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 #define SKIP_LATENCY (((u64)1) << BLK_STAT_RES_SHIFT)
  40
  41 static struct blkcg_policy blkcg_policy_throtl;
  42
  43 /* A workqueue to queue throttle related work */
  44 static struct workqueue_struct *kthrotld_workqueue;
  45
  46 /*
  47  * To implement hierarchical throttling, throtl_grps form a tree and bios
  48  * are dispatched upwards level by level until they reach the top and get
  49  * issued.  When dispatching bios from the children and local group at each
  50  * level, if the bios are dispatched into a single bio_list, there's a risk
  51  * of a local or child group which can queue many bios at once filling up
  52  * the list starving others.
  53  *
  54  * To avoid such starvation, dispatched bios are queued separately
  55  * according to where they came from.  When they are again dispatched to
  56  * the parent, they're popped in round-robin order so that no single source
  57  * hogs the dispatch window.
  58  *
  59  * throtl_qnode is used to keep the queued bios separated by their sources.
  60  * Bios are queued to throtl_qnode which in turn is queued to
  61  * throtl_service_queue and then dispatched in round-robin order.
  62  *
  63  * It's also used to track the reference counts on blkg's.  A qnode always
  64  * belongs to a throtl_grp and gets queued on itself or the parent, so
  65  * incrementing the reference of the associated throtl_grp when a qnode is
  66  * queued and decrementing when dequeued is enough to keep the whole blkg
  67  * tree pinned while bios are in flight.
  68  */
  69 struct throtl_qnode {
  70         struct list_head        node;           /* service_queue->queued[] */
  71         struct bio_list         bios;           /* queued bios */
  72         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  73 };
  74
  75 struct throtl_service_queue {
  76         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  77
  78         /*
  79          * Bios queued directly to this service_queue or dispatched from
  80          * children throtl_grp's.
  81          */
  82         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  83         unsigned int            nr_queued[2];   /* number of queued bios */
  84
  85         /*
  86          * RB tree of active children throtl_grp's, which are sorted by
  87          * their ->disptime.
  88          */
  89         struct rb_root          pending_tree;   /* RB tree of active tgs */
  90         struct rb_node          *first_pending; /* first node in the tree */
  91         unsigned int            nr_pending;     /* # queued in the tree */
  92         unsigned long           first_pending_disptime; /* disptime of the first tg */
  93         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  94 };
  95
  96 enum tg_state_flags {
  97         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  98         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  99 };
 100
 101 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
 102
 103 enum {
 104         LIMIT_LOW,
 105         LIMIT_MAX,
 106         LIMIT_CNT,
 107 };
 108
 109 struct throtl_grp {
 110         /* must be the first member */
 111         struct blkg_policy_data pd;
 112
 113         /* active throtl group service_queue member */
 114         struct rb_node rb_node;
 115
 116         /* throtl_data this group belongs to */
 117         struct throtl_data *td;
 118
 119         /* this group's service queue */
 120         struct throtl_service_queue service_queue;
 121
 122         /*
 123          * qnode_on_self is used when bios are directly queued to this
 124          * throtl_grp so that local bios compete fairly with bios
 125          * dispatched from children.  qnode_on_parent is used when bios are
 126          * dispatched from this throtl_grp into its parent and will compete
 127          * with the sibling qnode_on_parents and the parent's
 128          * qnode_on_self.
 129          */
 130         struct throtl_qnode qnode_on_self[2];
 131         struct throtl_qnode qnode_on_parent[2];
 132
 133         /*
 134          * Dispatch time in jiffies. This is the estimated time when group
 135          * will unthrottle and is ready to dispatch more bio. It is used as
 136          * key to sort active groups in service tree.
 137          */
 138         unsigned long disptime;
 139
 140         unsigned int flags;
 141
 142         /* are there any throtl rules between this group and td? */
 143         bool has_rules[2];
 144
 145         /* internally used bytes per second rate limits */
 146         uint64_t bps[2][LIMIT_CNT];
 147         /* user configured bps limits */
 148         uint64_t bps_conf[2][LIMIT_CNT];
 149
 150         /* internally used IOPS limits */
 151         unsigned int iops[2][LIMIT_CNT];
 152         /* user configured IOPS limits */
 153         unsigned int iops_conf[2][LIMIT_CNT];
 154
 155         /* Number of bytes disptached in current slice */
 156         uint64_t bytes_disp[2];
 157         /* Number of bio's dispatched in current slice */
 158         unsigned int io_disp[2];
 159
 160         unsigned long last_low_overflow_time[2];
 161
 162         uint64_t last_bytes_disp[2];
 163         unsigned int last_io_disp[2];
 164
 165         unsigned long last_check_time;
 166
 167         unsigned long latency_target; /* us */
 168         unsigned long latency_target_conf; /* us */
 169         /* When did we start a new slice */
 170         unsigned long slice_start[2];
 171         unsigned long slice_end[2];
 172
 173         unsigned long last_finish_time; /* ns / 1024 */
 174         unsigned long checked_last_finish_time; /* ns / 1024 */
 175         unsigned long avg_idletime; /* ns / 1024 */
 176         unsigned long idletime_threshold; /* us */
 177         unsigned long idletime_threshold_conf; /* us */
 178
 179         unsigned int bio_cnt; /* total bios */
 180         unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
 181         unsigned long bio_cnt_reset_time;
 182 };
 183
 184 /* We measure latency for request size from <= 4k to >= 1M */
 185 #define LATENCY_BUCKET_SIZE 9
 186
 187 struct latency_bucket {
 188         unsigned long total_latency; /* ns / 1024 */
 189         int samples;
 190 };
 191
 192 struct avg_latency_bucket {
 193         unsigned long latency; /* ns / 1024 */
 194         bool valid;
 195 };
 196
 197 struct throtl_data
 198 {
 199         /* service tree for active throtl groups */
 200         struct throtl_service_queue service_queue;
 201
 202         struct request_queue *queue;
 203
 204         /* Total Number of queued bios on READ and WRITE lists */
 205         unsigned int nr_queued[2];
 206
 207         unsigned int throtl_slice;
 208
 209         /* Work for dispatching throttled bios */
 210         struct work_struct dispatch_work;
 211         unsigned int limit_index;
 212         bool limit_valid[LIMIT_CNT];
 213
 214         unsigned long low_upgrade_time;
 215         unsigned long low_downgrade_time;
 216
 217         unsigned int scale;
 218
 219         struct latency_bucket tmp_buckets[LATENCY_BUCKET_SIZE];
 220         struct avg_latency_bucket avg_buckets[LATENCY_BUCKET_SIZE];
 221         struct latency_bucket __percpu *latency_buckets;
 222         unsigned long last_calculate_time;
 223         unsigned long filtered_latency;
 224
 225         bool track_bio_latency;
 226 };
 227
 228 static void throtl_pending_timer_fn(unsigned long arg);
 229
 230 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 231 {
 232         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 233 }
 234
 235 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 236 {
 237         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 238 }
 239
 240 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 241 {
 242         return pd_to_blkg(&tg->pd);
 243 }
 244
 245 /**
 246  * sq_to_tg - return the throl_grp the specified service queue belongs to
 247  * @sq: the throtl_service_queue of interest
 248  *
 249  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 250  * embedded in throtl_data, %NULL is returned.
 251  */
 252 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 253 {
 254         if (sq && sq->parent_sq)
 255                 return container_of(sq, struct throtl_grp, service_queue);
 256         else
 257                 return NULL;
 258 }
 259
 260 /**
 261  * sq_to_td - return throtl_data the specified service queue belongs to
 262  * @sq: the throtl_service_queue of interest
 263  *
 264  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 265  * Determine the associated throtl_data accordingly and return it.
 266  */
 267 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 268 {
 269         struct throtl_grp *tg = sq_to_tg(sq);
 270
 271         if (tg)
 272                 return tg->td;
 273         else
 274                 return container_of(sq, struct throtl_data, service_queue);
 275 }
 276
 277 /*
 278  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 279  * make the IO dispatch more smooth.
 280  * Scale up: linearly scale up according to lapsed time since upgrade. For
 281  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 282  *           limit hits .max limit
 283  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 284  */
 285 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 286 {
 287         /* arbitrary value to avoid too big scale */
 288         if (td->scale < 4096 && time_after_eq(jiffies,
 289             td->low_upgrade_time + td->scale * td->throtl_slice))
 290                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 291
 292         return low + (low >> 1) * td->scale;
 293 }
 294
 295 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 296 {
 297         struct blkcg_gq *blkg = tg_to_blkg(tg);
 298         struct throtl_data *td;
 299         uint64_t ret;
 300
 301         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 302                 return U64_MAX;
 303
 304         td = tg->td;
 305         ret = tg->bps[rw][td->limit_index];
 306         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 307                 /* intermediate node or iops isn't 0 */
 308                 if (!list_empty(&blkg->blkcg->css.children) ||
 309                     tg->iops[rw][td->limit_index])
 310                         return U64_MAX;
 311                 else
 312                         return MIN_THROTL_BPS;
 313         }
 314
 315         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 316             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 317                 uint64_t adjusted;
 318
 319                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 320                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 321         }
 322         return ret;
 323 }
 324
 325 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 326 {
 327         struct blkcg_gq *blkg = tg_to_blkg(tg);
 328         struct throtl_data *td;
 329         unsigned int ret;
 330
 331         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 332                 return UINT_MAX;
 333
 334         td = tg->td;
 335         ret = tg->iops[rw][td->limit_index];
 336         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 337                 /* intermediate node or bps isn't 0 */
 338                 if (!list_empty(&blkg->blkcg->css.children) ||
 339                     tg->bps[rw][td->limit_index])
 340                         return UINT_MAX;
 341                 else
 342                         return MIN_THROTL_IOPS;
 343         }
 344
 345         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 346             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 347                 uint64_t adjusted;
 348
 349                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 350                 if (adjusted > UINT_MAX)
 351                         adjusted = UINT_MAX;
 352                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 353         }
 354         return ret;
 355 }
 356
 357 #define request_bucket_index(sectors) \
 358         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 359
 360 /**
 361  * throtl_log - log debug message via blktrace
 362  * @sq: the service_queue being reported
 363  * @fmt: printf format string
 364  * @args: printf args
 365  *
 366  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 367  * throtl_grp; otherwise, just "throtl".
