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 <linux/blk-mq.h>
76 #include "blk-rq-qos.h"
80 #define DEFAULT_SCALE_COOKIE 1000000U
82 static struct blkcg_policy blkcg_policy_iolatency;
85 struct blk_iolatency {
87 struct timer_list timer;
90 * ->enabled is the master enable switch gating the throttling logic and
91 * inflight tracking. The number of cgroups which have iolat enabled is
92 * tracked in ->enable_cnt, and ->enable is flipped on/off accordingly
93 * from ->enable_work with the request_queue frozen. For details, See
94 * blkiolatency_enable_work_fn().
98 struct work_struct enable_work;
101 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
103 return container_of(rqos, struct blk_iolatency, rqos);
106 struct child_latency_info {
109 /* Last time we adjusted the scale of everybody. */
110 u64 last_scale_event;
112 /* The latency that we missed. */
115 /* Total io's from all of our children for the last summation. */
118 /* The guy who actually changed the latency numbers. */
119 struct iolatency_grp *scale_grp;
121 /* Cookie to tell if we need to scale up or down. */
122 atomic_t scale_cookie;
125 struct iolatency_grp {
126 struct blkg_policy_data pd;
127 struct blk_rq_stat __percpu *stats;
128 struct blk_iolatency *blkiolat;
129 struct rq_depth rq_depth;
130 struct rq_wait rq_wait;
131 atomic64_t window_start;
132 atomic_t scale_cookie;
136 /* total running average of our io latency. */
139 /* Our current number of IO's for the last summation. */
142 struct child_latency_info child_lat;
145 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
146 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
148 * These are the constants used to fake the fixed-point moving average
149 * calculation just like load average. The call to CALC_LOAD folds
150 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
151 * window size is bucketed to try to approximately calculate average
152 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
153 * elapse immediately. Note, windows only elapse with IO activity. Idle
154 * periods extend the most recent window.
156 #define BLKIOLATENCY_NR_EXP_FACTORS 5
157 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
158 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
159 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
160 2045, // exp(1/600) - 600 samples
161 2039, // exp(1/240) - 240 samples
162 2031, // exp(1/120) - 120 samples
163 2023, // exp(1/80) - 80 samples
164 2014, // exp(1/60) - 60 samples
167 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
169 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
172 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
174 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
177 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
179 return pd_to_blkg(&iolat->pd);
182 static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
183 wait_queue_entry_t *wait,
186 struct rq_wait *rqw = &iolat->rq_wait;
188 if (first_block && waitqueue_active(&rqw->wait) &&
189 rqw->wait.head.next != &wait->entry)
191 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
194 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
195 struct iolatency_grp *iolat,
196 spinlock_t *lock, bool issue_as_root,
201 struct rq_wait *rqw = &iolat->rq_wait;
202 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
204 bool first_block = true;
207 blkcg_schedule_throttle(rqos->q, use_memdelay);
210 * To avoid priority inversions we want to just take a slot if we are
211 * issuing as root. If we're being killed off there's no point in
212 * delaying things, we may have been killed by OOM so throttling may
213 * make recovery take even longer, so just let the IO's through so the
216 if (issue_as_root || fatal_signal_pending(current)) {
217 atomic_inc(&rqw->inflight);
221 if (iolatency_may_queue(iolat, &wait, first_block))
225 prepare_to_wait_exclusive(&rqw->wait, &wait,
226 TASK_UNINTERRUPTIBLE);
228 if (iolatency_may_queue(iolat, &wait, first_block))
233 spin_unlock_irq(lock);
241 finish_wait(&rqw->wait, &wait);
244 #define SCALE_DOWN_FACTOR 2
245 #define SCALE_UP_FACTOR 4
247 static inline unsigned long scale_amount(unsigned long qd, bool up)
249 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
253 * We scale the qd down faster than we scale up, so we need to use this helper
254 * to adjust the scale_cookie accordingly so we don't prematurely get
255 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
257 * Each group has their own local copy of the last scale cookie they saw, so if
258 * the global scale cookie goes up or down they know which way they need to go
259 * based on their last knowledge of it.
