block/blk-iolatency.c

   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
  82 static struct blkcg_policy blkcg_policy_iolatency;
  83 struct iolatency_grp;
  84
  85 struct blk_iolatency {
  86         struct rq_qos rqos;
  87         struct timer_list timer;
  88         atomic_t enabled;
  89 };
  90
  91 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
  92 {
  93         return container_of(rqos, struct blk_iolatency, rqos);
  94 }
  95
  96 static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
  97 {
  98         return atomic_read(&blkiolat->enabled) > 0;
  99 }
 100
 101 struct 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
 120 struct iolatency_grp {
 121         struct blkg_policy_data pd;
 122         struct blk_rq_stat __percpu *stats;
 123         struct blk_iolatency *blkiolat;
 124         struct rq_depth rq_depth;
 125         struct rq_wait rq_wait;
 126         atomic64_t window_start;
 127         atomic_t scale_cookie;
 128         u64 min_lat_nsec;
 129         u64 cur_win_nsec;
 130
 131         /* total running average of our io latency. */
 132         u64 lat_avg;
 133
 134         /* Our current number of IO's for the last summation. */
 135         u64 nr_samples;
 136
 137         struct child_latency_info child_lat;
 138 };
 139
 140 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
 141 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
 142 /*
 143  * These are the constants used to fake the fixed-point moving average
 144  * calculation just like load average.  The call to CALC_LOAD folds
 145  * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg.  The sampling
 146  * window size is bucketed to try to approximately calculate average
 147  * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
 148  * elapse immediately.  Note, windows only elapse with IO activity.  Idle
 149  * periods extend the most recent window.
 150  */
 151 #define BLKIOLATENCY_NR_EXP_FACTORS 5
 152 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
 153                                       (BLKIOLATENCY_NR_EXP_FACTORS - 1))
 154 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
 155         2045, // exp(1/600) - 600 samples
 156         2039, // exp(1/240) - 240 samples
 157         2031, // exp(1/120) - 120 samples
 158         2023, // exp(1/80)  - 80 samples
 159         2014, // exp(1/60)  - 60 samples
 160 };
 161
 162 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
 163 {
 164         return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
 165 }
 166
 167 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
 168 {
 169         return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
 170 }
 171
 172 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
 173 {
 174         return pd_to_blkg(&iolat->pd);
 175 }
 176
 177 static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
 178                                        wait_queue_entry_t *wait,
 179                                        bool first_block)
 180 {
 181         struct rq_wait *rqw = &iolat->rq_wait;
 182
 183         if (first_block && waitqueue_active(&rqw->wait) &&
 184             rqw->wait.head.next != &wait->entry)
 185                 return false;
 186         return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
 187 }
 188
 189 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
 190                                        struct iolatency_grp *iolat,
 191                                        spinlock_t *lock, bool issue_as_root,
 192                                        bool use_memdelay)
 193         __releases(lock)
 194         __acquires(lock)
 195 {
 196         struct rq_wait *rqw = &iolat->rq_wait;
 197         unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
 198         DEFINE_WAIT(wait);
 199         bool first_block = true;
 200
 201         if (use_delay)
 202                 blkcg_schedule_throttle(rqos->q, use_memdelay);
 203
 204         /*
 205          * To avoid priority inversions we want to just take a slot if we are
 206          * issuing as root.  If we're being killed off there's no point in
 207          * delaying things, we may have been killed by OOM so throttling may
 208          * make recovery take even longer, so just let the IO's through so the
 209          * task can go away.
 210          */
 211         if (issue_as_root || fatal_signal_pending(current)) {
 212                 atomic_inc(&rqw->inflight);
 213                 return;
 214         }
 215
 216         if (iolatency_may_queue(iolat, &wait, first_block))
 217                 return;
 218
 219         do {
 220                 prepare_to_wait_exclusive(&rqw->wait, &wait,
 221                                           TASK_UNINTERRUPTIBLE);
 222
 223                 if (iolatency_may_queue(iolat, &wait, first_block))
 224                         break;
 225                 first_block = false;
 226
 227                 if (lock) {
 228                         spin_unlock_irq(lock);
 229                         io_schedule();
 230                         spin_lock_irq(lock);
 231                 } else {
 232                         io_schedule();
 233                 }
 234         } while (1);
 235
 236         finish_wait(&rqw->wait, &wait);
 237 }
 238
 239 #define SCALE_DOWN_FACTOR 2
 240 #define SCALE_UP_FACTOR 4
 241
 242 static inline unsigned long scale_amount(unsigned long qd, bool up)
 243 {
 244         return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
 245 }
 246
 247 /*
 248  * We scale the qd down faster than we scale up, so we need to use this helper
 249  * to adjust the scale_cookie accordingly so we don't prematurely get
 250  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
 251  *
 252  * Each group has their own local copy of the last scale cookie they saw, so if
 253  * the global scale cookie goes up or down they know which way they need to go
 254  * based on their last knowledge of it.
