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 #include "blk-cgroup-rwstat.h"
  16
  17 /* Max dispatch from a group in 1 round */
  18 static int throtl_grp_quantum = 8;
  19
  20 /* Total max dispatch from all groups in one round */
  21 static int throtl_quantum = 32;
  22
  23 /* Throttling is performed over a slice and after that slice is renewed */
  24 #define DFL_THROTL_SLICE_HD (HZ / 10)
  25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  26 #define MAX_THROTL_SLICE (HZ)
  27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  28 #define MIN_THROTL_BPS (320 * 1024)
  29 #define MIN_THROTL_IOPS (10)
  30 #define DFL_LATENCY_TARGET (-1L)
  31 #define DFL_IDLE_THRESHOLD (0)
  32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  33 #define LATENCY_FILTERED_SSD (0)
  34 /*
  35  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  36  * help determine if its IO is impacted by others, hence we ignore the IO
  37  */
  38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  39
  40 static struct blkcg_policy blkcg_policy_throtl;
  41
  42 /* A workqueue to queue throttle related work */
  43 static struct workqueue_struct *kthrotld_workqueue;
  44
  45 /*
  46  * To implement hierarchical throttling, throtl_grps form a tree and bios
  47  * are dispatched upwards level by level until they reach the top and get
  48  * issued.  When dispatching bios from the children and local group at each
  49  * level, if the bios are dispatched into a single bio_list, there's a risk
  50  * of a local or child group which can queue many bios at once filling up
  51  * the list starving others.
  52  *
  53  * To avoid such starvation, dispatched bios are queued separately
  54  * according to where they came from.  When they are again dispatched to
  55  * the parent, they're popped in round-robin order so that no single source
  56  * hogs the dispatch window.
  57  *
  58  * throtl_qnode is used to keep the queued bios separated by their sources.
  59  * Bios are queued to throtl_qnode which in turn is queued to
  60  * throtl_service_queue and then dispatched in round-robin order.
  61  *
  62  * It's also used to track the reference counts on blkg's.  A qnode always
  63  * belongs to a throtl_grp and gets queued on itself or the parent, so
  64  * incrementing the reference of the associated throtl_grp when a qnode is
  65  * queued and decrementing when dequeued is enough to keep the whole blkg
  66  * tree pinned while bios are in flight.
  67  */
  68 struct throtl_qnode {
  69         struct list_head        node;           /* service_queue->queued[] */
  70         struct bio_list         bios;           /* queued bios */
  71         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  72 };
  73
  74 struct throtl_service_queue {
  75         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  76
  77         /*
  78          * Bios queued directly to this service_queue or dispatched from
  79          * children throtl_grp's.
  80          */
  81         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  82         unsigned int            nr_queued[2];   /* number of queued bios */
  83
  84         /*
  85          * RB tree of active children throtl_grp's, which are sorted by
  86          * their ->disptime.
  87          */
  88         struct rb_root_cached   pending_tree;   /* RB tree of active tgs */
  89         unsigned int            nr_pending;     /* # queued in the tree */
  90         unsigned long           first_pending_disptime; /* disptime of the first tg */
  91         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  92 };
  93
  94 enum tg_state_flags {
  95         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  96         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  97 };
  98
  99 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
 100
 101 enum {
 102         LIMIT_LOW,
 103         LIMIT_MAX,
 104         LIMIT_CNT,
 105 };
 106
 107 struct throtl_grp {
 108         /* must be the first member */
 109         struct blkg_policy_data pd;
 110
 111         /* active throtl group service_queue member */
 112         struct rb_node rb_node;
 113
 114         /* throtl_data this group belongs to */
 115         struct throtl_data *td;
 116
 117         /* this group's service queue */
 118         struct throtl_service_queue service_queue;
 119
 120         /*
 121          * qnode_on_self is used when bios are directly queued to this
 122          * throtl_grp so that local bios compete fairly with bios
 123          * dispatched from children.  qnode_on_parent is used when bios are
 124          * dispatched from this throtl_grp into its parent and will compete
 125          * with the sibling qnode_on_parents and the parent's
 126          * qnode_on_self.
 127          */
 128         struct throtl_qnode qnode_on_self[2];
 129         struct throtl_qnode qnode_on_parent[2];
 130
 131         /*
 132          * Dispatch time in jiffies. This is the estimated time when group
 133          * will unthrottle and is ready to dispatch more bio. It is used as
 134          * key to sort active groups in service tree.
 135          */
 136         unsigned long disptime;
 137
 138         unsigned int flags;
 139
 140         /* are there any throtl rules between this group and td? */
 141         bool has_rules[2];
 142
 143         /* internally used bytes per second rate limits */
 144         uint64_t bps[2][LIMIT_CNT];
 145         /* user configured bps limits */
 146         uint64_t bps_conf[2][LIMIT_CNT];
 147
 148         /* internally used IOPS limits */
 149         unsigned int iops[2][LIMIT_CNT];
 150         /* user configured IOPS limits */
 151         unsigned int iops_conf[2][LIMIT_CNT];
 152
 153         /* Number of bytes disptached in current slice */
 154         uint64_t bytes_disp[2];
 155         /* Number of bio's dispatched in current slice */
 156         unsigned int io_disp[2];
 157
 158         unsigned long last_low_overflow_time[2];
 159
 160         uint64_t last_bytes_disp[2];
 161         unsigned int last_io_disp[2];
 162
 163         unsigned long last_check_time;
 164
 165         unsigned long latency_target; /* us */
 166         unsigned long latency_target_conf; /* us */
 167         /* When did we start a new slice */
 168         unsigned long slice_start[2];
 169         unsigned long slice_end[2];
 170
 171         unsigned long last_finish_time; /* ns / 1024 */
 172         unsigned long checked_last_finish_time; /* ns / 1024 */
 173         unsigned long avg_idletime; /* ns / 1024 */
 174         unsigned long idletime_threshold; /* us */
 175         unsigned long idletime_threshold_conf; /* us */
 176
 177         unsigned int bio_cnt; /* total bios */
 178         unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
 179         unsigned long bio_cnt_reset_time;
 180
 181         struct blkg_rwstat stat_bytes;
 182         struct blkg_rwstat stat_ios;
 183 };
 184
 185 /* We measure latency for request size from <= 4k to >= 1M */
 186 #define LATENCY_BUCKET_SIZE 9
 187
 188 struct latency_bucket {
 189         unsigned long total_latency; /* ns / 1024 */
 190         int samples;
 191 };
 192
 193 struct avg_latency_bucket {
 194         unsigned long latency; /* ns / 1024 */
 195         bool valid;
 196 };
 197
 198 struct throtl_data
 199 {
 200         /* service tree for active throtl groups */
 201         struct throtl_service_queue service_queue;
 202
 203         struct request_queue *queue;
 204
 205         /* Total Number of queued bios on READ and WRITE lists */
 206         unsigned int nr_queued[2];
 207
 208         unsigned int throtl_slice;
 209
 210         /* Work for dispatching throttled bios */
 211         struct work_struct dispatch_work;
 212         unsigned int limit_index;
 213         bool limit_valid[LIMIT_CNT];
 214
 215         unsigned long low_upgrade_time;
 216         unsigned long low_downgrade_time;
 217
 218         unsigned int scale;
 219
 220         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
 221         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
 222         struct latency_bucket __percpu *latency_buckets[2];
 223         unsigned long last_calculate_time;
 224         unsigned long filtered_latency;
 225
 226         bool track_bio_latency;
 227 };
 228
 229 static void throtl_pending_timer_fn(struct timer_list *t);
 230
 231 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 232 {
 233         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 234 }
 235
 236 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 237 {
 238         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 239 }
 240
 241 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 242 {
 243         return pd_to_blkg(&tg->pd);
 244 }
 245
 246 /**
 247  * sq_to_tg - return the throl_grp the specified service queue belongs to
 248  * @sq: the throtl_service_queue of interest
 249  *
 250  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 251  * embedded in throtl_data, %NULL is returned.
 252  */
 253 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 254 {
 255         if (sq && sq->parent_sq)
 256                 return container_of(sq, struct throtl_grp, service_queue);
 257         else
 258                 return NULL;
 259 }
 260
 261 /**
 262  * sq_to_td - return throtl_data the specified service queue belongs to
 263  * @sq: the throtl_service_queue of interest
 264  *
 265  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 266  * Determine the associated throtl_data accordingly and return it.
 267  */
 268 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 269 {
 270         struct throtl_grp *tg = sq_to_tg(sq);
 271
 272         if (tg)
 273                 return tg->td;
 274         else
 275                 return container_of(sq, struct throtl_data, service_queue);
 276 }
 277
 278 /*
 279  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 280  * make the IO dispatch more smooth.