 368  */
 369 #define throtl_log(sq, fmt, args...)    do {                            \
 370         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 371         struct throtl_data *__td = sq_to_td((sq));                      \
 372                                                                         \
 373         (void)__td;                                                     \
 374         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 375                 break;                                                  \
 376         if ((__tg)) {                                                   \
 377                 blk_add_cgroup_trace_msg(__td->queue,                   \
 378                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 379         } else {                                                        \
 380                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 381         }                                                               \
 382 } while (0)
 383
 384 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 385 {
 386         /* assume it's one sector */
 387         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 388                 return 512;
 389         return bio->bi_iter.bi_size;
 390 }
 391
 392 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 393 {
 394         INIT_LIST_HEAD(&qn->node);
 395         bio_list_init(&qn->bios);
 396         qn->tg = tg;
 397 }
 398
 399 /**
 400  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 401  * @bio: bio being added
 402  * @qn: qnode to add bio to
 403  * @queued: the service_queue->queued[] list @qn belongs to
 404  *
 405  * Add @bio to @qn and put @qn on @queued if it's not already on.
 406  * @qn->tg's reference count is bumped when @qn is activated.  See the
 407  * comment on top of throtl_qnode definition for details.
 408  */
 409 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 410                                  struct list_head *queued)
 411 {
 412         bio_list_add(&qn->bios, bio);
 413         if (list_empty(&qn->node)) {
 414                 list_add_tail(&qn->node, queued);
 415                 blkg_get(tg_to_blkg(qn->tg));
 416         }
 417 }
 418
 419 /**
 420  * throtl_peek_queued - peek the first bio on a qnode list
 421  * @queued: the qnode list to peek
 422  */
 423 static struct bio *throtl_peek_queued(struct list_head *queued)
 424 {
 425         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 426         struct bio *bio;
 427
 428         if (list_empty(queued))
 429                 return NULL;
 430
 431         bio = bio_list_peek(&qn->bios);
 432         WARN_ON_ONCE(!bio);
 433         return bio;
 434 }
 435
 436 /**
 437  * throtl_pop_queued - pop the first bio form a qnode list
 438  * @queued: the qnode list to pop a bio from
 439  * @tg_to_put: optional out argument for throtl_grp to put
 440  *
 441  * Pop the first bio from the qnode list @queued.  After popping, the first
 442  * qnode is removed from @queued if empty or moved to the end of @queued so
 443  * that the popping order is round-robin.
 444  *
 445  * When the first qnode is removed, its associated throtl_grp should be put
 446  * too.  If @tg_to_put is NULL, this function automatically puts it;
 447  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 448  * responsible for putting it.
 449  */
 450 static struct bio *throtl_pop_queued(struct list_head *queued,
 451                                      struct throtl_grp **tg_to_put)
 452 {
 453         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 454         struct bio *bio;
 455
 456         if (list_empty(queued))
 457                 return NULL;
 458
 459         bio = bio_list_pop(&qn->bios);
 460         WARN_ON_ONCE(!bio);
 461
 462         if (bio_list_empty(&qn->bios)) {
 463                 list_del_init(&qn->node);
 464                 if (tg_to_put)
 465                         *tg_to_put = qn->tg;
 466                 else
 467                         blkg_put(tg_to_blkg(qn->tg));
 468         } else {
 469                 list_move_tail(&qn->node, queued);
 470         }
 471
 472         return bio;
 473 }
 474
 475 /* init a service_queue, assumes the caller zeroed it */
 476 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 477 {
 478         INIT_LIST_HEAD(&sq->queued[0]);
 479         INIT_LIST_HEAD(&sq->queued[1]);
 480         sq->pending_tree = RB_ROOT;
 481         setup_timer(&sq->pending_timer, throtl_pending_timer_fn,
 482                     (unsigned long)sq);
 483 }
 484
 485 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
 486 {
 487         struct throtl_grp *tg;
 488         int rw;
 489
 490         tg = kzalloc_node(sizeof(*tg), gfp, node);
 491         if (!tg)
 492                 return NULL;
 493
 494         throtl_service_queue_init(&tg->service_queue);
 495
 496         for (rw = READ; rw <= WRITE; rw++) {
 497                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 498                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 499         }
 500
 501         RB_CLEAR_NODE(&tg->rb_node);
 502         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 503         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 504         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 505         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 506         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 507         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 508         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 509         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 510         /* LIMIT_LOW will have default value 0 */
 511
 512         tg->latency_target = DFL_LATENCY_TARGET;
 513         tg->latency_target_conf = DFL_LATENCY_TARGET;
 514         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 515         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 516
 517         return &tg->pd;
 518 }
 519
 520 static void throtl_pd_init(struct blkg_policy_data *pd)
 521 {
 522         struct throtl_grp *tg = pd_to_tg(pd);
 523         struct blkcg_gq *blkg = tg_to_blkg(tg);
 524         struct throtl_data *td = blkg->q->td;
 525         struct throtl_service_queue *sq = &tg->service_queue;
 526
 527         /*
 528          * If on the default hierarchy, we switch to properly hierarchical
 529          * behavior where limits on a given throtl_grp are applied to the
 530          * whole subtree rather than just the group itself.  e.g. If 16M
 531          * read_bps limit is set on the root group, the whole system can't
 532          * exceed 16M for the device.
 533          *
 534          * If not on the default hierarchy, the broken flat hierarchy
 535          * behavior is retained where all throtl_grps are treated as if
 536          * they're all separate root groups right below throtl_data.
 537          * Limits of a group don't interact with limits of other groups
 538          * regardless of the position of the group in the hierarchy.
 539          */
 540         sq->parent_sq = &td->service_queue;
 541         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 542                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 543         tg->td = td;
 544 }
 545
 546 /*
 547  * Set has_rules[] if @tg or any of its parents have limits configured.
 548  * This doesn't require walking up to the top of the hierarchy as the
 549  * parent's has_rules[] is guaranteed to be correct.
 550  */
 551 static void tg_update_has_rules(struct throtl_grp *tg)
 552 {
 553         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 554         struct throtl_data *td = tg->td;
 555         int rw;
 556
 557         for (rw = READ; rw <= WRITE; rw++)
 558                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 559                         (td->limit_valid[td->limit_index] &&
 560                          (tg_bps_limit(tg, rw) != U64_MAX ||
 561                           tg_iops_limit(tg, rw) != UINT_MAX));
 562 }
 563
 564 static void throtl_pd_online(struct blkg_policy_data *pd)
 565 {
 566         struct throtl_grp *tg = pd_to_tg(pd);
 567         /*
 568          * We don't want new groups to escape the limits of its ancestors.
 569          * Update has_rules[] after a new group is brought online.