 255  */
 256 static void scale_cookie_change(struct blk_iolatency *blkiolat,
 257                                 struct child_latency_info *lat_info,
 258                                 bool up)
 259 {
 260         unsigned long qd = blk_queue_depth(blkiolat->rqos.q);
 261         unsigned long scale = scale_amount(qd, up);
 262         unsigned long old = atomic_read(&lat_info->scale_cookie);
 263         unsigned long max_scale = qd << 1;
 264         unsigned long diff = 0;
 265
 266         if (old < DEFAULT_SCALE_COOKIE)
 267                 diff = DEFAULT_SCALE_COOKIE - old;
 268
 269         if (up) {
 270                 if (scale + old > DEFAULT_SCALE_COOKIE)
 271                         atomic_set(&lat_info->scale_cookie,
 272                                    DEFAULT_SCALE_COOKIE);
 273                 else if (diff > qd)
 274                         atomic_inc(&lat_info->scale_cookie);
 275                 else
 276                         atomic_add(scale, &lat_info->scale_cookie);
 277         } else {
 278                 /*
 279                  * We don't want to dig a hole so deep that it takes us hours to
 280                  * dig out of it.  Just enough that we don't throttle/unthrottle
 281                  * with jagged workloads but can still unthrottle once pressure
 282                  * has sufficiently dissipated.
 283                  */
 284                 if (diff > qd) {
 285                         if (diff < max_scale)
 286                                 atomic_dec(&lat_info->scale_cookie);
 287                 } else {
 288                         atomic_sub(scale, &lat_info->scale_cookie);
 289                 }
 290         }
 291 }
 292
 293 /*
 294  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
 295  * queue depth at a time so we don't get wild swings and hopefully dial in to
 296  * fairer distribution of the overall queue depth.
 297  */
 298 static void scale_change(struct iolatency_grp *iolat, bool up)
 299 {
 300         unsigned long qd = blk_queue_depth(iolat->blkiolat->rqos.q);
 301         unsigned long scale = scale_amount(qd, up);
 302         unsigned long old = iolat->rq_depth.max_depth;
 303         bool changed = false;
 304
 305         if (old > qd)
 306                 old = qd;
 307
 308         if (up) {
 309                 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
 310                         return;
 311
 312                 if (old < qd) {
 313                         changed = true;
 314                         old += scale;
 315                         old = min(old, qd);
 316                         iolat->rq_depth.max_depth = old;
 317                         wake_up_all(&iolat->rq_wait.wait);
 318                 }
 319         } else if (old > 1) {
 320                 old >>= 1;
 321                 changed = true;
 322                 iolat->rq_depth.max_depth = max(old, 1UL);
 323         }
 324 }
 325
 326 /* Check our parent and see if the scale cookie has changed. */
 327 static void check_scale_change(struct iolatency_grp *iolat)
 328 {
 329         struct iolatency_grp *parent;
 330         struct child_latency_info *lat_info;
 331         unsigned int cur_cookie;
 332         unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
 333         u64 scale_lat;
 334         unsigned int old;
 335         int direction = 0;
 336
 337         if (lat_to_blkg(iolat)->parent == NULL)
 338                 return;
 339
 340         parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
 341         if (!parent)
 342                 return;
 343
 344         lat_info = &parent->child_lat;
 345         cur_cookie = atomic_read(&lat_info->scale_cookie);
 346         scale_lat = READ_ONCE(lat_info->scale_lat);
 347
 348         if (cur_cookie < our_cookie)
 349                 direction = -1;
 350         else if (cur_cookie > our_cookie)
 351                 direction = 1;
 352         else
 353                 return;
 354
 355         old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
 356
 357         /* Somebody beat us to the punch, just bail. */
 358         if (old != our_cookie)
 359                 return;
 360
 361         if (direction < 0 && iolat->min_lat_nsec) {
 362                 u64 samples_thresh;
 363
 364                 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
 365                         return;
 366
 367                 /*
 368                  * Sometimes high priority groups are their own worst enemy, so
 369                  * instead of taking it out on some poor other group that did 5%
 370                  * or less of the IO's for the last summation just skip this
 371                  * scale down event.