 281  * Scale up: linearly scale up according to lapsed time since upgrade. For
 282  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 283  *           limit hits .max limit
 284  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 285  */
 286 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 287 {
 288         /* arbitrary value to avoid too big scale */
 289         if (td->scale < 4096 && time_after_eq(jiffies,
 290             td->low_upgrade_time + td->scale * td->throtl_slice))
 291                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 292
 293         return low + (low >> 1) * td->scale;
 294 }
 295
 296 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 297 {
 298         struct blkcg_gq *blkg = tg_to_blkg(tg);
 299         struct throtl_data *td;
 300         uint64_t ret;
 301
 302         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 303                 return U64_MAX;
 304
 305         td = tg->td;
 306         ret = tg->bps[rw][td->limit_index];
 307         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 308                 /* intermediate node or iops isn't 0 */
 309                 if (!list_empty(&blkg->blkcg->css.children) ||
 310                     tg->iops[rw][td->limit_index])
 311                         return U64_MAX;
 312                 else
 313                         return MIN_THROTL_BPS;
 314         }
 315
 316         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 317             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 318                 uint64_t adjusted;
 319
 320                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 321                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 322         }
 323         return ret;
 324 }
 325
 326 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 327 {
 328         struct blkcg_gq *blkg = tg_to_blkg(tg);
 329         struct throtl_data *td;
 330         unsigned int ret;
 331
 332         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 333                 return UINT_MAX;
 334
 335         td = tg->td;
 336         ret = tg->iops[rw][td->limit_index];
 337         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 338                 /* intermediate node or bps isn't 0 */
 339                 if (!list_empty(&blkg->blkcg->css.children) ||
 340                     tg->bps[rw][td->limit_index])
 341                         return UINT_MAX;
 342                 else
 343                         return MIN_THROTL_IOPS;
 344         }
 345
 346         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 347             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 348                 uint64_t adjusted;
 349
 350                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 351                 if (adjusted > UINT_MAX)
 352                         adjusted = UINT_MAX;
 353                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 354         }
 355         return ret;
 356 }
 357
 358 #define request_bucket_index(sectors) \
 359         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 360
 361 /**
 362  * throtl_log - log debug message via blktrace
 363  * @sq: the service_queue being reported
 364  * @fmt: printf format string
 365  * @args: printf args
 366  *
 367  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 368  * throtl_grp; otherwise, just "throtl".
 369  */
 370 #define throtl_log(sq, fmt, args...)    do {                            \
 371         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 372         struct throtl_data *__td = sq_to_td((sq));                      \
 373                                                                         \
 374         (void)__td;                                                     \
 375         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 376                 break;                                                  \
 377         if ((__tg)) {                                                   \
 378                 blk_add_cgroup_trace_msg(__td->queue,                   \
 379                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 380         } else {                                                        \
 381                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 382         }                                                               \
 383 } while (0)
 384
 385 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 386 {
 387         /* assume it's one sector */
 388         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 389                 return 512;
 390         return bio->bi_iter.bi_size;
 391 }
 392
 393 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 394 {
 395         INIT_LIST_HEAD(&qn->node);
 396         bio_list_init(&qn->bios);
 397         qn->tg = tg;
 398 }
 399
 400 /**
 401  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 402  * @bio: bio being added
 403  * @qn: qnode to add bio to
 404  * @queued: the service_queue->queued[] list @qn belongs to
 405  *
 406  * Add @bio to @qn and put @qn on @queued if it's not already on.
 407  * @qn->tg's reference count is bumped when @qn is activated.  See the
 408  * comment on top of throtl_qnode definition for details.
 409  */
 410 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 411                                  struct list_head *queued)
 412 {
 413         bio_list_add(&qn->bios, bio);
 414         if (list_empty(&qn->node)) {
 415                 list_add_tail(&qn->node, queued);
 416                 blkg_get(tg_to_blkg(qn->tg));
 417         }
 418 }
 419
 420 /**
 421  * throtl_peek_queued - peek the first bio on a qnode list
 422  * @queued: the qnode list to peek
 423  */
 424 static struct bio *throtl_peek_queued(struct list_head *queued)
 425 {
 426         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 427         struct bio *bio;
 428
 429         if (list_empty(queued))
 430                 return NULL;
 431
 432         bio = bio_list_peek(&qn->bios);
 433         WARN_ON_ONCE(!bio);
 434         return bio;
 435 }
 436
 437 /**
 438  * throtl_pop_queued - pop the first bio form a qnode list
 439  * @queued: the qnode list to pop a bio from
 440  * @tg_to_put: optional out argument for throtl_grp to put
 441  *
 442  * Pop the first bio from the qnode list @queued.  After popping, the first
 443  * qnode is removed from @queued if empty or moved to the end of @queued so
 444  * that the popping order is round-robin.
 445  *
 446  * When the first qnode is removed, its associated throtl_grp should be put
 447  * too.  If @tg_to_put is NULL, this function automatically puts it;
 448  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 449  * responsible for putting it.
 450  */
 451 static struct bio *throtl_pop_queued(struct list_head *queued,
 452                                      struct throtl_grp **tg_to_put)
 453 {
 454         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 455         struct bio *bio;
 456
 457         if (list_empty(queued))
 458                 return NULL;
 459
 460         bio = bio_list_pop(&qn->bios);
 461         WARN_ON_ONCE(!bio);
 462
 463         if (bio_list_empty(&qn->bios)) {
 464                 list_del_init(&qn->node);
 465                 if (tg_to_put)
 466                         *tg_to_put = qn->tg;
 467                 else
 468                         blkg_put(tg_to_blkg(qn->tg));
 469         } else {
 470                 list_move_tail(&qn->node, queued);
 471         }
 472
 473         return bio;
 474 }
 475
 476 /* init a service_queue, assumes the caller zeroed it */
 477 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 478 {
 479         INIT_LIST_HEAD(&sq->queued[0]);
 480         INIT_LIST_HEAD(&sq->queued[1]);
 481         sq->pending_tree = RB_ROOT_CACHED;
 482         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 483 }
 484
 485 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
 486                                                 struct request_queue *q,
 487                                                 struct blkcg *blkcg)
 488 {
 489         struct throtl_grp *tg;
 490         int rw;
 491
 492         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
 493         if (!tg)
 494                 return NULL;
 495
 496         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 497                 goto err_free_tg;
 498
 499         if (blkg_rwstat_init(&tg->stat_ios, gfp))
 500                 goto err_exit_stat_bytes;
 501
 502         throtl_service_queue_init(&tg->service_queue);
 503
 504         for (rw = READ; rw <= WRITE; rw++) {
 505                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 506                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 507         }
 508
 509         RB_CLEAR_NODE(&tg->rb_node);
 510         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 511         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 512         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 513         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 514         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 515         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 516         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 517         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 518         /* LIMIT_LOW will have default value 0 */
 519
 520         tg->latency_target = DFL_LATENCY_TARGET;
 521         tg->latency_target_conf = DFL_LATENCY_TARGET;
 522         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 523         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 524
 525         return &tg->pd;
 526
 527 err_exit_stat_bytes:
 528         blkg_rwstat_exit(&tg->stat_bytes);
 529 err_free_tg:
 530         kfree(tg);
 531         return NULL;
 532 }
 533
 534 static void throtl_pd_init(struct blkg_policy_data *pd)
 535 {
 536         struct throtl_grp *tg = pd_to_tg(pd);
 537         struct blkcg_gq *blkg = tg_to_blkg(tg);
 538         struct throtl_data *td = blkg->q->td;
 539         struct throtl_service_queue *sq = &tg->service_queue;
 540
 541         /*
 542          * If on the default hierarchy, we switch to properly hierarchical
 543          * behavior where limits on a given throtl_grp are applied to the
 544          * whole subtree rather than just the group itself.  e.g. If 16M
 545          * read_bps limit is set on the root group, the whole system can't
 546          * exceed 16M for the device.
 547          *
 548          * If not on the default hierarchy, the broken flat hierarchy
 549          * behavior is retained where all throtl_grps are treated as if
 550          * they're all separate root groups right below throtl_data.
 551          * Limits of a group don't interact with limits of other groups
 552          * regardless of the position of the group in the hierarchy.
 553          */
 554         sq->parent_sq = &td->service_queue;
 555         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 556                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 557         tg->td = td;
 558 }
 559
 560 /*
 561  * Set has_rules[] if @tg or any of its parents have limits configured.
 562  * This doesn't require walking up to the top of the hierarchy as the
 563  * parent's has_rules[] is guaranteed to be correct.
