block/blk-throttle.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Interface for controlling IO bandwidth on a request queue
   4  *
   5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blktrace_api.h>
  13 #include <linux/blk-cgroup.h>
  14 #include "blk.h"
  15
  16 /* Max dispatch from a group in 1 round */
  17 static int throtl_grp_quantum = 8;
  18
  19 /* Total max dispatch from all groups in one round */
  20 static int throtl_quantum = 32;
  21
  22 /* Throttling is performed over a slice and after that slice is renewed */
  23 #define DFL_THROTL_SLICE_HD (HZ / 10)
  24 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  25 #define MAX_THROTL_SLICE (HZ)
  26 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  27 #define MIN_THROTL_BPS (320 * 1024)
  28 #define MIN_THROTL_IOPS (10)
  29 #define DFL_LATENCY_TARGET (-1L)
  30 #define DFL_IDLE_THRESHOLD (0)
  31 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  32 #define LATENCY_FILTERED_SSD (0)
  33 /*
  34  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  35  * help determine if its IO is impacted by others, hence we ignore the IO
  36  */
  37 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  38
  39 static struct blkcg_policy blkcg_policy_throtl;
  40
  41 /* A workqueue to queue throttle related work */
  42 static struct workqueue_struct *kthrotld_workqueue;
  43
  44 /*
  45  * To implement hierarchical throttling, throtl_grps form a tree and bios
  46  * are dispatched upwards level by level until they reach the top and get
  47  * issued.  When dispatching bios from the children and local group at each
  48  * level, if the bios are dispatched into a single bio_list, there's a risk
  49  * of a local or child group which can queue many bios at once filling up
  50  * the list starving others.
  51  *
  52  * To avoid such starvation, dispatched bios are queued separately
  53  * according to where they came from.  When they are again dispatched to
  54  * the parent, they're popped in round-robin order so that no single source
  55  * hogs the dispatch window.
  56  *
  57  * throtl_qnode is used to keep the queued bios separated by their sources.
  58  * Bios are queued to throtl_qnode which in turn is queued to
  59  * throtl_service_queue and then dispatched in round-robin order.
  60  *
  61  * It's also used to track the reference counts on blkg's.  A qnode always
  62  * belongs to a throtl_grp and gets queued on itself or the parent, so
  63  * incrementing the reference of the associated throtl_grp when a qnode is
  64  * queued and decrementing when dequeued is enough to keep the whole blkg
  65  * tree pinned while bios are in flight.
  66  */
  67 struct throtl_qnode {
  68         struct list_head        node;           /* service_queue->queued[] */
  69         struct bio_list         bios;           /* queued bios */
  70         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  71 };
  72
  73 struct throtl_service_queue {
  74         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  75
  76         /*
  77          * Bios queued directly to this service_queue or dispatched from
  78          * children throtl_grp's.
  79          */
  80         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  81         unsigned int            nr_queued[2];   /* number of queued bios */
  82
  83         /*
  84          * RB tree of active children throtl_grp's, which are sorted by
  85          * their ->disptime.
  86          */
  87         struct rb_root_cached   pending_tree;   /* RB tree of active tgs */
  88         unsigned int            nr_pending;     /* # queued in the tree */
  89         unsigned long           first_pending_disptime; /* disptime of the first tg */
  90         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  91 };
  92
  93 enum tg_state_flags {
  94         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  95         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  96 };
  97
  98 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  99
 100 enum {
 101         LIMIT_LOW,
 102         LIMIT_MAX,
 103         LIMIT_CNT,
 104 };
 105
 106 struct throtl_grp {
 107         /* must be the first member */
 108         struct blkg_policy_data pd;
 109
 110         /* active throtl group service_queue member */
 111         struct rb_node rb_node;
 112
 113         /* throtl_data this group belongs to */
 114         struct throtl_data *td;
 115
 116         /* this group's service queue */
 117         struct throtl_service_queue service_queue;
 118
 119         /*
 120          * qnode_on_self is used when bios are directly queued to this
 121          * throtl_grp so that local bios compete fairly with bios
 122          * dispatched from children.  qnode_on_parent is used when bios are
 123          * dispatched from this throtl_grp into its parent and will compete
 124          * with the sibling qnode_on_parents and the parent's
 125          * qnode_on_self.
 126          */
 127         struct throtl_qnode qnode_on_self[2];
 128         struct throtl_qnode qnode_on_parent[2];
 129
 130         /*
 131          * Dispatch time in jiffies. This is the estimated time when group
 132          * will unthrottle and is ready to dispatch more bio. It is used as
 133          * key to sort active groups in service tree.
 134          */
 135         unsigned long disptime;
 136
 137         unsigned int flags;
 138
 139         /* are there any throtl rules between this group and td? */
 140         bool has_rules[2];
 141
 142         /* internally used bytes per second rate limits */
 143         uint64_t bps[2][LIMIT_CNT];
 144         /* user configured bps limits */
 145         uint64_t bps_conf[2][LIMIT_CNT];
 146
 147         /* internally used IOPS limits */
 148         unsigned int iops[2][LIMIT_CNT];
 149         /* user configured IOPS limits */
 150         unsigned int iops_conf[2][LIMIT_CNT];
 151
 152         /* Number of bytes disptached in current slice */
 153         uint64_t bytes_disp[2];
 154         /* Number of bio's dispatched in current slice */
 155         unsigned int io_disp[2];
 156
 157         unsigned long last_low_overflow_time[2];
 158
 159         uint64_t last_bytes_disp[2];
 160         unsigned int last_io_disp[2];
 161
 162         unsigned long last_check_time;
 163
 164         unsigned long latency_target; /* us */
 165         unsigned long latency_target_conf; /* us */
 166         /* When did we start a new slice */
 167         unsigned long slice_start[2];
 168         unsigned long slice_end[2];
 169
 170         unsigned long last_finish_time; /* ns / 1024 */
 171         unsigned long checked_last_finish_time; /* ns / 1024 */
 172         unsigned long avg_idletime; /* ns / 1024 */
 173         unsigned long idletime_threshold; /* us */
 174         unsigned long idletime_threshold_conf; /* us */
 175
 176         unsigned int bio_cnt; /* total bios */
 177         unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
 178         unsigned long bio_cnt_reset_time;
 179 };
 180
 181 /* We measure latency for request size from <= 4k to >= 1M */
 182 #define LATENCY_BUCKET_SIZE 9
 183
 184 struct latency_bucket {
 185         unsigned long total_latency; /* ns / 1024 */
 186         int samples;
 187 };
 188
 189 struct avg_latency_bucket {
 190         unsigned long latency; /* ns / 1024 */
 191         bool valid;
 192 };
 193
 194 struct throtl_data
 195 {
 196         /* service tree for active throtl groups */
 197         struct throtl_service_queue service_queue;
 198
 199         struct request_queue *queue;
 200
 201         /* Total Number of queued bios on READ and WRITE lists */
 202         unsigned int nr_queued[2];
 203
 204         unsigned int throtl_slice;
 205
 206         /* Work for dispatching throttled bios */
 207         struct work_struct dispatch_work;
 208         unsigned int limit_index;
 209         bool limit_valid[LIMIT_CNT];
 210
 211         unsigned long low_upgrade_time;
 212         unsigned long low_downgrade_time;
 213
 214         unsigned int scale;
 215
 216         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
 217         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
 218         struct latency_bucket __percpu *latency_buckets[2];
 219         unsigned long last_calculate_time;
 220         unsigned long filtered_latency;
 221
 222         bool track_bio_latency;
 223 };
 224
 225 static void throtl_pending_timer_fn(struct timer_list *t);
 226
 227 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 228 {
 229         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 230 }
 231
 232 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 233 {
 234         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 235 }
 236
 237 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 238 {
 239         return pd_to_blkg(&tg->pd);
 240 }
 241
 242 /**
 243  * sq_to_tg - return the throl_grp the specified service queue belongs to
 244  * @sq: the throtl_service_queue of interest
 245  *
 246  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 247  * embedded in throtl_data, %NULL is returned.
 248  */
 249 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 250 {
 251         if (sq && sq->parent_sq)
 252                 return container_of(sq, struct throtl_grp, service_queue);
 253         else
 254                 return NULL;
 255 }
 256
 257 /**
 258  * sq_to_td - return throtl_data the specified service queue belongs to
 259  * @sq: the throtl_service_queue of interest
 260  *
 261  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 262  * Determine the associated throtl_data accordingly and return it.
 263  */
 264 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 265 {
 266         struct throtl_grp *tg = sq_to_tg(sq);
 267
 268         if (tg)
 269                 return tg->td;
 270         else
 271                 return container_of(sq, struct throtl_data, service_queue);
 272 }
 273
 274 /*
 275  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 276  * make the IO dispatch more smooth.