 570          */
 571         tg_update_has_rules(tg);
 572 }
 573
 574 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 575 {
 576         struct cgroup_subsys_state *pos_css;
 577         struct blkcg_gq *blkg;
 578         bool low_valid = false;
 579
 580         rcu_read_lock();
 581         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 582                 struct throtl_grp *tg = blkg_to_tg(blkg);
 583
 584                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 585                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
 586                         low_valid = true;
 587         }
 588         rcu_read_unlock();
 589
 590         td->limit_valid[LIMIT_LOW] = low_valid;
 591 }
 592
 593 static void throtl_upgrade_state(struct throtl_data *td);
 594 static void throtl_pd_offline(struct blkg_policy_data *pd)
 595 {
 596         struct throtl_grp *tg = pd_to_tg(pd);
 597
 598         tg->bps[READ][LIMIT_LOW] = 0;
 599         tg->bps[WRITE][LIMIT_LOW] = 0;
 600         tg->iops[READ][LIMIT_LOW] = 0;
 601         tg->iops[WRITE][LIMIT_LOW] = 0;
 602
 603         blk_throtl_update_limit_valid(tg->td);
 604
 605         if (!tg->td->limit_valid[tg->td->limit_index])
 606                 throtl_upgrade_state(tg->td);
 607 }
 608
 609 static void throtl_pd_free(struct blkg_policy_data *pd)
 610 {
 611         struct throtl_grp *tg = pd_to_tg(pd);
 612
 613         del_timer_sync(&tg->service_queue.pending_timer);
 614         kfree(tg);
 615 }
 616
 617 static struct throtl_grp *
 618 throtl_rb_first(struct throtl_service_queue *parent_sq)
 619 {
 620         /* Service tree is empty */
 621         if (!parent_sq->nr_pending)
 622                 return NULL;
 623
 624         if (!parent_sq->first_pending)
 625                 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
 626
 627         if (parent_sq->first_pending)
 628                 return rb_entry_tg(parent_sq->first_pending);
 629
 630         return NULL;
 631 }
 632
 633 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 634 {
 635         rb_erase(n, root);
 636         RB_CLEAR_NODE(n);
 637 }
 638
 639 static void throtl_rb_erase(struct rb_node *n,
 640                             struct throtl_service_queue *parent_sq)
 641 {
 642         if (parent_sq->first_pending == n)
 643                 parent_sq->first_pending = NULL;
 644         rb_erase_init(n, &parent_sq->pending_tree);
 645         --parent_sq->nr_pending;
 646 }
 647
 648 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 649 {
 650         struct throtl_grp *tg;
 651
 652         tg = throtl_rb_first(parent_sq);
 653         if (!tg)
 654                 return;
 655
 656         parent_sq->first_pending_disptime = tg->disptime;
 657 }
 658
 659 static void tg_service_queue_add(struct throtl_grp *tg)
 660 {
 661         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 662         struct rb_node **node = &parent_sq->pending_tree.rb_node;
 663         struct rb_node *parent = NULL;
 664         struct throtl_grp *__tg;
 665         unsigned long key = tg->disptime;
 666         int left = 1;
 667
 668         while (*node != NULL) {
 669                 parent = *node;
 670                 __tg = rb_entry_tg(parent);
 671
 672                 if (time_before(key, __tg->disptime))
 673                         node = &parent->rb_left;
 674                 else {
 675                         node = &parent->rb_right;
 676                         left = 0;
 677                 }
 678         }
 679
 680         if (left)
 681                 parent_sq->first_pending = &tg->rb_node;
 682
 683         rb_link_node(&tg->rb_node, parent, node);
 684         rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
 685 }
 686
 687 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 688 {
 689         tg_service_queue_add(tg);
 690         tg->flags |= THROTL_TG_PENDING;
 691         tg->service_queue.parent_sq->nr_pending++;
 692 }
 693
 694 static void throtl_enqueue_tg(struct throtl_grp *tg)
 695 {
 696         if (!(tg->flags & THROTL_TG_PENDING))
 697                 __throtl_enqueue_tg(tg);
 698 }
 699
 700 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 701 {
 702         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 703         tg->flags &= ~THROTL_TG_PENDING;
 704 }
 705
 706 static void throtl_dequeue_tg(struct throtl_grp *tg)
 707 {
 708         if (tg->flags & THROTL_TG_PENDING)
 709                 __throtl_dequeue_tg(tg);
 710 }
 711
 712 /* Call with queue lock held */
 713 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 714                                           unsigned long expires)
 715 {
 716         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 717
 718         /*
 719          * Since we are adjusting the throttle limit dynamically, the sleep
 720          * time calculated according to previous limit might be invalid. It's
 721          * possible the cgroup sleep time is very long and no other cgroups
 722          * have IO running so notify the limit changes. Make sure the cgroup
 723          * doesn't sleep too long to avoid the missed notification.
 724          */
 725         if (time_after(expires, max_expire))
 726                 expires = max_expire;
 727         mod_timer(&sq->pending_timer, expires);
 728         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 729                    expires - jiffies, jiffies);
 730 }
 731
 732 /**
 733  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 734  * @sq: the service_queue to schedule dispatch for
 735  * @force: force scheduling
 736  *
 737  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 738  * dispatch time of the first pending child.  Returns %true if either timer
 739  * is armed or there's no pending child left.  %false if the current
 740  * dispatch window is still open and the caller should continue
 741  * dispatching.
 742  *
 743  * If @force is %true, the dispatch timer is always scheduled and this
 744  * function is guaranteed to return %true.  This is to be used when the
 745  * caller can't dispatch itself and needs to invoke pending_timer
 746  * unconditionally.  Note that forced scheduling is likely to induce short
 747  * delay before dispatch starts even if @sq->first_pending_disptime is not
 748  * in the future and thus shouldn't be used in hot paths.
 749  */
 750 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 751                                           bool force)
 752 {
 753         /* any pending children left? */
 754         if (!sq->nr_pending)
 755                 return true;
 756
 757         update_min_dispatch_time(sq);
 758
 759         /* is the next dispatch time in the future? */
 760         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 761                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 762                 return true;
 763         }
 764
 765         /* tell the caller to continue dispatching */
 766         return false;
 767 }
 768
 769 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 770                 bool rw, unsigned long start)
 771 {
 772         tg->bytes_disp[rw] = 0;
 773         tg->io_disp[rw] = 0;
 774
 775         /*
 776          * Previous slice has expired. We must have trimmed it after last
 777          * bio dispatch. That means since start of last slice, we never used
 778          * that bandwidth. Do try to make use of that bandwidth while giving
 779          * credit.
 780          */
 781         if (time_after_eq(start, tg->slice_start[rw]))
 782                 tg->slice_start[rw] = start;
 783
 784         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 785         throtl_log(&tg->service_queue,
 786                    "[%c] new slice with credit 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_start_new_slice(struct throtl_grp *tg, bool rw)
 792 {
 793         tg->bytes_disp[rw] = 0;
 794         tg->io_disp[rw] = 0;
 795         tg->slice_start[rw] = jiffies;
 796         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 797         throtl_log(&tg->service_queue,
 798                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 799                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 800                    tg->slice_end[rw], jiffies);
 801 }
 802
 803 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 804                                         unsigned long jiffy_end)
 805 {
 806         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 807 }
 808
 809 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 810                                        unsigned long jiffy_end)
 811 {
 812         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 813         throtl_log(&tg->service_queue,
 814                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 815                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 816                    tg->slice_end[rw], jiffies);
 817 }
 818
 819 /* Determine if previously allocated or extended slice is complete or not */
 820 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 821 {
 822         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 823                 return false;
 824
 825         return 1;
 826 }
 827
 828 /* Trim the used slices and adjust slice start accordingly */
 829 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 830 {
 831         unsigned long nr_slices, time_elapsed, io_trim;
 832         u64 bytes_trim, tmp;
 833
 834         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 835
 836         /*
 837          * If bps are unlimited (-1), then time slice don't get
 838          * renewed. Don't try to trim the slice if slice is used. A new
 839          * slice will start when appropriate.
 840          */
 841         if (throtl_slice_used(tg, rw))
 842                 return;
 843
 844         /*
 845          * A bio has been dispatched. Also adjust slice_end. It might happen
 846          * that initially cgroup limit was very low resulting in high
 847          * slice_end, but later limit was bumped up and bio was dispached
 848          * sooner, then we need to reduce slice_end. A high bogus slice_end
 849          * is bad because it does not allow new slice to start.
 850          */
 851
 852         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 853
 854         time_elapsed = jiffies - tg->slice_start[rw];
 855
 856         nr_slices = time_elapsed / tg->td->throtl_slice;
 857
 858         if (!nr_slices)
 859                 return;
 860         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 861         do_div(tmp, HZ);
 862         bytes_trim = tmp;
 863
 864         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 865                 HZ;
 866
 867         if (!bytes_trim && !io_trim)
 868                 return;
 869
 870         if (tg->bytes_disp[rw] >= bytes_trim)
 871                 tg->bytes_disp[rw] -= bytes_trim;
 872         else
 873                 tg->bytes_disp[rw] = 0;
 874
 875         if (tg->io_disp[rw] >= io_trim)
 876                 tg->io_disp[rw] -= io_trim;
 877         else
 878                 tg->io_disp[rw] = 0;
 879
 880         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 881
 882         throtl_log(&tg->service_queue,
 883                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 884                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 885                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 886 }
 887
 888 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 889                                   unsigned long *wait)
 890 {
 891         bool rw = bio_data_dir(bio);
 892         unsigned int io_allowed;
 893         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 894         u64 tmp;
 895
 896         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 897
 898         /* Slice has just started. Consider one slice interval */
 899         if (!jiffy_elapsed)
 900                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 901
 902         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 903
 904         /*
 905          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 906          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 907          * will allow dispatch after 1 second and after that slice should
 908          * have been trimmed.
 909          */
 910
 911         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 912         do_div(tmp, HZ);
 913
 914         if (tmp > UINT_MAX)
 915                 io_allowed = UINT_MAX;
 916         else
 917                 io_allowed = tmp;
 918
 919         if (tg->io_disp[rw] + 1 <= io_allowed) {
 920                 if (wait)
 921                         *wait = 0;
 922                 return true;
 923         }
 924
 925         /* Calc approx time to dispatch */
 926         jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
 927
 928         if (jiffy_wait > jiffy_elapsed)
 929                 jiffy_wait = jiffy_wait - jiffy_elapsed;
 930         else
 931                 jiffy_wait = 1;
 932
 933         if (wait)
 934                 *wait = jiffy_wait;
 935         return 0;
 936 }
 937
 938 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 939                                  unsigned long *wait)
 940 {
 941         bool rw = bio_data_dir(bio);
 942         u64 bytes_allowed, extra_bytes, tmp;
 943         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 944         unsigned int bio_size = throtl_bio_data_size(bio);
 945
 946         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 947
 948         /* Slice has just started. Consider one slice interval */
 949         if (!jiffy_elapsed)
 950                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 951
 952         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 953
 954         tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
 955         do_div(tmp, HZ);
 956         bytes_allowed = tmp;
 957
 958         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 959                 if (wait)
 960                         *wait = 0;
 961                 return true;
 962         }
 963
 964         /* Calc approx time to dispatch */
 965         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 966         jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
 967
 968         if (!jiffy_wait)
 969                 jiffy_wait = 1;
 970
 971         /*
 972          * This wait time is without taking into consideration the rounding
 973          * up we did. Add that time also.
