for the last summation. */
149 struct child_latency_info child_lat;
152 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
153 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
155 * These are the constants used to fake the fixed-point moving average
156 * calculation just like load average. The call to calc_load() folds
157 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
158 * window size is bucketed to try to approximately calculate average
159 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
160 * elapse immediately. Note, windows only elapse with IO activity. Idle
161 * periods extend the most recent window.
163 #define BLKIOLATENCY_NR_EXP_FACTORS 5
164 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
165 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
166 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
167 2045, // exp(1/600) - 600 samples
168 2039, // exp(1/240) - 240 samples
169 2031, // exp(1/120) - 120 samples
170 2023, // exp(1/80) - 80 samples
171 2014, // exp(1/60) - 60 samples
174 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
176 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
179 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
181 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
184 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
186 return pd_to_blkg(&iolat->pd);
189 static inline void latency_stat_init(struct iolatency_grp *iolat,
190 struct latency_stat *stat)
196 blk_rq_stat_init(&stat->rqs);
199 static inline void latency_stat_sum(struct iolatency_grp *iolat,
200 struct latency_stat *sum,
201 struct latency_stat *stat)
204 sum->ps.total += stat->ps.total;
205 sum->ps.missed += stat->ps.missed;
207 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
210 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
213 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
215 if (req_time >= iolat->min_lat_nsec)
219 blk_rq_stat_add(&stat->rqs, req_time);
223 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
224 struct latency_stat *stat)
227 u64 thresh = div64_u64(stat->ps.total, 10);
228 thresh = max(thresh, 1ULL);
229 return stat->ps.missed < thresh;
231 return stat->rqs.mean <= iolat->min_lat_nsec;
234 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
235 struct latency_stat *stat)
238 return stat->ps.total;
239 return stat->rqs.nr_samples;
242 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
243 struct latency_stat *stat)
251 * calc_load() takes in a number stored in fixed point representation.
252 * Because we are using this for IO time in ns, the values stored
253 * are significantly larger than the FIXED_1 denominator (2048).
254 * Therefore, rounding errors in the calculation are negligible and
257 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
258 div64_u64(iolat->cur_win_nsec,
259 BLKIOLATENCY_EXP_BUCKET_SIZE));
260 iolat->lat_avg = calc_load(iolat->lat_avg,
261 iolatency_exp_factors[exp_idx],
265 static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
266 wait_queue_entry_t *wait,
269 struct rq_wait *rqw = &iolat->rq_wait;
271 if (first_block && waitqueue_active(&rqw->wait) &&
272 rqw->wait.head.next != &wait->entry)
274 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
277 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
278 struct iolatency_grp *iolat,
282 struct rq_wait *rqw = &iolat->rq_wait;
283 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
285 bool first_block = true;
288 blkcg_schedule_throttle(rqos->q, use_memdelay);
291 * To avoid priority inversions we want to just take a slot if we are
292 * issuing as root. If we're being killed off there's no point in
293 * delaying things, we may have been killed by OOM so throttling may
294 * make recovery take even longer, so just let the IO's through so the
297 if (issue_as_root || fatal_signal_pending(current)) {
298 atomic_inc(&rqw->inflight);
302 if (iolatency_may_queue(iolat, &wait, first_block))
306 prepare_to_wait_exclusive(&rqw->wait, &wait,
307 TASK_UNINTERRUPTIBLE);
309 if (iolatency_may_queue(iolat, &wait, first_block))
315 finish_wait(&rqw->wait, &wait);
318 #define SCALE_DOWN_FACTOR 2
319 #define SCALE_UP_FACTOR 4
321 static inline unsigned long scale_amount(unsigned long qd, bool up)
323 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
327 * We scale the qd down faster than we scale up, so we need to use this helper
328 * to adjust the scale_cookie accordingly so we don't prematurely get
329 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
331 * Each group has their own local copy of the last scale cookie they saw, so if
332 * the global scale cookie goes up or down they know which way they need to go
333 * based on their last knowledge of it.
335 static void scale_cookie_change(struct blk_iolatency *blkiolat,
336 struct child_latency_info *lat_info,
339 unsigned long qd = blkiolat->rqos.q->nr_requests;
340 unsigned long scale = scale_amount(qd, up);
341 unsigned long old = atomic_read(&lat_info->scale_cookie);
342 unsigned long max_scale = qd << 1;
343 unsigned long diff = 0;
345 if (old < DEFAULT_SCALE_COOKIE)
346 diff = DEFAULT_SCALE_COOKIE - old;
349 if (scale + old > DEFAULT_SCALE_COOKIE)
350 atomic_set(&lat_info->scale_cookie,
351 DEFAULT_SCALE_COOKIE);
353 atomic_inc(&lat_info->scale_cookie);
355 atomic_add(scale, &lat_info->scale_cookie);
358 * We don't want to dig a hole so deep that it takes us hours to
359 * dig out of it. Just enough that we don't throttle/unthrottle
360 * with jagged workloads but can still unthrottle once pressure
361 * has sufficiently dissipated.