 277  * Scale up: linearly scale up according to lapsed time since upgrade. For
 278  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 279  *           limit hits .max limit
 280  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 281  */
 282 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 283 {
 284         /* arbitrary value to avoid too big scale */
 285         if (td->scale < 4096 && time_after_eq(jiffies,
 286             td->low_upgrade_time + td->scale * td->throtl_slice))
 287                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 288
 289         return low + (low >> 1) * td->scale;
 290 }
 291
 292 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 293 {
 294         struct blkcg_gq *blkg = tg_to_blkg(tg);
 295         struct throtl_data *td;
 296         uint64_t ret;
 297
 298         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 299                 return U64_MAX;
 300
 301         td = tg->td;
 302         ret = tg->bps[rw][td->limit_index];
 303         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 304                 /* intermediate node or iops isn't 0 */
 305                 if (!list_empty(&blkg->blkcg->css.children) ||
 306                     tg->iops[rw][td->limit_index])
 307                         return U64_MAX;
 308                 else
 309                         return MIN_THROTL_BPS;
 310         }
 311
 312         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 313             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 314                 uint64_t adjusted;
 315
 316                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 317                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 318         }
 319         return ret;
 320 }
 321
 322 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 323 {
 324         struct blkcg_gq *blkg = tg_to_blkg(tg);
 325         struct throtl_data *td;
 326         unsigned int ret;
 327
 328         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 329                 return UINT_MAX;
 330
 331         td = tg->td;
 332         ret = tg->iops[rw][td->limit_index];
 333         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 334                 /* intermediate node or bps isn't 0 */
 335                 if (!list_empty(&blkg->blkcg->css.children) ||
 336                     tg->bps[rw][td->limit_index])
 337                         return UINT_MAX;
 338                 else
 339                         return MIN_THROTL_IOPS;
 340         }
 341
 342         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 343             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 344                 uint64_t adjusted;
 345
 346                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 347                 if (adjusted > UINT_MAX)
 348                         adjusted = UINT_MAX;
 349                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 350         }
 351         return ret;
 352 }
 353
 354 #define request_bucket_index(sectors) \
 355         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 356
 357 /**
 358  * throtl_log - log debug message via blktrace
 359  * @sq: the service_queue being reported
 360  * @fmt: printf format string
 361  * @args: printf args
 362  *
 363  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 364  * throtl_grp; otherwise, just "throtl".
 365  */
 366 #define throtl_log(sq, fmt, args...)    do {                            \
 367         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 368         struct throtl_data *__td = sq_to_td((sq));                      \
 369                                                                         \
 370         (void)__td;                                                     \
 371         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 372                 break;                                                  \
 373         if ((__tg)) {                                                   \
 374                 blk_add_cgroup_trace_msg(__td->queue,                   \
 375                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 376         } else {                                                        \
 377                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 378         }                                                               \
 379 } while (0)
 380
 381 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 382 {
 383         /* assume it's one sector */
 384         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 385                 return 512;
 386         return bio->bi_iter.bi_size;
 387 }
 388
 389 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 390 {
 391         INIT_LIST_HEAD(&qn->node);
 392         bio_list_init(&qn->bios);
 393         qn->tg = tg;
 394 }
 395
 396 /**
 397  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 398  * @bio: bio being added
 399  * @qn: qnode to add bio to
 400  * @queued: the service_queue->queued[] list @qn belongs to
 401  *
 402  * Add @bio to @qn and put @qn on @queued if it's not already on.
 403  * @qn->tg's reference count is bumped when @qn is activated.  See the
 404  * comment on top of throtl_qnode definition for details.
 405  */
 406 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 407                                  struct list_head *queued)
 408 {
 409         bio_list_add(&qn->bios, bio);
 410         if (list_empty(&qn->node)) {
 411                 list_add_tail(&qn->node, queued);
 412                 blkg_get(tg_to_blkg(qn->tg));
 413         }
 414 }
 415
 416 /**
 417  * throtl_peek_queued - peek the first bio on a qnode list
 418  * @queued: the qnode list to peek
 419  */
 420 static struct bio *throtl_peek_queued(struct list_head *queued)
 421 {
 422         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 423         struct bio *bio;
 424
 425         if (list_empty(queued))
 426                 return NULL;
 427
 428         bio = bio_list_peek(&qn->bios);
 429         WARN_ON_ONCE(!bio);
 430         return bio;
 431 }
 432
 433 /**
 434  * throtl_pop_queued - pop the first bio form a qnode list
 435  * @queued: the qnode list to pop a bio from
 436  * @tg_to_put: optional out argument for throtl_grp to put
 437  *
 438  * Pop the first bio from the qnode list @queued.  After popping, the first
 439  * qnode is removed from @queued if empty or moved to the end of @queued so
 440  * that the popping order is round-robin.
 441  *
 442  * When the first qnode is removed, its associated throtl_grp should be put
 443  * too.  If @tg_to_put is NULL, this function automatically puts it;
 444  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 445  * responsible for putting it.
 446  */
 447 static struct bio *throtl_pop_queued(struct list_head *queued,
 448                                      struct throtl_grp **tg_to_put)
 449 {
 450         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 451         struct bio *bio;
 452
 453         if (list_empty(queued))
 454                 return NULL;
 455
 456         bio = bio_list_pop(&qn->bios);
 457         WARN_ON_ONCE(!bio);
 458
 459         if (bio_list_empty(&qn->bios)) {
 460                 list_del_init(&qn->node);
 461                 if (tg_to_put)
 462                         *tg_to_put = qn->tg;
 463                 else
 464                         blkg_put(tg_to_blkg(qn->tg));
 465         } else {
 466                 list_move_tail(&qn->node, queued);
 467         }
 468
 469         return bio;
 470 }
 471
 472 /* init a service_queue, assumes the caller zeroed it */
 473 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 474 {
 475         INIT_LIST_HEAD(&sq->queued[0]);
 476         INIT_LIST_HEAD(&sq->queued[1]);
 477         sq->pending_tree = RB_ROOT_CACHED;
 478         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 479 }
 480
 481 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
 482 {
 483         struct throtl_grp *tg;
 484         int rw;
 485
 486         tg = kzalloc_node(sizeof(*tg), gfp, node);
 487         if (!tg)
 488                 return NULL;
 489
 490         throtl_service_queue_init(&tg->service_queue);
 491
 492         for (rw = READ; rw <= WRITE; rw++) {
 493                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 494                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 495         }
 496
 497         RB_CLEAR_NODE(&tg->rb_node);
 498         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 499         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 500         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 501         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 502         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 503         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 504         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 505         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 506         /* LIMIT_LOW will have default value 0 */
 507
 508         tg->latency_target = DFL_LATENCY_TARGET;
 509         tg->latency_target_conf = DFL_LATENCY_TARGET;
 510         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 511         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 512
 513         return &tg->pd;
 514 }
 515
 516 static void throtl_pd_init(struct blkg_policy_data *pd)
 517 {
 518         struct throtl_grp *tg = pd_to_tg(pd);
 519         struct blkcg_gq *blkg = tg_to_blkg(tg);
 520         struct throtl_data *td = blkg->q->td;
 521         struct throtl_service_queue *sq = &tg->service_queue;
 522
 523         /*
 524          * If on the default hierarchy, we switch to properly hierarchical
 525          * behavior where limits on a given throtl_grp are applied to the
 526          * whole subtree rather than just the group itself.  e.g. If 16M
 527          * read_bps limit is set on the root group, the whole system can't
 528          * exceed 16M for the device.
 529          *
 530          * If not on the default hierarchy, the broken flat hierarchy
 531          * behavior is retained where all throtl_grps are treated as if
 532          * they're all separate root groups right below throtl_data.
 533          * Limits of a group don't interact with limits of other groups
 534          * regardless of the position of the group in the hierarchy.
 535          */
 536         sq->parent_sq = &td->service_queue;
 537         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 538                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 539         tg->td = td;
 540 }
 541
 542 /*
 543  * Set has_rules[] if @tg or any of its parents have limits configured.
 544  * This doesn't require walking up to the top of the hierarchy as the
 545  * parent's has_rules[] is guaranteed to be correct.
 546  */
 547 static void tg_update_has_rules(struct throtl_grp *tg)
 548 {
 549         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 550         struct throtl_data *td = tg->td;
 551         int rw;
 552
 553         for (rw = READ; rw <= WRITE; rw++)
 554                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 555                         (td->limit_valid[td->limit_index] &&
 556                          (tg_bps_limit(tg, rw) != U64_MAX ||
 557                           tg_iops_limit(tg, rw) != UINT_MAX));
 558 }
 559
 560 static void throtl_pd_online(struct blkg_policy_data *pd)
 561 {
 562         struct throtl_grp *tg = pd_to_tg(pd);
 563         /*
 564          * We don't want new groups to escape the limits of its ancestors.
 565          * Update has_rules[] after a new group is brought online.
