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                                        spinlock_t *lock, bool issue_as_root,
 280                                        bool use_memdelay)
 281         __releases(lock)
 282         __acquires(lock)
 283 {
 284         struct rq_wait *rqw = &iolat->rq_wait;
 285         unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
 286         DEFINE_WAIT(wait);
 287         bool first_block = true;
 288
 289         if (use_delay)
 290                 blkcg_schedule_throttle(rqos->q, use_memdelay);
 291
 292         /*
 293          * To avoid priority inversions we want to just take a slot if we are
 294          * issuing as root.  If we're being killed off there's no point in
 295          * delaying things, we may have been killed by OOM so throttling may
 296          * make recovery take even longer, so just let the IO's through so the
 297          * task can go away.
 298          */
 299         if (issue_as_root || fatal_signal_pending(current)) {
 300                 atomic_inc(&rqw->inflight);
 301                 return;
 302         }
 303
 304         if (iolatency_may_queue(iolat, &wait, first_block))
 305                 return;
 306
 307         do {
 308                 prepare_to_wait_exclusive(&rqw->wait, &wait,
 309                                           TASK_UNINTERRUPTIBLE);
 310
 311                 if (iolatency_may_queue(iolat, &wait, first_block))
 312                         break;
 313                 first_block = false;
 314
 315                 if (lock) {
 316                         spin_unlock_irq(lock);
 317                         io_schedule();
 318                         spin_lock_irq(lock);
 319                 } else {
 320                         io_schedule();
 321                 }
 322         } while (1);
 323
 324         finish_wait(&rqw->wait, &wait);
 325 }
 326
 327 #define SCALE_DOWN_FACTOR 2
 328 #define SCALE_UP_FACTOR 4
 329
 330 static inline unsigned long scale_amount(unsigned long qd, bool up)
 331 {
 332         return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
 333 }
 334
 335 /*
 336  * We scale the qd down faster than we scale up, so we need to use this helper
 337  * to adjust the scale_cookie accordingly so we don't prematurely get
 338  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
 339  *
 340  * Each group has their own local copy of the last scale cookie they saw, so if
 341  * the global scale cookie goes up or down they know which way they need to go
 342  * based on their last knowledge of it.
 343  */
 344 static void scale_cookie_change(struct blk_iolatency *blkiolat,
 345                                 struct child_latency_info *lat_info,
 346                                 bool up)
 347 {
 348         unsigned long qd = blkiolat->rqos.q->nr_requests;
 349         unsigned long scale = scale_amount(qd, up);
 350         unsigned long old = atomic_read(&lat_info->scale_cookie);
 351         unsigned long max_scale = qd << 1;
 352         unsigned long diff = 0;
 353
 354         if (old < DEFAULT_SCALE_COOKIE)
 355                 diff = DEFAULT_SCALE_COOKIE - old;
 356
 357         if (up) {
 358                 if (scale + old > DEFAULT_SCALE_COOKIE)
 359                         atomic_set(&lat_info->scale_cookie,
 360                                    DEFAULT_SCALE_COOKIE);
 361                 else if (diff > qd)
 362                         atomic_inc(&lat_info->scale_cookie);
 363                 else
 364                         atomic_add(scale, &lat_info->scale_cookie);
 365         } else {
 366                 /*
 367                  * We don't want to dig a hole so deep that it takes us hours to
 368                  * dig out of it.  Just enough that we don't throttle/unthrottle
 369                  * with jagged workloads but can still unthrottle once pressure
 370                  * has sufficiently dissipated.
 371                  */
 372                 if (diff > qd) {
 373                         if (diff < max_scale)
 374                                 atomic_dec(&lat_info->scale_cookie);
 375                 } else {
 376                         atomic_sub(scale, &lat_info->scale_cookie);
 377                 }
 378         }
 379 }
 380
 381 /*
 382  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
 383  * queue depth at a time so we don't get wild swings and hopefully dial in to
 384  * fairer distribution of the overall queue depth.