 564  */
 565 static void tg_update_has_rules(struct throtl_grp *tg)
 566 {
 567         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 568         struct throtl_data *td = tg->td;
 569         int rw;
 570
 571         for (rw = READ; rw <= WRITE; rw++)
 572                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 573                         (td->limit_valid[td->limit_index] &&
 574                          (tg_bps_limit(tg, rw) != U64_MAX ||
 575                           tg_iops_limit(tg, rw) != UINT_MAX));
 576 }
 577
 578 static void throtl_pd_online(struct blkg_policy_data *pd)
 579 {
 580         struct throtl_grp *tg = pd_to_tg(pd);
 581         /*
 582          * We don't want new groups to escape the limits of its ancestors.
 583          * Update has_rules[] after a new group is brought online.
 584          */
 585         tg_update_has_rules(tg);
 586 }
 587
 588 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 589 {
 590         struct cgroup_subsys_state *pos_css;
 591         struct blkcg_gq *blkg;
 592         bool low_valid = false;
 593
 594         rcu_read_lock();
 595         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 596                 struct throtl_grp *tg = blkg_to_tg(blkg);
 597
 598                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 599                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 600                         low_valid = true;
 601                         break;
 602                 }
 603         }
 604         rcu_read_unlock();
 605
 606         td->limit_valid[LIMIT_LOW] = low_valid;
 607 }
 608
 609 static void throtl_upgrade_state(struct throtl_data *td);
 610 static void throtl_pd_offline(struct blkg_policy_data *pd)
 611 {
 612         struct throtl_grp *tg = pd_to_tg(pd);
 613
 614         tg->bps[READ][LIMIT_LOW] = 0;
 615         tg->bps[WRITE][LIMIT_LOW] = 0;
 616         tg->iops[READ][LIMIT_LOW] = 0;
 617         tg->iops[WRITE][LIMIT_LOW] = 0;
 618
 619         blk_throtl_update_limit_valid(tg->td);
 620
 621         if (!tg->td->limit_valid[tg->td->limit_index])
 622                 throtl_upgrade_state(tg->td);
 623 }
 624
 625 static void throtl_pd_free(struct blkg_policy_data *pd)
 626 {
 627         struct throtl_grp *tg = pd_to_tg(pd);
 628
 629         del_timer_sync(&tg->service_queue.pending_timer);
 630         blkg_rwstat_exit(&tg->stat_bytes);
 631         blkg_rwstat_exit(&tg->stat_ios);
 632         kfree(tg);
 633 }
 634
 635 static struct throtl_grp *
 636 throtl_rb_first(struct throtl_service_queue *parent_sq)
 637 {
 638         struct rb_node *n;
 639         /* Service tree is empty */
 640         if (!parent_sq->nr_pending)
 641                 return NULL;
 642
 643         n = rb_first_cached(&parent_sq->pending_tree);
 644         WARN_ON_ONCE(!n);
 645         if (!n)
 646                 return NULL;
 647         return rb_entry_tg(n);
 648 }
 649
 650 static void throtl_rb_erase(struct rb_node *n,
 651                             struct throtl_service_queue *parent_sq)
 652 {
 653         rb_erase_cached(n, &parent_sq->pending_tree);
 654         RB_CLEAR_NODE(n);
 655         --parent_sq->nr_pending;
 656 }
 657
 658 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 659 {
 660         struct throtl_grp *tg;
 661
 662         tg = throtl_rb_first(parent_sq);
 663         if (!tg)
 664                 return;
 665
 666         parent_sq->first_pending_disptime = tg->disptime;
 667 }
 668
 669 static void tg_service_queue_add(struct throtl_grp *tg)
 670 {
 671         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 672         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 673         struct rb_node *parent = NULL;
 674         struct throtl_grp *__tg;
 675         unsigned long key = tg->disptime;
 676         bool leftmost = true;
 677
 678         while (*node != NULL) {
 679                 parent = *node;
 680                 __tg = rb_entry_tg(parent);
 681
 682                 if (time_before(key, __tg->disptime))
 683                         node = &parent->rb_left;
 684                 else {
 685                         node = &parent->rb_right;
 686                         leftmost = false;
 687                 }
 688         }
 689
 690         rb_link_node(&tg->rb_node, parent, node);
 691         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 692                                leftmost);
 693 }
 694
 695 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 696 {
 697         tg_service_queue_add(tg);
 698         tg->flags |= THROTL_TG_PENDING;
 699         tg->service_queue.parent_sq->nr_pending++;
 700 }
 701
 702 static void throtl_enqueue_tg(struct throtl_grp *tg)
 703 {
 704         if (!(tg->flags & THROTL_TG_PENDING))
 705                 __throtl_enqueue_tg(tg);
 706 }
 707
 708 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 709 {
 710         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 711         tg->flags &= ~THROTL_TG_PENDING;
 712 }
 713
 714 static void throtl_dequeue_tg(struct throtl_grp *tg)
 715 {
 716         if (tg->flags & THROTL_TG_PENDING)
 717                 __throtl_dequeue_tg(tg);
 718 }
 719
 720 /* Call with queue lock held */
 721 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 722                                           unsigned long expires)
 723 {
 724         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 725
 726         /*
 727          * Since we are adjusting the throttle limit dynamically, the sleep
 728          * time calculated according to previous limit might be invalid. It's
 729          * possible the cgroup sleep time is very long and no other cgroups
 730          * have IO running so notify the limit changes. Make sure the cgroup
 731          * doesn't sleep too long to avoid the missed notification.
 732          */
 733         if (time_after(expires, max_expire))
 734                 expires = max_expire;
 735         mod_timer(&sq->pending_timer, expires);
 736         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 737                    expires - jiffies, jiffies);
 738 }
 739
 740 /**
 741  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 742  * @sq: the service_queue to schedule dispatch for
 743  * @force: force scheduling
 744  *
 745  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 746  * dispatch time of the first pending child.  Returns %true if either timer
 747  * is armed or there's no pending child left.  %false if the current
 748  * dispatch window is still open and the caller should continue
 749  * dispatching.
 750  *
 751  * If @force is %true, the dispatch timer is always scheduled and this
 752  * function is guaranteed to return %true.  This is to be used when the
 753  * caller can't dispatch itself and needs to invoke pending_timer
 754  * unconditionally.  Note that forced scheduling is likely to induce short
 755  * delay before dispatch starts even if @sq->first_pending_disptime is not
 756  * in the future and thus shouldn't be used in hot paths.
 757  */
 758 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 759                                           bool force)
 760 {
 761         /* any pending children left? */
 762         if (!sq->nr_pending)
 763                 return true;
 764
 765         update_min_dispatch_time(sq);
 766
 767         /* is the next dispatch time in the future? */
 768         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 769                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 770                 return true;
 771         }
 772
 773         /* tell the caller to continue dispatching */
 774         return false;
 775 }
 776
 777 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 778                 bool rw, unsigned long start)
 779 {
 780         tg->bytes_disp[rw] = 0;
 781         tg->io_disp[rw] = 0;
 782
 783         /*
 784          * Previous slice has expired. We must have trimmed it after last
 785          * bio dispatch. That means since start of last slice, we never used
 786          * that bandwidth. Do try to make use of that bandwidth while giving
 787          * credit.
 788          */
 789         if (time_after_eq(start, tg->slice_start[rw]))
 790                 tg->slice_start[rw] = start;
 791
 792         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 793         throtl_log(&tg->service_queue,
 794                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 795                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 796                    tg->slice_end[rw], jiffies);
 797 }
 798
 799 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 800 {
 801         tg->bytes_disp[rw] = 0;
 802         tg->io_disp[rw] = 0;
 803         tg->slice_start[rw] = jiffies;
 804         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 805         throtl_log(&tg->service_queue,
 806                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 807                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 808                    tg->slice_end[rw], jiffies);
 809 }
 810
 811 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 812                                         unsigned long jiffy_end)
 813 {
 814         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 815 }
 816
 817 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 818                                        unsigned long jiffy_end)
 819 {
 820         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 821         throtl_log(&tg->service_queue,
 822                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 823                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 824                    tg->slice_end[rw], jiffies);
 825 }
 826
 827 /* Determine if previously allocated or extended slice is complete or not */
 828 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 829 {
 830         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 831                 return false;
 832
 833         return true;
 834 }
 835
 836 /* Trim the used slices and adjust slice start accordingly */
 837 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 838 {
 839         unsigned long nr_slices, time_elapsed, io_trim;
 840         u64 bytes_trim, tmp;
 841
 842         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 843
 844         /*
 845          * If bps are unlimited (-1), then time slice don't get
 846          * renewed. Don't try to trim the slice if slice is used. A new
 847          * slice will start when appropriate.
