1/*
2 * Block rq-qos base io controller
3 *
4 * This works similar to wbt with a few exceptions
5 *
6 * - It's bio based, so the latency covers the whole block layer in addition to
7 * the actual io.
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.
15 *
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.
21 *
22 * Consider the following
23 *
24 * root blkg
25 * / \
26 * fast (target=5ms) slow (target=10ms)
27 * / \ / \
28 * a b normal(15ms) unloved
29 *
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.
34 *
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.
39 *
40 * There are 2 ways we throttle IO.
41 *
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.
46 *
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
53 *
54 * total_time += min_lat_nsec - actual_io_completion
55 *
56 * and then at throttle time will do
57 *
58 * throttle_time = min(total_time, NSEC_PER_SEC)
59 *
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.
63 *
64 * Copyright (C) 2018 Josef Bacik
65 */
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"
77#include "blk-stat.h"
78#include "blk.h"
79
80#define DEFAULT_SCALE_COOKIE 1000000U
81
82static struct blkcg_policy blkcg_policy_iolatency;
83struct iolatency_grp;
84
85struct blk_iolatency {
86 struct rq_qos rqos;
87 struct timer_list timer;
88 atomic_t enabled;
89};
90
91static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
92{
93 return container_of(rqos, struct blk_iolatency, rqos);
94}
95
96static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
97{
98 return atomic_read(&blkiolat->enabled) > 0;
99}
100
101struct child_latency_info {
102 spinlock_t lock;
103
104 /* Last time we adjusted the scale of everybody. */
105 u64 last_scale_event;
106
107 /* The latency that we missed. */
108 u64 scale_lat;
109
110 /* Total io's from all of our children for the last summation. */
111 u64 nr_samples;
112
113 /* The guy who actually changed the latency numbers. */
114 struct iolatency_grp *scale_grp;
115
116 /* Cookie to tell if we need to scale up or down. */
117 atomic_t scale_cookie;
118};
119
120struct percentile_stats {
121 u64 total;
122 u64 missed;
123};
124
125struct latency_stat {
126 union {
127 struct percentile_stats ps;
128 struct blk_rq_stat rqs;
129 };
130};
131
132struct iolatency_grp {
133 struct blkg_policy_data pd;
134 struct latency_stat __percpu *stats;
135 struct latency_stat cur_stat;
136 struct blk_iolatency *blkiolat;
137 struct rq_depth rq_depth;
138 struct rq_wait rq_wait;
139 atomic64_t window_start;
140 atomic_t scale_cookie;
141 u64 min_lat_nsec;
142 u64 cur_win_nsec;
143
144 /* total running average of our io latency. */
145 u64 lat_avg;
146
147 /* Our current number of IO's for the last summation. */
148 u64 nr_samples;
149
150 bool ssd;
151 struct child_latency_info child_lat;
152};
153
154#define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
155#define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
156/*
157 * These are the constants used to fake the fixed-point moving average
158 * calculation just like load average. The call to calc_load() folds
159 * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
160 * window size is bucketed to try to approximately calculate average
161 * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
162 * elapse immediately. Note, windows only elapse with IO activity. Idle
163 * periods extend the most recent window.
