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 CALC_LOAD(iolat->lat_avg, iolatency_exp_factors[exp_idx], stat->rqs.mean);
263 static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
264 wait_queue_entry_t *wait,
267 struct rq_wait *rqw = &iolat->rq_wait;
269 if (first_block && waitqueue_active(&rqw->wait) &&
270 rqw->wait.head.next != &wait->entry)
272 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
275 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
276 struct iolatency_grp *iolat,
277 spinlock_t *lock, bool issue_as_root,
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))
314 spin_unlock_irq(lock);
322 finish_wait(&rqw->wait, &wait);
325 #define SCALE_DOWN_FACTOR 2
326 #define SCALE_UP_FACTOR 4
328 static inline unsigned long scale_amount(unsigned long qd, bool up)
330 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
334 * We scale the qd down faster than we scale up, so we need to use this helper
335 * to adjust the scale_cookie accordingly so we don't prematurely get
336 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
338 * Each group has their own local copy of the last scale cookie they saw, so if
339 * the global scale cookie goes up or down they know which way they need to go
340 * based on their last knowledge of it.
342 static void scale_cookie_change(struct blk_iolatency *blkiolat,
343 struct child_latency_info *lat_info,
346 unsigned long qd = blkiolat->rqos.q->nr_requests;
347 unsigned long scale = scale_amount(qd, up);
348 unsigned long old = atomic_read(&lat_info->scale_cookie);
349 unsigned long max_scale = qd << 1;
350 unsigned long diff = 0;
352 if (old < DEFAULT_SCALE_COOKIE)
353 diff = DEFAULT_SCALE_COOKIE - old;
356 if (scale + old > DEFAULT_SCALE_COOKIE)
357 atomic_set(&lat_info->scale_cookie,
358 DEFAULT_SCALE_COOKIE);
360 atomic_inc(&lat_info->scale_cookie);
362 atomic_add(scale, &lat_info->scale_cookie);
365 * We don't want to dig a hole so deep that it takes us hours to
366 * dig out of it. Just enough that we don't throttle/unthrottle
367 * with jagged workloads but can still unthrottle once pressure
368 * has sufficiently dissipated.
371 if (diff < max_scale)
372 atomic_dec(&lat_info->scale_cookie);
374 atomic_sub(scale, &lat_info->scale_cookie);
380 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
381 * queue depth at a time so we don't get wild swings and hopefully dial in to
382 * fairer distribution of the overall queue depth.
384 static void scale_change(struct iolatency_grp *iolat, bool up)
386 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
387 unsigned long scale = scale_amount(qd, up);
388 unsigned long old = iolat->rq_depth.max_depth;
394 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
400 iolat->rq_depth.max_depth = old;
401 wake_up_all(&iolat->rq_wait.wait);
405 iolat->rq_depth.max_depth = max(old, 1UL);
409 /* Check our parent and see if the scale cookie has changed. */
410 static void check_scale_change(struct iolatency_grp *iolat)
412 struct iolatency_grp *parent;
413 struct child_latency_info *lat_info;
414 unsigned int cur_cookie;
415 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
420 if (lat_to_blkg(iolat)->parent == NULL)
423 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
427 lat_info = &parent->child_lat;
428 cur_cookie = atomic_read(&lat_info->scale_cookie);
429 scale_lat = READ_ONCE(lat_info->scale_lat);
431 if (cur_cookie < our_cookie)
433 else if (cur_cookie > our_cookie)
438 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
440 /* Somebody beat us to the punch, just bail. */
441 if (old != our_cookie)
444 if (direction < 0 && iolat->min_lat_nsec) {
447 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
451 * Sometimes high priority groups are their own worst enemy, so
452 * instead of taking it out on some poor other group that did 5%
453 * or less of the IO's for the last summation just skip this
456 samples_thresh = lat_info->nr_samples * 5;
457 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
458 if (iolat->nr_samples <= samples_thresh)
462 /* We're as low as we can go. */
463 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
464 blkcg_use_delay(lat_to_blkg(iolat));
468 /* We're back to the default cookie, unthrottle all the things. */
469 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
470 blkcg_clear_delay(lat_to_blkg(iolat));
471 iolat->rq_depth.max_depth = UINT_MAX;
472 wake_up_all(&iolat->rq_wait.wait);
476 scale_change(iolat, direction > 0);
479 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
482 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
483 struct blkcg_gq *blkg = bio->bi_blkg;
484 bool issue_as_root = bio_issue_as_root_blkg(bio);
486 if (!blk_iolatency_enabled(blkiolat))
489 while (blkg && blkg->parent) {
490 struct iolatency_grp *iolat = blkg_to_lat(blkg);
496 check_scale_change(iolat);
497 __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
498 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
501 if (!timer_pending(&blkiolat->timer))
502 mod_timer(&blkiolat->timer, jiffies + HZ);
505 static void iolatency_record_time(struct iolatency_grp *iolat,
506 struct bio_issue *issue, u64 now,
509 u64 start = bio_issue_time(issue);
513 * Have to do this so we are truncated to the correct time that our
514 * issue is truncated to.