 372                  */
 373                 samples_thresh = lat_info->nr_samples * 5;
 374                 samples_thresh = div64_u64(samples_thresh, 100);
 375                 if (iolat->nr_samples <= samples_thresh)
 376                         return;
 377         }
 378
 379         /* We're as low as we can go. */
 380         if (iolat->rq_depth.max_depth == 1 && direction < 0) {
 381                 blkcg_use_delay(lat_to_blkg(iolat));
 382                 return;
 383         }
 384
 385         /* We're back to the default cookie, unthrottle all the things. */
 386         if (cur_cookie == DEFAULT_SCALE_COOKIE) {
 387                 blkcg_clear_delay(lat_to_blkg(iolat));
 388                 iolat->rq_depth.max_depth = UINT_MAX;
 389                 wake_up_all(&iolat->rq_wait.wait);
 390                 return;
 391         }
 392
 393         scale_change(iolat, direction > 0);
 394 }
 395
 396 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
 397                                      spinlock_t *lock)
 398 {
 399         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 400         struct blkcg *blkcg;
 401         struct blkcg_gq *blkg;
 402         struct request_queue *q = rqos->q;
 403         bool issue_as_root = bio_issue_as_root_blkg(bio);
 404
 405         if (!blk_iolatency_enabled(blkiolat))
 406                 return;
 407
 408         rcu_read_lock();
 409         blkcg = bio_blkcg(bio);
 410         bio_associate_blkcg(bio, &blkcg->css);
 411         blkg = blkg_lookup(blkcg, q);
 412         if (unlikely(!blkg)) {
 413                 if (!lock)
 414                         spin_lock_irq(q->queue_lock);
 415                 blkg = blkg_lookup_create(blkcg, q);
 416                 if (IS_ERR(blkg))
 417                         blkg = NULL;
 418                 if (!lock)
 419                         spin_unlock_irq(q->queue_lock);
 420         }
 421         if (!blkg)
 422                 goto out;
 423
 424         bio_issue_init(&bio->bi_issue, bio_sectors(bio));
 425         bio_associate_blkg(bio, blkg);
 426 out:
 427         rcu_read_unlock();
 428         while (blkg && blkg->parent) {
 429                 struct iolatency_grp *iolat = blkg_to_lat(blkg);
 430                 if (!iolat) {
 431                         blkg = blkg->parent;
 432                         continue;
 433                 }
 434
 435                 check_scale_change(iolat);
 436                 __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
 437                                      (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
 438                 blkg = blkg->parent;
 439         }
 440         if (!timer_pending(&blkiolat->timer))
 441                 mod_timer(&blkiolat->timer, jiffies + HZ);
 442 }
 443
 444 static void iolatency_record_time(struct iolatency_grp *iolat,
 445                                   struct bio_issue *issue, u64 now,
 446                                   bool issue_as_root)
 447 {
 448         struct blk_rq_stat *rq_stat;
 449         u64 start = bio_issue_time(issue);
 450         u64 req_time;
 451
 452         /*
 453          * Have to do this so we are truncated to the correct time that our
 454          * issue is truncated to.
 455          */
 456         now = __bio_issue_time(now);
 457
 458         if (now <= start)
 459                 return;
 460
 461         req_time = now - start;
 462
 463         /*
 464          * We don't want to count issue_as_root bio's in the cgroups latency
 465          * statistics as it could skew the numbers downwards.
 466          */
 467         if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
 468                 u64 sub = iolat->min_lat_nsec;
 469                 if (req_time < sub)
 470                         blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
 471                 return;
 472         }
 473
 474         rq_stat = get_cpu_ptr(iolat->stats);
 475         blk_rq_stat_add(rq_stat, req_time);
 476         put_cpu_ptr(rq_stat);
 477 }
 478
 479 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
 480 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
 481
 482 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
 483 {
 484         struct blkcg_gq *blkg = lat_to_blkg(iolat);
 485         struct iolatency_grp *parent;
 486         struct child_latency_info *lat_info;
 487         struct blk_rq_stat stat;
 488         unsigned long flags;
 489         int cpu, exp_idx;
 490
 491         blk_rq_stat_init(&stat);
 492         preempt_disable();
 493         for_each_online_cpu(cpu) {
 494                 struct blk_rq_stat *s;
 495                 s = per_cpu_ptr(iolat->stats, cpu);
 496                 blk_rq_stat_sum(&stat, s);
 497                 blk_rq_stat_init(s);
 498         }
 499         preempt_enable();
 500
 501         parent = blkg_to_lat(blkg->parent);
 502         if (!parent)
 503                 return;
 504
 505         lat_info = &parent->child_lat;
 506
 507         /*
 508          * CALC_LOAD takes in a number stored in fixed point representation.