164 */
165#define BLKIOLATENCY_NR_EXP_FACTORS 5
166#define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
167 (BLKIOLATENCY_NR_EXP_FACTORS - 1))
168static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
169 2045, // exp(1/600) - 600 samples
170 2039, // exp(1/240) - 240 samples
171 2031, // exp(1/120) - 120 samples
172 2023, // exp(1/80) - 80 samples
173 2014, // exp(1/60) - 60 samples
174};
175
176static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
177{
178 return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
179}
180
181static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
182{
183 return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
184}
185
186static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
187{
188 return pd_to_blkg(&iolat->pd);
189}
190
191static inline void latency_stat_init(struct iolatency_grp *iolat,
192 struct latency_stat *stat)
193{
194 if (iolat->ssd) {
195 stat->ps.total = 0;
196 stat->ps.missed = 0;
197 } else
198 blk_rq_stat_init(&stat->rqs);
199}
200
201static inline void latency_stat_sum(struct iolatency_grp *iolat,
202 struct latency_stat *sum,
203 struct latency_stat *stat)
204{
205 if (iolat->ssd) {
206 sum->ps.total += stat->ps.total;
207 sum->ps.missed += stat->ps.missed;
208 } else
209 blk_rq_stat_sum(&sum->rqs, &stat->rqs);
210}
211
212static inline void latency_stat_record_time(struct iolatency_grp *iolat,
213 u64 req_time)
214{
215 struct latency_stat *stat = get_cpu_ptr(iolat->stats);
216 if (iolat->ssd) {
217 if (req_time >= iolat->min_lat_nsec)
218 stat->ps.missed++;
219 stat->ps.total++;
220 } else
221 blk_rq_stat_add(&stat->rqs, req_time);
222 put_cpu_ptr(stat);
223}
224
225static inline bool latency_sum_ok(struct iolatency_grp *iolat,
226 struct latency_stat *stat)
227{
228 if (iolat->ssd) {
229 u64 thresh = div64_u64(stat->ps.total, 10);
230 thresh = max(thresh, 1ULL);
231 return stat->ps.missed < thresh;
232 }
233 return stat->rqs.mean <= iolat->min_lat_nsec;
234}
235
236static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
237 struct latency_stat *stat)
238{
239 if (iolat->ssd)
240 return stat->ps.total;
241 return stat->rqs.nr_samples;
242}
243
244static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
245 struct latency_stat *stat)
246{
247 int exp_idx;
248
249 if (iolat->ssd)
250 return;
251
252 /*
253 * calc_load() takes in a number stored in fixed point representation.
254 * Because we are using this for IO time in ns, the values stored
255 * are significantly larger than the FIXED_1 denominator (2048).
256 * Therefore, rounding errors in the calculation are negligible and
257 * can be ignored.
258 */
259 exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
260 div64_u64(iolat->cur_win_nsec,
261 BLKIOLATENCY_EXP_BUCKET_SIZE));
262 iolat->lat_avg = calc_load(iolat->lat_avg,
263 iolatency_exp_factors[exp_idx],
264 stat->rqs.mean);
265}
266
267static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
268{
269 atomic_dec(&rqw->inflight);
270 wake_up(&rqw->wait);
271}
272
273static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
274{
275 struct iolatency_grp *iolat = private_data;
276 return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
277}
278
279static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
280 struct iolatency_grp *iolat,
281 bool issue_as_root,
282 bool use_memdelay)
283{
284 struct rq_wait *rqw = &iolat->rq_wait;
285 unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
286
287 if (use_delay)
288 blkcg_schedule_throttle(rqos->q, use_memdelay);
289
290 /*
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
295 * task can go away.
296 */
297 if (issue_as_root || fatal_signal_pending(current)) {
298 atomic_inc(&rqw->inflight);
299 return;
300 }
301
302 rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
303}
304
305#define SCALE_DOWN_FACTOR 2
306#define SCALE_UP_FACTOR 4
307
308static inline unsigned long scale_amount(unsigned long qd, bool up)
309{
310 return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
311}
312
313/*
314 * We scale the qd down faster than we scale up, so we need to use this helper
315 * to adjust the scale_cookie accordingly so we don't prematurely get
316 * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
317 *
318 * Each group has their own local copy of the last scale cookie they saw, so if
319 * the global scale cookie goes up or down they know which way they need to go
320 * based on their last knowledge of it.
321 */
322static void scale_cookie_change(struct blk_iolatency *blkiolat,
323 struct child_latency_info *lat_info,
324 bool up)
325{
326 unsigned long qd = blkiolat->rqos.q->nr_requests;
327 unsigned long scale = scale_amount(qd, up);
328 unsigned long old = atomic_read(&lat_info->scale_cookie);
329 unsigned long max_scale = qd << 1;
330 unsigned long diff = 0;
331
332 if (old < DEFAULT_SCALE_COOKIE)
333 diff = DEFAULT_SCALE_COOKIE - old;
334
335 if (up) {
336 if (scale + old > DEFAULT_SCALE_COOKIE)
337 atomic_set(&lat_info->scale_cookie,
338 DEFAULT_SCALE_COOKIE);
339 else if (diff > qd)
340 atomic_inc(&lat_info->scale_cookie);
341 else
342 atomic_add(scale, &lat_info->scale_cookie);
343 } else {
344 /*
345 * We don't want to dig a hole so deep that it takes us hours to
346 * dig out of it. Just enough that we don't throttle/unthrottle
347 * with jagged workloads but can still unthrottle once pressure
348 * has sufficiently dissipated.