 385  */
 386 static void scale_change(struct iolatency_grp *iolat, bool up)
 387 {
 388         unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
 389         unsigned long scale = scale_amount(qd, up);
 390         unsigned long old = iolat->rq_depth.max_depth;
 391
 392         if (old > qd)
 393                 old = qd;
 394
 395         if (up) {
 396                 if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
 397                         return;
 398
 399                 if (old < qd) {
 400                         old += scale;
 401                         old = min(old, qd);
 402                         iolat->rq_depth.max_depth = old;
 403                         wake_up_all(&iolat->rq_wait.wait);
 404                 }
 405         } else {
 406                 old >>= 1;
 407                 iolat->rq_depth.max_depth = max(old, 1UL);
 408         }
 409 }
 410
 411 /* Check our parent and see if the scale cookie has changed. */
 412 static void check_scale_change(struct iolatency_grp *iolat)
 413 {
 414         struct iolatency_grp *parent;
 415         struct child_latency_info *lat_info;
 416         unsigned int cur_cookie;
 417         unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
 418         u64 scale_lat;
 419         unsigned int old;
 420         int direction = 0;
 421
 422         if (lat_to_blkg(iolat)->parent == NULL)
 423                 return;
 424
 425         parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
 426         if (!parent)
 427                 return;
 428
 429         lat_info = &parent->child_lat;
 430         cur_cookie = atomic_read(&lat_info->scale_cookie);
 431         scale_lat = READ_ONCE(lat_info->scale_lat);
 432
 433         if (cur_cookie < our_cookie)
 434                 direction = -1;
 435         else if (cur_cookie > our_cookie)
 436                 direction = 1;
 437         else
 438                 return;
 439
 440         old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
 441
 442         /* Somebody beat us to the punch, just bail. */
 443         if (old != our_cookie)
 444                 return;
 445
 446         if (direction < 0 && iolat->min_lat_nsec) {
 447                 u64 samples_thresh;
 448
 449                 if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
 450                         return;
 451
 452                 /*
 453                  * Sometimes high priority groups are their own worst enemy, so
 454                  * instead of taking it out on some poor other group that did 5%
 455                  * or less of the IO's for the last summation just skip this
 456                  * scale down event.
 457                  */
 458                 samples_thresh = lat_info->nr_samples * 5;
 459                 samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
 460                 if (iolat->nr_samples <= samples_thresh)
 461                         return;
 462         }
 463
 464         /* We're as low as we can go. */
 465         if (iolat->rq_depth.max_depth == 1 && direction < 0) {
 466                 blkcg_use_delay(lat_to_blkg(iolat));
 467                 return;
 468         }
 469
 470         /* We're back to the default cookie, unthrottle all the things. */
 471         if (cur_cookie == DEFAULT_SCALE_COOKIE) {
 472                 blkcg_clear_delay(lat_to_blkg(iolat));
 473                 iolat->rq_depth.max_depth = UINT_MAX;
 474                 wake_up_all(&iolat->rq_wait.wait);
 475                 return;
 476         }
 477
 478         scale_change(iolat, direction > 0);
 479 }
 480
 481 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
 482                                      spinlock_t *lock)
 483 {
 484         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 485         struct blkcg_gq *blkg = bio->bi_blkg;
 486         bool issue_as_root = bio_issue_as_root_blkg(bio);
 487
 488         if (!blk_iolatency_enabled(blkiolat))
 489                 return;
 490
 491         while (blkg && blkg->parent) {
 492                 struct iolatency_grp *iolat = blkg_to_lat(blkg);
 493                 if (!iolat) {
 494                         blkg = blkg->parent;
 495                         continue;
 496                 }
 497
 498                 check_scale_change(iolat);
 499                 __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
 500                                      (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
 501                 blkg = blkg->parent;
 502         }
 503         if (!timer_pending(&blkiolat->timer))
 504                 mod_timer(&blkiolat->timer, jiffies + HZ);
 505 }
 506
 507 static void iolatency_record_time(struct iolatency_grp *iolat,
 508                                   struct bio_issue *issue, u64 now,
 509                                   bool issue_as_root)
 510 {
 511         u64 start = bio_issue_time(issue);
 512         u64 req_time;
 513
 514         /*
 515          * Have to do this so we are truncated to the correct time that our
 516          * issue is truncated to.
 517          */
 518         now = __bio_issue_time(now);
 519
 520         if (now <= start)
 521                 return;
 522
 523         req_time = now - start;
 524
 525         /*
 526          * We don't want to count issue_as_root bio's in the cgroups latency
 527          * statistics as it could skew the numbers downwards.