 974          */
 975         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 976         if (wait)
 977                 *wait = jiffy_wait;
 978         return 0;
 979 }
 980
 981 /*
 982  * Returns whether one can dispatch a bio or not. Also returns approx number
 983  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 984  */
 985 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 986                             unsigned long *wait)
 987 {
 988         bool rw = bio_data_dir(bio);
 989         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 990
 991         /*
 992          * Currently whole state machine of group depends on first bio
 993          * queued in the group bio list. So one should not be calling
 994          * this function with a different bio if there are other bios
 995          * queued.
 996          */
 997         BUG_ON(tg->service_queue.nr_queued[rw] &&
 998                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 999
1000         /* If tg->bps = -1, then BW is unlimited */
1001         if (tg_bps_limit(tg, rw) == U64_MAX &&
1002             tg_iops_limit(tg, rw) == UINT_MAX) {
1003                 if (wait)
1004                         *wait = 0;
1005                 return true;
1006         }
1007
1008         /*
1009          * If previous slice expired, start a new one otherwise renew/extend
1010          * existing slice to make sure it is at least throtl_slice interval
1011          * long since now. New slice is started only for empty throttle group.
1012          * If there is queued bio, that means there should be an active
1013          * slice and it should be extended instead.
1014          */
1015         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1016                 throtl_start_new_slice(tg, rw);
1017         else {
1018                 if (time_before(tg->slice_end[rw],
1019                     jiffies + tg->td->throtl_slice))
1020                         throtl_extend_slice(tg, rw,
1021                                 jiffies + tg->td->throtl_slice);
1022         }
1023
1024         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1025             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1026                 if (wait)
1027                         *wait = 0;
1028                 return 1;
1029         }
1030
1031         max_wait = max(bps_wait, iops_wait);
1032
1033         if (wait)
1034                 *wait = max_wait;
1035
1036         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1037                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1038
1039         return 0;
1040 }
1041
1042 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1043 {
1044         bool rw = bio_data_dir(bio);
1045         unsigned int bio_size = throtl_bio_data_size(bio);
1046
1047         /* Charge the bio to the group */
1048         tg->bytes_disp[rw] += bio_size;
1049         tg->io_disp[rw]++;
1050         tg->last_bytes_disp[rw] += bio_size;
1051         tg->last_io_disp[rw]++;
1052
1053         /*
1054          * BIO_THROTTLED is used to prevent the same bio to be throttled
1055          * more than once as a throttled bio will go through blk-throtl the
1056          * second time when it eventually gets issued.  Set it when a bio
1057          * is being charged to a tg.
1058          */
1059         if (!bio_flagged(bio, BIO_THROTTLED))
1060                 bio_set_flag(bio, BIO_THROTTLED);
1061 }
1062
1063 /**
1064  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1065  * @bio: bio to add
1066  * @qn: qnode to use
1067  * @tg: the target throtl_grp
1068  *
1069  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1070  * tg->qnode_on_self[] is used.
1071  */
1072 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1073                               struct throtl_grp *tg)
1074 {
1075         struct throtl_service_queue *sq = &tg->service_queue;
1076         bool rw = bio_data_dir(bio);
1077
1078         if (!qn)
1079                 qn = &tg->qnode_on_self[rw];
1080
1081         /*
1082          * If @tg doesn't currently have any bios queued in the same
1083          * direction, queueing @bio can change when @tg should be
1084          * dispatched.  Mark that @tg was empty.  This is automatically
1085          * cleaered on the next tg_update_disptime().
1086          */
1087         if (!sq->nr_queued[rw])
1088                 tg->flags |= THROTL_TG_WAS_EMPTY;
1089
1090         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1091
1092         sq->nr_queued[rw]++;
1093         throtl_enqueue_tg(tg);
1094 }
1095
1096 static void tg_update_disptime(struct throtl_grp *tg)
1097 {
1098         struct throtl_service_queue *sq = &tg->service_queue;
1099         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1100         struct bio *bio;
1101
1102         bio = throtl_peek_queued(&sq->queued[READ]);
1103         if (bio)
1104                 tg_may_dispatch(tg, bio, &read_wait);
1105
1106         bio = throtl_peek_queued(&sq->queued[WRITE]);
1107         if (bio)
1108                 tg_may_dispatch(tg, bio, &write_wait);
1109
1110         min_wait = min(read_wait, write_wait);
1111         disptime = jiffies + min_wait;
1112
1113         /* Update dispatch time */
1114         throtl_dequeue_tg(tg);
1115         tg->disptime = disptime;
1116         throtl_enqueue_tg(tg);
1117
1118         /* see throtl_add_bio_tg() */
1119         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1120 }
1121
1122 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1123                                         struct throtl_grp *parent_tg, bool rw)
1124 {
1125         if (throtl_slice_used(parent_tg, rw)) {
1126                 throtl_start_new_slice_with_credit(parent_tg, rw,
1127                                 child_tg->slice_start[rw]);
1128         }
1129
1130 }
1131
1132 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1133 {
1134         struct throtl_service_queue *sq = &tg->service_queue;
1135         struct throtl_service_queue *parent_sq = sq->parent_sq;
1136         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1137         struct throtl_grp *tg_to_put = NULL;
1138         struct bio *bio;
1139
1140         /*
1141          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1142          * from @tg may put its reference and @parent_sq might end up
1143          * getting released prematurely.  Remember the tg to put and put it
1144          * after @bio is transferred to @parent_sq.
1145          */
1146         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1147         sq->nr_queued[rw]--;
1148
1149         throtl_charge_bio(tg, bio);
1150
1151         /*
1152          * If our parent is another tg, we just need to transfer @bio to
1153          * the parent using throtl_add_bio_tg().  If our parent is
1154          * @td->service_queue, @bio is ready to be issued.  Put it on its
1155          * bio_lists[] and decrease total number queued.  The caller is
1156          * responsible for issuing these bios.
1157          */
1158         if (parent_tg) {
1159                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1160                 start_parent_slice_with_credit(tg, parent_tg, rw);
1161         } else {
1162                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1163                                      &parent_sq->queued[rw]);
1164                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1165                 tg->td->nr_queued[rw]--;
1166         }
1167
1168         throtl_trim_slice(tg, rw);
1169
1170         if (tg_to_put)
1171                 blkg_put(tg_to_blkg(tg_to_put));
1172 }
1173
1174 static int throtl_dispatch_tg(struct throtl_grp *tg)
1175 {
1176         struct throtl_service_queue *sq = &tg->service_queue;
1177         unsigned int nr_reads = 0, nr_writes = 0;
1178         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1179         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1180         struct bio *bio;
1181
1182         /* Try to dispatch 75% READS and 25% WRITES */
1183
1184         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1185                tg_may_dispatch(tg, bio, NULL)) {
1186
1187                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1188                 nr_reads++;
1189
1190                 if (nr_reads >= max_nr_reads)
1191                         break;
1192         }
1193
1194         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1195                tg_may_dispatch(tg, bio, NULL)) {
1196
1197                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1198                 nr_writes++;
1199
1200                 if (nr_writes >= max_nr_writes)
1201                         break;
1202         }
1203
1204         return nr_reads + nr_writes;
1205 }
1206
1207 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1208 {
1209         unsigned int nr_disp = 0;
1210
1211         while (1) {
1212                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1213                 struct throtl_service_queue *sq = &tg->service_queue;
1214
1215                 if (!tg)
1216                         break;
1217
1218                 if (time_before(jiffies, tg->disptime))
1219                         break;
1220
1221                 throtl_dequeue_tg(tg);
1222
1223                 nr_disp += throtl_dispatch_tg(tg);
1224
1225                 if (sq->nr_queued[0] || sq->nr_queued[1])
1226                         tg_update_disptime(tg);
1227
1228                 if (nr_disp >= throtl_quantum)
1229                         break;
1230         }
1231
1232         return nr_disp;
1233 }
1234
1235 static bool throtl_can_upgrade(struct throtl_data *td,
1236         struct throtl_grp *this_tg);
1237 /**
1238  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1239  * @arg: the throtl_service_queue being serviced
1240  *
1241  * This timer is armed when a child throtl_grp with active bio's become
1242  * pending and queued on the service_queue's pending_tree and expires when
1243  * the first child throtl_grp should be dispatched.  This function
1244  * dispatches bio's from the children throtl_grps to the parent
1245  * service_queue.
1246  *
1247  * If the parent's parent is another throtl_grp, dispatching is propagated
1248  * by either arming its pending_timer or repeating dispatch directly.  If
1249  * the top-level service_tree is reached, throtl_data->dispatch_work is
1250  * kicked so that the ready bio's are issued.
1251  */
1252 static void throtl_pending_timer_fn(unsigned long arg)
1253 {
1254         struct throtl_service_queue *sq = (void *)arg;
1255         struct throtl_grp *tg = sq_to_tg(sq);
1256         struct throtl_data *td = sq_to_td(sq);
1257         struct request_queue *q = td->queue;
1258         struct throtl_service_queue *parent_sq;
1259         bool dispatched;
1260         int ret;
1261
1262         spin_lock_irq(q->queue_lock);
1263         if (throtl_can_upgrade(td, NULL))
1264                 throtl_upgrade_state(td);
1265
1266 again:
1267         parent_sq = sq->parent_sq;
1268         dispatched = false;
1269
1270         while (true) {
1271                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1272                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1273                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1274
1275                 ret = throtl_select_dispatch(sq);
1276                 if (ret) {
1277                         throtl_log(sq, "bios disp=%u", ret);
1278                         dispatched = true;
1279                 }
1280
1281                 if (throtl_schedule_next_dispatch(sq, false))
1282                         break;
1283
1284                 /* this dispatch windows is still open, relax and repeat */
1285                 spin_unlock_irq(q->queue_lock);
1286                 cpu_relax();
1287                 spin_lock_irq(q->queue_lock);
1288         }
1289
1290         if (!dispatched)
1291                 goto out_unlock;
1292
1293         if (parent_sq) {
1294                 /* @parent_sq is another throl_grp, propagate dispatch */
1295                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1296                         tg_update_disptime(tg);
1297                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1298                                 /* window is already open, repeat dispatching */
1299                                 sq = parent_sq;
1300                                 tg = sq_to_tg(sq);
1301                                 goto again;
1302                         }
1303                 }
1304         } else {
1305                 /* reached the top-level, queue issueing */
1306                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1307         }
1308 out_unlock:
1309         spin_unlock_irq(q->queue_lock);
1310 }
1311
1312 /**
1313  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1314  * @work: work item being executed
1315  *
1316  * This function is queued for execution when bio's reach the bio_lists[]
1317  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1318  * function.