261 static void scale_cookie_change(struct blk_iolatency *blkiolat,
262 struct child_latency_info *lat_info,
265 unsigned long qd = blk_queue_depth(blkiolat->rqos.q);
266 unsigned long scale = scale_amount(qd, up);
267 unsigned long old = atomic_read(&lat_info->scale_cookie);
268 unsigned long max_scale = qd << 1;
269 unsigned long diff = 0;
271 if (old < DEFAULT_SCALE_COOKIE)
272 diff = DEFAULT_SCALE_COOKIE - old;
275 if (scale + old > DEFAULT_SCALE_COOKIE)
276 atomic_set(&lat_info->scale_cookie,
277 DEFAULT_SCALE_COOKIE);
279 atomic_inc(&lat_info->scale_cookie);
281 atomic_add(scale, &lat_info->scale_cookie);
284 * We don't want to dig a hole so deep that it takes us hours to
285 * dig out of it. Just enough that we don't throttle/unthrottle
286 * with jagged workloads but can still unthrottle once pressure
287 * has sufficiently dissipated.
290 if (diff < max_scale)
291 atomic_dec(&lat_info->scale_cookie);
293 atomic_sub(scale, &lat_info->scale_cookie);
299 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
300 * queue depth at a time so we don't get wild swings and hopefully dial in to
301 * fairer distribution of the overall queue depth.
303 static void scale_change(struct iolatency_grp *iolat, bool up)
305 unsigned long qd = blk_queue_depth(iolat->blkiolat->rqos.q);
306 unsigned long scale = scale_amount(qd, up);
307 unsigned long old = iolat->rq_depth.max_depth;
308 bool changed = false;
314 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
321 iolat->rq_depth.max_depth = old;
322 wake_up_all(&iolat->rq_wait.wait);
324 } else if (old > 1) {
327 iolat->rq_depth.max_depth = max(old, 1UL);
331 /* Check our parent and see if the scale cookie has changed. */
332 static void check_scale_change(struct iolatency_grp *iolat)
334 struct iolatency_grp *parent;
335 struct child_latency_info *lat_info;
336 unsigned int cur_cookie;
337 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
342 if (lat_to_blkg(iolat)->parent == NULL)
345 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
349 lat_info = &parent->child_lat;
350 cur_cookie = atomic_read(&lat_info->scale_cookie);
351 scale_lat = READ_ONCE(lat_info->scale_lat);
353 if (cur_cookie < our_cookie)
355 else if (cur_cookie > our_cookie)
360 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
362 /* Somebody beat us to the punch, just bail. */
363 if (old != our_cookie)
366 if (direction < 0 && iolat->min_lat_nsec) {
369 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
373 * Sometimes high priority groups are their own worst enemy, so
374 * instead of taking it out on some poor other group that did 5%
375 * or less of the IO's for the last summation just skip this
378 samples_thresh = lat_info->nr_samples * 5;
379 samples_thresh = div64_u64(samples_thresh, 100);
380 if (iolat->nr_samples <= samples_thresh)
384 /* We're as low as we can go. */
385 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
386 blkcg_use_delay(lat_to_blkg(iolat));
390 /* We're back to the default cookie, unthrottle all the things. */
391 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
392 blkcg_clear_delay(lat_to_blkg(iolat));
393 iolat->rq_depth.max_depth = UINT_MAX;
394 wake_up_all(&iolat->rq_wait.wait);
398 scale_change(iolat, direction > 0);
401 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
404 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
406 struct blkcg_gq *blkg;
407 struct request_queue *q = rqos->q;
408 bool issue_as_root = bio_issue_as_root_blkg(bio);
410 if (!blkiolat->enabled)
414 blkcg = bio_blkcg(bio);
415 bio_associate_blkcg(bio, &blkcg->css);
416 blkg = blkg_lookup(blkcg, q);
417 if (unlikely(!blkg)) {
419 spin_lock_irq(q->queue_lock);
420 blkg = blkg_lookup_create(blkcg, q);
424 spin_unlock_irq(q->queue_lock);
429 bio_issue_init(&bio->bi_issue, bio_sectors(bio));
430 bio_associate_blkg(bio, blkg);
433 while (blkg && blkg->parent) {
434 struct iolatency_grp *iolat = blkg_to_lat(blkg);
440 check_scale_change(iolat);
441 __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
442 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
445 if (!timer_pending(&blkiolat->timer))
446 mod_timer(&blkiolat->timer, jiffies + HZ);
449 static void iolatency_record_time(struct iolatency_grp *iolat,
450 struct bio_issue *issue, u64 now,
453 struct blk_rq_stat *rq_stat;
454 u64 start = bio_issue_time(issue);
458 * Have to do this so we are truncated to the correct time that our
459 * issue is truncated to.
