block/blk-throttle.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Interface for controlling IO bandwidth on a request queue
   4  *
   5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blktrace_api.h>
  13 #include <linux/blk-cgroup.h>
  14 #include "blk.h"
  15
  16 /* Max dispatch from a group in 1 round */
  17 static int throtl_grp_quantum = 8;
  18
  19 /* Total max dispatch from all groups in one round */
  20 static int throtl_quantum = 32;
  21
  22 /* Throttling is performed over a slice and after that slice is renewed */
  23 #define DFL_THROTL_SLICE_HD (HZ / 10)
  24 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  25 #define MAX_THROTL_SLICE (HZ)
  26 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  27 #define MIN_THROTL_BPS (320 * 1024)
  28 #define MIN_THROTL_IOPS (10)
  29 #define DFL_LATENCY_TARGET (-1L)
  30 #define DFL_IDLE_THRESHOLD (0)
  31 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  32 #define LATENCY_FILTERED_SSD (0)
  33 /*
  34  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  35  * help determine if its IO is impacted by others, hence we ignore the IO
  36  */
  37 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  38
  39 static struct blkcg_policy blkcg_policy_throtl;
  40
  41 /* A workqueue to queue throttle related work */
  42 static struct workqueue_struct *kthrotld_workqueue;
  43
  44 /*
  45  * To implement hierarchical throttling, throtl_grps form a tree and bios
  46  * are dispatched upwards level by level until they reach the top and get
  47  * issued.  When dispatching bios from the children and local group at each
  48  * level, if the bios are dispatched into a single bio_list, there's a risk
  49  * of a local or child group which can queue many bios at once filling up
  50  * the list starving others.
  51  *
  52  * To avoid such starvation, dispatched bios are queued separately
  53  * according to where they came from.  When they are again dispatched to
  54  * the parent, they're popped in round-robin order so that no single source
  55  * hogs the dispatch window.
  56  *
  57  * throtl_qnode is used to keep the queued bios separated by their sources.
  58  * Bios are queued to throtl_qnode which in turn is queued to
  59  * throtl_service_queue and then dispatched in round-robin order.
  60  *
  61  * It's also used to track the reference counts on blkg's.  A qnode always
  62  * belongs to a throtl_grp and gets queued on itself or the parent, so
  63  * incrementing the reference of the associated throtl_grp when a qnode is
  64  * queued and decrementing when dequeued is enough to keep the whole blkg
  65  * tree pinned while bios are in flight.
  66  */
  67 struct throtl_qnode {
  68         struct list_head        node;           /* service_queue->queued[] */
  69         struct bio_list         bios;           /* queued bios */
  70         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  71 };
  72
  73 struct throtl_service_queue {
  74         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  75
  76         /*
  77          * Bios queued directly to this service_queue or dispatched from
  78          * children throtl_grp's.
  79          */
  80         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  81         unsigned int            nr_queued[2];   /* number of queued bios */
  82
  83         /*
  84          * RB tree of active children throtl_grp's, which are sorted by
  85          * their ->disptime.
  86          */
  87         struct rb_root          pending_tree;   /* RB tree of active tgs */
  88         struct rb_node          *first_pending; /* first node in the tree */
  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
 182 /* We measure latency for request size from <= 4k to >= 1M */
 183 #define LATENCY_BUCKET_SIZE 9
 184
 185 struct latency_bucket {
 186         unsigned long total_latency; /* ns / 1024 */
 187         int samples;
 188 };
 189
 190 struct avg_latency_bucket {
 191         unsigned long latency; /* ns / 1024 */
 192         bool valid;
 193 };
 194
 195 struct throtl_data
 196 {
 197         /* service tree for active throtl groups */
 198         struct throtl_service_queue service_queue;
 199
 200         struct request_queue *queue;
 201
 202         /* Total Number of queued bios on READ and WRITE lists */
 203         unsigned int nr_queued[2];
 204
 205         unsigned int throtl_slice;
 206
 207         /* Work for dispatching throttled bios */
 208         struct work_struct dispatch_work;
 209         unsigned int limit_index;
 210         bool limit_valid[LIMIT_CNT];
 211
 212         unsigned long low_upgrade_time;
 213         unsigned long low_downgrade_time;
 214
 215         unsigned int scale;
 216
 217         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
 218         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
 219         struct latency_bucket __percpu *latency_buckets[2];
 220         unsigned long last_calculate_time;
 221         unsigned long filtered_latency;
 222
 223         bool track_bio_latency;
 224 };
 225
 226 static void throtl_pending_timer_fn(struct timer_list *t);
 227
 228 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 229 {
 230         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 231 }
 232
 233 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 234 {
 235         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 236 }
 237
 238 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 239 {
 240         return pd_to_blkg(&tg->pd);
 241 }
 242
 243 /**
 244  * sq_to_tg - return the throl_grp the specified service queue belongs to
 245  * @sq: the throtl_service_queue of interest
 246  *
 247  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 248  * embedded in throtl_data, %NULL is returned.
 249  */
 250 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 251 {
 252         if (sq && sq->parent_sq)
 253                 return container_of(sq, struct throtl_grp, service_queue);
 254         else
 255                 return NULL;
 256 }
 257
 258 /**
 259  * sq_to_td - return throtl_data the specified service queue belongs to
 260  * @sq: the throtl_service_queue of interest
 261  *
 262  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 263  * Determine the associated throtl_data accordingly and return it.
 264  */
 265 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 266 {
 267         struct throtl_grp *tg = sq_to_tg(sq);
 268
 269         if (tg)
 270                 return tg->td;
 271         else
 272                 return container_of(sq, struct throtl_data, service_queue);
 273 }
 274
 275 /*
 276  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 277  * make the IO dispatch more smooth.
 278  * Scale up: linearly scale up according to lapsed time since upgrade. For
 279  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 280  *           limit hits .max limit
 281  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 282  */
 283 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 284 {
 285         /* arbitrary value to avoid too big scale */
 286         if (td->scale < 4096 && time_after_eq(jiffies,
 287             td->low_upgrade_time + td->scale * td->throtl_slice))
 288                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 289
 290         return low + (low >> 1) * td->scale;
 291 }
 292
 293 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 294 {
 295         struct blkcg_gq *blkg = tg_to_blkg(tg);
 296         struct throtl_data *td;
 297         uint64_t ret;
 298
 299         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 300                 return U64_MAX;
 301
 302         td = tg->td;
 303         ret = tg->bps[rw][td->limit_index];
 304         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 305                 /* intermediate node or iops isn't 0 */
 306                 if (!list_empty(&blkg->blkcg->css.children) ||
 307                     tg->iops[rw][td->limit_index])
 308                         return U64_MAX;
 309                 else
 310                         return MIN_THROTL_BPS;
 311         }
 312
 313         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 314             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 315                 uint64_t adjusted;
 316
 317                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 318                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 319         }
 320         return ret;
 321 }
 322
 323 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 324 {
 325         struct blkcg_gq *blkg = tg_to_blkg(tg);
 326         struct throtl_data *td;
 327         unsigned int ret;
 328
 329         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 330                 return UINT_MAX;
 331
 332         td = tg->td;
 333         ret = tg->iops[rw][td->limit_index];
 334         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 335                 /* intermediate node or bps isn't 0 */
 336                 if (!list_empty(&blkg->blkcg->css.children) ||
 337                     tg->bps[rw][td->limit_index])
 338                         return UINT_MAX;
 339                 else
 340                         return MIN_THROTL_IOPS;
 341         }
 342
 343         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 344             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 345                 uint64_t adjusted;
 346
 347                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 348                 if (adjusted > UINT_MAX)
 349                         adjusted = UINT_MAX;
 350                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 351         }
 352         return ret;
 353 }
 354
 355 #define request_bucket_index(sectors) \
 356         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 357
 358 /**
 359  * throtl_log - log debug message via blktrace
 360  * @sq: the service_queue being reported
 361  * @fmt: printf format string
 362  * @args: printf args
 363  *
 364  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 365  * throtl_grp; otherwise, just "throtl".
 366  */
 367 #define throtl_log(sq, fmt, args...)    do {                            \
 368         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 369         struct throtl_data *__td = sq_to_td((sq));                      \
 370                                                                         \
 371         (void)__td;                                                     \
 372         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 373                 break;                                                  \
 374         if ((__tg)) {                                                   \
 375                 blk_add_cgroup_trace_msg(__td->queue,                   \
 376                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 377         } else {                                                        \
 378                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 379         }                                                               \
 380 } while (0)
 381
 382 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 383 {
 384         /* assume it's one sector */
 385         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 386                 return 512;
 387         return bio->bi_iter.bi_size;
 388 }
 389
 390 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 391 {
 392         INIT_LIST_HEAD(&qn->node);
 393         bio_list_init(&qn->bios);
 394         qn->tg = tg;
 395 }
 396
 397 /**
 398  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 399  * @bio: bio being added
 400  * @qn: qnode to add bio to
 401  * @queued: the service_queue->queued[] list @qn belongs to
 402  *
 403  * Add @bio to @qn and put @qn on @queued if it's not already on.
 404  * @qn->tg's reference count is bumped when @qn is activated.  See the
 405  * comment on top of throtl_qnode definition for details.