1319  */
1320 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1321 {
1322         struct throtl_data *td = container_of(work, struct throtl_data,
1323                                               dispatch_work);
1324         struct throtl_service_queue *td_sq = &td->service_queue;
1325         struct request_queue *q = td->queue;
1326         struct bio_list bio_list_on_stack;
1327         struct bio *bio;
1328         struct blk_plug plug;
1329         int rw;
1330
1331         bio_list_init(&bio_list_on_stack);
1332
1333         spin_lock_irq(q->queue_lock);
1334         for (rw = READ; rw <= WRITE; rw++)
1335                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1336                         bio_list_add(&bio_list_on_stack, bio);
1337         spin_unlock_irq(q->queue_lock);
1338
1339         if (!bio_list_empty(&bio_list_on_stack)) {
1340                 blk_start_plug(&plug);
1341                 while((bio = bio_list_pop(&bio_list_on_stack)))
1342                         generic_make_request(bio);
1343                 blk_finish_plug(&plug);
1344         }
1345 }
1346
1347 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1348                               int off)
1349 {
1350         struct throtl_grp *tg = pd_to_tg(pd);
1351         u64 v = *(u64 *)((void *)tg + off);
1352
1353         if (v == U64_MAX)
1354                 return 0;
1355         return __blkg_prfill_u64(sf, pd, v);
1356 }
1357
1358 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1359                                int off)
1360 {
1361         struct throtl_grp *tg = pd_to_tg(pd);
1362         unsigned int v = *(unsigned int *)((void *)tg + off);
1363
1364         if (v == UINT_MAX)
1365                 return 0;
1366         return __blkg_prfill_u64(sf, pd, v);
1367 }
1368
1369 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1370 {
1371         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1372                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1373         return 0;
1374 }
1375
1376 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1377 {
1378         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1379                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1380         return 0;
1381 }
1382
1383 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1384 {
1385         struct throtl_service_queue *sq = &tg->service_queue;
1386         struct cgroup_subsys_state *pos_css;
1387         struct blkcg_gq *blkg;
1388
1389         throtl_log(&tg->service_queue,
1390                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1391                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1392                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1393
1394         /*
1395          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1396          * considered to have rules if either the tg itself or any of its
1397          * ancestors has rules.  This identifies groups without any
1398          * restrictions in the whole hierarchy and allows them to bypass
1399          * blk-throttle.
1400          */
1401         blkg_for_each_descendant_pre(blkg, pos_css,
1402                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1403                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1404                 struct throtl_grp *parent_tg;
1405
1406                 tg_update_has_rules(this_tg);
1407                 /* ignore root/second level */
1408                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1409                     !blkg->parent->parent)
1410                         continue;
1411                 parent_tg = blkg_to_tg(blkg->parent);
1412                 /*
1413                  * make sure all children has lower idle time threshold and
1414                  * higher latency target
1415                  */
1416                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1417                                 parent_tg->idletime_threshold);
1418                 this_tg->latency_target = max(this_tg->latency_target,
1419                                 parent_tg->latency_target);
1420         }
1421
1422         /*
1423          * We're already holding queue_lock and know @tg is valid.  Let's
1424          * apply the new config directly.
1425          *
1426          * Restart the slices for both READ and WRITES. It might happen
1427          * that a group's limit are dropped suddenly and we don't want to
1428          * account recently dispatched IO with new low rate.
1429          */
1430         throtl_start_new_slice(tg, 0);
1431         throtl_start_new_slice(tg, 1);
1432
1433         if (tg->flags & THROTL_TG_PENDING) {
1434                 tg_update_disptime(tg);
1435                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1436         }
1437 }
1438
1439 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1440                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1441 {
1442         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1443         struct blkg_conf_ctx ctx;
1444         struct throtl_grp *tg;
1445         int ret;
1446         u64 v;
1447
1448         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1449         if (ret)
1450                 return ret;
1451
1452         ret = -EINVAL;
1453         if (sscanf(ctx.body, "%llu", &v) != 1)
1454                 goto out_finish;
1455         if (!v)
1456                 v = U64_MAX;
1457
1458         tg = blkg_to_tg(ctx.blkg);
1459
1460         if (is_u64)
1461                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1462         else
1463                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1464
1465         tg_conf_updated(tg, false);
1466         ret = 0;
1467 out_finish:
1468         blkg_conf_finish(&ctx);
1469         return ret ?: nbytes;
1470 }
1471
1472 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1473                                char *buf, size_t nbytes, loff_t off)
1474 {
1475         return tg_set_conf(of, buf, nbytes, off, true);
1476 }
1477
1478 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1479                                 char *buf, size_t nbytes, loff_t off)
1480 {
1481         return tg_set_conf(of, buf, nbytes, off, false);
1482 }
1483
1484 static struct cftype throtl_legacy_files[] = {
1485         {
1486                 .name = "throttle.read_bps_device",
1487                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1488                 .seq_show = tg_print_conf_u64,
1489                 .write = tg_set_conf_u64,
1490         },
1491         {
1492                 .name = "throttle.write_bps_device",
1493                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1494                 .seq_show = tg_print_conf_u64,
1495                 .write = tg_set_conf_u64,
1496         },
1497         {
1498                 .name = "throttle.read_iops_device",
1499                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1500                 .seq_show = tg_print_conf_uint,
1501                 .write = tg_set_conf_uint,
1502         },
1503         {
1504                 .name = "throttle.write_iops_device",
1505                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1506                 .seq_show = tg_print_conf_uint,
1507                 .write = tg_set_conf_uint,
1508         },
1509         {
1510                 .name = "throttle.io_service_bytes",
1511                 .private = (unsigned long)&blkcg_policy_throtl,
1512                 .seq_show = blkg_print_stat_bytes,
1513         },
1514         {
1515                 .name = "throttle.io_serviced",
1516                 .private = (unsigned long)&blkcg_policy_throtl,
1517                 .seq_show = blkg_print_stat_ios,
1518         },
1519         { }     /* terminate */
1520 };
1521
1522 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1523                          int off)
1524 {
1525         struct throtl_grp *tg = pd_to_tg(pd);
1526         const char *dname = blkg_dev_name(pd->blkg);
1527         char bufs[4][21] = { "max", "max", "max", "max" };
1528         u64 bps_dft;
1529         unsigned int iops_dft;
1530         char idle_time[26] = "";
1531         char latency_time[26] = "";
1532
1533         if (!dname)
1534                 return 0;
1535
1536         if (off == LIMIT_LOW) {
1537                 bps_dft = 0;
1538                 iops_dft = 0;
1539         } else {
1540                 bps_dft = U64_MAX;
1541                 iops_dft = UINT_MAX;
1542         }
1543
1544         if (tg->bps_conf[READ][off] == bps_dft &&
1545             tg->bps_conf[WRITE][off] == bps_dft &&
1546             tg->iops_conf[READ][off] == iops_dft &&
1547             tg->iops_conf[WRITE][off] == iops_dft &&
1548             (off != LIMIT_LOW ||
1549              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1550               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1551                 return 0;
1552
1553         if (tg->bps_conf[READ][off] != U64_MAX)
1554                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1555                         tg->bps_conf[READ][off]);
1556         if (tg->bps_conf[WRITE][off] != U64_MAX)
1557                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1558                         tg->bps_conf[WRITE][off]);
1559         if (tg->iops_conf[READ][off] != UINT_MAX)
1560                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1561                         tg->iops_conf[READ][off]);
1562         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1563                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1564                         tg->iops_conf[WRITE][off]);
1565         if (off == LIMIT_LOW) {
1566                 if (tg->idletime_threshold_conf == ULONG_MAX)
1567                         strcpy(idle_time, " idle=max");
1568                 else
1569                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1570                                 tg->idletime_threshold_conf);
1571
1572                 if (tg->latency_target_conf == ULONG_MAX)
1573                         strcpy(latency_time, " latency=max");
1574                 else
1575                         snprintf(latency_time, sizeof(latency_time),
1576                                 " latency=%lu", tg->latency_target_conf);
1577         }
1578
1579         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1580                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1581                    latency_time);
1582         return 0;
1583 }
1584
1585 static int tg_print_limit(struct seq_file *sf, void *v)
1586 {
1587         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1588                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1589         return 0;
1590 }
1591
1592 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1593                           char *buf, size_t nbytes, loff_t off)
1594 {
1595         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1596         struct blkg_conf_ctx ctx;
1597         struct throtl_grp *tg;
1598         u64 v[4];
1599         unsigned long idle_time;
1600         unsigned long latency_time;
1601         int ret;
1602         int index = of_cft(of)->private;
1603
1604         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1605         if (ret)
1606                 return ret;
1607