349 */
350 if (diff > qd) {
351 if (diff < max_scale)
352 atomic_dec(&lat_info->scale_cookie);
353 } else {
354 atomic_sub(scale, &lat_info->scale_cookie);
355 }
356 }
357}
358
359/*
360 * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the
361 * queue depth at a time so we don't get wild swings and hopefully dial in to
362 * fairer distribution of the overall queue depth.
363 */
364static void scale_change(struct iolatency_grp *iolat, bool up)
365{
366 unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
367 unsigned long scale = scale_amount(qd, up);
368 unsigned long old = iolat->rq_depth.max_depth;
369
370 if (old > qd)
371 old = qd;
372
373 if (up) {
374 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
375 return;
376
377 if (old < qd) {
378 old += scale;
379 old = min(old, qd);
380 iolat->rq_depth.max_depth = old;
381 wake_up_all(&iolat->rq_wait.wait);
382 }
383 } else {
384 old >>= 1;
385 iolat->rq_depth.max_depth = max(old, 1UL);
386 }
387}
388
389/* Check our parent and see if the scale cookie has changed. */
390static void check_scale_change(struct iolatency_grp *iolat)
391{
392 struct iolatency_grp *parent;
393 struct child_latency_info *lat_info;
394 unsigned int cur_cookie;
395 unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
396 u64 scale_lat;
397 unsigned int old;
398 int direction = 0;
399
400 if (lat_to_blkg(iolat)->parent == NULL)
401 return;
402
403 parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
404 if (!parent)
405 return;
406
407 lat_info = &parent->child_lat;
408 cur_cookie = atomic_read(&lat_info->scale_cookie);
409 scale_lat = READ_ONCE(lat_info->scale_lat);
410
411 if (cur_cookie < our_cookie)
412 direction = -1;
413 else if (cur_cookie > our_cookie)
414 direction = 1;
415 else
416 return;
417
418 old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
419
420 /* Somebody beat us to the punch, just bail. */
421 if (old != our_cookie)
422 return;
423
424 if (direction < 0 && iolat->min_lat_nsec) {
425 u64 samples_thresh;
426
427 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
428 return;
429
430 /*
431 * Sometimes high priority groups are their own worst enemy, so
432 * instead of taking it out on some poor other group that did 5%
433 * or less of the IO's for the last summation just skip this
434 * scale down event.
435 */
436 samples_thresh = lat_info->nr_samples * 5;
437 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
438 if (iolat->nr_samples <= samples_thresh)
439 return;
440 }
441
442 /* We're as low as we can go. */
443 if (iolat->rq_depth.max_depth == 1 && direction < 0) {
444 blkcg_use_delay(lat_to_blkg(iolat));
445 return;
446 }
447
448 /* We're back to the default cookie, unthrottle all the things. */
449 if (cur_cookie == DEFAULT_SCALE_COOKIE) {
450 blkcg_clear_delay(lat_to_blkg(iolat));
451 iolat->rq_depth.max_depth = UINT_MAX;
452 wake_up_all(&iolat->rq_wait.wait);
453 return;
454 }
455
456 scale_change(iolat, direction > 0);
457}
458
459static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
460{
461 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
462 struct blkcg_gq *blkg = bio->bi_blkg;
463 bool issue_as_root = bio_issue_as_root_blkg(bio);
464
465 if (!blk_iolatency_enabled(blkiolat))
466 return;
467
468 while (blkg && blkg->parent) {
469 struct iolatency_grp *iolat = blkg_to_lat(blkg);
470 if (!iolat) {
471 blkg = blkg->parent;
472 continue;
473 }
474
475 check_scale_change(iolat);
476 __blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
477 (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
478 blkg = blkg->parent;
479 }
480 if (!timer_pending(&blkiolat->timer))
481 mod_timer(&blkiolat->timer, jiffies + HZ);
482}
483
484static void iolatency_record_time(struct iolatency_grp *iolat,
485 struct bio_issue *issue, u64 now,
486 bool issue_as_root)
487{
488 u64 start = bio_issue_time(issue);
489 u64 req_time;
490
491 /*
492 * Have to do this so we are truncated to the correct time that our
493 * issue is truncated to.
494 */
495 now = __bio_issue_time(now);
496
497 if (now <= start)
498 return;
499
500 req_time = now - start;
501
502 /*
503 * We don't want to count issue_as_root bio's in the cgroups latency
504 * statistics as it could skew the numbers downwards.