 848          */
 849         if (throtl_slice_used(tg, rw))
 850                 return;
 851
 852         /*
 853          * A bio has been dispatched. Also adjust slice_end. It might happen
 854          * that initially cgroup limit was very low resulting in high
 855          * slice_end, but later limit was bumped up and bio was dispached
 856          * sooner, then we need to reduce slice_end. A high bogus slice_end
 857          * is bad because it does not allow new slice to start.
 858          */
 859
 860         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 861
 862         time_elapsed = jiffies - tg->slice_start[rw];
 863
 864         nr_slices = time_elapsed / tg->td->throtl_slice;
 865
 866         if (!nr_slices)
 867                 return;
 868         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 869         do_div(tmp, HZ);
 870         bytes_trim = tmp;
 871
 872         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 873                 HZ;
 874
 875         if (!bytes_trim && !io_trim)
 876                 return;
 877
 878         if (tg->bytes_disp[rw] >= bytes_trim)
 879                 tg->bytes_disp[rw] -= bytes_trim;
 880         else
 881                 tg->bytes_disp[rw] = 0;
 882
 883         if (tg->io_disp[rw] >= io_trim)
 884                 tg->io_disp[rw] -= io_trim;
 885         else
 886                 tg->io_disp[rw] = 0;
 887
 888         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 889
 890         throtl_log(&tg->service_queue,
 891                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 892                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 893                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 894 }
 895
 896 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 897                                   unsigned long *wait)
 898 {
 899         bool rw = bio_data_dir(bio);
 900         unsigned int io_allowed;
 901         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 902         u64 tmp;
 903
 904         jiffy_elapsed = jiffies - tg->slice_start[rw];
 905
 906         /* Round up to the next throttle slice, wait time must be nonzero */
 907         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 908
 909         /*
 910          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 911          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 912          * will allow dispatch after 1 second and after that slice should
 913          * have been trimmed.
 914          */
 915
 916         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 917         do_div(tmp, HZ);
 918
 919         if (tmp > UINT_MAX)
 920                 io_allowed = UINT_MAX;
 921         else
 922                 io_allowed = tmp;
 923
 924         if (tg->io_disp[rw] + 1 <= io_allowed) {
 925                 if (wait)
 926                         *wait = 0;
 927                 return true;
 928         }
 929
 930         /* Calc approx time to dispatch */
 931         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 932
 933         if (wait)
 934                 *wait = jiffy_wait;
 935         return false;
 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 false;
 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 true;
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 false;
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;
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                 sq = &tg->service_queue;
1226                 if (sq->nr_queued[0] || sq->nr_queued[1])
1227                         tg_update_disptime(tg);
1228
1229                 if (nr_disp >= throtl_quantum)
1230                         break;
1231         }
1232
1233         return nr_disp;
1234 }
1235
1236 static bool throtl_can_upgrade(struct throtl_data *td,
1237         struct throtl_grp *this_tg);
1238 /**
1239  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1240  * @t: the pending_timer member of the throtl_service_queue being serviced
1241  *
1242  * This timer is armed when a child throtl_grp with active bio's become
1243  * pending and queued on the service_queue's pending_tree and expires when
1244  * the first child throtl_grp should be dispatched.  This function
1245  * dispatches bio's from the children throtl_grps to the parent
1246  * service_queue.
1247  *
1248  * If the parent's parent is another throtl_grp, dispatching is propagated
1249  * by either arming its pending_timer or repeating dispatch directly.  If
1250  * the top-level service_tree is reached, throtl_data->dispatch_work is
1251  * kicked so that the ready bio's are issued.
1252  */
1253 static void throtl_pending_timer_fn(struct timer_list *t)
1254 {
1255         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1256         struct throtl_grp *tg = sq_to_tg(sq);
1257         struct throtl_data *td = sq_to_td(sq);
1258         struct request_queue *q = td->queue;
1259         struct throtl_service_queue *parent_sq;
1260         bool dispatched;
1261         int ret;
1262
1263         spin_lock_irq(&q->queue_lock);
1264         if (throtl_can_upgrade(td, NULL))
1265                 throtl_upgrade_state(td);
1266
1267 again:
1268         parent_sq = sq->parent_sq;
1269         dispatched = false;
1270
1271         while (true) {
1272                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1273                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1274                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1275
1276                 ret = throtl_select_dispatch(sq);
1277                 if (ret) {
1278                         throtl_log(sq, "bios disp=%u", ret);
1279                         dispatched = true;
1280                 }
1281
1282                 if (throtl_schedule_next_dispatch(sq, false))
1283                         break;
1284
1285                 /* this dispatch windows is still open, relax and repeat */
1286                 spin_unlock_irq(&q->queue_lock);
1287                 cpu_relax();
1288                 spin_lock_irq(&q->queue_lock);
1289         }
1290
1291         if (!dispatched)
1292                 goto out_unlock;
1293
1294         if (parent_sq) {
1295                 /* @parent_sq is another throl_grp, propagate dispatch */
1296                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1297                         tg_update_disptime(tg);
1298                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1299                                 /* window is already open, repeat dispatching */
1300                                 sq = parent_sq;
1301                                 tg = sq_to_tg(sq);
1302                                 goto again;
1303                         }
1304                 }
1305         } else {
1306                 /* reached the top-level, queue issueing */
1307                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1308         }
1309 out_unlock:
1310         spin_unlock_irq(&q->queue_lock);
1311 }
1312
1313 /**
1314  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1315  * @work: work item being executed
1316  *
1317  * This function is queued for execution when bio's reach the bio_lists[]
1318  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1319  * function.
1320  */
1321 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1322 {
1323         struct throtl_data *td = container_of(work, struct throtl_data,
1324                                               dispatch_work);
1325         struct throtl_service_queue *td_sq = &td->service_queue;
1326         struct request_queue *q = td->queue;
1327         struct bio_list bio_list_on_stack;
1328         struct bio *bio;
1329         struct blk_plug plug;
1330         int rw;
1331
1332         bio_list_init(&bio_list_on_stack);
1333
1334         spin_lock_irq(&q->queue_lock);
1335         for (rw = READ; rw <= WRITE; rw++)
1336                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1337                         bio_list_add(&bio_list_on_stack, bio);
1338         spin_unlock_irq(&q->queue_lock);
1339
1340         if (!bio_list_empty(&bio_list_on_stack)) {
1341                 blk_start_plug(&plug);
1342                 while((bio = bio_list_pop(&bio_list_on_stack)))
1343                         generic_make_request(bio);
1344                 blk_finish_plug(&plug);
1345         }
1346 }
1347
1348 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1349                               int off)
1350 {
1351         struct throtl_grp *tg = pd_to_tg(pd);
1352         u64 v = *(u64 *)((void *)tg + off);
1353
1354         if (v == U64_MAX)
1355                 return 0;
1356         return __blkg_prfill_u64(sf, pd, v);
1357 }
1358
1359 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1360                                int off)
1361 {
1362         struct throtl_grp *tg = pd_to_tg(pd);
1363         unsigned int v = *(unsigned int *)((void *)tg + off);
1364
1365         if (v == UINT_MAX)
1366                 return 0;
1367         return __blkg_prfill_u64(sf, pd, v);
1368 }
1369
1370 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1371 {
1372         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1373                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1374         return 0;
1375 }
1376
1377 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1378 {
1379         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1380                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1381         return 0;
1382 }
1383
1384 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1385 {
1386         struct throtl_service_queue *sq = &tg->service_queue;
1387         struct cgroup_subsys_state *pos_css;
1388         struct blkcg_gq *blkg;
1389
1390         throtl_log(&tg->service_queue,
1391                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1392                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1393                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1394
1395         /*
1396          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1397          * considered to have rules if either the tg itself or any of its
1398          * ancestors has rules.  This identifies groups without any
1399          * restrictions in the whole hierarchy and allows them to bypass
1400          * blk-throttle.
1401          */
1402         blkg_for_each_descendant_pre(blkg, pos_css,
1403                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1404                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1405                 struct throtl_grp *parent_tg;
1406
1407                 tg_update_has_rules(this_tg);
1408                 /* ignore root/second level */
1409                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1410                     !blkg->parent->parent)
1411                         continue;
1412                 parent_tg = blkg_to_tg(blkg->parent);
1413                 /*
1414                  * make sure all children has lower idle time threshold and
1415                  * higher latency target
1416                  */
1417                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1418                                 parent_tg->idletime_threshold);
1419                 this_tg->latency_target = max(this_tg->latency_target,
1420                                 parent_tg->latency_target);
1421         }
1422
1423         /*
1424          * We're already holding queue_lock and know @tg is valid.  Let's
1425          * apply the new config directly.