1608         tg = blkg_to_tg(ctx.blkg);
1609
1610         v[0] = tg->bps_conf[READ][index];
1611         v[1] = tg->bps_conf[WRITE][index];
1612         v[2] = tg->iops_conf[READ][index];
1613         v[3] = tg->iops_conf[WRITE][index];
1614
1615         idle_time = tg->idletime_threshold_conf;
1616         latency_time = tg->latency_target_conf;
1617         while (true) {
1618                 char tok[27];   /* wiops=18446744073709551616 */
1619                 char *p;
1620                 u64 val = U64_MAX;
1621                 int len;
1622
1623                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1624                         break;
1625                 if (tok[0] == '\0')
1626                         break;
1627                 ctx.body += len;
1628
1629                 ret = -EINVAL;
1630                 p = tok;
1631                 strsep(&p, "=");
1632                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1633                         goto out_finish;
1634
1635                 ret = -ERANGE;
1636                 if (!val)
1637                         goto out_finish;
1638
1639                 ret = -EINVAL;
1640                 if (!strcmp(tok, "rbps"))
1641                         v[0] = val;
1642                 else if (!strcmp(tok, "wbps"))
1643                         v[1] = val;
1644                 else if (!strcmp(tok, "riops"))
1645                         v[2] = min_t(u64, val, UINT_MAX);
1646                 else if (!strcmp(tok, "wiops"))
1647                         v[3] = min_t(u64, val, UINT_MAX);
1648                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1649                         idle_time = val;
1650                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1651                         latency_time = val;
1652                 else
1653                         goto out_finish;
1654         }
1655
1656         tg->bps_conf[READ][index] = v[0];
1657         tg->bps_conf[WRITE][index] = v[1];
1658         tg->iops_conf[READ][index] = v[2];
1659         tg->iops_conf[WRITE][index] = v[3];
1660
1661         if (index == LIMIT_MAX) {
1662                 tg->bps[READ][index] = v[0];
1663                 tg->bps[WRITE][index] = v[1];
1664                 tg->iops[READ][index] = v[2];
1665                 tg->iops[WRITE][index] = v[3];
1666         }
1667         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1668                 tg->bps_conf[READ][LIMIT_MAX]);
1669         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1670                 tg->bps_conf[WRITE][LIMIT_MAX]);
1671         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1672                 tg->iops_conf[READ][LIMIT_MAX]);
1673         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1674                 tg->iops_conf[WRITE][LIMIT_MAX]);
1675         tg->idletime_threshold_conf = idle_time;
1676         tg->latency_target_conf = latency_time;
1677
1678         /* force user to configure all settings for low limit  */
1679         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1680               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1681             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1682             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1683                 tg->bps[READ][LIMIT_LOW] = 0;
1684                 tg->bps[WRITE][LIMIT_LOW] = 0;
1685                 tg->iops[READ][LIMIT_LOW] = 0;
1686                 tg->iops[WRITE][LIMIT_LOW] = 0;
1687                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1688                 tg->latency_target = DFL_LATENCY_TARGET;
1689         } else if (index == LIMIT_LOW) {
1690                 tg->idletime_threshold = tg->idletime_threshold_conf;
1691                 tg->latency_target = tg->latency_target_conf;
1692         }
1693
1694         blk_throtl_update_limit_valid(tg->td);
1695         if (tg->td->limit_valid[LIMIT_LOW]) {
1696                 if (index == LIMIT_LOW)
1697                         tg->td->limit_index = LIMIT_LOW;
1698         } else
1699                 tg->td->limit_index = LIMIT_MAX;
1700         tg_conf_updated(tg, index == LIMIT_LOW &&
1701                 tg->td->limit_valid[LIMIT_LOW]);
1702         ret = 0;
1703 out_finish:
1704         blkg_conf_finish(&ctx);
1705         return ret ?: nbytes;
1706 }
1707
1708 static struct cftype throtl_files[] = {
1709 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1710         {
1711                 .name = "low",
1712                 .flags = CFTYPE_NOT_ON_ROOT,
1713                 .seq_show = tg_print_limit,
1714                 .write = tg_set_limit,
1715                 .private = LIMIT_LOW,
1716         },
1717 #endif
1718         {
1719                 .name = "max",
1720                 .flags = CFTYPE_NOT_ON_ROOT,
1721                 .seq_show = tg_print_limit,
1722                 .write = tg_set_limit,
1723                 .private = LIMIT_MAX,
1724         },
1725         { }     /* terminate */
1726 };
1727
1728 static void throtl_shutdown_wq(struct request_queue *q)
1729 {
1730         struct throtl_data *td = q->td;
1731
1732         cancel_work_sync(&td->dispatch_work);
1733 }
1734
1735 static struct blkcg_policy blkcg_policy_throtl = {
1736         .dfl_cftypes            = throtl_files,
1737         .legacy_cftypes         = throtl_legacy_files,
1738
1739         .pd_alloc_fn            = throtl_pd_alloc,
1740         .pd_init_fn             = throtl_pd_init,
1741         .pd_online_fn           = throtl_pd_online,
1742         .pd_offline_fn          = throtl_pd_offline,
1743         .pd_free_fn             = throtl_pd_free,
1744 };
1745
1746 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1747 {
1748         unsigned long rtime = jiffies, wtime = jiffies;
1749
1750         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1751                 rtime = tg->last_low_overflow_time[READ];
1752         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1753                 wtime = tg->last_low_overflow_time[WRITE];
1754         return min(rtime, wtime);
1755 }
1756
1757 /* tg should not be an intermediate node */
1758 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1759 {
1760         struct throtl_service_queue *parent_sq;
1761         struct throtl_grp *parent = tg;
1762         unsigned long ret = __tg_last_low_overflow_time(tg);
1763
1764         while (true) {
1765                 parent_sq = parent->service_queue.parent_sq;
1766                 parent = sq_to_tg(parent_sq);
1767                 if (!parent)
1768                         break;
1769
1770                 /*
1771                  * The parent doesn't have low limit, it always reaches low
1772                  * limit. Its overflow time is useless for children
1773                  */
1774                 if (!parent->bps[READ][LIMIT_LOW] &&
1775                     !parent->iops[READ][LIMIT_LOW] &&
1776                     !parent->bps[WRITE][LIMIT_LOW] &&
1777                     !parent->iops[WRITE][LIMIT_LOW])
1778                         continue;
1779                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1780                         ret = __tg_last_low_overflow_time(parent);
1781         }
1782         return ret;
1783 }
1784
1785 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1786 {
1787         /*
1788          * cgroup is idle if:
1789          * - single idle is too long, longer than a fixed value (in case user
1790          *   configure a too big threshold) or 4 times of idletime threshold
1791          * - average think time is more than threshold
1792          * - IO latency is largely below threshold
1793          */
1794         unsigned long time;
1795         bool ret;
1796
1797         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1798         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1799               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1800               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1801               tg->avg_idletime > tg->idletime_threshold ||
1802               (tg->latency_target && tg->bio_cnt &&
1803                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1804         throtl_log(&tg->service_queue,
1805                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1806                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1807                 tg->bio_cnt, ret, tg->td->scale);
1808         return ret;
1809 }
1810
1811 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1812 {
1813         struct throtl_service_queue *sq = &tg->service_queue;
1814         bool read_limit, write_limit;
1815
1816         /*
1817          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1818          * reaches), it's ok to upgrade to next limit
1819          */
1820         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1821         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1822         if (!read_limit && !write_limit)
1823                 return true;
1824         if (read_limit && sq->nr_queued[READ] &&
1825             (!write_limit || sq->nr_queued[WRITE]))
1826                 return true;
1827         if (write_limit && sq->nr_queued[WRITE] &&
1828             (!read_limit || sq->nr_queued[READ]))
1829                 return true;
1830
1831         if (time_after_eq(jiffies,
1832                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1833             throtl_tg_is_idle(tg))
1834                 return true;
1835         return false;
1836 }
1837
1838 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1839 {
1840         while (true) {
1841                 if (throtl_tg_can_upgrade(tg))
1842                         return true;
1843                 tg = sq_to_tg(tg->service_queue.parent_sq);
1844                 if (!tg || !tg_to_blkg(tg)->parent)
1845                         return false;
1846         }
1847         return false;
1848 }
1849
1850 static bool throtl_can_upgrade(struct throtl_data *td,
1851         struct throtl_grp *this_tg)
1852 {
1853         struct cgroup_subsys_state *pos_css;
1854         struct blkcg_gq *blkg;
1855
1856         if (td->limit_index != LIMIT_LOW)
1857                 return false;
1858
1859         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1860                 return false;
1861
1862         rcu_read_lock();
1863         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1864                 struct throtl_grp *tg = blkg_to_tg(blkg);
1865
1866                 if (tg == this_tg)
1867                         continue;
1868                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1869                         continue;
1870                 if (!throtl_hierarchy_can_upgrade(tg)) {
1871                         rcu_read_unlock();
1872                         return false;
1873                 }
1874         }
1875         rcu_read_unlock();
1876         return true;
1877 }
1878
1879 static void throtl_upgrade_check(struct throtl_grp *tg)
1880 {
1881         unsigned long now = jiffies;
1882
1883         if (tg->td->limit_index != LIMIT_LOW)
1884                 return;
1885
1886         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1887                 return;
1888
1889         tg->last_check_time = now;
1890
1891         if (!time_after_eq(now,
1892              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1893                 return;
1894
1895         if (throtl_can_upgrade(tg->td, NULL))