516 now = __bio_issue_time(now);
521 req_time = now - start;
524 * We don't want to count issue_as_root bio's in the cgroups latency
525 * statistics as it could skew the numbers downwards.
527 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
528 u64 sub = iolat->min_lat_nsec;
530 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
534 latency_stat_record_time(iolat, req_time);
537 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
538 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
540 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
542 struct blkcg_gq *blkg = lat_to_blkg(iolat);
543 struct iolatency_grp *parent;
544 struct child_latency_info *lat_info;
545 struct latency_stat stat;
549 latency_stat_init(iolat, &stat);
551 for_each_online_cpu(cpu) {
552 struct latency_stat *s;
553 s = per_cpu_ptr(iolat->stats, cpu);
554 latency_stat_sum(iolat, &stat, s);
555 latency_stat_init(iolat, s);
559 parent = blkg_to_lat(blkg->parent);
563 lat_info = &parent->child_lat;
565 iolat_update_total_lat_avg(iolat, &stat);
567 /* Everything is ok and we don't need to adjust the scale. */
568 if (latency_sum_ok(iolat, &stat) &&
569 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
572 /* Somebody beat us to the punch, just bail. */
573 spin_lock_irqsave(&lat_info->lock, flags);
575 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
576 lat_info->nr_samples -= iolat->nr_samples;
577 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
578 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
580 if ((lat_info->last_scale_event >= now ||
581 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
584 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
585 latency_sum_ok(iolat, &stat)) {
586 if (latency_stat_samples(iolat, &iolat->cur_stat) <
587 BLKIOLATENCY_MIN_GOOD_SAMPLES)
589 if (lat_info->scale_grp == iolat) {
590 lat_info->last_scale_event = now;
591 scale_cookie_change(iolat->blkiolat, lat_info, true);
593 } else if (lat_info->scale_lat == 0 ||
594 lat_info->scale_lat >= iolat->min_lat_nsec) {
595 lat_info->last_scale_event = now;
596 if (!lat_info->scale_grp ||
597 lat_info->scale_lat > iolat->min_lat_nsec) {
598 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
599 lat_info->scale_grp = iolat;
601 scale_cookie_change(iolat->blkiolat, lat_info, false);
603 latency_stat_init(iolat, &iolat->cur_stat);
605 spin_unlock_irqrestore(&lat_info->lock, flags);
608 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
610 struct blkcg_gq *blkg;
612 struct iolatency_grp *iolat;
614 u64 now = ktime_to_ns(ktime_get());
615 bool issue_as_root = bio_issue_as_root_blkg(bio);
616 bool enabled = false;
622 iolat = blkg_to_lat(bio->bi_blkg);
626 enabled = blk_iolatency_enabled(iolat->blkiolat);
627 while (blkg && blkg->parent) {
628 iolat = blkg_to_lat(blkg);
633 rqw = &iolat->rq_wait;
635 atomic_dec(&rqw->inflight);
636 if (!enabled || iolat->min_lat_nsec == 0)
638 iolatency_record_time(iolat, &bio->bi_issue, now,
640 window_start = atomic64_read(&iolat->window_start);
641 if (now > window_start &&
642 (now - window_start) >= iolat->cur_win_nsec) {
643 if (atomic64_cmpxchg(&iolat->window_start,
644 window_start, now) == window_start)
645 iolatency_check_latencies(iolat, now);
653 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
655 struct blkcg_gq *blkg;
658 while (blkg && blkg->parent) {
660 struct iolatency_grp *iolat;
662 iolat = blkg_to_lat(blkg);
666 rqw = &iolat->rq_wait;
667 atomic_dec(&rqw->inflight);
674 static void blkcg_iolatency_exit(struct rq_qos *rqos)
676 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
678 del_timer_sync(&blkiolat->timer);
679 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
683 static struct rq_qos_ops blkcg_iolatency_ops = {
684 .throttle = blkcg_iolatency_throttle,
685 .cleanup = blkcg_iolatency_cleanup,