505 */
506 if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
507 u64 sub = iolat->min_lat_nsec;
508 if (req_time < sub)
509 blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
510 return;
511 }
512
513 latency_stat_record_time(iolat, req_time);
514}
515
516#define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
517#define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
518
519static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
520{
521 struct blkcg_gq *blkg = lat_to_blkg(iolat);
522 struct iolatency_grp *parent;
523 struct child_latency_info *lat_info;
524 struct latency_stat stat;
525 unsigned long flags;
526 int cpu;
527
528 latency_stat_init(iolat, &stat);
529 preempt_disable();
530 for_each_online_cpu(cpu) {
531 struct latency_stat *s;
532 s = per_cpu_ptr(iolat->stats, cpu);
533 latency_stat_sum(iolat, &stat, s);
534 latency_stat_init(iolat, s);
535 }
536 preempt_enable();
537
538 parent = blkg_to_lat(blkg->parent);
539 if (!parent)
540 return;
541
542 lat_info = &parent->child_lat;
543
544 iolat_update_total_lat_avg(iolat, &stat);
545
546 /* Everything is ok and we don't need to adjust the scale. */
547 if (latency_sum_ok(iolat, &stat) &&
548 atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
549 return;
550
551 /* Somebody beat us to the punch, just bail. */
552 spin_lock_irqsave(&lat_info->lock, flags);
553
554 latency_stat_sum(iolat, &iolat->cur_stat, &stat);
555 lat_info->nr_samples -= iolat->nr_samples;
556 lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
557 iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
558
559 if ((lat_info->last_scale_event >= now ||
560 now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
561 goto out;
562
563 if (latency_sum_ok(iolat, &iolat->cur_stat) &&
564 latency_sum_ok(iolat, &stat)) {
565 if (latency_stat_samples(iolat, &iolat->cur_stat) <
566 BLKIOLATENCY_MIN_GOOD_SAMPLES)
567 goto out;
568 if (lat_info->scale_grp == iolat) {
569 lat_info->last_scale_event = now;
570 scale_cookie_change(iolat->blkiolat, lat_info, true);
571 }
572 } else if (lat_info->scale_lat == 0 ||
573 lat_info->scale_lat >= iolat->min_lat_nsec) {
574 lat_info->last_scale_event = now;
575 if (!lat_info->scale_grp ||
576 lat_info->scale_lat > iolat->min_lat_nsec) {
577 WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
578 lat_info->scale_grp = iolat;
579 }
580 scale_cookie_change(iolat->blkiolat, lat_info, false);
581 }
582 latency_stat_init(iolat, &iolat->cur_stat);
583out:
584 spin_unlock_irqrestore(&lat_info->lock, flags);
585}
586
587static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
588{
589 struct blkcg_gq *blkg;
590 struct rq_wait *rqw;
591 struct iolatency_grp *iolat;
592 u64 window_start;
593 u64 now = ktime_to_ns(ktime_get());
594 bool issue_as_root = bio_issue_as_root_blkg(bio);
595 bool enabled = false;
596 int inflight = 0;
597
598 blkg = bio->bi_blkg;
599 if (!blkg || !bio_flagged(bio, BIO_TRACKED))
600 return;
601
602 iolat = blkg_to_lat(bio->bi_blkg);
603 if (!iolat)
604 return;
605
606 enabled = blk_iolatency_enabled(iolat->blkiolat);
607 if (!enabled)
608 return;
609
610 while (blkg && blkg->parent) {
611 iolat = blkg_to_lat(blkg);
612 if (!iolat) {
613 blkg = blkg->parent;
614 continue;
615 }
616 rqw = &iolat->rq_wait;
617
618 inflight = atomic_dec_return(&rqw->inflight);
619 WARN_ON_ONCE(inflight < 0);
620 if (iolat->min_lat_nsec == 0)
621 goto next;
622 iolatency_record_time(iolat, &bio->bi_issue, now,
623 issue_as_root);