1896                 throtl_upgrade_state(tg->td);
1897 }
1898
1899 static void throtl_upgrade_state(struct throtl_data *td)
1900 {
1901         struct cgroup_subsys_state *pos_css;
1902         struct blkcg_gq *blkg;
1903
1904         throtl_log(&td->service_queue, "upgrade to max");
1905         td->limit_index = LIMIT_MAX;
1906         td->low_upgrade_time = jiffies;
1907         td->scale = 0;
1908         rcu_read_lock();
1909         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1910                 struct throtl_grp *tg = blkg_to_tg(blkg);
1911                 struct throtl_service_queue *sq = &tg->service_queue;
1912
1913                 tg->disptime = jiffies - 1;
1914                 throtl_select_dispatch(sq);
1915                 throtl_schedule_next_dispatch(sq, true);
1916         }
1917         rcu_read_unlock();
1918         throtl_select_dispatch(&td->service_queue);
1919         throtl_schedule_next_dispatch(&td->service_queue, true);
1920         queue_work(kthrotld_workqueue, &td->dispatch_work);
1921 }
1922
1923 static void throtl_downgrade_state(struct throtl_data *td, int new)
1924 {
1925         td->scale /= 2;
1926
1927         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1928         if (td->scale) {
1929                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1930                 return;
1931         }
1932
1933         td->limit_index = new;
1934         td->low_downgrade_time = jiffies;
1935 }
1936
1937 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1938 {
1939         struct throtl_data *td = tg->td;
1940         unsigned long now = jiffies;
1941
1942         /*
1943          * If cgroup is below low limit, consider downgrade and throttle other
1944          * cgroups
1945          */
1946         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1947             time_after_eq(now, tg_last_low_overflow_time(tg) +
1948                                         td->throtl_slice) &&
1949             (!throtl_tg_is_idle(tg) ||
1950              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1951                 return true;
1952         return false;
1953 }
1954
1955 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1956 {
1957         while (true) {
1958                 if (!throtl_tg_can_downgrade(tg))
1959                         return false;
1960                 tg = sq_to_tg(tg->service_queue.parent_sq);
1961                 if (!tg || !tg_to_blkg(tg)->parent)
1962                         break;
1963         }
1964         return true;
1965 }
1966
1967 static void throtl_downgrade_check(struct throtl_grp *tg)
1968 {
1969         uint64_t bps;
1970         unsigned int iops;
1971         unsigned long elapsed_time;
1972         unsigned long now = jiffies;
1973
1974         if (tg->td->limit_index != LIMIT_MAX ||
1975             !tg->td->limit_valid[LIMIT_LOW])
1976                 return;
1977         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1978                 return;
1979         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1980                 return;
1981
1982         elapsed_time = now - tg->last_check_time;
1983         tg->last_check_time = now;
1984
1985         if (time_before(now, tg_last_low_overflow_time(tg) +
1986                         tg->td->throtl_slice))
1987                 return;
1988
1989         if (tg->bps[READ][LIMIT_LOW]) {
1990                 bps = tg->last_bytes_disp[READ] * HZ;
1991                 do_div(bps, elapsed_time);
1992                 if (bps >= tg->bps[READ][LIMIT_LOW])
1993                         tg->last_low_overflow_time[READ] = now;
1994         }
1995
1996         if (tg->bps[WRITE][LIMIT_LOW]) {
1997                 bps = tg->last_bytes_disp[WRITE] * HZ;
1998                 do_div(bps, elapsed_time);
1999                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2000                         tg->last_low_overflow_time[WRITE] = now;
2001         }
2002
2003         if (tg->iops[READ][LIMIT_LOW]) {
2004                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2005                 if (iops >= tg->iops[READ][LIMIT_LOW])
2006                         tg->last_low_overflow_time[READ] = now;
2007         }
2008
2009         if (tg->iops[WRITE][LIMIT_LOW]) {
2010                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2011                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2012                         tg->last_low_overflow_time[WRITE] = now;
2013         }
2014
2015         /*
2016          * If cgroup is below low limit, consider downgrade and throttle other
2017          * cgroups
2018          */
2019         if (throtl_hierarchy_can_downgrade(tg))
2020                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2021
2022         tg->last_bytes_disp[READ] = 0;
2023         tg->last_bytes_disp[WRITE] = 0;
2024         tg->last_io_disp[READ] = 0;
2025         tg->last_io_disp[WRITE] = 0;
2026 }
2027
2028 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2029 {
2030         unsigned long now = ktime_get_ns() >> 10;
2031         unsigned long last_finish_time = tg->last_finish_time;
2032
2033         if (now <= last_finish_time || last_finish_time == 0 ||
2034             last_finish_time == tg->checked_last_finish_time)
2035                 return;
2036
2037         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2038         tg->checked_last_finish_time = last_finish_time;
2039 }
2040
2041 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2042 static void throtl_update_latency_buckets(struct throtl_data *td)
2043 {
2044         struct avg_latency_bucket avg_latency[LATENCY_BUCKET_SIZE];
2045         int i, cpu;
2046         unsigned long last_latency = 0;
2047         unsigned long latency;
2048
2049         if (!blk_queue_nonrot(td->queue))
2050                 return;
2051         if (time_before(jiffies, td->last_calculate_time + HZ))
2052                 return;
2053         td->last_calculate_time = jiffies;
2054
2055         memset(avg_latency, 0, sizeof(avg_latency));
2056         for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2057                 struct latency_bucket *tmp = &td->tmp_buckets[i];
2058
2059                 for_each_possible_cpu(cpu) {
2060                         struct latency_bucket *bucket;
2061
2062                         /* this isn't race free, but ok in practice */
2063                         bucket = per_cpu_ptr(td->latency_buckets, cpu);
2064                         tmp->total_latency += bucket[i].total_latency;
2065                         tmp->samples += bucket[i].samples;
2066                         bucket[i].total_latency = 0;
2067                         bucket[i].samples = 0;
2068                 }
2069
2070                 if (tmp->samples >= 32) {
2071                         int samples = tmp->samples;
2072
2073                         latency = tmp->total_latency;
2074
2075                         tmp->total_latency = 0;
2076                         tmp->samples = 0;
2077                         latency /= samples;
2078                         if (latency == 0)
2079                                 continue;
2080                         avg_latency[i].latency = latency;
2081                 }
2082         }
2083
2084         for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2085                 if (!avg_latency[i].latency) {
2086                         if (td->avg_buckets[i].latency < last_latency)
2087                                 td->avg_buckets[i].latency = last_latency;
2088                         continue;
2089                 }
2090
2091                 if (!td->avg_buckets[i].valid)
2092                         latency = avg_latency[i].latency;
2093                 else
2094                         latency = (td->avg_buckets[i].latency * 7 +
2095                                 avg_latency[i].latency) >> 3;
2096
2097                 td->avg_buckets[i].latency = max(latency, last_latency);
2098                 td->avg_buckets[i].valid = true;
2099                 last_latency = td->avg_buckets[i].latency;
2100         }
2101
2102         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2103                 throtl_log(&td->service_queue,
2104                         "Latency bucket %d: latency=%ld, valid=%d", i,
2105                         td->avg_buckets[i].latency, td->avg_buckets[i].valid);
2106 }
2107 #else
2108 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2109 {
2110 }
2111 #endif
2112
2113 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2114 {
2115 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2116         if (bio->bi_css) {
2117                 if (bio->bi_cg_private)
2118                         blkg_put(tg_to_blkg(bio->bi_cg_private));
2119                 bio->bi_cg_private = tg;
2120                 blkg_get(tg_to_blkg(tg));
2121         }
2122         blk_stat_set_issue(&bio->bi_issue_stat, bio_sectors(bio));
2123 #endif
2124 }
2125
2126 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2127                     struct bio *bio)
2128 {
2129         struct throtl_qnode *qn = NULL;
2130         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2131         struct throtl_service_queue *sq;
2132         bool rw = bio_data_dir(bio);
2133         bool throttled = false;
2134         struct throtl_data *td = tg->td;
2135
2136         WARN_ON_ONCE(!rcu_read_lock_held());
2137
2138         /* see throtl_charge_bio() */
2139         if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2140                 goto out;
2141
2142         spin_lock_irq(q->queue_lock);
2143
2144         throtl_update_latency_buckets(td);
2145
2146         if (unlikely(blk_queue_bypass(q)))
2147                 goto out_unlock;
2148
2149         blk_throtl_assoc_bio(tg, bio);
2150         blk_throtl_update_idletime(tg);
2151
2152         sq = &tg->service_queue;
2153
2154 again:
2155         while (true) {
2156                 if (tg->last_low_overflow_time[rw] == 0)
2157                         tg->last_low_overflow_time[rw] = jiffies;
2158                 throtl_downgrade_check(tg);
2159                 throtl_upgrade_check(tg);
2160                 /* throtl is FIFO - if bios are already queued, should queue */
2161                 if (sq->nr_queued[rw])
2162                         break;
2163
2164                 /* if above limits, break to queue */
2165                 if (!tg_may_dispatch(tg, bio, NULL)) {
2166                         tg->last_low_overflow_time[rw] = jiffies;
2167                         if (throtl_can_upgrade(td, tg)) {
2168                                 throtl_upgrade_state(td);
2169                                 goto again;
2170                         }
2171                         break;
2172                 }
2173
2174                 /* within limits, let's charge and dispatch directly */
2175                 throtl_charge_bio(tg, bio);
2176
2177                 /*
2178                  * We need to trim slice even when bios are not being queued
2179                  * otherwise it might happen that a bio is not queued for
2180                  * a long time and slice keeps on extending and trim is not
2181                  * called for a long time. Now if limits are reduced suddenly
2182                  * we take into account all the IO dispatched so far at new
2183                  * low rate and * newly queued IO gets a really long dispatch
2184                  * time.