 528          */
 529         if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
 530                 u64 sub = iolat->min_lat_nsec;
 531                 if (req_time < sub)
 532                         blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
 533                 return;
 534         }
 535
 536         latency_stat_record_time(iolat, req_time);
 537 }
 538
 539 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
 540 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
 541
 542 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
 543 {
 544         struct blkcg_gq *blkg = lat_to_blkg(iolat);
 545         struct iolatency_grp *parent;
 546         struct child_latency_info *lat_info;
 547         struct latency_stat stat;
 548         unsigned long flags;
 549         int cpu;
 550
 551         latency_stat_init(iolat, &stat);
 552         preempt_disable();
 553         for_each_online_cpu(cpu) {
 554                 struct latency_stat *s;
 555                 s = per_cpu_ptr(iolat->stats, cpu);
 556                 latency_stat_sum(iolat, &stat, s);
 557                 latency_stat_init(iolat, s);
 558         }
 559         preempt_enable();
 560
 561         parent = blkg_to_lat(blkg->parent);
 562         if (!parent)
 563                 return;
 564
 565         lat_info = &parent->child_lat;
 566
 567         iolat_update_total_lat_avg(iolat, &stat);
 568
 569         /* Everything is ok and we don't need to adjust the scale. */
 570         if (latency_sum_ok(iolat, &stat) &&
 571             atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
 572                 return;
 573
 574         /* Somebody beat us to the punch, just bail. */
 575         spin_lock_irqsave(&lat_info->lock, flags);
 576
 577         latency_stat_sum(iolat, &iolat->cur_stat, &stat);
 578         lat_info->nr_samples -= iolat->nr_samples;
 579         lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
 580         iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
 581
 582         if ((lat_info->last_scale_event >= now ||
 583             now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
 584                 goto out;
 585
 586         if (latency_sum_ok(iolat, &iolat->cur_stat) &&
 587             latency_sum_ok(iolat, &stat)) {
 588                 if (latency_stat_samples(iolat, &iolat->cur_stat) <
 589                     BLKIOLATENCY_MIN_GOOD_SAMPLES)
 590                         goto out;
 591                 if (lat_info->scale_grp == iolat) {
 592                         lat_info->last_scale_event = now;
 593                         scale_cookie_change(iolat->blkiolat, lat_info, true);
 594                 }
 595         } else if (lat_info->scale_lat == 0 ||
 596                    lat_info->scale_lat >= iolat->min_lat_nsec) {
 597                 lat_info->last_scale_event = now;
 598                 if (!lat_info->scale_grp ||
 599                     lat_info->scale_lat > iolat->min_lat_nsec) {
 600                         WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
 601                         lat_info->scale_grp = iolat;
 602                 }
 603                 scale_cookie_change(iolat->blkiolat, lat_info, false);
 604         }
 605         latency_stat_init(iolat, &iolat->cur_stat);
 606 out:
 607         spin_unlock_irqrestore(&lat_info->lock, flags);
 608 }
 609
 610 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
 611 {
 612         struct blkcg_gq *blkg;
 613         struct rq_wait *rqw;
 614         struct iolatency_grp *iolat;
 615         u64 window_start;
 616         u64 now = ktime_to_ns(ktime_get());
 617         bool issue_as_root = bio_issue_as_root_blkg(bio);
 618         bool enabled = false;
 619
 620         blkg = bio->bi_blkg;
 621         if (!blkg)
 622                 return;
 623
 624         iolat = blkg_to_lat(bio->bi_blkg);
 625         if (!iolat)
 626                 return;
 627
 628         enabled = blk_iolatency_enabled(iolat->blkiolat);
 629         while (blkg && blkg->parent) {
 630                 iolat = blkg_to_lat(blkg);
 631                 if (!iolat) {
 632                         blkg = blkg->parent;
 633                         continue;
 634                 }
 635                 rqw = &iolat->rq_wait;
 636
 637                 atomic_dec(&rqw->inflight);
 638                 if (!enabled || iolat->min_lat_nsec == 0)
 639                         goto next;
 640                 iolatency_record_time(iolat, &bio->bi_issue, now,
 641                                       issue_as_root);
 642                 window_start = atomic64_read(&iolat->window_start);
 643                 if (now > window_start &&
 644                     (now - window_start) >= iolat->cur_win_nsec) {
 645                         if (atomic64_cmpxchg(&iolat->window_start,
 646                                         window_start, now) == window_start)