624 window_start = atomic64_read(&iolat->window_start);
625 if (now > window_start &&
626 (now - window_start) >= iolat->cur_win_nsec) {
627 if (atomic64_cmpxchg(&iolat->window_start,
628 window_start, now) == window_start)
629 iolatency_check_latencies(iolat, now);
630 }
631next:
632 wake_up(&rqw->wait);
633 blkg = blkg->parent;
634 }
635}
636
637static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
638{
639 struct blkcg_gq *blkg;
640
641 blkg = bio->bi_blkg;
642 while (blkg && blkg->parent) {
643 struct rq_wait *rqw;
644 struct iolatency_grp *iolat;
645
646 iolat = blkg_to_lat(blkg);
647 if (!iolat)
648 goto next;
649
650 rqw = &iolat->rq_wait;
651 atomic_dec(&rqw->inflight);
652 wake_up(&rqw->wait);
653next:
654 blkg = blkg->parent;
655 }
656}
657
658static void blkcg_iolatency_exit(struct rq_qos *rqos)
659{
660 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
661
662 del_timer_sync(&blkiolat->timer);
663 blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
664 kfree(blkiolat);
665}
666
667static struct rq_qos_ops blkcg_iolatency_ops = {
668 .throttle = blkcg_iolatency_throttle,
669 .cleanup = blkcg_iolatency_cleanup,
670 .done_bio = blkcg_iolatency_done_bio,
671 .exit = blkcg_iolatency_exit,
672};
673
674static void blkiolatency_timer_fn(struct timer_list *t)
675{
676 struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
677 struct blkcg_gq *blkg;
678 struct cgroup_subsys_state *pos_css;
679 u64 now = ktime_to_ns(ktime_get());
680
681 rcu_read_lock();
682 blkg_for_each_descendant_pre(blkg, pos_css,
683 blkiolat->rqos.q->root_blkg) {
684 struct iolatency_grp *iolat;
685 struct child_latency_info *lat_info;
686 unsigned long flags;
687 u64 cookie;
688
689 /*
690 * We could be exiting, don't access the pd unless we have a
691 * ref on the blkg.
692 */
693 if (!blkg_tryget(blkg))
694 continue;
695
696 iolat = blkg_to_lat(blkg);
697 if (!iolat)
698 goto next;
699
700 lat_info = &iolat->child_lat;
701 cookie = atomic_read(&lat_info->scale_cookie);
702
703 if (cookie >= DEFAULT_SCALE_COOKIE)
704 goto next;
705
706 spin_lock_irqsave(&lat_info->lock, flags);
707 if (lat_info->last_scale_event >= now)
708 goto next_lock;
709
710 /*
711 * We scaled down but don't have a scale_grp, scale up and carry
712 * on.
713 */
714 if (lat_info->scale_grp == NULL) {
715 scale_cookie_change(iolat->blkiolat, lat_info, true);
716 goto next_lock;
717 }
718
719 /*
720 * It's been 5 seconds since our last scale event, clear the
721 * scale grp in case the group that needed the scale down isn't
722 * doing any IO currently.
723 */
724 if (now - lat_info->last_scale_event >=
725 ((u64)NSEC_PER_SEC * 5))
726 lat_info->scale_grp = NULL;
727next_lock:
728 spin_unlock_irqrestore(&lat_info->lock, flags);
729next:
730 blkg_put(blkg);
731 }
732 rcu_read_unlock();
733}
734
735int blk_iolatency_init(struct request_queue *q)
736{
737 struct blk_iolatency *blkiolat;
738 struct rq_qos *rqos;
739 int ret;
740
741 blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
742 if (!blkiolat)
743 return -ENOMEM;
744
745 rqos = &blkiolat->rqos;
746 rqos->id = RQ_QOS_CGROUP;
747 rqos->ops = &blkcg_iolatency_ops;
748 rqos->q = q;
749
750 rq_qos_add(q, rqos);
751
752 ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
753 if (ret) {
754 rq_qos_del(q, rqos);
755 kfree(blkiolat);
756 return ret;
757 }
758
759 timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
760
761 return 0;
762}
763
764/*
765 * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
766 * return 0.