364 if (diff < max_scale)
365 atomic_dec(&lat_info->scale_cookie);
367 atomic_sub(scale, &lat_info->scale_cookie);
373 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
374 * queue depth at a time so we don't get wild swings and hopefully dial in to
375 * fairer distribution of the overall queue depth.
377 static void scale_change(struct iolatency_grp *iolat, bool up)
379 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
380 unsigned long scale = scale_amount(qd, up);
381 unsigned long old = iolat->rq_depth.max_depth;
387 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
393 iolat->rq_depth.max_depth = old;
394 wake_up_all(&iolat->rq_wait.wait);
398 iolat->rq_depth.max_depth = max(old, 1UL);
402 /* Check our parent and see if the scale cookie has changed. */
403 static void check_scale_change(struct iolatency_grp *iolat)
405 struct iolatency_grp *parent;
406 struct child_latency_info *lat_info;
407 unsigned int cur_cookie;
408 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
413 if (lat_to_blkg(iolat)->parent == NULL)
416 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
420 lat_info = &parent->child_lat;
421 cur_cookie = atomic_read(&lat_info->scale_cookie);
422 scale_lat = READ_ONCE(lat_info->scale_lat);
424 if (cur_cookie < our_cookie)
426 else if (cur_cookie > our_cookie)
431 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
433 /* Somebody beat us to the punch, just bail. */
434 if (old != our_cookie)
437 if (direction < 0 && iolat->min_lat_nsec) {
440 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
444 * Sometimes high priority groups are their own worst enemy, so
445 * instead of taking it out on some poor other group that did 5%
446 * or less of the IO's for the last summation just skip this
449 samples_thresh = lat_info->nr_samples * 5;
450 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
451 if (iolat->nr_samples <= samples_thresh)
455 /* We're as low as we can go. */
456 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
457 blkcg_use_delay(lat_to_blkg(iolat));
461 /* We're back to the default cookie, unthrottle all the things. */
462 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
463 blkcg_clear_delay(lat_to_blkg(iolat));
464 iolat->rq_depth.max_depth = UINT_MAX;
465 wake_up_all(&iolat->rq_wait.wait);
469 scale_change(iolat, direction > 0);
472 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
474 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
475 struct blkcg_gq *blkg;
476 bool issue_as_root = bio_issue_as_root_blkg(bio);
478 if (!blk_iolatency_enabled(blkiolat))
481 bio_associate_blkg(bio);
483 bio_issue_init(&bio->bi_issue, bio_sectors(bio));
485 while (blkg && blkg->parent) {
486 struct iolatency_grp *iolat = blkg_to_lat(blkg);
492 check_scale_change(iolat);
493 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
494 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
497 if (!timer_pending(&blkiolat->timer))
498 mod_timer(&blkiolat->timer, jiffies + HZ);
501 static void iolatency_record_time(struct iolatency_grp *iolat,
502 struct bio_issue *issue, u64 now,
505 u64 start = bio_issue_time(issue);
509 * Have to do this so we are truncated to the correct time that our
510 * issue is truncated to.
512 now = __bio_issue_time(now);
517 req_time = now - start;
520 * We don't want to count issue_as_root bio's in the cgroups latency
521 * statistics as it could skew the numbers downwards.