461 now = __bio_issue_time(now);
466 req_time = now - start;
469 * We don't want to count issue_as_root bio's in the cgroups latency
470 * statistics as it could skew the numbers downwards.
472 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
473 u64 sub = iolat->min_lat_nsec;
475 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
479 rq_stat = get_cpu_ptr(iolat->stats);
480 blk_rq_stat_add(rq_stat, req_time);
481 put_cpu_ptr(rq_stat);
484 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
485 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
487 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
489 struct blkcg_gq *blkg = lat_to_blkg(iolat);
490 struct iolatency_grp *parent;
491 struct child_latency_info *lat_info;
492 struct blk_rq_stat stat;
496 blk_rq_stat_init(&stat);
498 for_each_online_cpu(cpu) {
499 struct blk_rq_stat *s;
500 s = per_cpu_ptr(iolat->stats, cpu);
501 blk_rq_stat_sum(&stat, s);
506 parent = blkg_to_lat(blkg->parent);
510 lat_info = &parent->child_lat;
513 * CALC_LOAD takes in a number stored in fixed point representation.
514 * Because we are using this for IO time in ns, the values stored
515 * are significantly larger than the FIXED_1 denominator (2048).
516 * Therefore, rounding errors in the calculation are negligible and
519 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
520 div64_u64(iolat->cur_win_nsec,
521 BLKIOLATENCY_EXP_BUCKET_SIZE));
522 CALC_LOAD(iolat->lat_avg, iolatency_exp_factors[exp_idx], stat.mean);
524 /* Everything is ok and we don't need to adjust the scale. */
525 if (stat.mean <= iolat->min_lat_nsec &&
526 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
529 /* Somebody beat us to the punch, just bail. */
530 spin_lock_irqsave(&lat_info->lock, flags);
531 lat_info->nr_samples -= iolat->nr_samples;
532 lat_info->nr_samples += stat.nr_samples;
533 iolat->nr_samples = stat.nr_samples;
535 if ((lat_info->last_scale_event >= now ||
536 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME) &&
537 lat_info->scale_lat <= iolat->min_lat_nsec)
540 if (stat.mean <= iolat->min_lat_nsec &&
541 stat.nr_samples >= BLKIOLATENCY_MIN_GOOD_SAMPLES) {
542 if (lat_info->scale_grp == iolat) {
543 lat_info->last_scale_event = now;
544 scale_cookie_change(iolat->blkiolat, lat_info, true);
546 } else if (stat.mean > iolat->min_lat_nsec) {
547 lat_info->last_scale_event = now;
548 if (!lat_info->scale_grp ||
549 lat_info->scale_lat > iolat->min_lat_nsec) {
550 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
551 lat_info->scale_grp = iolat;
553 scale_cookie_change(iolat->blkiolat, lat_info, false);
556 spin_unlock_irqrestore(&lat_info->lock, flags);
559 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
561 struct blkcg_gq *blkg;
563 struct iolatency_grp *iolat;
565 u64 now = ktime_to_ns(ktime_get());
566 bool issue_as_root = bio_issue_as_root_blkg(bio);
573 iolat = blkg_to_lat(bio->bi_blkg);
577 if (!iolat->blkiolat->enabled)
580 while (blkg && blkg->parent) {
581 iolat = blkg_to_lat(blkg);
586 rqw = &iolat->rq_wait;
588 inflight = atomic_dec_return(&rqw->inflight);
589 WARN_ON_ONCE(inflight < 0);
591 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
592 * submitted, so do not account for it.