 566          */
 567         tg_update_has_rules(tg);
 568 }
 569
 570 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 571 {
 572         struct cgroup_subsys_state *pos_css;
 573         struct blkcg_gq *blkg;
 574         bool low_valid = false;
 575
 576         rcu_read_lock();
 577         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 578                 struct throtl_grp *tg = blkg_to_tg(blkg);
 579
 580                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 581                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 582                         low_valid = true;
 583                         break;
 584                 }
 585         }
 586         rcu_read_unlock();
 587
 588         td->limit_valid[LIMIT_LOW] = low_valid;
 589 }
 590
 591 static void throtl_upgrade_state(struct throtl_data *td);
 592 static void throtl_pd_offline(struct blkg_policy_data *pd)
 593 {
 594         struct throtl_grp *tg = pd_to_tg(pd);
 595
 596         tg->bps[READ][LIMIT_LOW] = 0;
 597         tg->bps[WRITE][LIMIT_LOW] = 0;
 598         tg->iops[READ][LIMIT_LOW] = 0;
 599         tg->iops[WRITE][LIMIT_LOW] = 0;
 600
 601         blk_throtl_update_limit_valid(tg->td);
 602
 603         if (!tg->td->limit_valid[tg->td->limit_index])
 604                 throtl_upgrade_state(tg->td);
 605 }
 606
 607 static void throtl_pd_free(struct blkg_policy_data *pd)
 608 {
 609         struct throtl_grp *tg = pd_to_tg(pd);
 610
 611         del_timer_sync(&tg->service_queue.pending_timer);
 612         kfree(tg);
 613 }
 614
 615 static struct throtl_grp *
 616 throtl_rb_first(struct throtl_service_queue *parent_sq)
 617 {
 618         struct rb_node *n;
 619         /* Service tree is empty */
 620         if (!parent_sq->nr_pending)
 621                 return NULL;
 622
 623         n = rb_first_cached(&parent_sq->pending_tree);
 624         WARN_ON_ONCE(!n);
 625         if (!n)
 626                 return NULL;
 627         return rb_entry_tg(n);
 628 }
 629
 630 static void throtl_rb_erase(struct rb_node *n,
 631                             struct throtl_service_queue *parent_sq)
 632 {
 633         rb_erase_cached(n, &parent_sq->pending_tree);
 634         RB_CLEAR_NODE(n);
 635         --parent_sq->nr_pending;
 636 }
 637
 638 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 639 {
 640         struct throtl_grp *tg;
 641
 642         tg = throtl_rb_first(parent_sq);
 643         if (!tg)
 644                 return;
 645
 646         parent_sq->first_pending_disptime = tg->disptime;
 647 }
 648
 649 static void tg_service_queue_add(struct throtl_grp *tg)
 650 {
 651         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 652         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 653         struct rb_node *parent = NULL;
 654         struct throtl_grp *__tg;
 655         unsigned long key = tg->disptime;
 656         bool leftmost = true;
 657
 658         while (*node != NULL) {
 659                 parent = *node;
 660                 __tg = rb_entry_tg(parent);
 661
 662                 if (time_before(key, __tg->disptime))
 663                         node = &parent->rb_left;
 664                 else {
 665                         node = &parent->rb_right;
 666                         leftmost = false;
 667                 }
 668         }
 669
 670         rb_link_node(&tg->rb_node, parent, node);
 671         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 672                                leftmost);
 673 }
 674
 675 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 676 {
 677         tg_service_queue_add(tg);
 678         tg->flags |= THROTL_TG_PENDING;
 679         tg->service_queue.parent_sq->nr_pending++;
 680 }
 681
 682 static void throtl_enqueue_tg(struct throtl_grp *tg)
 683 {
 684         if (!(tg->flags & THROTL_TG_PENDING))
 685                 __throtl_enqueue_tg(tg);
 686 }
 687
 688 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 689 {
 690         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 691         tg->flags &= ~THROTL_TG_PENDING;
 692 }
 693
 694 static void throtl_dequeue_tg(struct throtl_grp *tg)
 695 {
 696         if (tg->flags & THROTL_TG_PENDING)
 697                 __throtl_dequeue_tg(tg);
 698 }
 699
 700 /* Call with queue lock held */
 701 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 702                                           unsigned long expires)
 703 {
 704         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 705
 706         /*
 707          * Since we are adjusting the throttle limit dynamically, the sleep
 708          * time calculated according to previous limit might be invalid. It's
 709          * possible the cgroup sleep time is very long and no other cgroups
 710          * have IO running so notify the limit changes. Make sure the cgroup
 711          * doesn't sleep too long to avoid the missed notification.
 712          */
 713         if (time_after(expires, max_expire))
 714                 expires = max_expire;
 715         mod_timer(&sq->pending_timer, expires);
 716         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 717                    expires - jiffies, jiffies);
 718 }
 719
 720 /**
 721  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 722  * @sq: the service_queue to schedule dispatch for
 723  * @force: force scheduling
 724  *
 725  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 726  * dispatch time of the first pending child.  Returns %true if either timer
 727  * is armed or there's no pending child left.  %false if the current
 728  * dispatch window is still open and the caller should continue
 729  * dispatching.
 730  *
 731  * If @force is %true, the dispatch timer is always scheduled and this
 732  * function is guaranteed to return %true.  This is to be used when the
 733  * caller can't dispatch itself and needs to invoke pending_timer
 734  * unconditionally.  Note that forced scheduling is likely to induce short
 735  * delay before dispatch starts even if @sq->first_pending_disptime is not
 736  * in the future and thus shouldn't be used in hot paths.
 737  */
 738 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 739                                           bool force)
 740 {
 741         /* any pending children left? */
 742         if (!sq->nr_pending)
 743                 return true;
 744
 745         update_min_dispatch_time(sq);
 746
 747         /* is the next dispatch time in the future? */
 748         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 749                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 750                 return true;
 751         }
 752
 753         /* tell the caller to continue dispatching */
 754         return false;
 755 }
 756
 757 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 758                 bool rw, unsigned long start)
 759 {
 760         tg->bytes_disp[rw] = 0;
 761         tg->io_disp[rw] = 0;
 762
 763         /*
 764          * Previous slice has expired. We must have trimmed it after last
 765          * bio dispatch. That means since start of last slice, we never used
 766          * that bandwidth. Do try to make use of that bandwidth while giving
 767          * credit.
 768          */
 769         if (time_after_eq(start, tg->slice_start[rw]))
 770                 tg->slice_start[rw] = start;
 771
 772         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 773         throtl_log(&tg->service_queue,
 774                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 775                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 776                    tg->slice_end[rw], jiffies);
 777 }
 778
 779 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 780 {
 781         tg->bytes_disp[rw] = 0;
 782         tg->io_disp[rw] = 0;
 783         tg->slice_start[rw] = jiffies;
 784         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 785         throtl_log(&tg->service_queue,
 786                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 787                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 788                    tg->slice_end[rw], jiffies);
 789 }
 790
 791 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 792                                         unsigned long jiffy_end)
 793 {
 794         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 795 }
 796
 797 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 798                                        unsigned long jiffy_end)
 799 {
 800         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 801         throtl_log(&tg->service_queue,
 802                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 803                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 804                    tg->slice_end[rw], jiffies);
 805 }
 806
 807 /* Determine if previously allocated or extended slice is complete or not */
 808 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 809 {
 810         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 811                 return false;
 812
 813         return true;
 814 }
 815
 816 /* Trim the used slices and adjust slice start accordingly */
 817 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 818 {
 819         unsigned long nr_slices, time_elapsed, io_trim;
 820         u64 bytes_trim, tmp;
 821
 822         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 823
 824         /*
 825          * If bps are unlimited (-1), then time slice don't get
 826          * renewed. Don't try to trim the slice if slice is used. A new
 827          * slice will start when appropriate.
 828          */
 829         if (throtl_slice_used(tg, rw))
 830                 return;
 831
 832         /*
 833          * A bio has been dispatched. Also adjust slice_end. It might happen
 834          * that initially cgroup limit was very low resulting in high
 835          * slice_end, but later limit was bumped up and bio was dispached
 836          * sooner, then we need to reduce slice_end. A high bogus slice_end
 837          * is bad because it does not allow new slice to start.
 838          */
 839
 840         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 841
 842         time_elapsed = jiffies - tg->slice_start[rw];
 843
 844         nr_slices = time_elapsed / tg->td->throtl_slice;
 845
 846         if (!nr_slices)
 847                 return;
 848         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 849         do_div(tmp, HZ);
 850         bytes_trim = tmp;
 851
 852         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 853                 HZ;
 854
 855         if (!bytes_trim && !io_trim)
 856                 return;
 857
 858         if (tg->bytes_disp[rw] >= bytes_trim)
 859                 tg->bytes_disp[rw] -= bytes_trim;
 860         else
 861                 tg->bytes_disp[rw] = 0;
 862
 863         if (tg->io_disp[rw] >= io_trim)
 864                 tg->io_disp[rw] -= io_trim;
 865         else
 866                 tg->io_disp[rw] = 0;
 867
 868         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 869
 870         throtl_log(&tg->service_queue,
 871                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 872                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 873                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 874 }
 875
 876 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 877                                   unsigned long *wait)
 878 {
 879         bool rw = bio_data_dir(bio);
 880         unsigned int io_allowed;
 881         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 882         u64 tmp;
 883
 884         jiffy_elapsed = jiffies - tg->slice_start[rw];
 885
 886         /* Round up to the next throttle slice, wait time must be nonzero */
 887         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 888
 889         /*
 890          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 891          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 892          * will allow dispatch after 1 second and after that slice should
 893          * have been trimmed.
 894          */
 895
 896         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 897         do_div(tmp, HZ);
 898
 899         if (tmp > UINT_MAX)
 900                 io_allowed = UINT_MAX;
 901         else
 902                 io_allowed = tmp;
 903
 904         if (tg->io_disp[rw] + 1 <= io_allowed) {
 905                 if (wait)
 906                         *wait = 0;
 907                 return true;
 908         }
 909
 910         /* Calc approx time to dispatch */
 911         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 912
 913         if (wait)
 914                 *wait = jiffy_wait;
 915         return false;
 916 }
 917
 918 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 919                                  unsigned long *wait)
 920 {
 921         bool rw = bio_data_dir(bio);
 922         u64 bytes_allowed, extra_bytes, tmp;
 923         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 924         unsigned int bio_size = throtl_bio_data_size(bio);
 925
 926         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 927
 928         /* Slice has just started. Consider one slice interval */
 929         if (!jiffy_elapsed)
 930                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 931
 932         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 933
 934         tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
 935         do_div(tmp, HZ);
 936         bytes_allowed = tmp;
 937
 938         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 939                 if (wait)
 940                         *wait = 0;
 941                 return true;
 942         }
 943
 944         /* Calc approx time to dispatch */
 945         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 946         jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
 947
 948         if (!jiffy_wait)
 949                 jiffy_wait = 1;
 950
 951         /*
 952          * This wait time is without taking into consideration the rounding
 953          * up we did. Add that time also.