1426          *
1427          * Restart the slices for both READ and WRITES. It might happen
1428          * that a group's limit are dropped suddenly and we don't want to
1429          * account recently dispatched IO with new low rate.
1430          */
1431         throtl_start_new_slice(tg, 0);
1432         throtl_start_new_slice(tg, 1);
1433
1434         if (tg->flags & THROTL_TG_PENDING) {
1435                 tg_update_disptime(tg);
1436                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1437         }
1438 }
1439
1440 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1441                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1442 {
1443         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1444         struct blkg_conf_ctx ctx;
1445         struct throtl_grp *tg;
1446         int ret;
1447         u64 v;
1448
1449         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1450         if (ret)
1451                 return ret;
1452
1453         ret = -EINVAL;
1454         if (sscanf(ctx.body, "%llu", &v) != 1)
1455                 goto out_finish;
1456         if (!v)
1457                 v = U64_MAX;
1458
1459         tg = blkg_to_tg(ctx.blkg);
1460
1461         if (is_u64)
1462                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1463         else
1464                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1465
1466         tg_conf_updated(tg, false);
1467         ret = 0;
1468 out_finish:
1469         blkg_conf_finish(&ctx);
1470         return ret ?: nbytes;
1471 }
1472
1473 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1474                                char *buf, size_t nbytes, loff_t off)
1475 {
1476         return tg_set_conf(of, buf, nbytes, off, true);
1477 }
1478
1479 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1480                                 char *buf, size_t nbytes, loff_t off)
1481 {
1482         return tg_set_conf(of, buf, nbytes, off, false);
1483 }
1484
1485 static int tg_print_rwstat(struct seq_file *sf, void *v)
1486 {
1487         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1488                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1489                           seq_cft(sf)->private, true);
1490         return 0;
1491 }
1492
1493 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1494                                       struct blkg_policy_data *pd, int off)
1495 {
1496         struct blkg_rwstat_sample sum;
1497
1498         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1499                                   &sum);
1500         return __blkg_prfill_rwstat(sf, pd, &sum);
1501 }
1502
1503 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1504 {
1505         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1506                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1507                           seq_cft(sf)->private, true);
1508         return 0;
1509 }
1510
1511 static struct cftype throtl_legacy_files[] = {
1512         {
1513                 .name = "throttle.read_bps_device",
1514                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1515                 .seq_show = tg_print_conf_u64,
1516                 .write = tg_set_conf_u64,
1517         },
1518         {
1519                 .name = "throttle.write_bps_device",
1520                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1521                 .seq_show = tg_print_conf_u64,
1522                 .write = tg_set_conf_u64,
1523         },
1524         {
1525                 .name = "throttle.read_iops_device",
1526                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1527                 .seq_show = tg_print_conf_uint,
1528                 .write = tg_set_conf_uint,
1529         },
1530         {
1531                 .name = "throttle.write_iops_device",
1532                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1533                 .seq_show = tg_print_conf_uint,
1534                 .write = tg_set_conf_uint,
1535         },
1536         {
1537                 .name = "throttle.io_service_bytes",
1538                 .private = offsetof(struct throtl_grp, stat_bytes),
1539                 .seq_show = tg_print_rwstat,
1540         },
1541         {
1542                 .name = "throttle.io_service_bytes_recursive",
1543                 .private = offsetof(struct throtl_grp, stat_bytes),
1544                 .seq_show = tg_print_rwstat_recursive,
1545         },
1546         {
1547                 .name = "throttle.io_serviced",
1548                 .private = offsetof(struct throtl_grp, stat_ios),
1549                 .seq_show = tg_print_rwstat,
1550         },
1551         {
1552                 .name = "throttle.io_serviced_recursive",
1553                 .private = offsetof(struct throtl_grp, stat_ios),
1554                 .seq_show = tg_print_rwstat_recursive,
1555         },
1556         { }     /* terminate */
1557 };
1558
1559 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1560                          int off)
1561 {
1562         struct throtl_grp *tg = pd_to_tg(pd);
1563         const char *dname = blkg_dev_name(pd->blkg);
1564         char bufs[4][21] = { "max", "max", "max", "max" };
1565         u64 bps_dft;
1566         unsigned int iops_dft;
1567         char idle_time[26] = "";
1568         char latency_time[26] = "";
1569
1570         if (!dname)
1571                 return 0;
1572
1573         if (off == LIMIT_LOW) {
1574                 bps_dft = 0;
1575                 iops_dft = 0;
1576         } else {
1577                 bps_dft = U64_MAX;
1578                 iops_dft = UINT_MAX;
1579         }
1580
1581         if (tg->bps_conf[READ][off] == bps_dft &&
1582             tg->bps_conf[WRITE][off] == bps_dft &&
1583             tg->iops_conf[READ][off] == iops_dft &&
1584             tg->iops_conf[WRITE][off] == iops_dft &&
1585             (off != LIMIT_LOW ||
1586              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1587               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1588                 return 0;
1589
1590         if (tg->bps_conf[READ][off] != U64_MAX)
1591                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1592                         tg->bps_conf[READ][off]);
1593         if (tg->bps_conf[WRITE][off] != U64_MAX)
1594                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1595                         tg->bps_conf[WRITE][off]);
1596         if (tg->iops_conf[READ][off] != UINT_MAX)
1597                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1598                         tg->iops_conf[READ][off]);
1599         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1600                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1601                         tg->iops_conf[WRITE][off]);
1602         if (off == LIMIT_LOW) {
1603                 if (tg->idletime_threshold_conf == ULONG_MAX)
1604                         strcpy(idle_time, " idle=max");
1605                 else
1606                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1607                                 tg->idletime_threshold_conf);
1608
1609                 if (tg->latency_target_conf == ULONG_MAX)
1610                         strcpy(latency_time, " latency=max");
1611                 else
1612                         snprintf(latency_time, sizeof(latency_time),
1613                                 " latency=%lu", tg->latency_target_conf);
1614         }
1615
1616         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1617                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1618                    latency_time);
1619         return 0;
1620 }
1621
1622 static int tg_print_limit(struct seq_file *sf, void *v)
1623 {
1624         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1625                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1626         return 0;
1627 }
1628
1629 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1630                           char *buf, size_t nbytes, loff_t off)
1631 {
1632         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1633         struct blkg_conf_ctx ctx;
1634         struct throtl_grp *tg;
1635         u64 v[4];
1636         unsigned long idle_time;
1637         unsigned long latency_time;
1638         int ret;
1639         int index = of_cft(of)->private;
1640
1641         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1642         if (ret)
1643                 return ret;
1644
1645         tg = blkg_to_tg(ctx.blkg);
1646
1647         v[0] = tg->bps_conf[READ][index];
1648         v[1] = tg->bps_conf[WRITE][index];
1649         v[2] = tg->iops_conf[READ][index];
1650         v[3] = tg->iops_conf[WRITE][index];
1651
1652         idle_time = tg->idletime_threshold_conf;
1653         latency_time = tg->latency_target_conf;
1654         while (true) {
1655                 char tok[27];   /* wiops=18446744073709551616 */
1656                 char *p;
1657                 u64 val = U64_MAX;
1658                 int len;
1659
1660                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1661                         break;
1662                 if (tok[0] == '\0')
1663                         break;
1664                 ctx.body += len;
1665
1666                 ret = -EINVAL;
1667                 p = tok;
1668                 strsep(&p, "=");
1669                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1670                         goto out_finish;
1671
1672                 ret = -ERANGE;
1673                 if (!val)
1674                         goto out_finish;
1675
1676                 ret = -EINVAL;
1677                 if (!strcmp(tok, "rbps"))
1678                         v[0] = val;
1679                 else if (!strcmp(tok, "wbps"))
1680                         v[1] = val;
1681                 else if (!strcmp(tok, "riops"))
1682                         v[2] = min_t(u64, val, UINT_MAX);
1683                 else if (!strcmp(tok, "wiops"))
1684                         v[3] = min_t(u64, val, UINT_MAX);
1685                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1686                         idle_time = val;
1687                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1688                         latency_time = val;
1689                 else
1690                         goto out_finish;
1691         }
1692
1693         tg->bps_conf[READ][index] = v[0];
1694         tg->bps_conf[WRITE][index] = v[1];
1695         tg->iops_conf[READ][index] = v[2];
1696         tg->iops_conf[WRITE][index] = v[3];
1697
1698         if (index == LIMIT_MAX) {
1699                 tg->bps[READ][index] = v[0];
1700                 tg->bps[WRITE][index] = v[1];