 647                                 iolatency_check_latencies(iolat, now);
 648                 }
 649 next:
 650                 wake_up(&rqw->wait);
 651                 blkg = blkg->parent;
 652         }
 653 }
 654
 655 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
 656 {
 657         struct blkcg_gq *blkg;
 658
 659         blkg = bio->bi_blkg;
 660         while (blkg && blkg->parent) {
 661                 struct rq_wait *rqw;
 662                 struct iolatency_grp *iolat;
 663
 664                 iolat = blkg_to_lat(blkg);
 665                 if (!iolat)
 666                         goto next;
 667
 668                 rqw = &iolat->rq_wait;
 669                 atomic_dec(&rqw->inflight);
 670                 wake_up(&rqw->wait);
 671 next:
 672                 blkg = blkg->parent;
 673         }
 674 }
 675
 676 static void blkcg_iolatency_exit(struct rq_qos *rqos)
 677 {
 678         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 679
 680         del_timer_sync(&blkiolat->timer);
 681         blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
 682         kfree(blkiolat);
 683 }
 684
 685 static struct rq_qos_ops blkcg_iolatency_ops = {
 686         .throttle = blkcg_iolatency_throttle,
 687         .cleanup = blkcg_iolatency_cleanup,
 688         .done_bio = blkcg_iolatency_done_bio,
 689         .exit = blkcg_iolatency_exit,
 690 };
 691
 692 static void blkiolatency_timer_fn(struct timer_list *t)
 693 {
 694         struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
 695         struct blkcg_gq *blkg;
 696         struct cgroup_subsys_state *pos_css;
 697         u64 now = ktime_to_ns(ktime_get());
 698
 699         rcu_read_lock();
 700         blkg_for_each_descendant_pre(blkg, pos_css,
 701                                      blkiolat->rqos.q->root_blkg) {
 702                 struct iolatency_grp *iolat;
 703                 struct child_latency_info *lat_info;
 704                 unsigned long flags;
 705                 u64 cookie;
 706
 707                 /*
 708                  * We could be exiting, don't access the pd unless we have a
 709                  * ref on the blkg.
 710                  */
 711                 if (!blkg_tryget(blkg))
 712                         continue;
 713
 714                 iolat = blkg_to_lat(blkg);
 715                 if (!iolat)
 716                         goto next;
 717
 718                 lat_info = &iolat->child_lat;
 719                 cookie = atomic_read(&lat_info->scale_cookie);
 720
 721                 if (cookie >= DEFAULT_SCALE_COOKIE)
 722                         goto next;
 723
 724                 spin_lock_irqsave(&lat_info->lock, flags);
 725                 if (lat_info->last_scale_event >= now)
 726                         goto next_lock;
 727
 728                 /*
 729                  * We scaled down but don't have a scale_grp, scale up and carry
 730                  * on.
 731                  */
 732                 if (lat_info->scale_grp == NULL) {
 733                         scale_cookie_change(iolat->blkiolat, lat_info, true);
 734                         goto next_lock;
 735                 }
 736
 737                 /*
 738                  * It's been 5 seconds since our last scale event, clear the
 739                  * scale grp in case the group that needed the scale down isn't
 740                  * doing any IO currently.
 741                  */
 742                 if (now - lat_info->last_scale_event >=
 743                     ((u64)NSEC_PER_SEC * 5))
 744                         lat_info->scale_grp = NULL;
 745 next_lock:
 746                 spin_unlock_irqrestore(&lat_info->lock, flags);
 747 next:
 748                 blkg_put(blkg);
 749         }
 750         rcu_read_unlock();
 751 }
 752
 753 int blk_iolatency_init(struct request_queue *q)
 754 {
 755         struct blk_iolatency *blkiolat;
 756         struct rq_qos *rqos;
 757         int ret;
 758
 759         blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
 760         if (!blkiolat)
 761                 return -ENOMEM;
 762
 763         rqos = &blkiolat->rqos;
 764         rqos->id = RQ_QOS_CGROUP;
 765         rqos->ops = &blkcg_iolatency_ops;
 766         rqos->q = q;
 767
 768         rq_qos_add(q, rqos);
 769
 770         ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
 771         if (ret) {
 772                 rq_qos_del(q, rqos);
 773                 kfree(blkiolat);
 774                 return ret;
 775         }
 776
 777         timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
 778
 779         return 0;
 780 }
 781
 782 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
 783 {
 784         struct iolatency_grp *iolat = blkg_to_lat(blkg);
 785         struct blk_iolatency *blkiolat = iolat->blkiolat;
 786         u64 oldval = iolat->min_lat_nsec;
