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