1701                 tg->iops[READ][index] = v[2];
1702                 tg->iops[WRITE][index] = v[3];
1703         }
1704         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1705                 tg->bps_conf[READ][LIMIT_MAX]);
1706         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1707                 tg->bps_conf[WRITE][LIMIT_MAX]);
1708         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1709                 tg->iops_conf[READ][LIMIT_MAX]);
1710         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1711                 tg->iops_conf[WRITE][LIMIT_MAX]);
1712         tg->idletime_threshold_conf = idle_time;
1713         tg->latency_target_conf = latency_time;
1714
1715         /* force user to configure all settings for low limit  */
1716         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1717               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1718             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1719             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1720                 tg->bps[READ][LIMIT_LOW] = 0;
1721                 tg->bps[WRITE][LIMIT_LOW] = 0;
1722                 tg->iops[READ][LIMIT_LOW] = 0;
1723                 tg->iops[WRITE][LIMIT_LOW] = 0;
1724                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1725                 tg->latency_target = DFL_LATENCY_TARGET;
1726         } else if (index == LIMIT_LOW) {
1727                 tg->idletime_threshold = tg->idletime_threshold_conf;
1728                 tg->latency_target = tg->latency_target_conf;
1729         }
1730
1731         blk_throtl_update_limit_valid(tg->td);
1732         if (tg->td->limit_valid[LIMIT_LOW]) {
1733                 if (index == LIMIT_LOW)
1734                         tg->td->limit_index = LIMIT_LOW;
1735         } else
1736                 tg->td->limit_index = LIMIT_MAX;
1737         tg_conf_updated(tg, index == LIMIT_LOW &&
1738                 tg->td->limit_valid[LIMIT_LOW]);
1739         ret = 0;
1740 out_finish:
1741         blkg_conf_finish(&ctx);
1742         return ret ?: nbytes;
1743 }
1744
1745 static struct cftype throtl_files[] = {
1746 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1747         {
1748                 .name = "low",
1749                 .flags = CFTYPE_NOT_ON_ROOT,
1750                 .seq_show = tg_print_limit,
1751                 .write = tg_set_limit,
1752                 .private = LIMIT_LOW,
1753         },
1754 #endif
1755         {
1756                 .name = "max",
1757                 .flags = CFTYPE_NOT_ON_ROOT,
1758                 .seq_show = tg_print_limit,
1759                 .write = tg_set_limit,
1760                 .private = LIMIT_MAX,
1761         },
1762         { }     /* terminate */
1763 };
1764
1765 static void throtl_shutdown_wq(struct request_queue *q)
1766 {
1767         struct throtl_data *td = q->td;
1768
1769         cancel_work_sync(&td->dispatch_work);
1770 }
1771
1772 static struct blkcg_policy blkcg_policy_throtl = {
1773         .dfl_cftypes            = throtl_files,
1774         .legacy_cftypes         = throtl_legacy_files,
1775
1776         .pd_alloc_fn            = throtl_pd_alloc,
1777         .pd_init_fn             = throtl_pd_init,
1778         .pd_online_fn           = throtl_pd_online,
1779         .pd_offline_fn          = throtl_pd_offline,
1780         .pd_free_fn             = throtl_pd_free,
1781 };
1782
1783 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1784 {
1785         unsigned long rtime = jiffies, wtime = jiffies;
1786
1787         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1788                 rtime = tg->last_low_overflow_time[READ];
1789         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1790                 wtime = tg->last_low_overflow_time[WRITE];
1791         return min(rtime, wtime);
1792 }
1793
1794 /* tg should not be an intermediate node */
1795 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1796 {
1797         struct throtl_service_queue *parent_sq;
1798         struct throtl_grp *parent = tg;
1799         unsigned long ret = __tg_last_low_overflow_time(tg);
1800
1801         while (true) {
1802                 parent_sq = parent->service_queue.parent_sq;
1803                 parent = sq_to_tg(parent_sq);
1804                 if (!parent)
1805                         break;
1806
1807                 /*
1808                  * The parent doesn't have low limit, it always reaches low
1809                  * limit. Its overflow time is useless for children
1810                  */
1811                 if (!parent->bps[READ][LIMIT_LOW] &&
1812                     !parent->iops[READ][LIMIT_LOW] &&
1813                     !parent->bps[WRITE][LIMIT_LOW] &&
1814                     !parent->iops[WRITE][LIMIT_LOW])
1815                         continue;
1816                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1817                         ret = __tg_last_low_overflow_time(parent);
1818         }
1819         return ret;
1820 }
1821
1822 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1823 {
1824         /*
1825          * cgroup is idle if:
1826          * - single idle is too long, longer than a fixed value (in case user
1827          *   configure a too big threshold) or 4 times of idletime threshold
1828          * - average think time is more than threshold
1829          * - IO latency is largely below threshold
1830          */
1831         unsigned long time;
1832         bool ret;
1833
1834         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1835         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1836               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1837               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1838               tg->avg_idletime > tg->idletime_threshold ||
1839               (tg->latency_target && tg->bio_cnt &&
1840                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1841         throtl_log(&tg->service_queue,
1842                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1843                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1844                 tg->bio_cnt, ret, tg->td->scale);
1845         return ret;
1846 }
1847
1848 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1849 {
1850         struct throtl_service_queue *sq = &tg->service_queue;
1851         bool read_limit, write_limit;
1852
1853         /*
1854          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1855          * reaches), it's ok to upgrade to next limit
1856          */
1857         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1858         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1859         if (!read_limit && !write_limit)
1860                 return true;
1861         if (read_limit && sq->nr_queued[READ] &&
1862             (!write_limit || sq->nr_queued[WRITE]))
1863                 return true;
1864         if (write_limit && sq->nr_queued[WRITE] &&
1865             (!read_limit || sq->nr_queued[READ]))
1866                 return true;
1867
1868         if (time_after_eq(jiffies,
1869                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1870             throtl_tg_is_idle(tg))
1871                 return true;
1872         return false;
1873 }
1874
1875 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1876 {
1877         while (true) {
1878                 if (throtl_tg_can_upgrade(tg))
1879                         return true;
1880                 tg = sq_to_tg(tg->service_queue.parent_sq);
1881                 if (!tg || !tg_to_blkg(tg)->parent)
1882                         return false;
1883         }
1884         return false;
1885 }
1886
1887 static bool throtl_can_upgrade(struct throtl_data *td,
1888         struct throtl_grp *this_tg)
1889 {
1890         struct cgroup_subsys_state *pos_css;
1891         struct blkcg_gq *blkg;
1892
1893         if (td->limit_index != LIMIT_LOW)
1894                 return false;
1895
1896         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1897                 return false;
1898
1899         rcu_read_lock();
1900         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1901                 struct throtl_grp *tg = blkg_to_tg(blkg);
1902
1903                 if (tg == this_tg)
1904                         continue;
1905                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1906                         continue;
1907                 if (!throtl_hierarchy_can_upgrade(tg)) {
1908                         rcu_read_unlock();
1909                         return false;
1910                 }
1911         }
1912         rcu_read_unlock();
1913         return true;
1914 }
1915
1916 static void throtl_upgrade_check(struct throtl_grp *tg)
1917 {
1918         unsigned long now = jiffies;
1919
1920         if (tg->td->limit_index != LIMIT_LOW)
1921                 return;
1922
1923         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1924                 return;
1925
1926         tg->last_check_time = now;
1927
1928         if (!time_after_eq(now,
1929              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1930                 return;
1931
1932         if (throtl_can_upgrade(tg->td, NULL))
1933                 throtl_upgrade_state(tg->td);
1934 }
1935
1936 static void throtl_upgrade_state(struct throtl_data *td)
1937 {
1938         struct cgroup_subsys_state *pos_css;
1939         struct blkcg_gq *blkg;
1940
1941         throtl_log(&td->service_queue, "upgrade to max");
1942         td->limit_index = LIMIT_MAX;
1943         td->low_upgrade_time = jiffies;
1944         td->scale = 0;
1945         rcu_read_lock();
1946         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1947                 struct throtl_grp *tg = blkg_to_tg(blkg);
1948                 struct throtl_service_queue *sq = &tg->service_queue;
1949
1950                 tg->disptime = jiffies - 1;
1951                 throtl_select_dispatch(sq);
1952                 throtl_schedule_next_dispatch(sq, true);
1953         }
1954         rcu_read_unlock();
1955         throtl_select_dispatch(&td->service_queue);
1956         throtl_schedule_next_dispatch(&td->service_queue, true);
1957         queue_work(kthrotld_workqueue, &td->dispatch_work);
1958 }
1959
1960 static void throtl_downgrade_state(struct throtl_data *td, int new)
1961 {
1962         td->scale /= 2;
1963
1964         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1965         if (td->scale) {
1966                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1967                 return;
1968         }
1969
1970         td->limit_index = new;
1971         td->low_downgrade_time = jiffies;
1972 }
1973
1974 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1975 {