 954          */
 955         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 956         if (wait)
 957                 *wait = jiffy_wait;
 958         return false;
 959 }
 960
 961 /*
 962  * Returns whether one can dispatch a bio or not. Also returns approx number
 963  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 964  */
 965 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 966                             unsigned long *wait)
 967 {
 968         bool rw = bio_data_dir(bio);
 969         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 970
 971         /*
 972          * Currently whole state machine of group depends on first bio
 973          * queued in the group bio list. So one should not be calling
 974          * this function with a different bio if there are other bios
 975          * queued.
 976          */
 977         BUG_ON(tg->service_queue.nr_queued[rw] &&
 978                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 979
 980         /* If tg->bps = -1, then BW is unlimited */
 981         if (tg_bps_limit(tg, rw) == U64_MAX &&
 982             tg_iops_limit(tg, rw) == UINT_MAX) {
 983                 if (wait)
 984                         *wait = 0;
 985                 return true;
 986         }
 987
 988         /*
 989          * If previous slice expired, start a new one otherwise renew/extend
 990          * existing slice to make sure it is at least throtl_slice interval
 991          * long since now. New slice is started only for empty throttle group.
 992          * If there is queued bio, that means there should be an active
 993          * slice and it should be extended instead.
 994          */
 995         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 996                 throtl_start_new_slice(tg, rw);
 997         else {
 998                 if (time_before(tg->slice_end[rw],
 999                     jiffies + tg->td->throtl_slice))
1000                         throtl_extend_slice(tg, rw,
1001                                 jiffies + tg->td->throtl_slice);
1002         }
1003
1004         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1005             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1006                 if (wait)
1007                         *wait = 0;
1008                 return true;
1009         }
1010
1011         max_wait = max(bps_wait, iops_wait);
1012
1013         if (wait)
1014                 *wait = max_wait;
1015
1016         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1017                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1018
1019         return false;
1020 }
1021
1022 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1023 {
1024         bool rw = bio_data_dir(bio);
1025         unsigned int bio_size = throtl_bio_data_size(bio);
1026
1027         /* Charge the bio to the group */
1028         tg->bytes_disp[rw] += bio_size;
1029         tg->io_disp[rw]++;
1030         tg->last_bytes_disp[rw] += bio_size;
1031         tg->last_io_disp[rw]++;
1032
1033         /*
1034          * BIO_THROTTLED is used to prevent the same bio to be throttled
1035          * more than once as a throttled bio will go through blk-throtl the
1036          * second time when it eventually gets issued.  Set it when a bio
1037          * is being charged to a tg.
1038          */
1039         if (!bio_flagged(bio, BIO_THROTTLED))
1040                 bio_set_flag(bio, BIO_THROTTLED);
1041 }
1042
1043 /**
1044  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1045  * @bio: bio to add
1046  * @qn: qnode to use
1047  * @tg: the target throtl_grp
1048  *
1049  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1050  * tg->qnode_on_self[] is used.
1051  */
1052 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1053                               struct throtl_grp *tg)
1054 {
1055         struct throtl_service_queue *sq = &tg->service_queue;
1056         bool rw = bio_data_dir(bio);
1057
1058         if (!qn)
1059                 qn = &tg->qnode_on_self[rw];
1060
1061         /*
1062          * If @tg doesn't currently have any bios queued in the same
1063          * direction, queueing @bio can change when @tg should be
1064          * dispatched.  Mark that @tg was empty.  This is automatically
1065          * cleaered on the next tg_update_disptime().
1066          */
1067         if (!sq->nr_queued[rw])
1068                 tg->flags |= THROTL_TG_WAS_EMPTY;
1069
1070         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1071
1072         sq->nr_queued[rw]++;
1073         throtl_enqueue_tg(tg);
1074 }
1075
1076 static void tg_update_disptime(struct throtl_grp *tg)
1077 {
1078         struct throtl_service_queue *sq = &tg->service_queue;
1079         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1080         struct bio *bio;
1081
1082         bio = throtl_peek_queued(&sq->queued[READ]);
1083         if (bio)
1084                 tg_may_dispatch(tg, bio, &read_wait);
1085
1086         bio = throtl_peek_queued(&sq->queued[WRITE]);
1087         if (bio)
1088                 tg_may_dispatch(tg, bio, &write_wait);
1089
1090         min_wait = min(read_wait, write_wait);
1091         disptime = jiffies + min_wait;
1092
1093         /* Update dispatch time */
1094         throtl_dequeue_tg(tg);
1095         tg->disptime = disptime;
1096         throtl_enqueue_tg(tg);
1097
1098         /* see throtl_add_bio_tg() */
1099         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1100 }
1101
1102 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1103                                         struct throtl_grp *parent_tg, bool rw)
1104 {
1105         if (throtl_slice_used(parent_tg, rw)) {
1106                 throtl_start_new_slice_with_credit(parent_tg, rw,
1107                                 child_tg->slice_start[rw]);
1108         }
1109
1110 }
1111
1112 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1113 {
1114         struct throtl_service_queue *sq = &tg->service_queue;
1115         struct throtl_service_queue *parent_sq = sq->parent_sq;
1116         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1117         struct throtl_grp *tg_to_put = NULL;
1118         struct bio *bio;
1119
1120         /*
1121          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1122          * from @tg may put its reference and @parent_sq might end up
1123          * getting released prematurely.  Remember the tg to put and put it
1124          * after @bio is transferred to @parent_sq.
1125          */
1126         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1127         sq->nr_queued[rw]--;
1128
1129         throtl_charge_bio(tg, bio);
1130
1131         /*
1132          * If our parent is another tg, we just need to transfer @bio to
1133          * the parent using throtl_add_bio_tg().  If our parent is
1134          * @td->service_queue, @bio is ready to be issued.  Put it on its
1135          * bio_lists[] and decrease total number queued.  The caller is
1136          * responsible for issuing these bios.
1137          */
1138         if (parent_tg) {
1139                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1140                 start_parent_slice_with_credit(tg, parent_tg, rw);
1141         } else {
1142                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1143                                      &parent_sq->queued[rw]);
1144                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1145                 tg->td->nr_queued[rw]--;
1146         }
1147
1148         throtl_trim_slice(tg, rw);
1149
1150         if (tg_to_put)
1151                 blkg_put(tg_to_blkg(tg_to_put));
1152 }
1153
1154 static int throtl_dispatch_tg(struct throtl_grp *tg)
1155 {
1156         struct throtl_service_queue *sq = &tg->service_queue;
1157         unsigned int nr_reads = 0, nr_writes = 0;
1158         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1159         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1160         struct bio *bio;
1161
1162         /* Try to dispatch 75% READS and 25% WRITES */
1163
1164         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1165                tg_may_dispatch(tg, bio, NULL)) {
1166
1167                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1168                 nr_reads++;
1169
1170                 if (nr_reads >= max_nr_reads)
1171                         break;
1172         }
1173
1174         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1175                tg_may_dispatch(tg, bio, NULL)) {
1176
1177                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1178                 nr_writes++;
1179
1180                 if (nr_writes >= max_nr_writes)
1181                         break;
1182         }
1183
1184         return nr_reads + nr_writes;
1185 }
1186
1187 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1188 {
1189         unsigned int nr_disp = 0;
1190
1191         while (1) {
1192                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1193                 struct throtl_service_queue *sq;
1194
1195                 if (!tg)
1196                         break;
1197
1198                 if (time_before(jiffies, tg->disptime))
1199                         break;
1200
1201                 throtl_dequeue_tg(tg);
1202
1203                 nr_disp += throtl_dispatch_tg(tg);
1204
1205                 sq = &tg->service_queue;
1206                 if (sq->nr_queued[0] || sq->nr_queued[1])
1207                         tg_update_disptime(tg);
1208
1209                 if (nr_disp >= throtl_quantum)
1210                         break;
1211         }
1212
1213         return nr_disp;
1214 }
1215
1216 static bool throtl_can_upgrade(struct throtl_data *td,
1217         struct throtl_grp *this_tg);
1218 /**
1219  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1220  * @t: the pending_timer member of the throtl_service_queue being serviced
1221  *
1222  * This timer is armed when a child throtl_grp with active bio's become
1223  * pending and queued on the service_queue's pending_tree and expires when
1224  * the first child throtl_grp should be dispatched.  This function
1225  * dispatches bio's from the children throtl_grps to the parent
1226  * service_queue.
1227  *
1228  * If the parent's parent is another throtl_grp, dispatching is propagated
1229  * by either arming its pending_timer or repeating dispatch directly.  If
1230  * the top-level service_tree is reached, throtl_data->dispatch_work is
1231  * kicked so that the ready bio's are issued.