 406  */
 407 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 408                                  struct list_head *queued)
 409 {
 410         bio_list_add(&qn->bios, bio);
 411         if (list_empty(&qn->node)) {
 412                 list_add_tail(&qn->node, queued);
 413                 blkg_get(tg_to_blkg(qn->tg));
 414         }
 415 }
 416
 417 /**
 418  * throtl_peek_queued - peek the first bio on a qnode list
 419  * @queued: the qnode list to peek
 420  */
 421 static struct bio *throtl_peek_queued(struct list_head *queued)
 422 {
 423         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 424         struct bio *bio;
 425
 426         if (list_empty(queued))
 427                 return NULL;
 428
 429         bio = bio_list_peek(&qn->bios);
 430         WARN_ON_ONCE(!bio);
 431         return bio;
 432 }
 433
 434 /**
 435  * throtl_pop_queued - pop the first bio form a qnode list
 436  * @queued: the qnode list to pop a bio from
 437  * @tg_to_put: optional out argument for throtl_grp to put
 438  *
 439  * Pop the first bio from the qnode list @queued.  After popping, the first
 440  * qnode is removed from @queued if empty or moved to the end of @queued so
 441  * that the popping order is round-robin.
 442  *
 443  * When the first qnode is removed, its associated throtl_grp should be put
 444  * too.  If @tg_to_put is NULL, this function automatically puts it;
 445  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 446  * responsible for putting it.
 447  */
 448 static struct bio *throtl_pop_queued(struct list_head *queued,
 449                                      struct throtl_grp **tg_to_put)
 450 {
 451         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 452         struct bio *bio;
 453
 454         if (list_empty(queued))
 455                 return NULL;
 456
 457         bio = bio_list_pop(&qn->bios);
 458         WARN_ON_ONCE(!bio);
 459
 460         if (bio_list_empty(&qn->bios)) {
 461                 list_del_init(&qn->node);
 462                 if (tg_to_put)
 463                         *tg_to_put = qn->tg;
 464                 else
 465                         blkg_put(tg_to_blkg(qn->tg));
 466         } else {
 467                 list_move_tail(&qn->node, queued);
 468         }
 469
 470         return bio;
 471 }
 472
 473 /* init a service_queue, assumes the caller zeroed it */
 474 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 475 {
 476         INIT_LIST_HEAD(&sq->queued[0]);
 477         INIT_LIST_HEAD(&sq->queued[1]);
 478         sq->pending_tree = RB_ROOT;
 479         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 480 }
 481
 482 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp, int node)
 483 {
 484         struct throtl_grp *tg;
 485         int rw;
 486
 487         tg = kzalloc_node(sizeof(*tg), gfp, node);
 488         if (!tg)
 489                 return NULL;
 490
 491         throtl_service_queue_init(&tg->service_queue);
 492
 493         for (rw = READ; rw <= WRITE; rw++) {
 494                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 495                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 496         }
 497
 498         RB_CLEAR_NODE(&tg->rb_node);
 499         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 500         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 501         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 502         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 503         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 504         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 505         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 506         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 507         /* LIMIT_LOW will have default value 0 */
 508
 509         tg->latency_target = DFL_LATENCY_TARGET;
 510         tg->latency_target_conf = DFL_LATENCY_TARGET;
 511         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 512         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 513
 514         return &tg->pd;
 515 }
 516
 517 static void throtl_pd_init(struct blkg_policy_data *pd)
 518 {
 519         struct throtl_grp *tg = pd_to_tg(pd);
 520         struct blkcg_gq *blkg = tg_to_blkg(tg);
 521         struct throtl_data *td = blkg->q->td;
 522         struct throtl_service_queue *sq = &tg->service_queue;
 523
 524         /*
 525          * If on the default hierarchy, we switch to properly hierarchical
 526          * behavior where limits on a given throtl_grp are applied to the
 527          * whole subtree rather than just the group itself.  e.g. If 16M
 528          * read_bps limit is set on the root group, the whole system can't
 529          * exceed 16M for the device.
 530          *
 531          * If not on the default hierarchy, the broken flat hierarchy
 532          * behavior is retained where all throtl_grps are treated as if
 533          * they're all separate root groups right below throtl_data.
 534          * Limits of a group don't interact with limits of other groups
 535          * regardless of the position of the group in the hierarchy.
 536          */
 537         sq->parent_sq = &td->service_queue;
 538         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 539                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 540         tg->td = td;
 541 }
 542
 543 /*
 544  * Set has_rules[] if @tg or any of its parents have limits configured.
 545  * This doesn't require walking up to the top of the hierarchy as the
 546  * parent's has_rules[] is guaranteed to be correct.
 547  */
 548 static void tg_update_has_rules(struct throtl_grp *tg)
 549 {
 550         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 551         struct throtl_data *td = tg->td;
 552         int rw;
 553
 554         for (rw = READ; rw <= WRITE; rw++)
 555                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 556                         (td->limit_valid[td->limit_index] &&
 557                          (tg_bps_limit(tg, rw) != U64_MAX ||
 558                           tg_iops_limit(tg, rw) != UINT_MAX));
 559 }
 560
 561 static void throtl_pd_online(struct blkg_policy_data *pd)
 562 {
 563         struct throtl_grp *tg = pd_to_tg(pd);
 564         /*
 565          * We don't want new groups to escape the limits of its ancestors.
 566          * Update has_rules[] after a new group is brought online.
 567          */
 568         tg_update_has_rules(tg);
 569 }
 570
 571 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 572 {
 573         struct cgroup_subsys_state *pos_css;
 574         struct blkcg_gq *blkg;
 575         bool low_valid = false;
 576
 577         rcu_read_lock();
 578         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 579                 struct throtl_grp *tg = blkg_to_tg(blkg);
 580
 581                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 582                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 583                         low_valid = true;
 584                         break;
 585                 }
 586         }
 587         rcu_read_unlock();
 588
 589         td->limit_valid[LIMIT_LOW] = low_valid;
 590 }
 591
 592 static void throtl_upgrade_state(struct throtl_data *td);
 593 static void throtl_pd_offline(struct blkg_policy_data *pd)
 594 {
 595         struct throtl_grp *tg = pd_to_tg(pd);
 596
 597         tg->bps[READ][LIMIT_LOW] = 0;
 598         tg->bps[WRITE][LIMIT_LOW] = 0;
 599         tg->iops[READ][LIMIT_LOW] = 0;
 600         tg->iops[WRITE][LIMIT_LOW] = 0;
 601
 602         blk_throtl_update_limit_valid(tg->td);
 603
 604         if (!tg->td->limit_valid[tg->td->limit_index])
 605                 throtl_upgrade_state(tg->td);
 606 }
 607
 608 static void throtl_pd_free(struct blkg_policy_data *pd)
 609 {
 610         struct throtl_grp *tg = pd_to_tg(pd);
 611
 612         del_timer_sync(&tg->service_queue.pending_timer);
 613         kfree(tg);
 614 }
 615
 616 static struct throtl_grp *
 617 throtl_rb_first(struct throtl_service_queue *parent_sq)
 618 {
 619         /* Service tree is empty */
 620         if (!parent_sq->nr_pending)
 621                 return NULL;
 622
 623         if (!parent_sq->first_pending)
 624                 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
 625
 626         if (parent_sq->first_pending)
 627                 return rb_entry_tg(parent_sq->first_pending);
 628
 629         return NULL;
 630 }
 631
 632 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 633 {
 634         rb_erase(n, root);
 635         RB_CLEAR_NODE(n);
 636 }
 637
 638 static void throtl_rb_erase(struct rb_node *n,
 639                             struct throtl_service_queue *parent_sq)
 640 {
 641         if (parent_sq->first_pending == n)
 642                 parent_sq->first_pending = NULL;
 643         rb_erase_init(n, &parent_sq->pending_tree);
 644         --parent_sq->nr_pending;
 645 }
 646
 647 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 648 {
 649         struct throtl_grp *tg;
 650
 651         tg = throtl_rb_first(parent_sq);
 652         if (!tg)
 653                 return;
 654
 655         parent_sq->first_pending_disptime = tg->disptime;
 656 }
 657
 658 static void tg_service_queue_add(struct throtl_grp *tg)
 659 {
 660         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 661         struct rb_node **node = &parent_sq->pending_tree.rb_node;
 662         struct rb_node *parent = NULL;
 663         struct throtl_grp *__tg;
 664         unsigned long key = tg->disptime;
 665         int left = 1;
 666
 667         while (*node != NULL) {
 668                 parent = *node;
 669                 __tg = rb_entry_tg(parent);
 670
 671                 if (time_before(key, __tg->disptime))
 672                         node = &parent->rb_left;
 673                 else {
 674                         node = &parent->rb_right;
 675                         left = 0;
 676                 }
 677         }
 678
 679         if (left)
 680                 parent_sq->first_pending = &tg->rb_node;
 681
 682         rb_link_node(&tg->rb_node, parent, node);
 683         rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
 684 }
 685
 686 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 687 {
 688         tg_service_queue_add(tg);
 689         tg->flags |= THROTL_TG_PENDING;
 690         tg->service_queue.parent_sq->nr_pending++;
 691 }
 692
 693 static void throtl_enqueue_tg(struct throtl_grp *tg)
 694 {
 695         if (!(tg->flags & THROTL_TG_PENDING))
 696                 __throtl_enqueue_tg(tg);
 697 }
 698
 699 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 700 {
 701         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 702         tg->flags &= ~THROTL_TG_PENDING;
 703 }
 704
 705 static void throtl_dequeue_tg(struct throtl_grp *tg)
 706 {
 707         if (tg->flags & THROTL_TG_PENDING)
 708                 __throtl_dequeue_tg(tg);
 709 }
 710
 711 /* Call with queue lock held */
 712 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 713                                           unsigned long expires)
 714 {
 715         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 716
 717         /*
 718          * Since we are adjusting the throttle limit dynamically, the sleep
 719          * time calculated according to previous limit might be invalid. It's
 720          * possible the cgroup sleep time is very long and no other cgroups
 721          * have IO running so notify the limit changes. Make sure the cgroup
 722          * doesn't sleep too long to avoid the missed notification.