1976         struct throtl_data *td = tg->td;
1977         unsigned long now = jiffies;
1978
1979         /*
1980          * If cgroup is below low limit, consider downgrade and throttle other
1981          * cgroups
1982          */
1983         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1984             time_after_eq(now, tg_last_low_overflow_time(tg) +
1985                                         td->throtl_slice) &&
1986             (!throtl_tg_is_idle(tg) ||
1987              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1988                 return true;
1989         return false;
1990 }
1991
1992 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1993 {
1994         while (true) {
1995                 if (!throtl_tg_can_downgrade(tg))
1996                         return false;
1997                 tg = sq_to_tg(tg->service_queue.parent_sq);
1998                 if (!tg || !tg_to_blkg(tg)->parent)
1999                         break;
2000         }
2001         return true;
2002 }
2003
2004 static void throtl_downgrade_check(struct throtl_grp *tg)
2005 {
2006         uint64_t bps;
2007         unsigned int iops;
2008         unsigned long elapsed_time;
2009         unsigned long now = jiffies;
2010
2011         if (tg->td->limit_index != LIMIT_MAX ||
2012             !tg->td->limit_valid[LIMIT_LOW])
2013                 return;
2014         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2015                 return;
2016         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2017                 return;
2018
2019         elapsed_time = now - tg->last_check_time;
2020         tg->last_check_time = now;
2021
2022         if (time_before(now, tg_last_low_overflow_time(tg) +
2023                         tg->td->throtl_slice))
2024                 return;
2025
2026         if (tg->bps[READ][LIMIT_LOW]) {
2027                 bps = tg->last_bytes_disp[READ] * HZ;
2028                 do_div(bps, elapsed_time);
2029                 if (bps >= tg->bps[READ][LIMIT_LOW])
2030                         tg->last_low_overflow_time[READ] = now;
2031         }
2032
2033         if (tg->bps[WRITE][LIMIT_LOW]) {
2034                 bps = tg->last_bytes_disp[WRITE] * HZ;
2035                 do_div(bps, elapsed_time);
2036                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2037                         tg->last_low_overflow_time[WRITE] = now;
2038         }
2039
2040         if (tg->iops[READ][LIMIT_LOW]) {
2041                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2042                 if (iops >= tg->iops[READ][LIMIT_LOW])
2043                         tg->last_low_overflow_time[READ] = now;
2044         }
2045
2046         if (tg->iops[WRITE][LIMIT_LOW]) {
2047                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2048                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2049                         tg->last_low_overflow_time[WRITE] = now;
2050         }
2051
2052         /*
2053          * If cgroup is below low limit, consider downgrade and throttle other
2054          * cgroups
2055          */
2056         if (throtl_hierarchy_can_downgrade(tg))
2057                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2058
2059         tg->last_bytes_disp[READ] = 0;
2060         tg->last_bytes_disp[WRITE] = 0;
2061         tg->last_io_disp[READ] = 0;
2062         tg->last_io_disp[WRITE] = 0;
2063 }
2064
2065 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2066 {
2067         unsigned long now = ktime_get_ns() >> 10;
2068         unsigned long last_finish_time = tg->last_finish_time;
2069
2070         if (now <= last_finish_time || last_finish_time == 0 ||
2071             last_finish_time == tg->checked_last_finish_time)
2072                 return;
2073
2074         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2075         tg->checked_last_finish_time = last_finish_time;
2076 }
2077
2078 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2079 static void throtl_update_latency_buckets(struct throtl_data *td)
2080 {
2081         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2082         int i, cpu, rw;
2083         unsigned long last_latency[2] = { 0 };
2084         unsigned long latency[2];
2085
2086         if (!blk_queue_nonrot(td->queue))
2087                 return;
2088         if (time_before(jiffies, td->last_calculate_time + HZ))
2089                 return;
2090         td->last_calculate_time = jiffies;
2091
2092         memset(avg_latency, 0, sizeof(avg_latency));
2093         for (rw = READ; rw <= WRITE; rw++) {
2094                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2095                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2096
2097                         for_each_possible_cpu(cpu) {
2098                                 struct latency_bucket *bucket;
2099
2100                                 /* this isn't race free, but ok in practice */
2101                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2102                                         cpu);
2103                                 tmp->total_latency += bucket[i].total_latency;
2104                                 tmp->samples += bucket[i].samples;
2105                                 bucket[i].total_latency = 0;
2106                                 bucket[i].samples = 0;
2107                         }
2108
2109                         if (tmp->samples >= 32) {
2110                                 int samples = tmp->samples;
2111
2112                                 latency[rw] = tmp->total_latency;
2113
2114                                 tmp->total_latency = 0;
2115                                 tmp->samples = 0;
2116                                 latency[rw] /= samples;
2117                                 if (latency[rw] == 0)
2118                                         continue;
2119                                 avg_latency[rw][i].latency = latency[rw];
2120                         }
2121                 }
2122         }
2123
2124         for (rw = READ; rw <= WRITE; rw++) {
2125                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2126                         if (!avg_latency[rw][i].latency) {
2127                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2128                                         td->avg_buckets[rw][i].latency =
2129                                                 last_latency[rw];
2130                                 continue;
2131                         }
2132
2133                         if (!td->avg_buckets[rw][i].valid)
2134                                 latency[rw] = avg_latency[rw][i].latency;
2135                         else
2136                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2137                                         avg_latency[rw][i].latency) >> 3;
2138
2139                         td->avg_buckets[rw][i].latency = max(latency[rw],
2140                                 last_latency[rw]);
2141                         td->avg_buckets[rw][i].valid = true;
2142                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2143                 }
2144         }
2145
2146         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2147                 throtl_log(&td->service_queue,
2148                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2149                         "write latency=%ld, write valid=%d", i,
2150                         td->avg_buckets[READ][i].latency,
2151                         td->avg_buckets[READ][i].valid,
2152                         td->avg_buckets[WRITE][i].latency,
2153                         td->avg_buckets[WRITE][i].valid);
2154 }
2155 #else
2156 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2157 {
2158 }
2159 #endif
2160
2161 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2162                     struct bio *bio)
2163 {
2164         struct throtl_qnode *qn = NULL;
2165         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2166         struct throtl_service_queue *sq;
2167         bool rw = bio_data_dir(bio);
2168         bool throttled = false;
2169         struct throtl_data *td = tg->td;
2170
2171         WARN_ON_ONCE(!rcu_read_lock_held());
2172
2173         /* see throtl_charge_bio() */
2174         if (bio_flagged(bio, BIO_THROTTLED))
2175                 goto out;
2176
2177         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2178                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2179                                 bio->bi_iter.bi_size);
2180                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2181         }
2182
2183         if (!tg->has_rules[rw])
2184                 goto out;
2185
2186         spin_lock_irq(&q->queue_lock);
2187
2188         throtl_update_latency_buckets(td);
2189
2190         blk_throtl_update_idletime(tg);
2191
2192         sq = &tg->service_queue;
2193
2194 again:
2195         while (true) {
2196                 if (tg->last_low_overflow_time[rw] == 0)
2197                         tg->last_low_overflow_time[rw] = jiffies;
2198                 throtl_downgrade_check(tg);
2199                 throtl_upgrade_check(tg);
2200                 /* throtl is FIFO - if bios are already queued, should queue */
2201                 if (sq->nr_queued[rw])
2202                         break;
2203
2204                 /* if above limits, break to queue */
2205                 if (!tg_may_dispatch(tg, bio, NULL)) {
2206                         tg->last_low_overflow_time[rw] = jiffies;
2207                         if (throtl_can_upgrade(td, tg)) {
2208                                 throtl_upgrade_state(td);
2209                                 goto again;
2210                         }
2211                         break;
2212                 }
2213
2214                 /* within limits, let's charge and dispatch directly */
2215                 throtl_charge_bio(tg, bio);
2216
2217                 /*
2218                  * We need to trim slice even when bios are not being queued
2219                  * otherwise it might happen that a bio is not queued for
2220                  * a long time and slice keeps on extending and trim is not
2221                  * called for a long time. Now if limits are reduced suddenly
2222                  * we take into account all the IO dispatched so far at new
2223                  * low rate and * newly queued IO gets a really long dispatch
2224                  * time.
2225                  *
2226                  * So keep on trimming slice even if bio is not queued.
2227                  */
2228                 throtl_trim_slice(tg, rw);
2229
2230                 /*
2231                  * @bio passed through this layer without being throttled.