1232  */
1233 static void throtl_pending_timer_fn(struct timer_list *t)
1234 {
1235         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1236         struct throtl_grp *tg = sq_to_tg(sq);
1237         struct throtl_data *td = sq_to_td(sq);
1238         struct request_queue *q = td->queue;
1239         struct throtl_service_queue *parent_sq;
1240         bool dispatched;
1241         int ret;
1242
1243         spin_lock_irq(&q->queue_lock);
1244         if (throtl_can_upgrade(td, NULL))
1245                 throtl_upgrade_state(td);
1246
1247 again:
1248         parent_sq = sq->parent_sq;
1249         dispatched = false;
1250
1251         while (true) {
1252                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1253                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1254                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1255
1256                 ret = throtl_select_dispatch(sq);
1257                 if (ret) {
1258                         throtl_log(sq, "bios disp=%u", ret);
1259                         dispatched = true;
1260                 }
1261
1262                 if (throtl_schedule_next_dispatch(sq, false))
1263                         break;
1264
1265                 /* this dispatch windows is still open, relax and repeat */
1266                 spin_unlock_irq(&q->queue_lock);
1267                 cpu_relax();
1268                 spin_lock_irq(&q->queue_lock);
1269         }
1270
1271         if (!dispatched)
1272                 goto out_unlock;
1273
1274         if (parent_sq) {
1275                 /* @parent_sq is another throl_grp, propagate dispatch */
1276                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1277                         tg_update_disptime(tg);
1278                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1279                                 /* window is already open, repeat dispatching */
1280                                 sq = parent_sq;
1281                                 tg = sq_to_tg(sq);
1282                                 goto again;
1283                         }
1284                 }
1285         } else {
1286                 /* reached the top-level, queue issueing */
1287                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1288         }
1289 out_unlock:
1290         spin_unlock_irq(&q->queue_lock);
1291 }
1292
1293 /**
1294  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1295  * @work: work item being executed
1296  *
1297  * This function is queued for execution when bio's reach the bio_lists[]
1298  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1299  * function.
1300  */
1301 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1302 {
1303         struct throtl_data *td = container_of(work, struct throtl_data,
1304                                               dispatch_work);
1305         struct throtl_service_queue *td_sq = &td->service_queue;
1306         struct request_queue *q = td->queue;
1307         struct bio_list bio_list_on_stack;
1308         struct bio *bio;
1309         struct blk_plug plug;
1310         int rw;
1311
1312         bio_list_init(&bio_list_on_stack);
1313
1314         spin_lock_irq(&q->queue_lock);
1315         for (rw = READ; rw <= WRITE; rw++)
1316                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1317                         bio_list_add(&bio_list_on_stack, bio);
1318         spin_unlock_irq(&q->queue_lock);
1319
1320         if (!bio_list_empty(&bio_list_on_stack)) {
1321                 blk_start_plug(&plug);
1322                 while((bio = bio_list_pop(&bio_list_on_stack)))
1323                         generic_make_request(bio);
1324                 blk_finish_plug(&plug);
1325         }
1326 }
1327
1328 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1329                               int off)
1330 {
1331         struct throtl_grp *tg = pd_to_tg(pd);
1332         u64 v = *(u64 *)((void *)tg + off);
1333
1334         if (v == U64_MAX)
1335                 return 0;
1336         return __blkg_prfill_u64(sf, pd, v);
1337 }
1338
1339 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1340                                int off)
1341 {
1342         struct throtl_grp *tg = pd_to_tg(pd);
1343         unsigned int v = *(unsigned int *)((void *)tg + off);
1344
1345         if (v == UINT_MAX)
1346                 return 0;
1347         return __blkg_prfill_u64(sf, pd, v);
1348 }
1349
1350 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1351 {
1352         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1353                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1354         return 0;
1355 }
1356
1357 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1358 {
1359         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1360                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1361         return 0;
1362 }
1363
1364 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1365 {
1366         struct throtl_service_queue *sq = &tg->service_queue;
1367         struct cgroup_subsys_state *pos_css;
1368         struct blkcg_gq *blkg;
1369
1370         throtl_log(&tg->service_queue,
1371                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1372                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1373                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1374
1375         /*
1376          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1377          * considered to have rules if either the tg itself or any of its
1378          * ancestors has rules.  This identifies groups without any
1379          * restrictions in the whole hierarchy and allows them to bypass
1380          * blk-throttle.
1381          */
1382         blkg_for_each_descendant_pre(blkg, pos_css,
1383                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1384                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1385                 struct throtl_grp *parent_tg;
1386
1387                 tg_update_has_rules(this_tg);
1388                 /* ignore root/second level */
1389                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1390                     !blkg->parent->parent)
1391                         continue;
1392                 parent_tg = blkg_to_tg(blkg->parent);
1393                 /*
1394                  * make sure all children has lower idle time threshold and
1395                  * higher latency target
1396                  */
1397                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1398                                 parent_tg->idletime_threshold);
1399                 this_tg->latency_target = max(this_tg->latency_target,
1400                                 parent_tg->latency_target);
1401         }
1402
1403         /*
1404          * We're already holding queue_lock and know @tg is valid.  Let's
1405          * apply the new config directly.
1406          *
1407          * Restart the slices for both READ and WRITES. It might happen
1408          * that a group's limit are dropped suddenly and we don't want to
1409          * account recently dispatched IO with new low rate.
1410          */
1411         throtl_start_new_slice(tg, 0);
1412         throtl_start_new_slice(tg, 1);
1413
1414         if (tg->flags & THROTL_TG_PENDING) {
1415                 tg_update_disptime(tg);
1416                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1417         }
1418 }
1419
1420 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1421                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1422 {
1423         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1424         struct blkg_conf_ctx ctx;
1425         struct throtl_grp *tg;
1426         int ret;
1427         u64 v;
1428
1429         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1430         if (ret)
1431                 return ret;
1432
1433         ret = -EINVAL;
1434         if (sscanf(ctx.body, "%llu", &v) != 1)
1435                 goto out_finish;
1436         if (!v)
1437                 v = U64_MAX;
1438
1439         tg = blkg_to_tg(ctx.blkg);
1440
1441         if (is_u64)
1442                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1443         else
1444                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1445
1446         tg_conf_updated(tg, false);
1447         ret = 0;
1448 out_finish:
1449         blkg_conf_finish(&ctx);
1450         return ret ?: nbytes;
1451 }
1452
1453 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1454                                char *buf, size_t nbytes, loff_t off)
1455 {
1456         return tg_set_conf(of, buf, nbytes, off, true);
1457 }
1458
1459 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1460                                 char *buf, size_t nbytes, loff_t off)
1461 {
1462         return tg_set_conf(of, buf, nbytes, off, false);
1463 }
1464
1465 static struct cftype throtl_legacy_files[] = {
1466         {
1467                 .name = "throttle.read_bps_device",
1468                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1469                 .seq_show = tg_print_conf_u64,
1470                 .write = tg_set_conf_u64,
1471         },
1472         {
1473                 .name = "throttle.write_bps_device",
1474                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1475                 .seq_show = tg_print_conf_u64,
1476                 .write = tg_set_conf_u64,
1477         },
1478         {
1479                 .name = "throttle.read_iops_device",
1480                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1481                 .seq_show = tg_print_conf_uint,
1482                 .write = tg_set_conf_uint,
1483         },
1484         {
1485                 .name = "throttle.write_iops_device",
1486                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1487                 .seq_show = tg_print_conf_uint,
1488                 .write = tg_set_conf_uint,
1489         },
1490         {
1491                 .name = "throttle.io_service_bytes",
1492                 .private = (unsigned long)&blkcg_policy_throtl,
1493                 .seq_show = blkg_print_stat_bytes,
1494         },
1495         {
1496                 .name = "throttle.io_service_bytes_recursive",
1497                 .private = (unsigned long)&blkcg_policy_throtl,
1498                 .seq_show = blkg_print_stat_bytes_recursive,
1499         },
1500         {
1501                 .name = "throttle.io_serviced",
1502                 .private = (unsigned long)&blkcg_policy_throtl,
1503                 .seq_show = blkg_print_stat_ios,
1504         },
1505         {
1506                 .name = "throttle.io_serviced_recursive",
1507                 .private = (unsigned long)&blkcg_policy_throtl,
1508                 .seq_show = blkg_print_stat_ios_recursive,
1509         },
1510         { }     /* terminate */
1511 };
1512
1513 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1514                          int off)
1515 {
1516         struct throtl_grp *tg = pd_to_tg(pd);
1517         const char *dname = blkg_dev_name(pd->blkg);
1518         char bufs[4][21] = { "max", "max", "max", "max" };
1519         u64 bps_dft;
1520         unsigned int iops_dft;
1521         char idle_time[26] = "";
1522         char latency_time[26] = "";
1523
1524         if (!dname)
1525                 return 0;
1526
1527         if (off == LIMIT_LOW) {
1528                 bps_dft = 0;
1529                 iops_dft = 0;
1530         } else {
1531                 bps_dft = U64_MAX;
1532                 iops_dft = UINT_MAX;
1533         }
1534
1535         if (tg->bps_conf[READ][off] == bps_dft &&
1536             tg->bps_conf[WRITE][off] == bps_dft &&
1537             tg->iops_conf[READ][off] == iops_dft &&
1538             tg->iops_conf[WRITE][off] == iops_dft &&
1539             (off != LIMIT_LOW ||
1540              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1541               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1542                 return 0;
1543
1544         if (tg->bps_conf[READ][off] != U64_MAX)
1545                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1546                         tg->bps_conf[READ][off]);
1547         if (tg->bps_conf[WRITE][off] != U64_MAX)
1548                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1549                         tg->bps_conf[WRITE][off]);
1550         if (tg->iops_conf[READ][off] != UINT_MAX)