2185                  *
2186                  * So keep on trimming slice even if bio is not queued.
2187                  */
2188                 throtl_trim_slice(tg, rw);
2189
2190                 /*
2191                  * @bio passed through this layer without being throttled.
2192                  * Climb up the ladder.  If we''re already at the top, it
2193                  * can be executed directly.
2194                  */
2195                 qn = &tg->qnode_on_parent[rw];
2196                 sq = sq->parent_sq;
2197                 tg = sq_to_tg(sq);
2198                 if (!tg)
2199                         goto out_unlock;
2200         }
2201
2202         /* out-of-limit, queue to @tg */
2203         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2204                    rw == READ ? 'R' : 'W',
2205                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2206                    tg_bps_limit(tg, rw),
2207                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2208                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2209
2210         tg->last_low_overflow_time[rw] = jiffies;
2211
2212         td->nr_queued[rw]++;
2213         throtl_add_bio_tg(bio, qn, tg);
2214         throttled = true;
2215
2216         /*
2217          * Update @tg's dispatch time and force schedule dispatch if @tg
2218          * was empty before @bio.  The forced scheduling isn't likely to
2219          * cause undue delay as @bio is likely to be dispatched directly if
2220          * its @tg's disptime is not in the future.
2221          */
2222         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2223                 tg_update_disptime(tg);
2224                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2225         }
2226
2227 out_unlock:
2228         spin_unlock_irq(q->queue_lock);
2229 out:
2230         /*
2231          * As multiple blk-throtls may stack in the same issue path, we
2232          * don't want bios to leave with the flag set.  Clear the flag if
2233          * being issued.
2234          */
2235         if (!throttled)
2236                 bio_clear_flag(bio, BIO_THROTTLED);
2237
2238 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2239         if (throttled || !td->track_bio_latency)
2240                 bio->bi_issue_stat.stat |= SKIP_LATENCY;
2241 #endif
2242         return throttled;
2243 }
2244
2245 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2246 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2247         int op, unsigned long time)
2248 {
2249         struct latency_bucket *latency;
2250         int index;
2251
2252         if (!td || td->limit_index != LIMIT_LOW || op != REQ_OP_READ ||
2253             !blk_queue_nonrot(td->queue))
2254                 return;
2255
2256         index = request_bucket_index(size);
2257
2258         latency = get_cpu_ptr(td->latency_buckets);
2259         latency[index].total_latency += time;
2260         latency[index].samples++;
2261         put_cpu_ptr(td->latency_buckets);
2262 }
2263
2264 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2265 {
2266         struct request_queue *q = rq->q;
2267         struct throtl_data *td = q->td;
2268
2269         throtl_track_latency(td, blk_stat_size(&rq->issue_stat),
2270                 req_op(rq), time_ns >> 10);
2271 }
2272
2273 void blk_throtl_bio_endio(struct bio *bio)
2274 {
2275         struct throtl_grp *tg;
2276         u64 finish_time_ns;
2277         unsigned long finish_time;
2278         unsigned long start_time;
2279         unsigned long lat;
2280
2281         tg = bio->bi_cg_private;
2282         if (!tg)
2283                 return;
2284         bio->bi_cg_private = NULL;
2285
2286         finish_time_ns = ktime_get_ns();
2287         tg->last_finish_time = finish_time_ns >> 10;
2288
2289         start_time = blk_stat_time(&bio->bi_issue_stat) >> 10;
2290         finish_time = __blk_stat_time(finish_time_ns) >> 10;
2291         if (!start_time || finish_time <= start_time) {
2292                 blkg_put(tg_to_blkg(tg));
2293                 return;
2294         }
2295
2296         lat = finish_time - start_time;
2297         /* this is only for bio based driver */
2298         if (!(bio->bi_issue_stat.stat & SKIP_LATENCY))
2299                 throtl_track_latency(tg->td, blk_stat_size(&bio->bi_issue_stat),
2300                         bio_op(bio), lat);
2301
2302         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2303                 int bucket;
2304                 unsigned int threshold;
2305
2306                 bucket = request_bucket_index(
2307                         blk_stat_size(&bio->bi_issue_stat));
2308                 threshold = tg->td->avg_buckets[bucket].latency +
2309                         tg->latency_target;
2310                 if (lat > threshold)
2311                         tg->bad_bio_cnt++;
2312                 /*
2313                  * Not race free, could get wrong count, which means cgroups
2314                  * will be throttled
2315                  */
2316                 tg->bio_cnt++;
2317         }
2318
2319         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2320                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2321                 tg->bio_cnt /= 2;
2322                 tg->bad_bio_cnt /= 2;
2323         }
2324
2325         blkg_put(tg_to_blkg(tg));
2326 }
2327 #endif
2328
2329 /*
2330  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2331  * return, @parent_sq is guaranteed to not have any active children tg's
2332  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2333  */
2334 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2335 {
2336         struct throtl_grp *tg;
2337
2338         while ((tg = throtl_rb_first(parent_sq))) {
2339                 struct throtl_service_queue *sq = &tg->service_queue;
2340                 struct bio *bio;
2341
2342                 throtl_dequeue_tg(tg);
2343
2344                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2345                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2346                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2347                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2348         }
2349 }
2350
2351 /**
2352  * blk_throtl_drain - drain throttled bios
2353  * @q: request_queue to drain throttled bios for
2354  *
2355  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2356  */
2357 void blk_throtl_drain(struct request_queue *q)
2358         __releases(q->queue_lock) __acquires(q->queue_lock)
2359 {
2360         struct throtl_data *td = q->td;
2361         struct blkcg_gq *blkg;
2362         struct cgroup_subsys_state *pos_css;
2363         struct bio *bio;
2364         int rw;
2365
2366         queue_lockdep_assert_held(q);
2367         rcu_read_lock();
2368
2369         /*
2370          * Drain each tg while doing post-order walk on the blkg tree, so
2371          * that all bios are propagated to td->service_queue.  It'd be
2372          * better to walk service_queue tree directly but blkg walk is
2373          * easier.
2374          */
2375         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2376                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2377
2378         /* finally, transfer bios from top-level tg's into the td */
2379         tg_drain_bios(&td->service_queue);
2380
2381         rcu_read_unlock();
2382         spin_unlock_irq(q->queue_lock);
2383
2384         /* all bios now should be in td->service_queue, issue them */
2385         for (rw = READ; rw <= WRITE; rw++)
2386                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2387                                                 NULL)))
2388                         generic_make_request(bio);
2389
2390         spin_lock_irq(q->queue_lock);
2391 }
2392
2393 int blk_throtl_init(struct request_queue *q)
2394 {
2395         struct throtl_data *td;
2396         int ret;
2397
2398         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2399         if (!td)
2400                 return -ENOMEM;
2401         td->latency_buckets = __alloc_percpu(sizeof(struct latency_bucket) *
2402                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2403         if (!td->latency_buckets) {
2404                 kfree(td);
2405                 return -ENOMEM;
2406         }
2407
2408         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2409         throtl_service_queue_init(&td->service_queue);
2410
2411         q->td = td;
2412         td->queue = q;
2413
2414         td->limit_valid[LIMIT_MAX] = true;
2415         td->limit_index = LIMIT_MAX;
2416         td->low_upgrade_time = jiffies;
2417         td->low_downgrade_time = jiffies;
2418
2419         /* activate policy */
2420         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2421         if (ret) {
2422                 free_percpu(td->latency_buckets);
2423                 kfree(td);
2424         }
2425         return ret;
2426 }
2427
2428 void blk_throtl_exit(struct request_queue *q)
2429 {
2430         BUG_ON(!q->td);
2431         throtl_shutdown_wq(q);
2432         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2433         free_percpu(q->td->latency_buckets);
2434         kfree(q->td);
2435 }
2436
2437 void blk_throtl_register_queue(struct request_queue *q)
2438 {
2439         struct throtl_data *td;
2440         int i;
2441
2442         td = q->td;
2443         BUG_ON(!td);
2444
2445         if (blk_queue_nonrot(q)) {
2446                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2447                 td->filtered_latency = LATENCY_FILTERED_SSD;
2448         } else {
2449                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2450                 td->filtered_latency = LATENCY_FILTERED_HD;
2451                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2452                         td->avg_buckets[i].latency = DFL_HD_BASELINE_LATENCY;
2453         }
2454 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2455         /* if no low limit, use previous default */
2456         td->throtl_slice = DFL_THROTL_SLICE_HD;
2457 #endif
2458
2459         td->track_bio_latency = !q->mq_ops && !q->request_fn;
2460         if (!td->track_bio_latency)
2461                 blk_stat_enable_accounting(q);
2462 }
2463
2464 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2465 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2466 {
2467         if (!q->td)
2468                 return -EINVAL;
2469         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2470 }
2471
2472 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2473         const char *page, size_t count)
2474 {
2475         unsigned long v;
2476         unsigned long t;
2477
2478         if (!q->td)
2479                 return -EINVAL;
2480         if (kstrtoul(page, 10, &v))
2481                 return -EINVAL;
2482         t = msecs_to_jiffies(v);
2483         if (t == 0 || t > MAX_THROTL_SLICE)
2484                 return -EINVAL;
2485         q->td->throtl_slice = t;
2486         return count;
2487 }
2488 #endif
2489
2490 static int __init throtl_init(void)
2491 {
2492         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2493         if (!kthrotld_workqueue)
2494                 panic("Failed to create kthrotld\n");
2495
2496         return blkcg_policy_register(&blkcg_policy_throtl);
2497 }
2498
2499 module_init(throtl_init);