 723          */
 724         if (time_after(expires, max_expire))
 725                 expires = max_expire;
 726         mod_timer(&sq->pending_timer, expires);
 727         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 728                    expires - jiffies, jiffies);
 729 }
 730
 731 /**
 732  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 733  * @sq: the service_queue to schedule dispatch for
 734  * @force: force scheduling
 735  *
 736  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 737  * dispatch time of the first pending child.  Returns %true if either timer
 738  * is armed or there's no pending child left.  %false if the current
 739  * dispatch window is still open and the caller should continue
 740  * dispatching.
 741  *
 742  * If @force is %true, the dispatch timer is always scheduled and this
 743  * function is guaranteed to return %true.  This is to be used when the
 744  * caller can't dispatch itself and needs to invoke pending_timer
 745  * unconditionally.  Note that forced scheduling is likely to induce short
 746  * delay before dispatch starts even if @sq->first_pending_disptime is not
 747  * in the future and thus shouldn't be used in hot paths.
 748  */
 749 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 750                                           bool force)
 751 {
 752         /* any pending children left? */
 753         if (!sq->nr_pending)
 754                 return true;
 755
 756         update_min_dispatch_time(sq);
 757
 758         /* is the next dispatch time in the future? */
 759         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 760                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 761                 return true;
 762         }
 763
 764         /* tell the caller to continue dispatching */
 765         return false;
 766 }
 767
 768 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 769                 bool rw, unsigned long start)
 770 {
 771         tg->bytes_disp[rw] = 0;
 772         tg->io_disp[rw] = 0;
 773
 774         /*
 775          * Previous slice has expired. We must have trimmed it after last
 776          * bio dispatch. That means since start of last slice, we never used
 777          * that bandwidth. Do try to make use of that bandwidth while giving
 778          * credit.
 779          */
 780         if (time_after_eq(start, tg->slice_start[rw]))
 781                 tg->slice_start[rw] = start;
 782
 783         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 784         throtl_log(&tg->service_queue,
 785                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 786                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 787                    tg->slice_end[rw], jiffies);
 788 }
 789
 790 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 791 {
 792         tg->bytes_disp[rw] = 0;
 793         tg->io_disp[rw] = 0;
 794         tg->slice_start[rw] = jiffies;
 795         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 796         throtl_log(&tg->service_queue,
 797                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 798                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 799                    tg->slice_end[rw], jiffies);
 800 }
 801
 802 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 803                                         unsigned long jiffy_end)
 804 {
 805         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 806 }
 807
 808 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 809                                        unsigned long jiffy_end)
 810 {
 811         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 812         throtl_log(&tg->service_queue,
 813                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 814                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 815                    tg->slice_end[rw], jiffies);
 816 }
 817
 818 /* Determine if previously allocated or extended slice is complete or not */
 819 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 820 {
 821         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 822                 return false;
 823
 824         return true;
 825 }
 826
 827 /* Trim the used slices and adjust slice start accordingly */
 828 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 829 {
 830         unsigned long nr_slices, time_elapsed, io_trim;
 831         u64 bytes_trim, tmp;
 832
 833         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 834
 835         /*
 836          * If bps are unlimited (-1), then time slice don't get
 837          * renewed. Don't try to trim the slice if slice is used. A new
 838          * slice will start when appropriate.
 839          */
 840         if (throtl_slice_used(tg, rw))
 841                 return;
 842
 843         /*
 844          * A bio has been dispatched. Also adjust slice_end. It might happen
 845          * that initially cgroup limit was very low resulting in high
 846          * slice_end, but later limit was bumped up and bio was dispached
 847          * sooner, then we need to reduce slice_end. A high bogus slice_end
 848          * is bad because it does not allow new slice to start.
 849          */
 850
 851         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 852
 853         time_elapsed = jiffies - tg->slice_start[rw];
 854
 855         nr_slices = time_elapsed / tg->td->throtl_slice;
 856
 857         if (!nr_slices)
 858                 return;
 859         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 860         do_div(tmp, HZ);
 861         bytes_trim = tmp;
 862
 863         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 864                 HZ;
 865
 866         if (!bytes_trim && !io_trim)
 867                 return;
 868
 869         if (tg->bytes_disp[rw] >= bytes_trim)
 870                 tg->bytes_disp[rw] -= bytes_trim;
 871         else
 872                 tg->bytes_disp[rw] = 0;
 873
 874         if (tg->io_disp[rw] >= io_trim)
 875                 tg->io_disp[rw] -= io_trim;
 876         else
 877                 tg->io_disp[rw] = 0;
 878
 879         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 880
 881         throtl_log(&tg->service_queue,
 882                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 883                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 884                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 885 }
 886
 887 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 888                                   unsigned long *wait)
 889 {
 890         bool rw = bio_data_dir(bio);
 891         unsigned int io_allowed;
 892         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 893         u64 tmp;
 894
 895         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 896
 897         /* Slice has just started. Consider one slice interval */
 898         if (!jiffy_elapsed)
 899                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 900
 901         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 902
 903         /*
 904          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 905          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 906          * will allow dispatch after 1 second and after that slice should
 907          * have been trimmed.
 908          */
 909
 910         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 911         do_div(tmp, HZ);
 912
 913         if (tmp > UINT_MAX)
 914                 io_allowed = UINT_MAX;
 915         else
 916                 io_allowed = tmp;
 917
 918         if (tg->io_disp[rw] + 1 <= io_allowed) {
 919                 if (wait)
 920                         *wait = 0;
 921                 return true;
 922         }
 923
 924         /* Calc approx time to dispatch */
 925         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 926
 927         if (wait)
 928                 *wait = jiffy_wait;
 929         return false;
 930 }
 931
 932 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 933                                  unsigned long *wait)
 934 {
 935         bool rw = bio_data_dir(bio);
 936         u64 bytes_allowed, extra_bytes, tmp;
 937         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 938         unsigned int bio_size = throtl_bio_data_size(bio);
 939
 940         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 941
 942         /* Slice has just started. Consider one slice interval */
 943         if (!jiffy_elapsed)
 944                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 945
 946         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 947
 948         tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
 949         do_div(tmp, HZ);
 950         bytes_allowed = tmp;
 951
 952         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 953                 if (wait)
 954                         *wait = 0;
 955                 return true;
 956         }
 957
 958         /* Calc approx time to dispatch */
 959         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 960         jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
 961
 962         if (!jiffy_wait)
 963                 jiffy_wait = 1;
 964
 965         /*
 966          * This wait time is without taking into consideration the rounding
 967          * up we did. Add that time also.
 968          */
 969         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 970         if (wait)
 971                 *wait = jiffy_wait;
 972         return false;
 973 }
 974
 975 /*
 976  * Returns whether one can dispatch a bio or not. Also returns approx number
 977  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 978  */
 979 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 980                             unsigned long *wait)
 981 {
 982         bool rw = bio_data_dir(bio);
 983         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 984
 985         /*
 986          * Currently whole state machine of group depends on first bio
 987          * queued in the group bio list. So one should not be calling
 988          * this function with a different bio if there are other bios
 989          * queued.
 990          */
 991         BUG_ON(tg->service_queue.nr_queued[rw] &&
 992                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 993
 994         /* If tg->bps = -1, then BW is unlimited */
 995         if (tg_bps_limit(tg, rw) == U64_MAX &&
 996             tg_iops_limit(tg, rw) == UINT_MAX) {
 997                 if (wait)
 998                         *wait = 0;
 999                 return true;
1000         }
1001
1002         /*
1003          * If previous slice expired, start a new one otherwise renew/extend
1004          * existing slice to make sure it is at least throtl_slice interval
1005          * long since now. New slice is started only for empty throttle group.
1006          * If there is queued bio, that means there should be an active
1007          * slice and it should be extended instead.