2232                  * Climb up the ladder.  If we''re already at the top, it
2233                  * can be executed directly.
2234                  */
2235                 qn = &tg->qnode_on_parent[rw];
2236                 sq = sq->parent_sq;
2237                 tg = sq_to_tg(sq);
2238                 if (!tg)
2239                         goto out_unlock;
2240         }
2241
2242         /* out-of-limit, queue to @tg */
2243         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2244                    rw == READ ? 'R' : 'W',
2245                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2246                    tg_bps_limit(tg, rw),
2247                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2248                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2249
2250         tg->last_low_overflow_time[rw] = jiffies;
2251
2252         td->nr_queued[rw]++;
2253         throtl_add_bio_tg(bio, qn, tg);
2254         throttled = true;
2255
2256         /*
2257          * Update @tg's dispatch time and force schedule dispatch if @tg
2258          * was empty before @bio.  The forced scheduling isn't likely to
2259          * cause undue delay as @bio is likely to be dispatched directly if
2260          * its @tg's disptime is not in the future.
2261          */
2262         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2263                 tg_update_disptime(tg);
2264                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2265         }
2266
2267 out_unlock:
2268         spin_unlock_irq(&q->queue_lock);
2269 out:
2270         bio_set_flag(bio, BIO_THROTTLED);
2271
2272 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2273         if (throttled || !td->track_bio_latency)
2274                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2275 #endif
2276         return throttled;
2277 }
2278
2279 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2280 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2281         int op, unsigned long time)
2282 {
2283         struct latency_bucket *latency;
2284         int index;
2285
2286         if (!td || td->limit_index != LIMIT_LOW ||
2287             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2288             !blk_queue_nonrot(td->queue))
2289                 return;
2290
2291         index = request_bucket_index(size);
2292
2293         latency = get_cpu_ptr(td->latency_buckets[op]);
2294         latency[index].total_latency += time;
2295         latency[index].samples++;
2296         put_cpu_ptr(td->latency_buckets[op]);
2297 }
2298
2299 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2300 {
2301         struct request_queue *q = rq->q;
2302         struct throtl_data *td = q->td;
2303
2304         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2305                              time_ns >> 10);
2306 }
2307
2308 void blk_throtl_bio_endio(struct bio *bio)
2309 {
2310         struct blkcg_gq *blkg;
2311         struct throtl_grp *tg;
2312         u64 finish_time_ns;
2313         unsigned long finish_time;
2314         unsigned long start_time;
2315         unsigned long lat;
2316         int rw = bio_data_dir(bio);
2317
2318         blkg = bio->bi_blkg;
2319         if (!blkg)
2320                 return;
2321         tg = blkg_to_tg(blkg);
2322
2323         finish_time_ns = ktime_get_ns();
2324         tg->last_finish_time = finish_time_ns >> 10;
2325
2326         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2327         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2328         if (!start_time || finish_time <= start_time)
2329                 return;
2330
2331         lat = finish_time - start_time;
2332         /* this is only for bio based driver */
2333         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2334                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2335                                      bio_op(bio), lat);
2336
2337         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2338                 int bucket;
2339                 unsigned int threshold;
2340
2341                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2342                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2343                         tg->latency_target;
2344                 if (lat > threshold)
2345                         tg->bad_bio_cnt++;
2346                 /*
2347                  * Not race free, could get wrong count, which means cgroups
2348                  * will be throttled
2349                  */
2350                 tg->bio_cnt++;
2351         }
2352
2353         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2354                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2355                 tg->bio_cnt /= 2;
2356                 tg->bad_bio_cnt /= 2;
2357         }
2358 }
2359 #endif
2360
2361 /*
2362  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2363  * return, @parent_sq is guaranteed to not have any active children tg's
2364  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2365  */
2366 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2367 {
2368         struct throtl_grp *tg;
2369
2370         while ((tg = throtl_rb_first(parent_sq))) {
2371                 struct throtl_service_queue *sq = &tg->service_queue;
2372                 struct bio *bio;
2373
2374                 throtl_dequeue_tg(tg);
2375
2376                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2377                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2378                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2379                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2380         }
2381 }
2382
2383 /**
2384  * blk_throtl_drain - drain throttled bios
2385  * @q: request_queue to drain throttled bios for
2386  *
2387  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2388  */
2389 void blk_throtl_drain(struct request_queue *q)
2390         __releases(&q->queue_lock) __acquires(&q->queue_lock)
2391 {
2392         struct throtl_data *td = q->td;
2393         struct blkcg_gq *blkg;
2394         struct cgroup_subsys_state *pos_css;
2395         struct bio *bio;
2396         int rw;
2397
2398         rcu_read_lock();
2399
2400         /*
2401          * Drain each tg while doing post-order walk on the blkg tree, so
2402          * that all bios are propagated to td->service_queue.  It'd be
2403          * better to walk service_queue tree directly but blkg walk is
2404          * easier.
2405          */
2406         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2407                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2408
2409         /* finally, transfer bios from top-level tg's into the td */
2410         tg_drain_bios(&td->service_queue);
2411
2412         rcu_read_unlock();
2413         spin_unlock_irq(&q->queue_lock);
2414
2415         /* all bios now should be in td->service_queue, issue them */
2416         for (rw = READ; rw <= WRITE; rw++)
2417                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2418                                                 NULL)))
2419                         generic_make_request(bio);
2420
2421         spin_lock_irq(&q->queue_lock);
2422 }
2423
2424 int blk_throtl_init(struct request_queue *q)
2425 {
2426         struct throtl_data *td;
2427         int ret;
2428
2429         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2430         if (!td)
2431                 return -ENOMEM;
2432         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2433                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2434         if (!td->latency_buckets[READ]) {
2435                 kfree(td);
2436                 return -ENOMEM;
2437         }
2438         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2439                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2440         if (!td->latency_buckets[WRITE]) {
2441                 free_percpu(td->latency_buckets[READ]);
2442                 kfree(td);
2443                 return -ENOMEM;
2444         }
2445
2446         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2447         throtl_service_queue_init(&td->service_queue);
2448
2449         q->td = td;
2450         td->queue = q;
2451
2452         td->limit_valid[LIMIT_MAX] = true;
2453         td->limit_index = LIMIT_MAX;
2454         td->low_upgrade_time = jiffies;
2455         td->low_downgrade_time = jiffies;
2456
2457         /* activate policy */
2458         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2459         if (ret) {
2460                 free_percpu(td->latency_buckets[READ]);
2461                 free_percpu(td->latency_buckets[WRITE]);
2462                 kfree(td);
2463         }
2464         return ret;
2465 }
2466
2467 void blk_throtl_exit(struct request_queue *q)
2468 {
2469         BUG_ON(!q->td);
2470         throtl_shutdown_wq(q);
2471         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2472         free_percpu(q->td->latency_buckets[READ]);
2473         free_percpu(q->td->latency_buckets[WRITE]);
2474         kfree(q->td);
2475 }
2476
2477 void blk_throtl_register_queue(struct request_queue *q)
2478 {
2479         struct throtl_data *td;
2480         int i;
2481
2482         td = q->td;
2483         BUG_ON(!td);
2484
2485         if (blk_queue_nonrot(q)) {
2486                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2487                 td->filtered_latency = LATENCY_FILTERED_SSD;
2488         } else {
2489                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2490                 td->filtered_latency = LATENCY_FILTERED_HD;
2491                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2492                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2493                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2494                 }
2495         }
2496 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2497         /* if no low limit, use previous default */
2498         td->throtl_slice = DFL_THROTL_SLICE_HD;
2499 #endif
2500
2501         td->track_bio_latency = !queue_is_mq(q);
2502         if (!td->track_bio_latency)
2503                 blk_stat_enable_accounting(q);
2504 }
2505
2506 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2507 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2508 {
2509         if (!q->td)
2510                 return -EINVAL;
2511         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2512 }
2513
2514 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2515         const char *page, size_t count)
2516 {
2517         unsigned long v;
2518         unsigned long t;
2519
2520         if (!q->td)
2521                 return -EINVAL;
2522         if (kstrtoul(page, 10, &v))
2523                 return -EINVAL;
2524         t = msecs_to_jiffies(v);
2525         if (t == 0 || t > MAX_THROTL_SLICE)
2526                 return -EINVAL;
2527         q->td->throtl_slice = t;
2528         return count;
2529 }
2530 #endif
2531
2532 static int __init throtl_init(void)
2533 {
2534         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2535         if (!kthrotld_workqueue)
2536                 panic("Failed to create kthrotld\n");
2537
2538         return blkcg_policy_register(&blkcg_policy_throtl);
2539 }
2540
2541 module_init(throtl_init);