 509          * Because we are using this for IO time in ns, the values stored
 510          * are significantly larger than the FIXED_1 denominator (2048).
 511          * Therefore, rounding errors in the calculation are negligible and
 512          * can be ignored.
 513          */
 514         exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
 515                         div64_u64(iolat->cur_win_nsec,
 516                                   BLKIOLATENCY_EXP_BUCKET_SIZE));
 517         CALC_LOAD(iolat->lat_avg, iolatency_exp_factors[exp_idx], stat.mean);
 518
 519         /* Everything is ok and we don't need to adjust the scale. */
 520         if (stat.mean <= iolat->min_lat_nsec &&
 521             atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
 522                 return;
 523
 524         /* Somebody beat us to the punch, just bail. */
 525         spin_lock_irqsave(&lat_info->lock, flags);
 526         lat_info->nr_samples -= iolat->nr_samples;
 527         lat_info->nr_samples += stat.nr_samples;
 528         iolat->nr_samples = stat.nr_samples;
 529
 530         if ((lat_info->last_scale_event >= now ||
 531             now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME) &&
 532             lat_info->scale_lat <= iolat->min_lat_nsec)
 533                 goto out;
 534
 535         if (stat.mean <= iolat->min_lat_nsec &&
 536             stat.nr_samples >= BLKIOLATENCY_MIN_GOOD_SAMPLES) {
 537                 if (lat_info->scale_grp == iolat) {
 538                         lat_info->last_scale_event = now;
 539                         scale_cookie_change(iolat->blkiolat, lat_info, true);
 540                 }
 541         } else if (stat.mean > iolat->min_lat_nsec) {
 542                 lat_info->last_scale_event = now;
 543                 if (!lat_info->scale_grp ||
 544                     lat_info->scale_lat > iolat->min_lat_nsec) {
 545                         WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
 546                         lat_info->scale_grp = iolat;
 547                 }
 548                 scale_cookie_change(iolat->blkiolat, lat_info, false);
 549         }
 550 out:
 551         spin_unlock_irqrestore(&lat_info->lock, flags);
 552 }
 553
 554 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
 555 {
 556         struct blkcg_gq *blkg;
 557         struct rq_wait *rqw;
 558         struct iolatency_grp *iolat;
 559         u64 window_start;
 560         u64 now = ktime_to_ns(ktime_get());
 561         bool issue_as_root = bio_issue_as_root_blkg(bio);
 562         bool enabled = false;
 563
 564         blkg = bio->bi_blkg;
 565         if (!blkg)
 566                 return;
 567
 568         iolat = blkg_to_lat(bio->bi_blkg);
 569         if (!iolat)
 570                 return;
 571
 572         enabled = blk_iolatency_enabled(iolat->blkiolat);
 573         if (!enabled)
 574                 return;
 575
 576         while (blkg && blkg->parent) {
 577                 iolat = blkg_to_lat(blkg);
 578                 if (!iolat) {
 579                         blkg = blkg->parent;
 580                         continue;
 581                 }
 582                 rqw = &iolat->rq_wait;
 583
 584                 atomic_dec(&rqw->inflight);
 585                 if (iolat->min_lat_nsec == 0)
 586                         goto next;
 587                 iolatency_record_time(iolat, &bio->bi_issue, now,
 588                                       issue_as_root);
 589                 window_start = atomic64_read(&iolat->window_start);
 590                 if (now > window_start &&
 591                     (now - window_start) >= iolat->cur_win_nsec) {
 592                         if (atomic64_cmpxchg(&iolat->window_start,
 593                                         window_start, now) == window_start)
 594                                 iolatency_check_latencies(iolat, now);
 595                 }
 596 next:
 597                 wake_up(&rqw->wait);
 598                 blkg = blkg->parent;
 599         }
 600 }
 601
 602 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
 603 {
 604         struct blkcg_gq *blkg;
 605
 606         blkg = bio->bi_blkg;
 607         while (blkg && blkg->parent) {
 608                 struct rq_wait *rqw;