1008          */
1009         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1010                 throtl_start_new_slice(tg, rw);
1011         else {
1012                 if (time_before(tg->slice_end[rw],
1013                     jiffies + tg->td->throtl_slice))
1014                         throtl_extend_slice(tg, rw,
1015                                 jiffies + tg->td->throtl_slice);
1016         }
1017
1018         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1019             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1020                 if (wait)
1021                         *wait = 0;
1022                 return true;
1023         }
1024
1025         max_wait = max(bps_wait, iops_wait);
1026
1027         if (wait)
1028                 *wait = max_wait;
1029
1030         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1031                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1032
1033         return false;
1034 }
1035
1036 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1037 {
1038         bool rw = bio_data_dir(bio);
1039         unsigned int bio_size = throtl_bio_data_size(bio);
1040
1041         /* Charge the bio to the group */
1042         tg->bytes_disp[rw] += bio_size;
1043         tg->io_disp[rw]++;
1044         tg->last_bytes_disp[rw] += bio_size;
1045         tg->last_io_disp[rw]++;
1046
1047         /*
1048          * BIO_THROTTLED is used to prevent the same bio to be throttled
1049          * more than once as a throttled bio will go through blk-throtl the
1050          * second time when it eventually gets issued.  Set it when a bio
1051          * is being charged to a tg.
1052          */
1053         if (!bio_flagged(bio, BIO_THROTTLED))
1054                 bio_set_flag(bio, BIO_THROTTLED);
1055 }
1056
1057 /**
1058  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1059  * @bio: bio to add
1060  * @qn: qnode to use
1061  * @tg: the target throtl_grp
1062  *
1063  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1064  * tg->qnode_on_self[] is used.
1065  */
1066 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1067                               struct throtl_grp *tg)
1068 {
1069         struct throtl_service_queue *sq = &tg->service_queue;
1070         bool rw = bio_data_dir(bio);
1071
1072         if (!qn)
1073                 qn = &tg->qnode_on_self[rw];
1074
1075         /*
1076          * If @tg doesn't currently have any bios queued in the same
1077          * direction, queueing @bio can change when @tg should be
1078          * dispatched.  Mark that @tg was empty.  This is automatically
1079          * cleaered on the next tg_update_disptime().
1080          */
1081         if (!sq->nr_queued[rw])
1082                 tg->flags |= THROTL_TG_WAS_EMPTY;
1083
1084         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1085
1086         sq->nr_queued[rw]++;
1087         throtl_enqueue_tg(tg);
1088 }
1089
1090 static void tg_update_disptime(struct throtl_grp *tg)
1091 {
1092         struct throtl_service_queue *sq = &tg->service_queue;
1093         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1094         struct bio *bio;
1095
1096         bio = throtl_peek_queued(&sq->queued[READ]);
1097         if (bio)
1098                 tg_may_dispatch(tg, bio, &read_wait);
1099
1100         bio = throtl_peek_queued(&sq->queued[WRITE]);
1101         if (bio)
1102                 tg_may_dispatch(tg, bio, &write_wait);
1103
1104         min_wait = min(read_wait, write_wait);
1105         disptime = jiffies + min_wait;
1106
1107         /* Update dispatch time */
1108         throtl_dequeue_tg(tg);
1109         tg->disptime = disptime;
1110         throtl_enqueue_tg(tg);
1111
1112         /* see throtl_add_bio_tg() */
1113         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1114 }
1115
1116 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1117                                         struct throtl_grp *parent_tg, bool rw)
1118 {
1119         if (throtl_slice_used(parent_tg, rw)) {
1120                 throtl_start_new_slice_with_credit(parent_tg, rw,
1121                                 child_tg->slice_start[rw]);
1122         }
1123
1124 }
1125
1126 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1127 {
1128         struct throtl_service_queue *sq = &tg->service_queue;
1129         struct throtl_service_queue *parent_sq = sq->parent_sq;
1130         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1131         struct throtl_grp *tg_to_put = NULL;
1132         struct bio *bio;
1133
1134         /*
1135          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1136          * from @tg may put its reference and @parent_sq might end up
1137          * getting released prematurely.  Remember the tg to put and put it
1138          * after @bio is transferred to @parent_sq.
1139          */
1140         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1141         sq->nr_queued[rw]--;
1142
1143         throtl_charge_bio(tg, bio);
1144
1145         /*
1146          * If our parent is another tg, we just need to transfer @bio to
1147          * the parent using throtl_add_bio_tg().  If our parent is
1148          * @td->service_queue, @bio is ready to be issued.  Put it on its
1149          * bio_lists[] and decrease total number queued.  The caller is
1150          * responsible for issuing these bios.
1151          */
1152         if (parent_tg) {
1153                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1154                 start_parent_slice_with_credit(tg, parent_tg, rw);
1155         } else {
1156                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1157                                      &parent_sq->queued[rw]);
1158                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1159                 tg->td->nr_queued[rw]--;
1160         }
1161
1162         throtl_trim_slice(tg, rw);
1163
1164         if (tg_to_put)
1165                 blkg_put(tg_to_blkg(tg_to_put));
1166 }
1167
1168 static int throtl_dispatch_tg(struct throtl_grp *tg)
1169 {
1170         struct throtl_service_queue *sq = &tg->service_queue;
1171         unsigned int nr_reads = 0, nr_writes = 0;
1172         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1173         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1174         struct bio *bio;
1175
1176         /* Try to dispatch 75% READS and 25% WRITES */
1177
1178         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1179                tg_may_dispatch(tg, bio, NULL)) {
1180
1181                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1182                 nr_reads++;
1183
1184                 if (nr_reads >= max_nr_reads)
1185                         break;
1186         }
1187
1188         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1189                tg_may_dispatch(tg, bio, NULL)) {
1190
1191                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1192                 nr_writes++;
1193
1194                 if (nr_writes >= max_nr_writes)
1195                         break;
1196         }
1197
1198         return nr_reads + nr_writes;
1199 }
1200
1201 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1202 {
1203         unsigned int nr_disp = 0;
1204
1205         while (1) {
1206                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1207                 struct throtl_service_queue *sq;
1208
1209                 if (!tg)
1210                         break;
1211
1212                 if (time_before(jiffies, tg->disptime))
1213                         break;
1214
1215                 throtl_dequeue_tg(tg);
1216
1217                 nr_disp += throtl_dispatch_tg(tg);
1218
1219                 sq = &tg->service_queue;
1220                 if (sq->nr_queued[0] || sq->nr_queued[1])
1221                         tg_update_disptime(tg);
1222
1223                 if (nr_disp >= throtl_quantum)
1224                         break;
1225         }
1226
1227         return nr_disp;
1228 }
1229
1230 static bool throtl_can_upgrade(struct throtl_data *td,
1231         struct throtl_grp *this_tg);
1232 /**
1233  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1234  * @arg: the throtl_service_queue being serviced
1235  *
1236  * This timer is armed when a child throtl_grp with active bio's become
1237  * pending and queued on the service_queue's pending_tree and expires when
1238  * the first child throtl_grp should be dispatched.  This function
1239  * dispatches bio's from the children throtl_grps to the parent
1240  * service_queue.
1241  *
1242  * If the parent's parent is another throtl_grp, dispatching is propagated
1243  * by either arming its pending_timer or repeating dispatch directly.  If
1244  * the top-level service_tree is reached, throtl_data->dispatch_work is
1245  * kicked so that the ready bio's are issued.
1246  */
1247 static void throtl_pending_timer_fn(struct timer_list *t)
1248 {
1249         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1250         struct throtl_grp *tg = sq_to_tg(sq);
1251         struct throtl_data *td = sq_to_td(sq);
1252         struct request_queue *q = td->queue;
1253         struct throtl_service_queue *parent_sq;
1254         bool dispatched;
1255         int ret;
1256
1257         spin_lock_irq(q->queue_lock);
1258         if (throtl_can_upgrade(td, NULL))
1259                 throtl_upgrade_state(td);
1260
1261 again:
1262         parent_sq = sq->parent_sq;
1263         dispatched = false;
1264
1265         while (true) {
1266                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1267                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1268                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1269
1270                 ret = throtl_select_dispatch(sq);
1271                 if (ret) {
1272                         throtl_log(sq, "bios disp=%u", ret);
1273                         dispatched = true;
1274                 }
1275
1276                 if (throtl_schedule_next_dispatch(sq, false))
1277                         break;
1278
1279                 /* this dispatch windows is still open, relax and repeat */
1280                 spin_unlock_irq(q->queue_lock);
1281                 cpu_relax();
1282                 spin_lock_irq(q->queue_lock);
1283         }
1284
1285         if (!dispatched)
1286                 goto out_unlock;
1287
1288         if (parent_sq) {
1289                 /* @parent_sq is another throl_grp, propagate dispatch */
1290                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1291                         tg_update_disptime(tg);
1292                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1293                                 /* window is already open, repeat dispatching */
1294                                 sq = parent_sq;
1295                                 tg = sq_to_tg(sq);
1296                                 goto again;
1297                         }
1298                 }
1299         } else {
1300                 /* reached the top-level, queue issueing */
1301                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1302         }
1303 out_unlock:
1304         spin_unlock_irq(q->queue_lock);
1305 }
1306
1307 /**
1308  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1309  * @work: work item being executed
1310  *
1311  * This function is queued for execution when bio's reach the bio_lists[]
1312  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1313  * function.