594 if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
595 iolatency_record_time(iolat, &bio->bi_issue, now,
597 window_start = atomic64_read(&iolat->window_start);
598 if (now > window_start &&
599 (now - window_start) >= iolat->cur_win_nsec) {
600 if (atomic64_cmpxchg(&iolat->window_start,
601 window_start, now) == window_start)
602 iolatency_check_latencies(iolat, now);
610 static void blkcg_iolatency_exit(struct rq_qos *rqos)
612 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
614 del_timer_sync(&blkiolat->timer);
615 flush_work(&blkiolat->enable_work);
616 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
620 static struct rq_qos_ops blkcg_iolatency_ops = {
621 .throttle = blkcg_iolatency_throttle,
622 .done_bio = blkcg_iolatency_done_bio,
623 .exit = blkcg_iolatency_exit,
626 static void blkiolatency_timer_fn(struct timer_list *t)
628 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
629 struct blkcg_gq *blkg;
630 struct cgroup_subsys_state *pos_css;
631 u64 now = ktime_to_ns(ktime_get());
634 blkg_for_each_descendant_pre(blkg, pos_css,
635 blkiolat->rqos.q->root_blkg) {
636 struct iolatency_grp *iolat;
637 struct child_latency_info *lat_info;
642 * We could be exiting, don't access the pd unless we have a
645 if (!blkg_try_get(blkg))
648 iolat = blkg_to_lat(blkg);
652 lat_info = &iolat->child_lat;
653 cookie = atomic_read(&lat_info->scale_cookie);
655 if (cookie >= DEFAULT_SCALE_COOKIE)
658 spin_lock_irqsave(&lat_info->lock, flags);
659 if (lat_info->last_scale_event >= now)
663 * We scaled down but don't have a scale_grp, scale up and carry
666 if (lat_info->scale_grp == NULL) {
667 scale_cookie_change(iolat->blkiolat, lat_info, true);
672 * It's been 5 seconds since our last scale event, clear the
673 * scale grp in case the group that needed the scale down isn't
674 * doing any IO currently.
676 if (now - lat_info->last_scale_event >=
677 ((u64)NSEC_PER_SEC * 5))
678 lat_info->scale_grp = NULL;
680 spin_unlock_irqrestore(&lat_info->lock, flags);
688 * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
689 * @work: enable_work of the blk_iolatency of interest
691 * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
692 * is relatively expensive as it involves walking up the hierarchy twice for
693 * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
694 * want to disable the in-flight tracking.
696 * We have to make sure that the counting is balanced - we don't want to leak
697 * the in-flight counts by disabling accounting in the completion path while IOs
698 * are in flight. This is achieved by ensuring that no IO is in flight by
699 * freezing the queue while flipping ->enabled. As this requires a sleepable
700 * context, ->enabled flipping is punted to this work function.
702 static void blkiolatency_enable_work_fn(struct work_struct *work)
704 struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
709 * There can only be one instance of this function running for @blkiolat
710 * and it's guaranteed to be executed at least once after the latest
711 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
714 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
715 * in blkcg_iolatency_exit().
717 enabled = atomic_read(&blkiolat->enable_cnt);
718 if (enabled != blkiolat->enabled) {
719 blk_mq_freeze_queue(blkiolat->rqos.q);
720 blkiolat->enabled = enabled;
721 blk_mq_unfreeze_queue(blkiolat->rqos.q);
725 int blk_iolatency_init(struct request_queue *q)
727 struct blk_iolatency *blkiolat;
731 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
735 rqos = &blkiolat->rqos;
736 rqos->id = RQ_QOS_CGROUP;
737 rqos->ops = &blkcg_iolatency_ops;
742 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
749 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
750 INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
755 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
757 struct iolatency_grp *iolat = blkg_to_lat(blkg);
758 struct blk_iolatency *blkiolat = iolat->blkiolat;
759 u64 oldval = iolat->min_lat_nsec;
761 iolat->min_lat_nsec = val;
762 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
763 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
764 BLKIOLATENCY_MAX_WIN_SIZE);
766 if (!oldval && val) {
767 if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
768 schedule_work(&blkiolat->enable_work);
770 if (oldval && !val) {
771 blkcg_clear_delay(blkg);
772 if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
773 schedule_work(&blkiolat->enable_work);
777 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
780 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
781 struct child_latency_info *lat_info;
785 lat_info = &iolat->child_lat;
786 spin_lock(&lat_info->lock);
787 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