 787
 788         iolat->min_lat_nsec = val;
 789         iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
 790         iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
 791                                     BLKIOLATENCY_MAX_WIN_SIZE);
 792
 793         if (!oldval && val)
 794                 atomic_inc(&blkiolat->enabled);
 795         if (oldval && !val)
 796                 atomic_dec(&blkiolat->enabled);
 797 }
 798
 799 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
 800 {
 801         if (blkg->parent) {
 802                 struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
 803                 struct child_latency_info *lat_info;
 804                 if (!iolat)
 805                         return;
 806
 807                 lat_info = &iolat->child_lat;
 808                 spin_lock(&lat_info->lock);
 809                 atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
 810                 lat_info->last_scale_event = 0;
 811                 lat_info->scale_grp = NULL;
 812                 lat_info->scale_lat = 0;
 813                 spin_unlock(&lat_info->lock);
 814         }
 815 }
 816
 817 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
 818                              size_t nbytes, loff_t off)
 819 {
 820         struct blkcg *blkcg = css_to_blkcg(of_css(of));
 821         struct blkcg_gq *blkg;
 822         struct blkg_conf_ctx ctx;
 823         struct iolatency_grp *iolat;
 824         char *p, *tok;
 825         u64 lat_val = 0;
 826         u64 oldval;
 827         int ret;
 828
 829         ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
 830         if (ret)
 831                 return ret;
 832
 833         iolat = blkg_to_lat(ctx.blkg);
 834         p = ctx.body;
 835
 836         ret = -EINVAL;
 837         while ((tok = strsep(&p, " "))) {
 838                 char key[16];
 839                 char val[21];   /* 18446744073709551616 */
 840
 841                 if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
 842                         goto out;
 843
 844                 if (!strcmp(key, "target")) {
 845                         u64 v;
 846
 847                         if (!strcmp(val, "max"))
 848                                 lat_val = 0;
 849                         else if (sscanf(val, "%llu", &v) == 1)
 850                                 lat_val = v * NSEC_PER_USEC;
 851                         else
 852                                 goto out;
 853                 } else {
 854                         goto out;
 855                 }
 856         }
 857
 858         /* Walk up the tree to see if our new val is lower than it should be. */
 859         blkg = ctx.blkg;
 860         oldval = iolat->min_lat_nsec;
 861
 862         iolatency_set_min_lat_nsec(blkg, lat_val);
 863         if (oldval != iolat->min_lat_nsec) {
 864                 iolatency_clear_scaling(blkg);
 865         }
 866
 867         ret = 0;
 868 out:
 869         blkg_conf_finish(&ctx);
 870         return ret ?: nbytes;
 871 }
 872
 873 static u64 iolatency_prfill_limit(struct seq_file *sf,
 874                                   struct blkg_policy_data *pd, int off)
 875 {
 876         struct iolatency_grp *iolat = pd_to_lat(pd);
 877         const char *dname = blkg_dev_name(pd->blkg);
 878
 879         if (!dname || !iolat->min_lat_nsec)
 880                 return 0;
 881         seq_printf(sf, "%s target=%llu\n",
 882                    dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
 883         return 0;
 884 }
 885
 886 static int iolatency_print_limit(struct seq_file *sf, void *v)
 887 {
 888         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 889                           iolatency_prfill_limit,
 890                           &blkcg_policy_iolatency, seq_cft(sf)->private, false);
 891         return 0;
 892 }
 893
 894 static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
 895                                  size_t size)
 896 {
 897         struct latency_stat stat;
 898         int cpu;
 899
 900         latency_stat_init(iolat, &stat);
 901         preempt_disable();
 902         for_each_online_cpu(cpu) {
 903                 struct latency_stat *s;
 904                 s = per_cpu_ptr(iolat->stats, cpu);
 905                 latency_stat_sum(iolat, &stat, s);
 906         }
 907         preempt_enable();
 908
 909         if (iolat->rq_depth.max_depth == UINT_MAX)
 910                 return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
 911                                  (unsigned long long)stat.ps.missed,
 912                                  (unsigned long long)stat.ps.total);
 913         return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
 914                          (unsigned long long)stat.ps.missed,
 915                          (unsigned long long)stat.ps.total,