1314  */
1315 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1316 {
1317         struct throtl_data *td = container_of(work, struct throtl_data,
1318                                               dispatch_work);
1319         struct throtl_service_queue *td_sq = &td->service_queue;
1320         struct request_queue *q = td->queue;
1321         struct bio_list bio_list_on_stack;
1322         struct bio *bio;
1323         struct blk_plug plug;
1324         int rw;
1325
1326         bio_list_init(&bio_list_on_stack);
1327
1328         spin_lock_irq(q->queue_lock);
1329         for (rw = READ; rw <= WRITE; rw++)
1330                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1331                         bio_list_add(&bio_list_on_stack, bio);
1332         spin_unlock_irq(q->queue_lock);
1333
1334         if (!bio_list_empty(&bio_list_on_stack)) {
1335                 blk_start_plug(&plug);
1336                 while((bio = bio_list_pop(&bio_list_on_stack)))
1337                         generic_make_request(bio);
1338                 blk_finish_plug(&plug);
1339         }
1340 }
1341
1342 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1343                               int off)
1344 {
1345         struct throtl_grp *tg = pd_to_tg(pd);
1346         u64 v = *(u64 *)((void *)tg + off);
1347
1348         if (v == U64_MAX)
1349                 return 0;
1350         return __blkg_prfill_u64(sf, pd, v);
1351 }
1352
1353 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1354                                int off)
1355 {
1356         struct throtl_grp *tg = pd_to_tg(pd);
1357         unsigned int v = *(unsigned int *)((void *)tg + off);
1358
1359         if (v == UINT_MAX)
1360                 return 0;
1361         return __blkg_prfill_u64(sf, pd, v);
1362 }
1363
1364 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1365 {
1366         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1367                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1368         return 0;
1369 }
1370
1371 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1372 {
1373         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1374                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1375         return 0;
1376 }
1377
1378 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1379 {
1380         struct throtl_service_queue *sq = &tg->service_queue;
1381         struct cgroup_subsys_state *pos_css;
1382         struct blkcg_gq *blkg;
1383
1384         throtl_log(&tg->service_queue,
1385                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1386                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1387                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1388
1389         /*
1390          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1391          * considered to have rules if either the tg itself or any of its
1392          * ancestors has rules.  This identifies groups without any
1393          * restrictions in the whole hierarchy and allows them to bypass
1394          * blk-throttle.
1395          */
1396         blkg_for_each_descendant_pre(blkg, pos_css,
1397                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1398                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1399                 struct throtl_grp *parent_tg;
1400
1401                 tg_update_has_rules(this_tg);
1402                 /* ignore root/second level */
1403                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1404                     !blkg->parent->parent)
1405                         continue;
1406                 parent_tg = blkg_to_tg(blkg->parent);
1407                 /*
1408                  * make sure all children has lower idle time threshold and
1409                  * higher latency target
1410                  */
1411                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1412                                 parent_tg->idletime_threshold);
1413                 this_tg->latency_target = max(this_tg->latency_target,
1414                                 parent_tg->latency_target);
1415         }
1416
1417         /*
1418          * We're already holding queue_lock and know @tg is valid.  Let's
1419          * apply the new config directly.
1420          *
1421          * Restart the slices for both READ and WRITES. It might happen
1422          * that a group's limit are dropped suddenly and we don't want to
1423          * account recently dispatched IO with new low rate.
1424          */
1425         throtl_start_new_slice(tg, 0);
1426         throtl_start_new_slice(tg, 1);
1427
1428         if (tg->flags & THROTL_TG_PENDING) {
1429                 tg_update_disptime(tg);
1430                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1431         }
1432 }
1433
1434 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1435                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1436 {
1437         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1438         struct blkg_conf_ctx ctx;
1439         struct throtl_grp *tg;
1440         int ret;
1441         u64 v;
1442
1443         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1444         if (ret)
1445                 return ret;
1446
1447         ret = -EINVAL;
1448         if (sscanf(ctx.body, "%llu", &v) != 1)
1449                 goto out_finish;
1450         if (!v)
1451                 v = U64_MAX;
1452
1453         tg = blkg_to_tg(ctx.blkg);
1454
1455         if (is_u64)
1456                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1457         else
1458                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1459
1460         tg_conf_updated(tg, false);
1461         ret = 0;
1462 out_finish:
1463         blkg_conf_finish(&ctx);
1464         return ret ?: nbytes;
1465 }
1466
1467 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1468                                char *buf, size_t nbytes, loff_t off)
1469 {
1470         return tg_set_conf(of, buf, nbytes, off, true);
1471 }
1472
1473 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1474                                 char *buf, size_t nbytes, loff_t off)
1475 {
1476         return tg_set_conf(of, buf, nbytes, off, false);
1477 }
1478
1479 static struct cftype throtl_legacy_files[] = {
1480         {
1481                 .name = "throttle.read_bps_device",
1482                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1483                 .seq_show = tg_print_conf_u64,
1484                 .write = tg_set_conf_u64,
1485         },
1486         {
1487                 .name = "throttle.write_bps_device",
1488                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1489                 .seq_show = tg_print_conf_u64,
1490                 .write = tg_set_conf_u64,
1491         },
1492         {
1493                 .name = "throttle.read_iops_device",
1494                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1495                 .seq_show = tg_print_conf_uint,
1496                 .write = tg_set_conf_uint,
1497         },
1498         {
1499                 .name = "throttle.write_iops_device",
1500                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1501                 .seq_show = tg_print_conf_uint,
1502                 .write = tg_set_conf_uint,
1503         },
1504         {
1505                 .name = "throttle.io_service_bytes",
1506                 .private = (unsigned long)&blkcg_policy_throtl,
1507                 .seq_show = blkg_print_stat_bytes,
1508         },
1509         {
1510                 .name = "throttle.io_service_bytes_recursive",
1511                 .private = (unsigned long)&blkcg_policy_throtl,
1512                 .seq_show = blkg_print_stat_bytes_recursive,
1513         },
1514         {
1515                 .name = "throttle.io_serviced",
1516                 .private = (unsigned long)&blkcg_policy_throtl,
1517                 .seq_show = blkg_print_stat_ios,
1518         },
1519         {
1520                 .name = "throttle.io_serviced_recursive",
1521                 .private = (unsigned long)&blkcg_policy_throtl,
1522                 .seq_show = blkg_print_stat_ios_recursive,
1523         },
1524         { }     /* terminate */
1525 };
1526
1527 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1528                          int off)
1529 {
1530         struct throtl_grp *tg = pd_to_tg(pd);
1531         const char *dname = blkg_dev_name(pd->blkg);
1532         char bufs[4][21] = { "max", "max", "max", "max" };
1533         u64 bps_dft;
1534         unsigned int iops_dft;
1535         char idle_time[26] = "";
1536         char latency_time[26] = "";
1537
1538         if (!dname)
1539                 return 0;
1540
1541         if (off == LIMIT_LOW) {
1542                 bps_dft = 0;
1543                 iops_dft = 0;
1544         } else {
1545                 bps_dft = U64_MAX;
1546                 iops_dft = UINT_MAX;
1547         }
1548
1549         if (tg->bps_conf[READ][off] == bps_dft &&
1550             tg->bps_conf[WRITE][off] == bps_dft &&
1551             tg->iops_conf[READ][off] == iops_dft &&
1552             tg->iops_conf[WRITE][off] == iops_dft &&
1553             (off != LIMIT_LOW ||
1554              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1555               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1556                 return 0;
1557
1558         if (tg->bps_conf[READ][off] != U64_MAX)
1559                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1560                         tg->bps_conf[READ][off]);
1561         if (tg->bps_conf[WRITE][off] != U64_MAX)
1562                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1563                         tg->bps_conf[WRITE][off]);
1564         if (tg->iops_conf[READ][off] != UINT_MAX)
1565                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1566                         tg->iops_conf[READ][off]);
1567         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1568                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1569                         tg->iops_conf[WRITE][off]);
1570         if (off == LIMIT_LOW) {
1571                 if (tg->idletime_threshold_conf == ULONG_MAX)
1572                         strcpy(idle_time, " idle=max");
1573                 else
1574                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1575                                 tg->idletime_threshold_conf);
1576
1577                 if (tg->latency_target_conf == ULONG_MAX)
1578                         strcpy(latency_time, " latency=max");
1579                 else
1580                         snprintf(latency_time, sizeof(latency_time),
1581                                 " latency=%lu", tg->latency_target_conf);
1582         }
1583
1584         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1585                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1586                    latency_time);
1587         return 0;
1588 }
1589
1590 static int tg_print_limit(struct seq_file *sf, void *v)
1591 {
1592         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1593                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1594         return 0;
1595 }
1596
1597 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1598                           char *buf, size_t nbytes, loff_t off)
1599 {
1600         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1601         struct blkg_conf_ctx ctx;
1602         struct throtl_grp *tg;
1603         u64 v[4];
1604         unsigned long idle_time;
1605         unsigned long latency_time;
1606         int ret;
1607         int index = of_cft(of)->private;
1608
1609         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1610         if (ret)