 609                 struct iolatency_grp *iolat;
 610
 611                 iolat = blkg_to_lat(blkg);
 612                 if (!iolat)
 613                         goto next;
 614
 615                 rqw = &iolat->rq_wait;
 616                 atomic_dec(&rqw->inflight);
 617                 wake_up(&rqw->wait);
 618 next:
 619                 blkg = blkg->parent;
 620         }
 621 }
 622
 623 static void blkcg_iolatency_exit(struct rq_qos *rqos)
 624 {
 625         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 626
 627         del_timer_sync(&blkiolat->timer);
 628         blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
 629         kfree(blkiolat);
 630 }
 631
 632 static struct rq_qos_ops blkcg_iolatency_ops = {
 633         .throttle = blkcg_iolatency_throttle,
 634         .cleanup = blkcg_iolatency_cleanup,
 635         .done_bio = blkcg_iolatency_done_bio,
 636         .exit = blkcg_iolatency_exit,
 637 };
 638
 639 static void blkiolatency_timer_fn(struct timer_list *t)
 640 {
 641         struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
 642         struct blkcg_gq *blkg;
 643         struct cgroup_subsys_state *pos_css;
 644         u64 now = ktime_to_ns(ktime_get());
 645
 646         rcu_read_lock();
 647         blkg_for_each_descendant_pre(blkg, pos_css,
 648                                      blkiolat->rqos.q->root_blkg) {
 649                 struct iolatency_grp *iolat;
 650                 struct child_latency_info *lat_info;
 651                 unsigned long flags;
 652                 u64 cookie;
 653
 654                 /*
 655                  * We could be exiting, don't access the pd unless we have a
 656                  * ref on the blkg.
 657                  */
 658                 if (!blkg_try_get(blkg))
 659                         continue;
 660
 661                 iolat = blkg_to_lat(blkg);
 662                 if (!iolat)
 663                         goto next;
 664
 665                 lat_info = &iolat->child_lat;
 666                 cookie = atomic_read(&lat_info->scale_cookie);
 667
 668                 if (cookie >= DEFAULT_SCALE_COOKIE)
 669                         goto next;
 670
 671                 spin_lock_irqsave(&lat_info->lock, flags);
 672                 if (lat_info->last_scale_event >= now)
 673                         goto next_lock;
 674
 675                 /*
 676                  * We scaled down but don't have a scale_grp, scale up and carry
 677                  * on.
 678                  */
 679                 if (lat_info->scale_grp == NULL) {
 680                         scale_cookie_change(iolat->blkiolat, lat_info, true);
 681                         goto next_lock;
 682                 }
 683
 684                 /*
 685                  * It's been 5 seconds since our last scale event, clear the
 686                  * scale grp in case the group that needed the scale down isn't
 687                  * doing any IO currently.
 688                  */
 689                 if (now - lat_info->last_scale_event >=
 690                     ((u64)NSEC_PER_SEC * 5))
 691                         lat_info->scale_grp = NULL;
 692 next_lock:
 693                 spin_unlock_irqrestore(&lat_info->lock, flags);
 694 next:
 695                 blkg_put(blkg);
 696         }
 697         rcu_read_unlock();
 698 }
 699
 700 int blk_iolatency_init(struct request_queue *q)
 701 {
 702         struct blk_iolatency *blkiolat;
 703         struct rq_qos *rqos;
 704         int ret;
 705
 706         blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
 707         if (!blkiolat)
 708                 return -ENOMEM;
 709
 710         rqos = &blkiolat->rqos;
 711         rqos->id = RQ_QOS_CGROUP;
 712         rqos->ops = &blkcg_iolatency_ops;
 713         rqos->q = q;
 714
 715         rq_qos_add(q, rqos);
 716
 717         ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
 718         if (ret) {
 719                 rq_qos_del(q, rqos);
 720                 kfree(blkiolat);
 721                 return ret;
 722         }
 723
 724         timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
 725
 726         return 0;
 727 }
 728
 729 /*
 730  * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
 731  * return 0.