767 */
768static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
769{
770 struct iolatency_grp *iolat = blkg_to_lat(blkg);
771 u64 oldval = iolat->min_lat_nsec;
772
773 iolat->min_lat_nsec = val;
774 iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
775 iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
776 BLKIOLATENCY_MAX_WIN_SIZE);
777
778 if (!oldval && val)
779 return 1;
780 if (oldval && !val)
781 return -1;
782 return 0;
783}
784
785static void iolatency_clear_scaling(struct blkcg_gq *blkg)
786{
787 if (blkg->parent) {
788 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
789 struct child_latency_info *lat_info;
790 if (!iolat)
791 return;
792
793 lat_info = &iolat->child_lat;
794 spin_lock(&lat_info->lock);
795 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
796 lat_info->last_scale_event = 0;
797 lat_info->scale_grp = NULL;
798 lat_info->scale_lat = 0;
799 spin_unlock(&lat_info->lock);
800 }
801}
802
803static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
804 size_t nbytes, loff_t off)
805{
806 struct blkcg *blkcg = css_to_blkcg(of_css(of));
807 struct blkcg_gq *blkg;
808 struct blkg_conf_ctx ctx;
809 struct iolatency_grp *iolat;
810 char *p, *tok;
811 u64 lat_val = 0;
812 u64 oldval;
813 int ret;
814 int enable = 0;
815
816 ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
817 if (ret)
818 return ret;
819
820 iolat = blkg_to_lat(ctx.blkg);
821 p = ctx.body;
822
823 ret = -EINVAL;
824 while ((tok = strsep(&p, " "))) {
825 char key[16];
826 char val[21]; /* 18446744073709551616 */
827
828 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
829 goto out;
830
831 if (!strcmp(key, "target")) {
832 u64 v;
833
834 if (!strcmp(val, "max"))
835 lat_val = 0;
836 else if (sscanf(val, "%llu", &v) == 1)
837 lat_val = v * NSEC_PER_USEC;
838 else
839 goto out;
840 } else {
841 goto out;
842 }
843 }
844
845 /* Walk up the tree to see if our new val is lower than it should be. */
846 blkg = ctx.blkg;
847 oldval = iolat->min_lat_nsec;
848
849 enable = iolatency_set_min_lat_nsec(blkg, lat_val);
850 if (enable) {
851 WARN_ON_ONCE(!blk_get_queue(blkg->q));
852 blkg_get(blkg);
853 }
854
855 if (oldval != iolat->min_lat_nsec) {
856 iolatency_clear_scaling(blkg);
857 }
858
859 ret = 0;
860out:
861 blkg_conf_finish(&ctx);
862 if (ret == 0 && enable) {
863 struct iolatency_grp *tmp = blkg_to_lat(blkg);
864 struct blk_iolatency *blkiolat = tmp->blkiolat;
865
866 blk_mq_freeze_queue(blkg->q);
867
868 if (enable == 1)
869 atomic_inc(&blkiolat->enabled);
870 else if (enable == -1)
871 atomic_dec(&blkiolat->enabled);
872 else
873 WARN_ON_ONCE(1);
874
875 blk_mq_unfreeze_queue(blkg->q);
876
877 blkg_put(blkg);
878 blk_put_queue(blkg->q);
879 }
880 return ret ?: nbytes;
881}
882
883static u64 iolatency_prfill_limit(struct seq_file *sf,
884 struct blkg_policy_data *pd, int off)
885{
886 struct iolatency_grp *iolat = pd_to_lat(pd);
887 const char *dname = blkg_dev_name(pd->blkg);
888
889 if (!dname || !iolat->min_lat_nsec)
890 return 0;
891 seq_printf(sf, "%s target=%llu\n",
892 dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
893 return 0;
894}
895
896static int iolatency_print_limit(struct seq_file *sf, void *v)
897{
898 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
899 iolatency_prfill_limit,
900 &blkcg_policy_iolatency, seq_cft(sf)->private, false);
901 return 0;
902}
903
904static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
905 size_t size)
906{
907 struct latency_stat stat;
908 int cpu;
909
910 latency_stat_init(iolat, &stat);
911 preempt_disable();
912 for_each_online_cpu(cpu) {
913 struct latency_stat *s;
914 s = per_cpu_ptr(iolat->stats, cpu);
915 latency_stat_sum(iolat, &stat, s);
916 }
917 preempt_enable();
918
919 if (iolat->rq_depth.max_depth == UINT_MAX)
920 return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
921 (unsigned long long)stat.ps.missed,
922 (unsigned long long)stat.ps.total);
923 return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
924 (unsigned long long)stat.ps.missed,
925 (unsigned long long)stat.ps.total,