523 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
524 u64 sub = iolat->min_lat_nsec;
526 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
530 latency_stat_record_time(iolat, req_time);
533 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
534 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
536 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
538 struct blkcg_gq *blkg = lat_to_blkg(iolat);
539 struct iolatency_grp *parent;
540 struct child_latency_info *lat_info;
541 struct latency_stat stat;
545 latency_stat_init(iolat, &stat);
547 for_each_online_cpu(cpu) {
548 struct latency_stat *s;
549 s = per_cpu_ptr(iolat->stats, cpu);
550 latency_stat_sum(iolat, &stat, s);
551 latency_stat_init(iolat, s);
555 parent = blkg_to_lat(blkg->parent);
559 lat_info = &parent->child_lat;
561 iolat_update_total_lat_avg(iolat, &stat);
563 /* Everything is ok and we don't need to adjust the scale. */
564 if (latency_sum_ok(iolat, &stat) &&
565 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
568 /* Somebody beat us to the punch, just bail. */
569 spin_lock_irqsave(&lat_info->lock, flags);
571 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
572 lat_info->nr_samples -= iolat->nr_samples;
573 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
574 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
576 if ((lat_info->last_scale_event >= now ||
577 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
580 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
581 latency_sum_ok(iolat, &stat)) {
582 if (latency_stat_samples(iolat, &iolat->cur_stat) <
583 BLKIOLATENCY_MIN_GOOD_SAMPLES)
585 if (lat_info->scale_grp == iolat) {
586 lat_info->last_scale_event = now;
587 scale_cookie_change(iolat->blkiolat, lat_info, true);
589 } else if (lat_info->scale_lat == 0 ||
590 lat_info->scale_lat >= iolat->min_lat_nsec) {
591 lat_info->last_scale_event = now;
592 if (!lat_info->scale_grp ||
593 lat_info->scale_lat > iolat->min_lat_nsec) {
594 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
595 lat_info->scale_grp = iolat;
597 scale_cookie_change(iolat->blkiolat, lat_info, false);
599 latency_stat_init(iolat, &iolat->cur_stat);
601 spin_unlock_irqrestore(&lat_info->lock, flags);
604 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
606 struct blkcg_gq *blkg;
608 struct iolatency_grp *iolat;
610 u64 now = ktime_to_ns(ktime_get());
611 bool issue_as_root = bio_issue_as_root_blkg(bio);
612 bool enabled = false;
618 iolat = blkg_to_lat(bio->bi_blkg);
622 enabled = blk_iolatency_enabled(iolat->blkiolat);
623 while (blkg && blkg->parent) {
624 iolat = blkg_to_lat(blkg);
629 rqw = &iolat->rq_wait;
631 atomic_dec(&rqw->inflight);
632 if (!enabled || iolat->min_lat_nsec == 0)
634 iolatency_record_time(iolat, &bio->bi_issue, now,
636 window_start = atomic64_read(&iolat->window_start);
637 if (now > window_start &&
638 (now - window_start) >= iolat->cur_win_nsec) {
639 if (atomic64_cmpxchg(&iolat->window_start,
640 window_start, now) == window_start)
641 iolatency_check_latencies(iolat, now);
649 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
651 struct blkcg_gq *blkg;
654 while (blkg && blkg->parent) {
656 struct iolatency_grp *iolat;
658 iolat = blkg_to_lat(blkg);
662 rqw = &iolat->rq_wait;
663 atomic_dec(&rqw->inflight);
670 static void blkcg_iolatency_exit(struct rq_qos *rqos)
672 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
674 del_timer_sync(&blkiolat->timer);
675 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
679 static struct rq_qos_ops blkcg_iolatency_ops = {
680 .throttle = blkcg_iolatency_throttle,
681 .cleanup = blkcg_iolatency_cleanup,
682 .done_bio = blkcg_iolatency_done_bio,
683 .exit = blkcg_iolatency_exit,
686 static void blkiolatency_timer_fn(struct timer_list *t)
688 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
689 struct blkcg_gq *blkg;
690 struct cgroup_subsys_state *pos_css;
691 u64 now = ktime_to_ns(ktime_get());
694 blkg_for_each_descendant_pre(blkg, pos_css,
695 blkiolat->rqos.q->root_blkg) {
696 struct iolatency_grp *iolat;
697 struct child_latency_info *lat_info;
702 * We could be exiting, don't access the pd unless we have a
705 if (!blkg_try_get(blkg))
708 iolat = blkg_to_lat(blkg);
712 lat_info = &iolat->child_lat;
713 cookie = atomic_read(&lat_info->scale_cookie);
715 if (cookie >= DEFAULT_SCALE_COOKIE)
718 spin_lock_irqsave(&lat_info->lock, flags);
719 if (lat_info->last_scale_event >= now)
723 * We scaled down but don't have a scale_grp, scale up and carry
726 if (lat_info->scale_grp == NULL) {
727 scale_cookie_change(iolat->blkiolat, lat_info, true);
732 * It's been 5 seconds since our last scale event, clear the
733 * scale grp in case the group that needed the scale down isn't
734 * doing any IO currently.