 732  */
 733 static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
 734 {
 735         struct iolatency_grp *iolat = blkg_to_lat(blkg);
 736         u64 oldval = iolat->min_lat_nsec;
 737
 738         iolat->min_lat_nsec = val;
 739         iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
 740         iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
 741                                     BLKIOLATENCY_MAX_WIN_SIZE);
 742
 743         if (!oldval && val)
 744                 return 1;
 745         if (oldval && !val)
 746                 return -1;
 747         return 0;
 748 }
 749
 750 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
 751 {
 752         if (blkg->parent) {
 753                 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
 754                 struct child_latency_info *lat_info;
 755                 if (!iolat)
 756                         return;
 757
 758                 lat_info = &iolat->child_lat;
 759                 spin_lock(&lat_info->lock);
 760                 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
 761                 lat_info->last_scale_event = 0;
 762                 lat_info->scale_grp = NULL;
 763                 lat_info->scale_lat = 0;
 764                 spin_unlock(&lat_info->lock);
 765         }
 766 }
 767
 768 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
 769                              size_t nbytes, loff_t off)
 770 {
 771         struct blkcg *blkcg = css_to_blkcg(of_css(of));
 772         struct blkcg_gq *blkg;
 773         struct blk_iolatency *blkiolat;
 774         struct blkg_conf_ctx ctx;
 775         struct iolatency_grp *iolat;
 776         char *p, *tok;
 777         u64 lat_val = 0;
 778         u64 oldval;
 779         int ret;
 780         int enable = 0;
 781
 782         ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
 783         if (ret)
 784                 return ret;
 785
 786         iolat = blkg_to_lat(ctx.blkg);
 787         blkiolat = iolat->blkiolat;
 788         p = ctx.body;
 789
 790         ret = -EINVAL;
 791         while ((tok = strsep(&p, " "))) {
 792                 char key[16];
 793                 char val[21];   /* 18446744073709551616 */
 794
 795                 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
 796                         goto out;
 797
 798                 if (!strcmp(key, "target")) {
 799                         u64 v;
 800
 801                         if (!strcmp(val, "max"))
 802                                 lat_val = 0;
 803                         else if (sscanf(val, "%llu", &v) == 1)
 804                                 lat_val = v * NSEC_PER_USEC;
 805                         else
 806                                 goto out;
 807                 } else {
 808                         goto out;
 809                 }
 810         }
 811
 812         /* Walk up the tree to see if our new val is lower than it should be. */
 813         blkg = ctx.blkg;
 814         oldval = iolat->min_lat_nsec;
 815
 816         enable = iolatency_set_min_lat_nsec(blkg, lat_val);
 817         if (enable) {
 818                 WARN_ON_ONCE(!blk_get_queue(blkg->q));
 819                 blkg_get(blkg);
 820         }
 821
 822         if (oldval != iolat->min_lat_nsec) {
 823                 iolatency_clear_scaling(blkg);
 824         }
 825
 826         ret = 0;
 827 out:
 828         blkg_conf_finish(&ctx);
 829         if (ret == 0 && enable) {
 830                 struct iolatency_grp *tmp = blkg_to_lat(blkg);
 831                 struct blk_iolatency *blkiolat = tmp->blkiolat;
 832
 833                 blk_mq_freeze_queue(blkg->q);
 834
 835                 if (enable == 1)
 836                         atomic_inc(&blkiolat->enabled);
 837                 else if (enable == -1)
 838                         atomic_dec(&blkiolat->enabled);
 839                 else
 840                         WARN_ON_ONCE(1);
 841
 842                 blk_mq_unfreeze_queue(blkg->q);
 843
 844                 blkg_put(blkg);
 845                 blk_put_queue(blkg->q);
 846         }
 847         return ret ?: nbytes;
 848 }
 849
 850 static u64 iolatency_prfill_limit(struct seq_file *sf,
 851                                   struct blkg_policy_data *pd, int off)
 852 {
 853         struct iolatency_grp *iolat = pd_to_lat(pd);
 854         const char *dname = blkg_dev_name(pd->blkg);
 855
 856         if (!dname || !iolat->min_lat_nsec)
 857                 return 0;
 858         seq_printf(sf, "%s target=%llu\n",
 859                    dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
 860         return 0;
 861 }
 862
 863 static int iolatency_print_limit(struct seq_file *sf, void *v)
 864 {
 865         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 866                           iolatency_prfill_limit,