1551                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1552                         tg->iops_conf[READ][off]);
1553         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1554                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1555                         tg->iops_conf[WRITE][off]);
1556         if (off == LIMIT_LOW) {
1557                 if (tg->idletime_threshold_conf == ULONG_MAX)
1558                         strcpy(idle_time, " idle=max");
1559                 else
1560                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1561                                 tg->idletime_threshold_conf);
1562
1563                 if (tg->latency_target_conf == ULONG_MAX)
1564                         strcpy(latency_time, " latency=max");
1565                 else
1566                         snprintf(latency_time, sizeof(latency_time),
1567                                 " latency=%lu", tg->latency_target_conf);
1568         }
1569
1570         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1571                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1572                    latency_time);
1573         return 0;
1574 }
1575
1576 static int tg_print_limit(struct seq_file *sf, void *v)
1577 {
1578         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1579                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1580         return 0;
1581 }
1582
1583 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1584                           char *buf, size_t nbytes, loff_t off)
1585 {
1586         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1587         struct blkg_conf_ctx ctx;
1588         struct throtl_grp *tg;
1589         u64 v[4];
1590         unsigned long idle_time;
1591         unsigned long latency_time;
1592         int ret;
1593         int index = of_cft(of)->private;
1594
1595         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1596         if (ret)
1597                 return ret;
1598
1599         tg = blkg_to_tg(ctx.blkg);
1600
1601         v[0] = tg->bps_conf[READ][index];
1602         v[1] = tg->bps_conf[WRITE][index];
1603         v[2] = tg->iops_conf[READ][index];
1604         v[3] = tg->iops_conf[WRITE][index];
1605
1606         idle_time = tg->idletime_threshold_conf;
1607         latency_time = tg->latency_target_conf;
1608         while (true) {
1609                 char tok[27];   /* wiops=18446744073709551616 */
1610                 char *p;
1611                 u64 val = U64_MAX;
1612                 int len;
1613
1614                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1615                         break;
1616                 if (tok[0] == '\0')
1617                         break;
1618                 ctx.body += len;
1619
1620                 ret = -EINVAL;
1621                 p = tok;
1622                 strsep(&p, "=");
1623                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1624                         goto out_finish;
1625
1626                 ret = -ERANGE;
1627                 if (!val)
1628                         goto out_finish;
1629
1630                 ret = -EINVAL;
1631                 if (!strcmp(tok, "rbps"))
1632                         v[0] = val;
1633                 else if (!strcmp(tok, "wbps"))
1634                         v[1] = val;
1635                 else if (!strcmp(tok, "riops"))
1636                         v[2] = min_t(u64, val, UINT_MAX);
1637                 else if (!strcmp(tok, "wiops"))
1638                         v[3] = min_t(u64, val, UINT_MAX);
1639                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1640                         idle_time = val;
1641                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1642                         latency_time = val;
1643                 else
1644                         goto out_finish;
1645         }
1646
1647         tg->bps_conf[READ][index] = v[0];
1648         tg->bps_conf[WRITE][index] = v[1];
1649         tg->iops_conf[READ][index] = v[2];
1650         tg->iops_conf[WRITE][index] = v[3];
1651
1652         if (index == LIMIT_MAX) {
1653                 tg->bps[READ][index] = v[0];
1654                 tg->bps[WRITE][index] = v[1];
1655                 tg->iops[READ][index] = v[2];
1656                 tg->iops[WRITE][index] = v[3];
1657         }
1658         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1659                 tg->bps_conf[READ][LIMIT_MAX]);
1660         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1661                 tg->bps_conf[WRITE][LIMIT_MAX]);
1662         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1663                 tg->iops_conf[READ][LIMIT_MAX]);
1664         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1665                 tg->iops_conf[WRITE][LIMIT_MAX]);
1666         tg->idletime_threshold_conf = idle_time;
1667         tg->latency_target_conf = latency_time;
1668
1669         /* force user to configure all settings for low limit  */
1670         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1671               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1672             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1673             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1674                 tg->bps[READ][LIMIT_LOW] = 0;
1675                 tg->bps[WRITE][LIMIT_LOW] = 0;
1676                 tg->iops[READ][LIMIT_LOW] = 0;
1677                 tg->iops[WRITE][LIMIT_LOW] = 0;
1678                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1679                 tg->latency_target = DFL_LATENCY_TARGET;
1680         } else if (index == LIMIT_LOW) {
1681                 tg->idletime_threshold = tg->idletime_threshold_conf;
1682                 tg->latency_target = tg->latency_target_conf;
1683         }
1684
1685         blk_throtl_update_limit_valid(tg->td);
1686         if (tg->td->limit_valid[LIMIT_LOW]) {
1687                 if (index == LIMIT_LOW)
1688                         tg->td->limit_index = LIMIT_LOW;
1689         } else
1690                 tg->td->limit_index = LIMIT_MAX;
1691         tg_conf_updated(tg, index == LIMIT_LOW &&
1692                 tg->td->limit_valid[LIMIT_LOW]);
1693         ret = 0;
1694 out_finish:
1695         blkg_conf_finish(&ctx);
1696         return ret ?: nbytes;
1697 }
1698
1699 static struct cftype throtl_files[] = {
1700 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1701         {
1702                 .name = "low",
1703                 .flags = CFTYPE_NOT_ON_ROOT,
1704                 .seq_show = tg_print_limit,
1705                 .write = tg_set_limit,
1706                 .private = LIMIT_LOW,
1707         },
1708 #endif
1709         {
1710                 .name = "max",
1711                 .flags = CFTYPE_NOT_ON_ROOT,
1712                 .seq_show = tg_print_limit,
1713                 .write = tg_set_limit,
1714                 .private = LIMIT_MAX,
1715         },
1716         { }     /* terminate */
1717 };
1718
1719 static void throtl_shutdown_wq(struct request_queue *q)
1720 {
1721         struct throtl_data *td = q->td;
1722
1723         cancel_work_sync(&td->dispatch_work);
1724 }
1725
1726 static struct blkcg_policy blkcg_policy_throtl = {
1727         .dfl_cftypes            = throtl_files,
1728         .legacy_cftypes         = throtl_legacy_files,
1729
1730         .pd_alloc_fn            = throtl_pd_alloc,
1731         .pd_init_fn             = throtl_pd_init,
1732         .pd_online_fn           = throtl_pd_online,
1733         .pd_offline_fn          = throtl_pd_offline,
1734         .pd_free_fn             = throtl_pd_free,
1735 };
1736
1737 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1738 {
1739         unsigned long rtime = jiffies, wtime = jiffies;
1740
1741         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1742                 rtime = tg->last_low_overflow_time[READ];
1743         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1744                 wtime = tg->last_low_overflow_time[WRITE];
1745         return min(rtime, wtime);
1746 }
1747
1748 /* tg should not be an intermediate node */
1749 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1750 {
1751         struct throtl_service_queue *parent_sq;
1752         struct throtl_grp *parent = tg;
1753         unsigned long ret = __tg_last_low_overflow_time(tg);
1754
1755         while (true) {
1756                 parent_sq = parent->service_queue.parent_sq;
1757                 parent = sq_to_tg(parent_sq);
1758                 if (!parent)
1759                         break;
1760
1761                 /*
1762                  * The parent doesn't have low limit, it always reaches low
1763                  * limit. Its overflow time is useless for children
1764                  */
1765                 if (!parent->bps[READ][LIMIT_LOW] &&
1766                     !parent->iops[READ][LIMIT_LOW] &&
1767                     !parent->bps[WRITE][LIMIT_LOW] &&
1768                     !parent->iops[WRITE][LIMIT_LOW])
1769                         continue;
1770                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1771                         ret = __tg_last_low_overflow_time(parent);
1772         }
1773         return ret;
1774 }
1775
1776 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1777 {
1778         /*
1779          * cgroup is idle if:
1780          * - single idle is too long, longer than a fixed value (in case user
1781          *   configure a too big threshold) or 4 times of idletime threshold
1782          * - average think time is more than threshold
1783          * - IO latency is largely below threshold
1784          */
1785         unsigned long time;
1786         bool ret;
1787
1788         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1789         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1790               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1791               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1792               tg->avg_idletime > tg->idletime_threshold ||
1793               (tg->latency_target && tg->bio_cnt &&
1794                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1795         throtl_log(&tg->service_queue,
1796                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1797                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1798                 tg->bio_cnt, ret, tg->td->scale);
1799         return ret;
1800 }
1801
1802 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1803 {
1804         struct throtl_service_queue *sq = &tg->service_queue;
1805         bool read_limit, write_limit;
1806
1807         /*
1808          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1809          * reaches), it's ok to upgrade to next limit
1810          */
1811         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1812         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1813         if (!read_limit && !write_limit)
1814                 return true;
1815         if (read_limit && sq->nr_queued[READ] &&
1816             (!write_limit || sq->nr_queued[WRITE]))
1817                 return true;
1818         if (write_limit && sq->nr_queued[WRITE] &&
1819             (!read_limit || sq->nr_queued[READ]))
1820                 return true;
1821
1822         if (time_after_eq(jiffies,
1823                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1824             throtl_tg_is_idle(tg))
1825                 return true;
1826         return false;
1827 }
1828
1829 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1830 {
1831         while (true) {
1832                 if (throtl_tg_can_upgrade(tg))
1833                         return true;
1834                 tg = sq_to_tg(tg->service_queue.parent_sq);
1835                 if (!tg || !tg_to_blkg(tg)->parent)
1836                         return false;
1837         }
1838         return false;
1839 }
1840
1841 static bool throtl_can_upgrade(struct throtl_data *td,
1842         struct throtl_grp *this_tg)
1843 {
1844         struct cgroup_subsys_state *pos_css;
1845         struct blkcg_gq *blkg;
1846
1847         if (td->limit_index != LIMIT_LOW)
1848                 return false;
1849
1850         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1851                 return false;
1852
1853         rcu_read_lock();
1854         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1855                 struct throtl_grp *tg = blkg_to_tg(blkg);
1856
1857                 if (tg == this_tg)
1858                         continue;