1611                 return ret;
1612
1613         tg = blkg_to_tg(ctx.blkg);
1614
1615         v[0] = tg->bps_conf[READ][index];
1616         v[1] = tg->bps_conf[WRITE][index];
1617         v[2] = tg->iops_conf[READ][index];
1618         v[3] = tg->iops_conf[WRITE][index];
1619
1620         idle_time = tg->idletime_threshold_conf;
1621         latency_time = tg->latency_target_conf;
1622         while (true) {
1623                 char tok[27];   /* wiops=18446744073709551616 */
1624                 char *p;
1625                 u64 val = U64_MAX;
1626                 int len;
1627
1628                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1629                         break;
1630                 if (tok[0] == '\0')
1631                         break;
1632                 ctx.body += len;
1633
1634                 ret = -EINVAL;
1635                 p = tok;
1636                 strsep(&p, "=");
1637                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1638                         goto out_finish;
1639
1640                 ret = -ERANGE;
1641                 if (!val)
1642                         goto out_finish;
1643
1644                 ret = -EINVAL;
1645                 if (!strcmp(tok, "rbps"))
1646                         v[0] = val;
1647                 else if (!strcmp(tok, "wbps"))
1648                         v[1] = val;
1649                 else if (!strcmp(tok, "riops"))
1650                         v[2] = min_t(u64, val, UINT_MAX);
1651                 else if (!strcmp(tok, "wiops"))
1652                         v[3] = min_t(u64, val, UINT_MAX);
1653                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1654                         idle_time = val;
1655                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1656                         latency_time = val;
1657                 else
1658                         goto out_finish;
1659         }
1660
1661         tg->bps_conf[READ][index] = v[0];
1662         tg->bps_conf[WRITE][index] = v[1];
1663         tg->iops_conf[READ][index] = v[2];
1664         tg->iops_conf[WRITE][index] = v[3];
1665
1666         if (index == LIMIT_MAX) {
1667                 tg->bps[READ][index] = v[0];
1668                 tg->bps[WRITE][index] = v[1];
1669                 tg->iops[READ][index] = v[2];
1670                 tg->iops[WRITE][index] = v[3];
1671         }
1672         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1673                 tg->bps_conf[READ][LIMIT_MAX]);
1674         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1675                 tg->bps_conf[WRITE][LIMIT_MAX]);
1676         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1677                 tg->iops_conf[READ][LIMIT_MAX]);
1678         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1679                 tg->iops_conf[WRITE][LIMIT_MAX]);
1680         tg->idletime_threshold_conf = idle_time;
1681         tg->latency_target_conf = latency_time;
1682
1683         /* force user to configure all settings for low limit  */
1684         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1685               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1686             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1687             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1688                 tg->bps[READ][LIMIT_LOW] = 0;
1689                 tg->bps[WRITE][LIMIT_LOW] = 0;
1690                 tg->iops[READ][LIMIT_LOW] = 0;
1691                 tg->iops[WRITE][LIMIT_LOW] = 0;
1692                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1693                 tg->latency_target = DFL_LATENCY_TARGET;
1694         } else if (index == LIMIT_LOW) {
1695                 tg->idletime_threshold = tg->idletime_threshold_conf;
1696                 tg->latency_target = tg->latency_target_conf;
1697         }
1698
1699         blk_throtl_update_limit_valid(tg->td);
1700         if (tg->td->limit_valid[LIMIT_LOW]) {
1701                 if (index == LIMIT_LOW)
1702                         tg->td->limit_index = LIMIT_LOW;
1703         } else
1704                 tg->td->limit_index = LIMIT_MAX;
1705         tg_conf_updated(tg, index == LIMIT_LOW &&
1706                 tg->td->limit_valid[LIMIT_LOW]);
1707         ret = 0;
1708 out_finish:
1709         blkg_conf_finish(&ctx);
1710         return ret ?: nbytes;
1711 }
1712
1713 static struct cftype throtl_files[] = {
1714 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1715         {
1716                 .name = "low",
1717                 .flags = CFTYPE_NOT_ON_ROOT,
1718                 .seq_show = tg_print_limit,
1719                 .write = tg_set_limit,
1720                 .private = LIMIT_LOW,
1721         },
1722 #endif
1723         {
1724                 .name = "max",
1725                 .flags = CFTYPE_NOT_ON_ROOT,
1726                 .seq_show = tg_print_limit,
1727                 .write = tg_set_limit,
1728                 .private = LIMIT_MAX,
1729         },
1730         { }     /* terminate */
1731 };
1732
1733 static void throtl_shutdown_wq(struct request_queue *q)
1734 {
1735         struct throtl_data *td = q->td;
1736
1737         cancel_work_sync(&td->dispatch_work);
1738 }
1739
1740 static struct blkcg_policy blkcg_policy_throtl = {
1741         .dfl_cftypes            = throtl_files,
1742         .legacy_cftypes         = throtl_legacy_files,
1743
1744         .pd_alloc_fn            = throtl_pd_alloc,
1745         .pd_init_fn             = throtl_pd_init,
1746         .pd_online_fn           = throtl_pd_online,
1747         .pd_offline_fn          = throtl_pd_offline,
1748         .pd_free_fn             = throtl_pd_free,
1749 };
1750
1751 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1752 {
1753         unsigned long rtime = jiffies, wtime = jiffies;
1754
1755         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1756                 rtime = tg->last_low_overflow_time[READ];
1757         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1758                 wtime = tg->last_low_overflow_time[WRITE];
1759         return min(rtime, wtime);
1760 }
1761
1762 /* tg should not be an intermediate node */
1763 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1764 {
1765         struct throtl_service_queue *parent_sq;
1766         struct throtl_grp *parent = tg;
1767         unsigned long ret = __tg_last_low_overflow_time(tg);
1768
1769         while (true) {
1770                 parent_sq = parent->service_queue.parent_sq;
1771                 parent = sq_to_tg(parent_sq);
1772                 if (!parent)
1773                         break;
1774
1775                 /*
1776                  * The parent doesn't have low limit, it always reaches low
1777                  * limit. Its overflow time is useless for children
1778                  */
1779                 if (!parent->bps[READ][LIMIT_LOW] &&
1780                     !parent->iops[READ][LIMIT_LOW] &&
1781                     !parent->bps[WRITE][LIMIT_LOW] &&
1782                     !parent->iops[WRITE][LIMIT_LOW])
1783                         continue;
1784                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1785                         ret = __tg_last_low_overflow_time(parent);
1786         }
1787         return ret;
1788 }
1789
1790 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1791 {
1792         /*
1793          * cgroup is idle if:
1794          * - single idle is too long, longer than a fixed value (in case user
1795          *   configure a too big threshold) or 4 times of idletime threshold
1796          * - average think time is more than threshold
1797          * - IO latency is largely below threshold
1798          */
1799         unsigned long time;
1800         bool ret;
1801
1802         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1803         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1804               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1805               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1806               tg->avg_idletime > tg->idletime_threshold ||
1807               (tg->latency_target && tg->bio_cnt &&
1808                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1809         throtl_log(&tg->service_queue,
1810                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1811                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1812                 tg->bio_cnt, ret, tg->td->scale);
1813         return ret;
1814 }
1815
1816 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1817 {
1818         struct throtl_service_queue *sq = &tg->service_queue;
1819         bool read_limit, write_limit;
1820
1821         /*
1822          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1823          * reaches), it's ok to upgrade to next limit
1824          */
1825         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1826         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1827         if (!read_limit && !write_limit)
1828                 return true;
1829         if (read_limit && sq->nr_queued[READ] &&
1830             (!write_limit || sq->nr_queued[WRITE]))
1831                 return true;
1832         if (write_limit && sq->nr_queued[WRITE] &&
1833             (!read_limit || sq->nr_queued[READ]))
1834                 return true;
1835
1836         if (time_after_eq(jiffies,
1837                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1838             throtl_tg_is_idle(tg))
1839                 return true;
1840         return false;
1841 }
1842
1843 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1844 {
1845         while (true) {
1846                 if (throtl_tg_can_upgrade(tg))
1847                         return true;
1848                 tg = sq_to_tg(tg->service_queue.parent_sq);
1849                 if (!tg || !tg_to_blkg(tg)->parent)
1850                         return false;
1851         }
1852         return false;
1853 }
1854
1855 static bool throtl_can_upgrade(struct throtl_data *td,
1856         struct throtl_grp *this_tg)
1857 {
1858         struct cgroup_subsys_state *pos_css;
1859         struct blkcg_gq *blkg;
1860
1861         if (td->limit_index != LIMIT_LOW)
1862                 return false;
1863
1864         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1865                 return false;
1866
1867         rcu_read_lock();
1868         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1869                 struct throtl_grp *tg = blkg_to_tg(blkg);
1870
1871                 if (tg == this_tg)
1872                         continue;
1873                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1874                         continue;
1875                 if (!throtl_hierarchy_can_upgrade(tg)) {
1876                         rcu_read_unlock();
1877                         return false;
1878                 }
1879         }
1880         rcu_read_unlock();
1881         return true;
1882 }
1883
1884 static void throtl_upgrade_check(struct throtl_grp *tg)
1885 {
1886         unsigned long now = jiffies;
1887
1888         if (tg->td->limit_index != LIMIT_LOW)
1889                 return;
1890
1891         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1892                 return;
1893
1894         tg->last_check_time = now;
1895
1896         if (!time_after_eq(now,
1897              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1898                 return;
1899
1900         if (throtl_can_upgrade(tg->td, NULL))
1901                 throtl_upgrade_state(tg->td);
1902 }
1903
1904 static void throtl_upgrade_state(struct throtl_data *td)
1905 {
1906         struct cgroup_subsys_state *pos_css;
1907         struct blkcg_gq *blkg;
1908
1909         throtl_log(&td->service_queue, "upgrade to max");
1910         td->limit_index = LIMIT_MAX;
1911         td->low_upgrade_time = jiffies;