686 .done_bio = blkcg_iolatency_done_bio,
687 .exit = blkcg_iolatency_exit,
690 static void blkiolatency_timer_fn(struct timer_list *t)
692 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
693 struct blkcg_gq *blkg;
694 struct cgroup_subsys_state *pos_css;
695 u64 now = ktime_to_ns(ktime_get());
698 blkg_for_each_descendant_pre(blkg, pos_css,
699 blkiolat->rqos.q->root_blkg) {
700 struct iolatency_grp *iolat;
701 struct child_latency_info *lat_info;
706 * We could be exiting, don't access the pd unless we have a
709 if (!blkg_tryget(blkg))
712 iolat = blkg_to_lat(blkg);
716 lat_info = &iolat->child_lat;
717 cookie = atomic_read(&lat_info->scale_cookie);
719 if (cookie >= DEFAULT_SCALE_COOKIE)
722 spin_lock_irqsave(&lat_info->lock, flags);
723 if (lat_info->last_scale_event >= now)
727 * We scaled down but don't have a scale_grp, scale up and carry
730 if (lat_info->scale_grp == NULL) {
731 scale_cookie_change(iolat->blkiolat, lat_info, true);
736 * It's been 5 seconds since our last scale event, clear the
737 * scale grp in case the group that needed the scale down isn't
738 * doing any IO currently.
740 if (now - lat_info->last_scale_event >=
741 ((u64)NSEC_PER_SEC * 5))
742 lat_info->scale_grp = NULL;
744 spin_unlock_irqrestore(&lat_info->lock, flags);
751 int blk_iolatency_init(struct request_queue *q)
753 struct blk_iolatency *blkiolat;
757 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
761 rqos = &blkiolat->rqos;
762 rqos->id = RQ_QOS_CGROUP;
763 rqos->ops = &blkcg_iolatency_ops;
768 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
775 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
780 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
782 struct iolatency_grp *iolat = blkg_to_lat(blkg);
783 struct blk_iolatency *blkiolat = iolat->blkiolat;
784 u64 oldval = iolat->min_lat_nsec;
786 iolat->min_lat_nsec = val;
787 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
788 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
789 BLKIOLATENCY_MAX_WIN_SIZE);
792 atomic_inc(&blkiolat->enabled);
794 atomic_dec(&blkiolat->enabled);
797 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
800 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
801 struct child_latency_info *lat_info;
805 lat_info = &iolat->child_lat;
806 spin_lock(&lat_info->lock);
807 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
808 lat_info->last_scale_event = 0;
809 lat_info->scale_grp = NULL;
810 lat_info->scale_lat = 0;
811 spin_unlock(&lat_info->lock);
815 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
816 size_t nbytes, loff_t off)
818 struct blkcg *blkcg = css_to_blkcg(of_css(of));
819 struct blkcg_gq *blkg;
820 struct blkg_conf_ctx ctx;
821 struct iolatency_grp *iolat;
827 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
831 iolat = blkg_to_lat(ctx.blkg);
835 while ((tok = strsep(&p, " "))) {
837 char val[21]; /* 18446744073709551616 */
839 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
842 if (!strcmp(key, "target")) {
845 if (!strcmp(val, "max"))
847 else if (sscanf(val, "%llu", &v) == 1)
848 lat_val = v * NSEC_PER_USEC;
856 /* Walk up the tree to see if our new val is lower than it should be. */
858 oldval = iolat->min_lat_nsec;
860 iolatency_set_min_lat_nsec(blkg, lat_val);
861 if (oldval != iolat->min_lat_nsec) {
862 iolatency_clear_scaling(blkg);
867 blkg_conf_finish(&ctx);
868 return ret ?: nbytes;
871 static u64 iolatency_prfill_limit(struct seq_file *sf,
872 struct blkg_policy_data *pd, int off)
874 struct iolatency_grp *iolat = pd_to_lat(pd);
875 const char *dname = blkg_dev_name(pd->blkg);
877 if (!dname || !iolat->min_lat_nsec)
879 seq_printf(sf, "%s target=%llu\n",
880 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
884 static int iolatency_print_limit(struct seq_file *sf, void *v)
886 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