1859                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1860                         continue;
1861                 if (!throtl_hierarchy_can_upgrade(tg)) {
1862                         rcu_read_unlock();
1863                         return false;
1864                 }
1865         }
1866         rcu_read_unlock();
1867         return true;
1868 }
1869
1870 static void throtl_upgrade_check(struct throtl_grp *tg)
1871 {
1872         unsigned long now = jiffies;
1873
1874         if (tg->td->limit_index != LIMIT_LOW)
1875                 return;
1876
1877         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1878                 return;
1879
1880         tg->last_check_time = now;
1881
1882         if (!time_after_eq(now,
1883              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1884                 return;
1885
1886         if (throtl_can_upgrade(tg->td, NULL))
1887                 throtl_upgrade_state(tg->td);
1888 }
1889
1890 static void throtl_upgrade_state(struct throtl_data *td)
1891 {
1892         struct cgroup_subsys_state *pos_css;
1893         struct blkcg_gq *blkg;
1894
1895         throtl_log(&td->service_queue, "upgrade to max");
1896         td->limit_index = LIMIT_MAX;
1897         td->low_upgrade_time = jiffies;
1898         td->scale = 0;
1899         rcu_read_lock();
1900         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1901                 struct throtl_grp *tg = blkg_to_tg(blkg);
1902                 struct throtl_service_queue *sq = &tg->service_queue;
1903
1904                 tg->disptime = jiffies - 1;
1905                 throtl_select_dispatch(sq);
1906                 throtl_schedule_next_dispatch(sq, true);
1907         }
1908         rcu_read_unlock();
1909         throtl_select_dispatch(&td->service_queue);
1910         throtl_schedule_next_dispatch(&td->service_queue, true);
1911         queue_work(kthrotld_workqueue, &td->dispatch_work);
1912 }
1913
1914 static void throtl_downgrade_state(struct throtl_data *td, int new)
1915 {
1916         td->scale /= 2;
1917
1918         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1919         if (td->scale) {
1920                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1921                 return;
1922         }
1923
1924         td->limit_index = new;
1925         td->low_downgrade_time = jiffies;
1926 }
1927
1928 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1929 {
1930         struct throtl_data *td = tg->td;
1931         unsigned long now = jiffies;
1932
1933         /*
1934          * If cgroup is below low limit, consider downgrade and throttle other
1935          * cgroups
1936          */
1937         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1938             time_after_eq(now, tg_last_low_overflow_time(tg) +
1939                                         td->throtl_slice) &&
1940             (!throtl_tg_is_idle(tg) ||
1941              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1942                 return true;
1943         return false;
1944 }
1945
1946 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1947 {
1948         while (true) {
1949                 if (!throtl_tg_can_downgrade(tg))
1950                         return false;
1951                 tg = sq_to_tg(tg->service_queue.parent_sq);
1952                 if (!tg || !tg_to_blkg(tg)->parent)
1953                         break;
1954         }
1955         return true;
1956 }
1957
1958 static void throtl_downgrade_check(struct throtl_grp *tg)
1959 {
1960         uint64_t bps;
1961         unsigned int iops;
1962         unsigned long elapsed_time;
1963         unsigned long now = jiffies;
1964
1965         if (tg->td->limit_index != LIMIT_MAX ||
1966             !tg->td->limit_valid[LIMIT_LOW])
1967                 return;
1968         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1969                 return;
1970         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1971                 return;
1972
1973         elapsed_time = now - tg->last_check_time;
1974         tg->last_check_time = now;
1975
1976         if (time_before(now, tg_last_low_overflow_time(tg) +
1977                         tg->td->throtl_slice))
1978                 return;
1979
1980         if (tg->bps[READ][LIMIT_LOW]) {
1981                 bps = tg->last_bytes_disp[READ] * HZ;
1982                 do_div(bps, elapsed_time);
1983                 if (bps >= tg->bps[READ][LIMIT_LOW])
1984                         tg->last_low_overflow_time[READ] = now;
1985         }
1986
1987         if (tg->bps[WRITE][LIMIT_LOW]) {
1988                 bps = tg->last_bytes_disp[WRITE] * HZ;
1989                 do_div(bps, elapsed_time);
1990                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1991                         tg->last_low_overflow_time[WRITE] = now;
1992         }
1993
1994         if (tg->iops[READ][LIMIT_LOW]) {
1995                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1996                 if (iops >= tg->iops[READ][LIMIT_LOW])
1997                         tg->last_low_overflow_time[READ] = now;
1998         }
1999
2000         if (tg->iops[WRITE][LIMIT_LOW]) {
2001                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2002                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2003                         tg->last_low_overflow_time[WRITE] = now;
2004         }
2005
2006         /*
2007          * If cgroup is below low limit, consider downgrade and throttle other
2008          * cgroups
2009          */
2010         if (throtl_hierarchy_can_downgrade(tg))
2011                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2012
2013         tg->last_bytes_disp[READ] = 0;
2014         tg->last_bytes_disp[WRITE] = 0;
2015         tg->last_io_disp[READ] = 0;
2016         tg->last_io_disp[WRITE] = 0;
2017 }
2018
2019 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2020 {
2021         unsigned long now = ktime_get_ns() >> 10;
2022         unsigned long last_finish_time = tg->last_finish_time;
2023
2024         if (now <= last_finish_time || last_finish_time == 0 ||
2025             last_finish_time == tg->checked_last_finish_time)
2026                 return;
2027
2028         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2029         tg->checked_last_finish_time = last_finish_time;
2030 }
2031
2032 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2033 static void throtl_update_latency_buckets(struct throtl_data *td)
2034 {
2035         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2036         int i, cpu, rw;
2037         unsigned long last_latency[2] = { 0 };
2038         unsigned long latency[2];
2039
2040         if (!blk_queue_nonrot(td->queue))
2041                 return;
2042         if (time_before(jiffies, td->last_calculate_time + HZ))
2043                 return;
2044         td->last_calculate_time = jiffies;
2045
2046         memset(avg_latency, 0, sizeof(avg_latency));
2047         for (rw = READ; rw <= WRITE; rw++) {
2048                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2049                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2050
2051                         for_each_possible_cpu(cpu) {
2052                                 struct latency_bucket *bucket;
2053
2054                                 /* this isn't race free, but ok in practice */
2055                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2056                                         cpu);
2057                                 tmp->total_latency += bucket[i].total_latency;
2058                                 tmp->samples += bucket[i].samples;
2059                                 bucket[i].total_latency = 0;
2060                                 bucket[i].samples = 0;
2061                         }
2062
2063                         if (tmp->samples >= 32) {
2064                                 int samples = tmp->samples;
2065
2066                                 latency[rw] = tmp->total_latency;
2067
2068                                 tmp->total_latency = 0;
2069                                 tmp->samples = 0;
2070                                 latency[rw] /= samples;
2071                                 if (latency[rw] == 0)
2072                                         continue;
2073                                 avg_latency[rw][i].latency = latency[rw];
2074                         }
2075                 }
2076         }
2077
2078         for (rw = READ; rw <= WRITE; rw++) {
2079                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2080                         if (!avg_latency[rw][i].latency) {
2081                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2082                                         td->avg_buckets[rw][i].latency =
2083                                                 last_latency[rw];
2084                                 continue;
2085                         }
2086
2087                         if (!td->avg_buckets[rw][i].valid)
2088                                 latency[rw] = avg_latency[rw][i].latency;
2089                         else
2090                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2091                                         avg_latency[rw][i].latency) >> 3;
2092
2093                         td->avg_buckets[rw][i].latency = max(latency[rw],
2094                                 last_latency[rw]);
2095                         td->avg_buckets[rw][i].valid = true;
2096                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2097                 }
2098         }
2099
2100         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2101                 throtl_log(&td->service_queue,
2102                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2103                         "write latency=%ld, write valid=%d", i,
2104                         td->avg_buckets[READ][i].latency,
2105                         td->avg_buckets[READ][i].valid,
2106                         td->avg_buckets[WRITE][i].latency,
2107                         td->avg_buckets[WRITE][i].valid);
2108 }
2109 #else
2110 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2111 {
2112 }
2113 #endif
2114
2115 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2116                     struct bio *bio)
2117 {
2118         struct throtl_qnode *qn = NULL;
2119         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2120         struct throtl_service_queue *sq;
2121         bool rw = bio_data_dir(bio);
2122         bool throttled = false;
2123         struct throtl_data *td = tg->td;
2124
2125         WARN_ON_ONCE(!rcu_read_lock_held());
2126
2127         /* see throtl_charge_bio() */
2128         if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2129                 goto out;
2130
2131         spin_lock_irq(&q->queue_lock);
2132
2133         throtl_update_latency_buckets(td);
2134
2135         blk_throtl_update_idletime(tg);
2136
2137         sq = &tg->service_queue;
2138
2139 again:
2140         while (true) {
2141                 if (tg->last_low_overflow_time[rw] == 0)
2142                         tg->last_low_overflow_time[rw] = jiffies;
2143                 throtl_downgrade_check(tg);
2144                 throtl_upgrade_check(tg);
2145                 /* throtl is FIFO - if bios are already queued, should queue */
2146                 if (sq->nr_queued[rw])
2147                         break;
2148
2149                 /* if above limits, break to queue */
2150                 if (!tg_may_dispatch(tg, bio, NULL)) {
2151                         tg->last_low_overflow_time[rw] = jiffies;
2152                         if (throtl_can_upgrade(td, tg)) {
2153                                 throtl_upgrade_state(td);
2154                                 goto again;
2155                         }
2156                         break;
2157                 }
2158
2159                 /* within limits, let's charge and dispatch directly */
2160                 throtl_charge_bio(tg, bio);
2161
2162                 /*
2163                  * We need to trim slice even when bios are not being queued
2164                  * otherwise it might happen that a bio is not queued for
2165                  * a long time and slice keeps on extending and trim is not
2166                  * called for a long time. Now if limits are reduced suddenly
2167                  * we take into account all the IO dispatched so far at new
2168                  * low rate and * newly queued IO gets a really long dispatch
2169                  * time.