 916                          iolat->rq_depth.max_depth);
 917 }
 918
 919 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
 920                                 size_t size)
 921 {
 922         struct iolatency_grp *iolat = pd_to_lat(pd);
 923         unsigned long long avg_lat;
 924         unsigned long long cur_win;
 925
 926         if (iolat->ssd)
 927                 return iolatency_ssd_stat(iolat, buf, size);
 928
 929         avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
 930         cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
 931         if (iolat->rq_depth.max_depth == UINT_MAX)
 932                 return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
 933                                  avg_lat, cur_win);
 934
 935         return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
 936                          iolat->rq_depth.max_depth, avg_lat, cur_win);
 937 }
 938
 939
 940 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
 941 {
 942         struct iolatency_grp *iolat;
 943
 944         iolat = kzalloc_node(sizeof(*iolat), gfp, node);
 945         if (!iolat)
 946                 return NULL;
 947         iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
 948                                        __alignof__(struct latency_stat), gfp);
 949         if (!iolat->stats) {
 950                 kfree(iolat);
 951                 return NULL;
 952         }
 953         return &iolat->pd;
 954 }
 955
 956 static void iolatency_pd_init(struct blkg_policy_data *pd)
 957 {
 958         struct iolatency_grp *iolat = pd_to_lat(pd);
 959         struct blkcg_gq *blkg = lat_to_blkg(iolat);
 960         struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
 961         struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
 962         u64 now = ktime_to_ns(ktime_get());
 963         int cpu;
 964
 965         if (blk_queue_nonrot(blkg->q))
 966                 iolat->ssd = true;
 967         else
 968                 iolat->ssd = false;
 969
 970         for_each_possible_cpu(cpu) {
 971                 struct latency_stat *stat;
 972                 stat = per_cpu_ptr(iolat->stats, cpu);
 973                 latency_stat_init(iolat, stat);
 974         }
 975
 976         latency_stat_init(iolat, &iolat->cur_stat);
 977         rq_wait_init(&iolat->rq_wait);
 978         spin_lock_init(&iolat->child_lat.lock);
 979         iolat->rq_depth.queue_depth = blkg->q->nr_requests;
 980         iolat->rq_depth.max_depth = UINT_MAX;
 981         iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
 982         iolat->blkiolat = blkiolat;
 983         iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
 984         atomic64_set(&iolat->window_start, now);
 985
 986         /*
 987          * We init things in list order, so the pd for the parent may not be
 988          * init'ed yet for whatever reason.
 989          */
 990         if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
 991                 struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
 992                 atomic_set(&iolat->scale_cookie,
 993                            atomic_read(&parent->child_lat.scale_cookie));
 994         } else {
 995                 atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
 996         }
 997
 998         atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
 999 }
1000
1001 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1002 {
1003         struct iolatency_grp *iolat = pd_to_lat(pd);
1004         struct blkcg_gq *blkg = lat_to_blkg(iolat);
1005
1006         iolatency_set_min_lat_nsec(blkg, 0);
1007         iolatency_clear_scaling(blkg);
1008 }
1009
1010 static void iolatency_pd_free(struct blkg_policy_data *pd)
1011 {
1012         struct iolatency_grp *iolat = pd_to_lat(pd);
1013         free_percpu(iolat->stats);
1014         kfree(iolat);
1015 }
1016
1017 static struct cftype iolatency_files[] = {
1018         {
1019                 .name = "latency",
1020                 .flags = CFTYPE_NOT_ON_ROOT,
1021                 .seq_show = iolatency_print_limit,
1022                 .write = iolatency_set_limit,
1023         },
1024         {}
1025 };
1026
1027 static struct blkcg_policy blkcg_policy_iolatency = {
1028         .dfl_cftypes    = iolatency_files,
1029         .pd_alloc_fn    = iolatency_pd_alloc,
1030         .pd_init_fn     = iolatency_pd_init,
1031         .pd_offline_fn  = iolatency_pd_offline,
1032         .pd_free_fn     = iolatency_pd_free,
1033         .pd_stat_fn     = iolatency_pd_stat,
1034 };
1035
1036 static int __init iolatency_init(void)
1037 {
1038         return blkcg_policy_register(&blkcg_policy_iolatency);
1039 }
1040
1041 static void __exit iolatency_exit(void)
1042 {
1043         return blkcg_policy_unregister(&blkcg_policy_iolatency);
1044 }
1045
1046 module_init(iolatency_init);
1047 module_exit(iolatency_exit);