1912         td->scale = 0;
1913         rcu_read_lock();
1914         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1915                 struct throtl_grp *tg = blkg_to_tg(blkg);
1916                 struct throtl_service_queue *sq = &tg->service_queue;
1917
1918                 tg->disptime = jiffies - 1;
1919                 throtl_select_dispatch(sq);
1920                 throtl_schedule_next_dispatch(sq, true);
1921         }
1922         rcu_read_unlock();
1923         throtl_select_dispatch(&td->service_queue);
1924         throtl_schedule_next_dispatch(&td->service_queue, true);
1925         queue_work(kthrotld_workqueue, &td->dispatch_work);
1926 }
1927
1928 static void throtl_downgrade_state(struct throtl_data *td, int new)
1929 {
1930         td->scale /= 2;
1931
1932         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1933         if (td->scale) {
1934                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1935                 return;
1936         }
1937
1938         td->limit_index = new;
1939         td->low_downgrade_time = jiffies;
1940 }
1941
1942 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1943 {
1944         struct throtl_data *td = tg->td;
1945         unsigned long now = jiffies;
1946
1947         /*
1948          * If cgroup is below low limit, consider downgrade and throttle other
1949          * cgroups
1950          */
1951         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1952             time_after_eq(now, tg_last_low_overflow_time(tg) +
1953                                         td->throtl_slice) &&
1954             (!throtl_tg_is_idle(tg) ||
1955              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1956                 return true;
1957         return false;
1958 }
1959
1960 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1961 {
1962         while (true) {
1963                 if (!throtl_tg_can_downgrade(tg))
1964                         return false;
1965                 tg = sq_to_tg(tg->service_queue.parent_sq);
1966                 if (!tg || !tg_to_blkg(tg)->parent)
1967                         break;
1968         }
1969         return true;
1970 }
1971
1972 static void throtl_downgrade_check(struct throtl_grp *tg)
1973 {
1974         uint64_t bps;
1975         unsigned int iops;
1976         unsigned long elapsed_time;
1977         unsigned long now = jiffies;
1978
1979         if (tg->td->limit_index != LIMIT_MAX ||
1980             !tg->td->limit_valid[LIMIT_LOW])
1981                 return;
1982         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1983                 return;
1984         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1985                 return;
1986
1987         elapsed_time = now - tg->last_check_time;
1988         tg->last_check_time = now;
1989
1990         if (time_before(now, tg_last_low_overflow_time(tg) +
1991                         tg->td->throtl_slice))
1992                 return;
1993
1994         if (tg->bps[READ][LIMIT_LOW]) {
1995                 bps = tg->last_bytes_disp[READ] * HZ;
1996                 do_div(bps, elapsed_time);
1997                 if (bps >= tg->bps[READ][LIMIT_LOW])
1998                         tg->last_low_overflow_time[READ] = now;
1999         }
2000
2001         if (tg->bps[WRITE][LIMIT_LOW]) {
2002                 bps = tg->last_bytes_disp[WRITE] * HZ;
2003                 do_div(bps, elapsed_time);
2004                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2005                         tg->last_low_overflow_time[WRITE] = now;
2006         }
2007
2008         if (tg->iops[READ][LIMIT_LOW]) {
2009                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2010                 if (iops >= tg->iops[READ][LIMIT_LOW])
2011                         tg->last_low_overflow_time[READ] = now;
2012         }
2013
2014         if (tg->iops[WRITE][LIMIT_LOW]) {
2015                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2016                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2017                         tg->last_low_overflow_time[WRITE] = now;
2018         }
2019
2020         /*
2021          * If cgroup is below low limit, consider downgrade and throttle other
2022          * cgroups
2023          */
2024         if (throtl_hierarchy_can_downgrade(tg))
2025                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2026
2027         tg->last_bytes_disp[READ] = 0;
2028         tg->last_bytes_disp[WRITE] = 0;
2029         tg->last_io_disp[READ] = 0;
2030         tg->last_io_disp[WRITE] = 0;
2031 }
2032
2033 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2034 {
2035         unsigned long now = ktime_get_ns() >> 10;
2036         unsigned long last_finish_time = tg->last_finish_time;
2037
2038         if (now <= last_finish_time || last_finish_time == 0 ||
2039             last_finish_time == tg->checked_last_finish_time)
2040                 return;
2041
2042         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2043         tg->checked_last_finish_time = last_finish_time;
2044 }
2045
2046 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2047 static void throtl_update_latency_buckets(struct throtl_data *td)
2048 {
2049         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2050         int i, cpu, rw;
2051         unsigned long last_latency[2] = { 0 };
2052         unsigned long latency[2];
2053
2054         if (!blk_queue_nonrot(td->queue))
2055                 return;
2056         if (time_before(jiffies, td->last_calculate_time + HZ))
2057                 return;
2058         td->last_calculate_time = jiffies;
2059
2060         memset(avg_latency, 0, sizeof(avg_latency));
2061         for (rw = READ; rw <= WRITE; rw++) {
2062                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2063                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2064
2065                         for_each_possible_cpu(cpu) {
2066                                 struct latency_bucket *bucket;
2067
2068                                 /* this isn't race free, but ok in practice */
2069                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2070                                         cpu);
2071                                 tmp->total_latency += bucket[i].total_latency;
2072                                 tmp->samples += bucket[i].samples;
2073                                 bucket[i].total_latency = 0;
2074                                 bucket[i].samples = 0;
2075                         }
2076
2077                         if (tmp->samples >= 32) {
2078                                 int samples = tmp->samples;
2079
2080                                 latency[rw] = tmp->total_latency;
2081
2082                                 tmp->total_latency = 0;
2083                                 tmp->samples = 0;
2084                                 latency[rw] /= samples;
2085                                 if (latency[rw] == 0)
2086                                         continue;
2087                                 avg_latency[rw][i].latency = latency[rw];
2088                         }
2089                 }
2090         }
2091
2092         for (rw = READ; rw <= WRITE; rw++) {
2093                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2094                         if (!avg_latency[rw][i].latency) {
2095                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2096                                         td->avg_buckets[rw][i].latency =
2097                                                 last_latency[rw];
2098                                 continue;
2099                         }
2100
2101                         if (!td->avg_buckets[rw][i].valid)
2102                                 latency[rw] = avg_latency[rw][i].latency;
2103                         else
2104                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2105                                         avg_latency[rw][i].latency) >> 3;
2106
2107                         td->avg_buckets[rw][i].latency = max(latency[rw],
2108                                 last_latency[rw]);
2109                         td->avg_buckets[rw][i].valid = true;
2110                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2111                 }
2112         }
2113
2114         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2115                 throtl_log(&td->service_queue,
2116                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2117                         "write latency=%ld, write valid=%d", i,
2118                         td->avg_buckets[READ][i].latency,
2119                         td->avg_buckets[READ][i].valid,
2120                         td->avg_buckets[WRITE][i].latency,
2121                         td->avg_buckets[WRITE][i].valid);
2122 }
2123 #else
2124 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2125 {
2126 }
2127 #endif
2128
2129 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2130 {
2131 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2132         /* fallback to root_blkg if we fail to get a blkg ref */
2133         if (bio->bi_css && (bio_associate_blkg(bio, tg_to_blkg(tg)) == -ENODEV))
2134                 bio_associate_blkg(bio, bio->bi_disk->queue->root_blkg);
2135         bio_issue_init(&bio->bi_issue, bio_sectors(bio));
2136 #endif
2137 }
2138
2139 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2140                     struct bio *bio)
2141 {
2142         struct throtl_qnode *qn = NULL;
2143         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2144         struct throtl_service_queue *sq;
2145         bool rw = bio_data_dir(bio);
2146         bool throttled = false;
2147         struct throtl_data *td = tg->td;
2148
2149         WARN_ON_ONCE(!rcu_read_lock_held());
2150
2151         /* see throtl_charge_bio() */
2152         if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2153                 goto out;
2154
2155         spin_lock_irq(q->queue_lock);
2156
2157         throtl_update_latency_buckets(td);
2158
2159         if (unlikely(blk_queue_bypass(q)))
2160                 goto out_unlock;
2161
2162         blk_throtl_assoc_bio(tg, bio);
2163         blk_throtl_update_idletime(tg);
2164
2165         sq = &tg->service_queue;
2166
2167 again:
2168         while (true) {
2169                 if (tg->last_low_overflow_time[rw] == 0)
2170                         tg->last_low_overflow_time[rw] = jiffies;
2171                 throtl_downgrade_check(tg);
2172                 throtl_upgrade_check(tg);
2173                 /* throtl is FIFO - if bios are already queued, should queue */
2174                 if (sq->nr_queued[rw])
2175                         break;
2176
2177                 /* if above limits, break to queue */
2178                 if (!tg_may_dispatch(tg, bio, NULL)) {
2179                         tg->last_low_overflow_time[rw] = jiffies;
2180                         if (throtl_can_upgrade(td, tg)) {
2181                                 throtl_upgrade_state(td);
2182                                 goto again;
2183                         }
2184                         break;
2185                 }
2186
2187                 /* within limits, let's charge and dispatch directly */
2188                 throtl_charge_bio(tg, bio);
2189
2190                 /*
2191                  * We need to trim slice even when bios are not being queued
2192                  * otherwise it might happen that a bio is not queued for
2193                  * a long time and slice keeps on extending and trim is not
2194                  * called for a long time. Now if limits are reduced suddenly
2195                  * we take into account all the IO dispatched so far at new
2196                  * low rate and * newly queued IO gets a really long dispatch
2197                  * time.