2170                  *
2171                  * So keep on trimming slice even if bio is not queued.
2172                  */
2173                 throtl_trim_slice(tg, rw);
2174
2175                 /*
2176                  * @bio passed through this layer without being throttled.
2177                  * Climb up the ladder.  If we''re already at the top, it
2178                  * can be executed directly.
2179                  */
2180                 qn = &tg->qnode_on_parent[rw];
2181                 sq = sq->parent_sq;
2182                 tg = sq_to_tg(sq);
2183                 if (!tg)
2184                         goto out_unlock;
2185         }
2186
2187         /* out-of-limit, queue to @tg */
2188         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2189                    rw == READ ? 'R' : 'W',
2190                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2191                    tg_bps_limit(tg, rw),
2192                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2193                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2194
2195         tg->last_low_overflow_time[rw] = jiffies;
2196
2197         td->nr_queued[rw]++;
2198         throtl_add_bio_tg(bio, qn, tg);
2199         throttled = true;
2200
2201         /*
2202          * Update @tg's dispatch time and force schedule dispatch if @tg
2203          * was empty before @bio.  The forced scheduling isn't likely to
2204          * cause undue delay as @bio is likely to be dispatched directly if
2205          * its @tg's disptime is not in the future.
2206          */
2207         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2208                 tg_update_disptime(tg);
2209                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2210         }
2211
2212 out_unlock:
2213         spin_unlock_irq(&q->queue_lock);
2214 out:
2215         bio_set_flag(bio, BIO_THROTTLED);
2216
2217 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2218         if (throttled || !td->track_bio_latency)
2219                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2220 #endif
2221         return throttled;
2222 }
2223
2224 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2225 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2226         int op, unsigned long time)
2227 {
2228         struct latency_bucket *latency;
2229         int index;
2230
2231         if (!td || td->limit_index != LIMIT_LOW ||
2232             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2233             !blk_queue_nonrot(td->queue))
2234                 return;
2235
2236         index = request_bucket_index(size);
2237
2238         latency = get_cpu_ptr(td->latency_buckets[op]);
2239         latency[index].total_latency += time;
2240         latency[index].samples++;
2241         put_cpu_ptr(td->latency_buckets[op]);
2242 }
2243
2244 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2245 {
2246         struct request_queue *q = rq->q;
2247         struct throtl_data *td = q->td;
2248
2249         throtl_track_latency(td, rq->throtl_size, req_op(rq), time_ns >> 10);
2250 }
2251
2252 void blk_throtl_bio_endio(struct bio *bio)
2253 {
2254         struct blkcg_gq *blkg;
2255         struct throtl_grp *tg;
2256         u64 finish_time_ns;
2257         unsigned long finish_time;
2258         unsigned long start_time;
2259         unsigned long lat;
2260         int rw = bio_data_dir(bio);
2261
2262         blkg = bio->bi_blkg;
2263         if (!blkg)
2264                 return;
2265         tg = blkg_to_tg(blkg);
2266
2267         finish_time_ns = ktime_get_ns();
2268         tg->last_finish_time = finish_time_ns >> 10;
2269
2270         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2271         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2272         if (!start_time || finish_time <= start_time)
2273                 return;
2274
2275         lat = finish_time - start_time;
2276         /* this is only for bio based driver */
2277         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2278                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2279                                      bio_op(bio), lat);
2280
2281         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2282                 int bucket;
2283                 unsigned int threshold;
2284
2285                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2286                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2287                         tg->latency_target;
2288                 if (lat > threshold)
2289                         tg->bad_bio_cnt++;
2290                 /*
2291                  * Not race free, could get wrong count, which means cgroups
2292                  * will be throttled
2293                  */
2294                 tg->bio_cnt++;
2295         }
2296
2297         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2298                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2299                 tg->bio_cnt /= 2;
2300                 tg->bad_bio_cnt /= 2;
2301         }
2302 }
2303 #endif
2304
2305 /*
2306  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2307  * return, @parent_sq is guaranteed to not have any active children tg's
2308  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2309  */
2310 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2311 {
2312         struct throtl_grp *tg;
2313
2314         while ((tg = throtl_rb_first(parent_sq))) {
2315                 struct throtl_service_queue *sq = &tg->service_queue;
2316                 struct bio *bio;
2317
2318                 throtl_dequeue_tg(tg);
2319
2320                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2321                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2322                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2323                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2324         }
2325 }
2326
2327 /**
2328  * blk_throtl_drain - drain throttled bios
2329  * @q: request_queue to drain throttled bios for
2330  *
2331  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2332  */
2333 void blk_throtl_drain(struct request_queue *q)
2334         __releases(&q->queue_lock) __acquires(&q->queue_lock)
2335 {
2336         struct throtl_data *td = q->td;
2337         struct blkcg_gq *blkg;
2338         struct cgroup_subsys_state *pos_css;
2339         struct bio *bio;
2340         int rw;
2341
2342         rcu_read_lock();
2343
2344         /*
2345          * Drain each tg while doing post-order walk on the blkg tree, so
2346          * that all bios are propagated to td->service_queue.  It'd be
2347          * better to walk service_queue tree directly but blkg walk is
2348          * easier.
2349          */
2350         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2351                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2352
2353         /* finally, transfer bios from top-level tg's into the td */
2354         tg_drain_bios(&td->service_queue);
2355
2356         rcu_read_unlock();
2357         spin_unlock_irq(&q->queue_lock);
2358
2359         /* all bios now should be in td->service_queue, issue them */
2360         for (rw = READ; rw <= WRITE; rw++)
2361                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2362                                                 NULL)))
2363                         generic_make_request(bio);
2364
2365         spin_lock_irq(&q->queue_lock);
2366 }
2367
2368 int blk_throtl_init(struct request_queue *q)
2369 {
2370         struct throtl_data *td;
2371         int ret;
2372
2373         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2374         if (!td)
2375                 return -ENOMEM;
2376         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2377                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2378         if (!td->latency_buckets[READ]) {
2379                 kfree(td);
2380                 return -ENOMEM;
2381         }
2382         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2383                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2384         if (!td->latency_buckets[WRITE]) {
2385                 free_percpu(td->latency_buckets[READ]);
2386                 kfree(td);
2387                 return -ENOMEM;
2388         }
2389
2390         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2391         throtl_service_queue_init(&td->service_queue);
2392
2393         q->td = td;
2394         td->queue = q;
2395
2396         td->limit_valid[LIMIT_MAX] = true;
2397         td->limit_index = LIMIT_MAX;
2398         td->low_upgrade_time = jiffies;
2399         td->low_downgrade_time = jiffies;
2400
2401         /* activate policy */
2402         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2403         if (ret) {
2404                 free_percpu(td->latency_buckets[READ]);
2405                 free_percpu(td->latency_buckets[WRITE]);
2406                 kfree(td);
2407         }
2408         return ret;
2409 }
2410
2411 void blk_throtl_exit(struct request_queue *q)
2412 {
2413         BUG_ON(!q->td);
2414         throtl_shutdown_wq(q);
2415         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2416         free_percpu(q->td->latency_buckets[READ]);
2417         free_percpu(q->td->latency_buckets[WRITE]);
2418         kfree(q->td);
2419 }
2420
2421 void blk_throtl_register_queue(struct request_queue *q)
2422 {
2423         struct throtl_data *td;
2424         int i;
2425
2426         td = q->td;
2427         BUG_ON(!td);
2428
2429         if (blk_queue_nonrot(q)) {
2430                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2431                 td->filtered_latency = LATENCY_FILTERED_SSD;
2432         } else {
2433                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2434                 td->filtered_latency = LATENCY_FILTERED_HD;
2435                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2436                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2437                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2438                 }
2439         }
2440 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2441         /* if no low limit, use previous default */
2442         td->throtl_slice = DFL_THROTL_SLICE_HD;
2443 #endif
2444
2445         td->track_bio_latency = !queue_is_mq(q);
2446         if (!td->track_bio_latency)
2447                 blk_stat_enable_accounting(q);
2448 }
2449
2450 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2451 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2452 {
2453         if (!q->td)
2454                 return -EINVAL;
2455         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2456 }
2457
2458 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2459         const char *page, size_t count)
2460 {
2461         unsigned long v;
2462         unsigned long t;
2463
2464         if (!q->td)
2465                 return -EINVAL;
2466         if (kstrtoul(page, 10, &v))
2467                 return -EINVAL;
2468         t = msecs_to_jiffies(v);
2469         if (t == 0 || t > MAX_THROTL_SLICE)
2470                 return -EINVAL;
2471         q->td->throtl_slice = t;
2472         return count;
2473 }
2474 #endif
2475
2476 static int __init throtl_init(void)
2477 {
2478         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2479         if (!kthrotld_workqueue)
2480                 panic("Failed to create kthrotld\n");
2481
2482         return blkcg_policy_register(&blkcg_policy_throtl);
2483 }
2484
2485 module_init(throtl_init);