926 iolat->rq_depth.max_depth);
927}
928
929static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
930 size_t size)
931{
932 struct iolatency_grp *iolat = pd_to_lat(pd);
933 unsigned long long avg_lat;
934 unsigned long long cur_win;
935
936 if (iolat->ssd)
937 return iolatency_ssd_stat(iolat, buf, size);
938
939 avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
940 cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
941 if (iolat->rq_depth.max_depth == UINT_MAX)
942 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
943 avg_lat, cur_win);
944
945 return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
946 iolat->rq_depth.max_depth, avg_lat, cur_win);
947}
948
949
950static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
951{
952 struct iolatency_grp *iolat;
953
954 iolat = kzalloc_node(sizeof(*iolat), gfp, node);
955 if (!iolat)
956 return NULL;
957 iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
958 __alignof__(struct latency_stat), gfp);
959 if (!iolat->stats) {
960 kfree(iolat);
961 return NULL;
962 }
963 return &iolat->pd;
964}
965
966static void iolatency_pd_init(struct blkg_policy_data *pd)
967{
968 struct iolatency_grp *iolat = pd_to_lat(pd);
969 struct blkcg_gq *blkg = lat_to_blkg(iolat);
970 struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
971 struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
972 u64 now = ktime_to_ns(ktime_get());
973 int cpu;
974
975 if (blk_queue_nonrot(blkg->q))
976 iolat->ssd = true;
977 else
978 iolat->ssd = false;
979
980 for_each_possible_cpu(cpu) {
981 struct latency_stat *stat;
982 stat = per_cpu_ptr(iolat->stats, cpu);
983 latency_stat_init(iolat, stat);
984 }
985
986 latency_stat_init(iolat, &iolat->cur_stat);
987 rq_wait_init(&iolat->rq_wait);
988 spin_lock_init(&iolat->child_lat.lock);
989 iolat->rq_depth.queue_depth = blkg->q->nr_requests;
990 iolat->rq_depth.max_depth = UINT_MAX;
991 iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
992 iolat->blkiolat = blkiolat;
993 iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
994 atomic64_set(&iolat->window_start, now);
995
996 /*
997 * We init things in list order, so the pd for the parent may not be
998 * init'ed yet for whatever reason.
999 */
1000 if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1001 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1002 atomic_set(&iolat->scale_cookie,
1003 atomic_read(&parent->child_lat.scale_cookie));
1004 } else {
1005 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1006 }
1007
1008 atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1009}
1010
1011static void iolatency_pd_offline(struct blkg_policy_data *pd)
1012{
1013 struct iolatency_grp *iolat = pd_to_lat(pd);
1014 struct blkcg_gq *blkg = lat_to_blkg(iolat);
1015 struct blk_iolatency *blkiolat = iolat->blkiolat;
1016 int ret;
1017
1018 ret = iolatency_set_min_lat_nsec(blkg, 0);
1019 if (ret == 1)
1020 atomic_inc(&blkiolat->enabled);
1021 if (ret == -1)
1022 atomic_dec(&blkiolat->enabled);
1023 iolatency_clear_scaling(blkg);
1024}
1025
1026static void iolatency_pd_free(struct blkg_policy_data *pd)
1027{
1028 struct iolatency_grp *iolat = pd_to_lat(pd);
1029 free_percpu(iolat->stats);
1030 kfree(iolat);
1031}
1032
1033static struct cftype iolatency_files[] = {
1034 {
1035 .name = "latency",
1036 .flags = CFTYPE_NOT_ON_ROOT,
1037 .seq_show = iolatency_print_limit,
1038 .write = iolatency_set_limit,
1039 },
1040 {}
1041};
1042
1043static struct blkcg_policy blkcg_policy_iolatency = {
1044 .dfl_cftypes = iolatency_files,
1045 .pd_alloc_fn = iolatency_pd_alloc,
1046 .pd_init_fn = iolatency_pd_init,
1047 .pd_offline_fn = iolatency_pd_offline,
1048 .pd_free_fn = iolatency_pd_free,
1049 .pd_stat_fn = iolatency_pd_stat,
1050};
1051
1052static int __init iolatency_init(void)
1053{
1054 return blkcg_policy_register(&blkcg_policy_iolatency);
1055}
1056
1057static void __exit iolatency_exit(void)
1058{
1059 return blkcg_policy_unregister(&blkcg_policy_iolatency);
1060}
1061
1062module_init(iolatency_init);
1063module_exit(iolatency_exit);
1064