736 if (now - lat_info->last_scale_event >=
737 ((u64)NSEC_PER_SEC * 5))
738 lat_info->scale_grp = NULL;
740 spin_unlock_irqrestore(&lat_info->lock, flags);
747 int blk_iolatency_init(struct request_queue *q)
749 struct blk_iolatency *blkiolat;
753 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
757 rqos = &blkiolat->rqos;
758 rqos->id = RQ_QOS_CGROUP;
759 rqos->ops = &blkcg_iolatency_ops;
764 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
771 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
776 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
778 struct iolatency_grp *iolat = blkg_to_lat(blkg);
779 struct blk_iolatency *blkiolat = iolat->blkiolat;
780 u64 oldval = iolat->min_lat_nsec;
782 iolat->min_lat_nsec = val;
783 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
784 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
785 BLKIOLATENCY_MAX_WIN_SIZE);
788 atomic_inc(&blkiolat->enabled);
790 atomic_dec(&blkiolat->enabled);
793 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
796 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
797 struct child_latency_info *lat_info;
801 lat_info = &iolat->child_lat;
802 spin_lock(&lat_info->lock);
803 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
804 lat_info->last_scale_event = 0;
805 lat_info->scale_grp = NULL;
806 lat_info->scale_lat = 0;
807 spin_unlock(&lat_info->lock);
811 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
812 size_t nbytes, loff_t off)
814 struct blkcg *blkcg = css_to_blkcg(of_css(of));
815 struct blkcg_gq *blkg;
816 struct blkg_conf_ctx ctx;
817 struct iolatency_grp *iolat;
823 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
827 iolat = blkg_to_lat(ctx.blkg);
831 while ((tok = strsep(&p, " "))) {
833 char val[21]; /* 18446744073709551616 */
835 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
838 if (!strcmp(key, "target")) {
841 if (!strcmp(val, "max"))
843 else if (sscanf(val, "%llu", &v) == 1)
844 lat_val = v * NSEC_PER_USEC;
852 /* Walk up the tree to see if our new val is lower than it should be. */
854 oldval = iolat->min_lat_nsec;
856 iolatency_set_min_lat_nsec(blkg, lat_val);
857 if (oldval != iolat->min_lat_nsec) {
858 iolatency_clear_scaling(blkg);
863 blkg_conf_finish(&ctx);
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;
921 return iolatency_ssd_stat(iolat, buf, size);
923 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
924 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
925 if (iolat->rq_depth.max_depth == UINT_MAX)
926 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
929 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
930 iolat->rq_depth.max_depth, avg_lat, cur_win);
934 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
936 struct iolatency_grp *iolat;
938 iolat = kzalloc_node(sizeof(*iolat), gfp, node);
941 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
942 __alignof__(struct latency_stat), gfp);
950 static void iolatency_pd_init(struct blkg_policy_data *pd)
952 struct iolatency_grp *iolat = pd_to_lat(pd);
953 struct blkcg_gq *blkg = lat_to_blkg(iolat);
954 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
955 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
956 u64 now = ktime_to_ns(ktime_get());
959 if (blk_queue_nonrot(blkg->q))
964 for_each_possible_cpu(cpu) {
965 struct latency_stat *stat;
966 stat = per_cpu_ptr(iolat->stats, cpu);
967 latency_stat_init(iolat, stat);
970 latency_stat_init(iolat, &iolat->cur_stat);
971 rq_wait_init(&iolat->rq_wait);
972 spin_lock_init(&iolat->child_lat.lock);
973 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
974 iolat->rq_depth.max_depth = UINT_MAX;
975 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
976 iolat->blkiolat = blkiolat;
977 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
978 atomic64_set(&iolat->window_start, now);
981 * We init things in list order, so the pd for the parent may not be
982 * init'ed yet for whatever reason.
984 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
985 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
986 atomic_set(&iolat->scale_cookie,
987 atomic_read(&parent->child_lat.scale_cookie));
989 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
992 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
995 static void iolatency_pd_offline(struct blkg_policy_data *pd)
997 struct iolatency_grp *iolat = pd_to_lat(pd);
998 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1000 iolatency_set_min_lat_nsec(blkg, 0);
1001 iolatency_clear_scaling(blkg);
1004 static void iolatency_pd_free(struct blkg_policy_data *pd)
1006 struct iolatency_grp *iolat = pd_to_lat(pd);
1007 free_percpu(iolat->stats);
1011 static struct cftype iolatency_files[] = {
1014 .flags = CFTYPE_NOT_ON_ROOT,
1015 .seq_show = iolatency_print_limit,
1016 .write = iolatency_set_limit,
1021 static struct blkcg_policy blkcg_policy_iolatency = {
1022 .dfl_cftypes = iolatency_files,
1023 .pd_alloc_fn = iolatency_pd_alloc,
1024 .pd_init_fn = iolatency_pd_init,
1025 .pd_offline_fn = iolatency_pd_offline,
1026 .pd_free_fn = iolatency_pd_free,
1027 .pd_stat_fn = iolatency_pd_stat,
1030 static int __init iolatency_init(void)
1032 return blkcg_policy_register(&blkcg_policy_iolatency);
1035 static void __exit iolatency_exit(void)
1037 return blkcg_policy_unregister(&blkcg_policy_iolatency);
1040 module_init(iolatency_init);
1041 module_exit(iolatency_exit);