788 lat_info->last_scale_event = 0;
789 lat_info->scale_grp = NULL;
790 lat_info->scale_lat = 0;
791 spin_unlock(&lat_info->lock);
795 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
796 size_t nbytes, loff_t off)
798 struct blkcg *blkcg = css_to_blkcg(of_css(of));
799 struct blkcg_gq *blkg;
800 struct blk_iolatency *blkiolat;
801 struct blkg_conf_ctx ctx;
802 struct iolatency_grp *iolat;
808 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
812 iolat = blkg_to_lat(ctx.blkg);
813 blkiolat = iolat->blkiolat;
817 while ((tok = strsep(&p, " "))) {
819 char val[21]; /* 18446744073709551616 */
821 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
824 if (!strcmp(key, "target")) {
827 if (!strcmp(val, "max"))
829 else if (sscanf(val, "%llu", &v) == 1)
830 lat_val = v * NSEC_PER_USEC;
838 /* Walk up the tree to see if our new val is lower than it should be. */
840 oldval = iolat->min_lat_nsec;
842 iolatency_set_min_lat_nsec(blkg, lat_val);
843 if (oldval != iolat->min_lat_nsec)
844 iolatency_clear_scaling(blkg);
847 blkg_conf_finish(&ctx);
848 return ret ?: nbytes;
851 static u64 iolatency_prfill_limit(struct seq_file *sf,
852 struct blkg_policy_data *pd, int off)
854 struct iolatency_grp *iolat = pd_to_lat(pd);
855 const char *dname = blkg_dev_name(pd->blkg);
857 if (!dname || !iolat->min_lat_nsec)
859 seq_printf(sf, "%s target=%llu\n",
860 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
864 static int iolatency_print_limit(struct seq_file *sf, void *v)
866 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
867 iolatency_prfill_limit,
868 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
872 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
875 struct iolatency_grp *iolat = pd_to_lat(pd);
876 unsigned long long avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
877 unsigned long long cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
879 if (iolat->rq_depth.max_depth == UINT_MAX)
880 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
883 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
884 iolat->rq_depth.max_depth, avg_lat, cur_win);
888 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
890 struct iolatency_grp *iolat;
892 iolat = kzalloc_node(sizeof(*iolat), gfp, node);
895 iolat->stats = __alloc_percpu_gfp(sizeof(struct blk_rq_stat),
896 __alignof__(struct blk_rq_stat), gfp);
904 static void iolatency_pd_init(struct blkg_policy_data *pd)
906 struct iolatency_grp *iolat = pd_to_lat(pd);
907 struct blkcg_gq *blkg = lat_to_blkg(iolat);
908 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
909 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
910 u64 now = ktime_to_ns(ktime_get());
913 for_each_possible_cpu(cpu) {
914 struct blk_rq_stat *stat;
915 stat = per_cpu_ptr(iolat->stats, cpu);
916 blk_rq_stat_init(stat);
919 rq_wait_init(&iolat->rq_wait);
920 spin_lock_init(&iolat->child_lat.lock);
921 iolat->rq_depth.queue_depth = blk_queue_depth(blkg->q);
922 iolat->rq_depth.max_depth = UINT_MAX;
923 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
924 iolat->blkiolat = blkiolat;
925 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
926 atomic64_set(&iolat->window_start, now);
929 * We init things in list order, so the pd for the parent may not be
930 * init'ed yet for whatever reason.
932 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
933 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
934 atomic_set(&iolat->scale_cookie,
935 atomic_read(&parent->child_lat.scale_cookie));
937 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
940 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
943 static void iolatency_pd_offline(struct blkg_policy_data *pd)
945 struct iolatency_grp *iolat = pd_to_lat(pd);
946 struct blkcg_gq *blkg = lat_to_blkg(iolat);
948 iolatency_set_min_lat_nsec(blkg, 0);
949 iolatency_clear_scaling(blkg);
952 static void iolatency_pd_free(struct blkg_policy_data *pd)
954 struct iolatency_grp *iolat = pd_to_lat(pd);
955 free_percpu(iolat->stats);
959 static struct cftype iolatency_files[] = {
962 .flags = CFTYPE_NOT_ON_ROOT,
963 .seq_show = iolatency_print_limit,
964 .write = iolatency_set_limit,
969 static struct blkcg_policy blkcg_policy_iolatency = {
970 .dfl_cftypes = iolatency_files,
971 .pd_alloc_fn = iolatency_pd_alloc,
972 .pd_init_fn = iolatency_pd_init,
973 .pd_offline_fn = iolatency_pd_offline,
974 .pd_free_fn = iolatency_pd_free,
975 .pd_stat_fn = iolatency_pd_stat,
978 static int __init iolatency_init(void)
980 return blkcg_policy_register(&blkcg_policy_iolatency);
983 static void __exit iolatency_exit(void)
985 return blkcg_policy_unregister(&blkcg_policy_iolatency);
988 module_init(iolatency_init);
989 module_exit(iolatency_exit);