2198                  *
2199                  * So keep on trimming slice even if bio is not queued.
2200                  */
2201                 throtl_trim_slice(tg, rw);
2202
2203                 /*
2204                  * @bio passed through this layer without being throttled.
2205                  * Climb up the ladder.  If we''re already at the top, it
2206                  * can be executed directly.
2207                  */
2208                 qn = &tg->qnode_on_parent[rw];
2209                 sq = sq->parent_sq;
2210                 tg = sq_to_tg(sq);
2211                 if (!tg)
2212                         goto out_unlock;
2213         }
2214
2215         /* out-of-limit, queue to @tg */
2216         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2217                    rw == READ ? 'R' : 'W',
2218                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2219                    tg_bps_limit(tg, rw),
2220                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2221                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2222
2223         tg->last_low_overflow_time[rw] = jiffies;
2224
2225         td->nr_queued[rw]++;
2226         throtl_add_bio_tg(bio, qn, tg);
2227         throttled = true;
2228
2229         /*
2230          * Update @tg's dispatch time and force schedule dispatch if @tg
2231          * was empty before @bio.  The forced scheduling isn't likely to
2232          * cause undue delay as @bio is likely to be dispatched directly if
2233          * its @tg's disptime is not in the future.
2234          */
2235         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2236                 tg_update_disptime(tg);
2237                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2238         }
2239
2240 out_unlock:
2241         spin_unlock_irq(q->queue_lock);
2242 out:
2243         bio_set_flag(bio, BIO_THROTTLED);
2244
2245 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2246         if (throttled || !td->track_bio_latency)
2247                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2248 #endif
2249         return throttled;
2250 }
2251
2252 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2253 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2254         int op, unsigned long time)
2255 {
2256         struct latency_bucket *latency;
2257         int index;
2258
2259         if (!td || td->limit_index != LIMIT_LOW ||
2260             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2261             !blk_queue_nonrot(td->queue))
2262                 return;
2263
2264         index = request_bucket_index(size);
2265
2266         latency = get_cpu_ptr(td->latency_buckets[op]);
2267         latency[index].total_latency += time;
2268         latency[index].samples++;
2269         put_cpu_ptr(td->latency_buckets[op]);
2270 }
2271
2272 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2273 {
2274         struct request_queue *q = rq->q;
2275         struct throtl_data *td = q->td;
2276
2277         throtl_track_latency(td, rq->throtl_size, req_op(rq), time_ns >> 10);
2278 }
2279
2280 void blk_throtl_bio_endio(struct bio *bio)
2281 {
2282         struct blkcg_gq *blkg;
2283         struct throtl_grp *tg;
2284         u64 finish_time_ns;
2285         unsigned long finish_time;
2286         unsigned long start_time;
2287         unsigned long lat;
2288         int rw = bio_data_dir(bio);
2289
2290         blkg = bio->bi_blkg;
2291         if (!blkg)
2292                 return;
2293         tg = blkg_to_tg(blkg);
2294
2295         finish_time_ns = ktime_get_ns();
2296         tg->last_finish_time = finish_time_ns >> 10;
2297
2298         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2299         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2300         if (!start_time || finish_time <= start_time)
2301                 return;
2302
2303         lat = finish_time - start_time;
2304         /* this is only for bio based driver */
2305         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2306                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2307                                      bio_op(bio), lat);
2308
2309         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2310                 int bucket;
2311                 unsigned int threshold;
2312
2313                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2314                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2315                         tg->latency_target;
2316                 if (lat > threshold)
2317                         tg->bad_bio_cnt++;
2318                 /*
2319                  * Not race free, could get wrong count, which means cgroups
2320                  * will be throttled
2321                  */
2322                 tg->bio_cnt++;
2323         }
2324
2325         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2326                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2327                 tg->bio_cnt /= 2;
2328                 tg->bad_bio_cnt /= 2;
2329         }
2330 }
2331 #endif
2332
2333 /*
2334  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2335  * return, @parent_sq is guaranteed to not have any active children tg's
2336  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2337  */
2338 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2339 {
2340         struct throtl_grp *tg;
2341
2342         while ((tg = throtl_rb_first(parent_sq))) {
2343                 struct throtl_service_queue *sq = &tg->service_queue;
2344                 struct bio *bio;
2345
2346                 throtl_dequeue_tg(tg);
2347
2348                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2349                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2350                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2351                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2352         }
2353 }
2354
2355 /**
2356  * blk_throtl_drain - drain throttled bios
2357  * @q: request_queue to drain throttled bios for
2358  *
2359  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2360  */
2361 void blk_throtl_drain(struct request_queue *q)
2362         __releases(q->queue_lock) __acquires(q->queue_lock)
2363 {
2364         struct throtl_data *td = q->td;
2365         struct blkcg_gq *blkg;
2366         struct cgroup_subsys_state *pos_css;
2367         struct bio *bio;
2368         int rw;
2369
2370         queue_lockdep_assert_held(q);
2371         rcu_read_lock();
2372
2373         /*
2374          * Drain each tg while doing post-order walk on the blkg tree, so
2375          * that all bios are propagated to td->service_queue.  It'd be
2376          * better to walk service_queue tree directly but blkg walk is
2377          * easier.
2378          */
2379         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2380                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2381
2382         /* finally, transfer bios from top-level tg's into the td */
2383         tg_drain_bios(&td->service_queue);
2384
2385         rcu_read_unlock();
2386         spin_unlock_irq(q->queue_lock);
2387
2388         /* all bios now should be in td->service_queue, issue them */
2389         for (rw = READ; rw <= WRITE; rw++)
2390                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2391                                                 NULL)))
2392                         generic_make_request(bio);
2393
2394         spin_lock_irq(q->queue_lock);
2395 }
2396
2397 int blk_throtl_init(struct request_queue *q)
2398 {
2399         struct throtl_data *td;
2400         int ret;
2401
2402         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2403         if (!td)
2404                 return -ENOMEM;
2405         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2406                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2407         if (!td->latency_buckets[READ]) {
2408                 kfree(td);
2409                 return -ENOMEM;
2410         }
2411         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2412                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2413         if (!td->latency_buckets[WRITE]) {
2414                 free_percpu(td->latency_buckets[READ]);
2415                 kfree(td);
2416                 return -ENOMEM;
2417         }
2418
2419         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2420         throtl_service_queue_init(&td->service_queue);
2421
2422         q->td = td;
2423         td->queue = q;
2424
2425         td->limit_valid[LIMIT_MAX] = true;
2426         td->limit_index = LIMIT_MAX;
2427         td->low_upgrade_time = jiffies;
2428         td->low_downgrade_time = jiffies;
2429
2430         /* activate policy */
2431         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2432         if (ret) {
2433                 free_percpu(td->latency_buckets[READ]);
2434                 free_percpu(td->latency_buckets[WRITE]);
2435                 kfree(td);
2436         }
2437         return ret;
2438 }
2439
2440 void blk_throtl_exit(struct request_queue *q)
2441 {
2442         BUG_ON(!q->td);
2443         throtl_shutdown_wq(q);
2444         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2445         free_percpu(q->td->latency_buckets[READ]);
2446         free_percpu(q->td->latency_buckets[WRITE]);
2447         kfree(q->td);
2448 }
2449
2450 void blk_throtl_register_queue(struct request_queue *q)
2451 {
2452         struct throtl_data *td;
2453         int i;
2454
2455         td = q->td;
2456         BUG_ON(!td);
2457
2458         if (blk_queue_nonrot(q)) {
2459                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2460                 td->filtered_latency = LATENCY_FILTERED_SSD;
2461         } else {
2462                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2463                 td->filtered_latency = LATENCY_FILTERED_HD;
2464                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2465                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2466                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2467                 }
2468         }
2469 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2470         /* if no low limit, use previous default */
2471         td->throtl_slice = DFL_THROTL_SLICE_HD;
2472 #endif
2473
2474         td->track_bio_latency = !queue_is_rq_based(q);
2475         if (!td->track_bio_latency)
2476                 blk_stat_enable_accounting(q);
2477 }
2478
2479 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2480 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2481 {
2482         if (!q->td)
2483                 return -EINVAL;
2484         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2485 }
2486
2487 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2488         const char *page, size_t count)
2489 {
2490         unsigned long v;
2491         unsigned long t;
2492
2493         if (!q->td)
2494                 return -EINVAL;
2495         if (kstrtoul(page, 10, &v))
2496                 return -EINVAL;
2497         t = msecs_to_jiffies(v);
2498         if (t == 0 || t > MAX_THROTL_SLICE)
2499                 return -EINVAL;
2500         q->td->throtl_slice = t;
2501         return count;
2502 }
2503 #endif
2504
2505 static int __init throtl_init(void)
2506 {
2507         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2508         if (!kthrotld_workqueue)
2509                 panic("Failed to create kthrotld\n");
2510
2511         return blkcg_policy_register(&blkcg_policy_throtl);
2512 }
2513
2514 module_init(throtl_init);