887 iolatency_prfill_limit,
888 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
892 static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
895 struct latency_stat stat;
898 latency_stat_init(iolat, &stat);
900 for_each_online_cpu(cpu) {
901 struct latency_stat *s;
902 s = per_cpu_ptr(iolat->stats, cpu);
903 latency_stat_sum(iolat, &stat, s);
907 if (iolat->rq_depth.max_depth == UINT_MAX)
908 return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
909 (unsigned long long)stat.ps.missed,
910 (unsigned long long)stat.ps.total);
911 return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
912 (unsigned long long)stat.ps.missed,
913 (unsigned long long)stat.ps.total,
914 iolat->rq_depth.max_depth);
917 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
920 struct iolatency_grp *iolat = pd_to_lat(pd);
921 unsigned long long avg_lat;
922 unsigned long long cur_win;
925 return iolatency_ssd_stat(iolat, buf, size);
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",
933 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
934 iolat->rq_depth.max_depth, avg_lat, cur_win);
938 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
940 struct iolatency_grp *iolat;
942 iolat = kzalloc_node(sizeof(*iolat), gfp, node);
945 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
946 __alignof__(struct latency_stat), gfp);
954 static void iolatency_pd_init(struct blkg_policy_data *pd)
956 struct iolatency_grp *iolat = pd_to_lat(pd);
957 struct blkcg_gq *blkg = lat_to_blkg(iolat);
958 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
959 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
960 u64 now = ktime_to_ns(ktime_get());
963 if (blk_queue_nonrot(blkg->q))
968 for_each_possible_cpu(cpu) {
969 struct latency_stat *stat;
970 stat = per_cpu_ptr(iolat->stats, cpu);
971 latency_stat_init(iolat, stat);
974 latency_stat_init(iolat, &iolat->cur_stat);
975 rq_wait_init(&iolat->rq_wait);
976 spin_lock_init(&iolat->child_lat.lock);
977 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
978 iolat->rq_depth.max_depth = UINT_MAX;
979 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
980 iolat->blkiolat = blkiolat;
981 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
982 atomic64_set(&iolat->window_start, now);
985 * We init things in list order, so the pd for the parent may not be
986 * init'ed yet for whatever reason.
988 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
989 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
990 atomic_set(&iolat->scale_cookie,
991 atomic_read(&parent->child_lat.scale_cookie));
993 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
996 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
999 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1001 struct iolatency_grp *iolat = pd_to_lat(pd);
1002 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1004 iolatency_set_min_lat_nsec(blkg, 0);
1005 iolatency_clear_scaling(blkg);
1008 static void iolatency_pd_free(struct blkg_policy_data *pd)
1010 struct iolatency_grp *iolat = pd_to_lat(pd);
1011 free_percpu(iolat->stats);
1015 static struct cftype iolatency_files[] = {
1018 .flags = CFTYPE_NOT_ON_ROOT,
1019 .seq_show = iolatency_print_limit,
1020 .write = iolatency_set_limit,
1025 static struct blkcg_policy blkcg_policy_iolatency = {
1026 .dfl_cftypes = iolatency_files,
1027 .pd_alloc_fn = iolatency_pd_alloc,
1028 .pd_init_fn = iolatency_pd_init,
1029 .pd_offline_fn = iolatency_pd_offline,
1030 .pd_free_fn = iolatency_pd_free,
1031 .pd_stat_fn = iolatency_pd_stat,
1034 static int __init iolatency_init(void)
1036 return blkcg_policy_register(&blkcg_policy_iolatency);
1039 static void __exit iolatency_exit(void)
1041 return blkcg_policy_unregister(&blkcg_policy_iolatency);
1044 module_init(iolatency_init);
1045 module_exit(iolatency_exit);