 867                           &blkcg_policy_iolatency, seq_cft(sf)->private, false);
 868         return 0;
 869 }
 870
 871 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
 872                                 size_t size)
 873 {
 874         struct iolatency_grp *iolat = pd_to_lat(pd);
 875         unsigned long long avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
 876         unsigned long long cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
 877
 878         if (iolat->rq_depth.max_depth == UINT_MAX)
 879                 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
 880                                  avg_lat, cur_win);
 881
 882         return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
 883                          iolat->rq_depth.max_depth, avg_lat, cur_win);
 884 }
 885
 886
 887 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
 888 {
 889         struct iolatency_grp *iolat;
 890
 891         iolat = kzalloc_node(sizeof(*iolat), gfp, node);
 892         if (!iolat)
 893                 return NULL;
 894         iolat->stats = __alloc_percpu_gfp(sizeof(struct blk_rq_stat),
 895                                        __alignof__(struct blk_rq_stat), gfp);
 896         if (!iolat->stats) {
 897                 kfree(iolat);
 898                 return NULL;
 899         }
 900         return &iolat->pd;
 901 }
 902
 903 static void iolatency_pd_init(struct blkg_policy_data *pd)
 904 {
 905         struct iolatency_grp *iolat = pd_to_lat(pd);
 906         struct blkcg_gq *blkg = lat_to_blkg(iolat);
 907         struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
 908         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 909         u64 now = ktime_to_ns(ktime_get());
 910         int cpu;
 911
 912         for_each_possible_cpu(cpu) {
 913                 struct blk_rq_stat *stat;
 914                 stat = per_cpu_ptr(iolat->stats, cpu);
 915                 blk_rq_stat_init(stat);
 916         }
 917
 918         rq_wait_init(&iolat->rq_wait);
 919         spin_lock_init(&iolat->child_lat.lock);
 920         iolat->rq_depth.queue_depth = blk_queue_depth(blkg->q);
 921         iolat->rq_depth.max_depth = UINT_MAX;
 922         iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
 923         iolat->blkiolat = blkiolat;
 924         iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
 925         atomic64_set(&iolat->window_start, now);
 926
 927         /*
 928          * We init things in list order, so the pd for the parent may not be
 929          * init'ed yet for whatever reason.
 930          */
 931         if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
 932                 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
 933                 atomic_set(&iolat->scale_cookie,
 934                            atomic_read(&parent->child_lat.scale_cookie));
 935         } else {
 936                 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
 937         }
 938
 939         atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
 940 }
 941
 942 static void iolatency_pd_offline(struct blkg_policy_data *pd)
 943 {
 944         struct iolatency_grp *iolat = pd_to_lat(pd);
 945         struct blkcg_gq *blkg = lat_to_blkg(iolat);
 946         struct blk_iolatency *blkiolat = iolat->blkiolat;
 947         int ret;
 948
 949         ret = iolatency_set_min_lat_nsec(blkg, 0);
 950         if (ret == 1)
 951                 atomic_inc(&blkiolat->enabled);
 952         if (ret == -1)
 953                 atomic_dec(&blkiolat->enabled);
 954         iolatency_clear_scaling(blkg);
 955 }
 956
 957 static void iolatency_pd_free(struct blkg_policy_data *pd)
 958 {
 959         struct iolatency_grp *iolat = pd_to_lat(pd);
 960         free_percpu(iolat->stats);
 961         kfree(iolat);
 962 }
 963
 964 static struct cftype iolatency_files[] = {
 965         {
 966                 .name = "latency",
 967                 .flags = CFTYPE_NOT_ON_ROOT,
 968                 .seq_show = iolatency_print_limit,
 969                 .write = iolatency_set_limit,
 970         },
 971         {}
 972 };
 973
 974 static struct blkcg_policy blkcg_policy_iolatency = {
 975         .dfl_cftypes    = iolatency_files,
 976         .pd_alloc_fn    = iolatency_pd_alloc,
 977         .pd_init_fn     = iolatency_pd_init,
 978         .pd_offline_fn  = iolatency_pd_offline,
 979         .pd_free_fn     = iolatency_pd_free,
 980         .pd_stat_fn     = iolatency_pd_stat,
 981 };
 982
 983 static int __init iolatency_init(void)
 984 {
 985         return blkcg_policy_register(&blkcg_policy_iolatency);
 986 }
 987
 988 static void __exit iolatency_exit(void)
 989 {
 990         return blkcg_policy_unregister(&blkcg_policy_iolatency);
 991 }
 992
 993 module_init(iolatency_init);
 994 module_exit(iolatency_exit);