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         }
 585         rcu_read_unlock();
 586
 587         td->limit_valid[LIMIT_LOW] = low_valid;
 588 }
 589
 590 static void throtl_upgrade_state(struct throtl_data *td);
 591 static void throtl_pd_offline(struct blkg_policy_data *pd)
 592 {
 593         struct throtl_grp *tg = pd_to_tg(pd);
 594
 595         tg->bps[READ][LIMIT_LOW] = 0;
 596         tg->bps[WRITE][LIMIT_LOW] = 0;
 597         tg->iops[READ][LIMIT_LOW] = 0;
 598         tg->iops[WRITE][LIMIT_LOW] = 0;
 599
 600         blk_throtl_update_limit_valid(tg->td);
 601
 602         if (!tg->td->limit_valid[tg->td->limit_index])
 603                 throtl_upgrade_state(tg->td);
 604 }
 605
 606 static void throtl_pd_free(struct blkg_policy_data *pd)
 607 {
 608         struct throtl_grp *tg = pd_to_tg(pd);
 609
 610         del_timer_sync(&tg->service_queue.pending_timer);
 611         kfree(tg);
 612 }
 613
 614 static struct throtl_grp *
 615 throtl_rb_first(struct throtl_service_queue *parent_sq)
 616 {
 617         /* Service tree is empty */
 618         if (!parent_sq->nr_pending)
 619                 return NULL;
 620
 621         if (!parent_sq->first_pending)
 622                 parent_sq->first_pending = rb_first(&parent_sq->pending_tree);
 623
 624         if (parent_sq->first_pending)
 625                 return rb_entry_tg(parent_sq->first_pending);
 626
 627         return NULL;
 628 }
 629
 630 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
 631 {
 632         rb_erase(n, root);
 633         RB_CLEAR_NODE(n);
 634 }
 635
 636 static void throtl_rb_erase(struct rb_node *n,
 637                             struct throtl_service_queue *parent_sq)
 638 {
 639         if (parent_sq->first_pending == n)
 640                 parent_sq->first_pending = NULL;
 641         rb_erase_init(n, &parent_sq->pending_tree);
 642         --parent_sq->nr_pending;
 643 }
 644
 645 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 646 {
 647         struct throtl_grp *tg;
 648
 649         tg = throtl_rb_first(parent_sq);
 650         if (!tg)
 651                 return;
 652
 653         parent_sq->first_pending_disptime = tg->disptime;
 654 }
 655
 656 static void tg_service_queue_add(struct throtl_grp *tg)
 657 {
 658         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 659         struct rb_node **node = &parent_sq->pending_tree.rb_node;
 660         struct rb_node *parent = NULL;
 661         struct throtl_grp *__tg;
 662         unsigned long key = tg->disptime;
 663         int left = 1;
 664
 665         while (*node != NULL) {
 666                 parent = *node;
 667                 __tg = rb_entry_tg(parent);
 668
 669                 if (time_before(key, __tg->disptime))
 670                         node = &parent->rb_left;
 671                 else {
 672                         node = &parent->rb_right;
 673                         left = 0;
 674                 }
 675         }
 676
 677         if (left)
 678                 parent_sq->first_pending = &tg->rb_node;
 679
 680         rb_link_node(&tg->rb_node, parent, node);
 681         rb_insert_color(&tg->rb_node, &parent_sq->pending_tree);
 682 }
 683
 684 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 685 {
 686         tg_service_queue_add(tg);
 687         tg->flags |= THROTL_TG_PENDING;
 688         tg->service_queue.parent_sq->nr_pending++;
 689 }
 690
 691 static void throtl_enqueue_tg(struct throtl_grp *tg)
 692 {
 693         if (!(tg->flags & THROTL_TG_PENDING))
 694                 __throtl_enqueue_tg(tg);
 695 }
 696
 697 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 698 {
 699         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 700         tg->flags &= ~THROTL_TG_PENDING;
 701 }
 702
 703 static void throtl_dequeue_tg(struct throtl_grp *tg)
 704 {
 705         if (tg->flags & THROTL_TG_PENDING)
 706                 __throtl_dequeue_tg(tg);
 707 }
 708
 709 /* Call with queue lock held */
 710 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 711                                           unsigned long expires)
 712 {
 713         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 714
 715         /*
 716          * Since we are adjusting the throttle limit dynamically, the sleep
 717          * time calculated according to previous limit might be invalid. It's
 718          * possible the cgroup sleep time is very long and no other cgroups
 719          * have IO running so notify the limit changes. Make sure the cgroup
 720          * doesn't sleep too long to avoid the missed notification.
 721          */
 722         if (time_after(expires, max_expire))
 723                 expires = max_expire;
 724         mod_timer(&sq->pending_timer, expires);
 725         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 726                    expires - jiffies, jiffies);
 727 }
 728
 729 /**
 730  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 731  * @sq: the service_queue to schedule dispatch for
 732  * @force: force scheduling
 733  *
 734  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 735  * dispatch time of the first pending child.  Returns %true if either timer
 736  * is armed or there's no pending child left.  %false if the current
 737  * dispatch window is still open and the caller should continue
 738  * dispatching.
 739  *
 740  * If @force is %true, the dispatch timer is always scheduled and this
 741  * function is guaranteed to return %true.  This is to be used when the
 742  * caller can't dispatch itself and needs to invoke pending_timer
 743  * unconditionally.  Note that forced scheduling is likely to induce short
 744  * delay before dispatch starts even if @sq->first_pending_disptime is not
 745  * in the future and thus shouldn't be used in hot paths.
 746  */
 747 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 748                                           bool force)
 749 {
 750         /* any pending children left? */
 751         if (!sq->nr_pending)
 752                 return true;
 753
 754         update_min_dispatch_time(sq);
 755
 756         /* is the next dispatch time in the future? */
 757         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 758                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 759                 return true;
 760         }
 761
 762         /* tell the caller to continue dispatching */
 763         return false;
 764 }
 765
 766 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 767                 bool rw, unsigned long start)
 768 {
 769         tg->bytes_disp[rw] = 0;
 770         tg->io_disp[rw] = 0;
 771
 772         /*
 773          * Previous slice has expired. We must have trimmed it after last
 774          * bio dispatch. That means since start of last slice, we never used
 775          * that bandwidth. Do try to make use of that bandwidth while giving
 776          * credit.
 777          */
 778         if (time_after_eq(start, tg->slice_start[rw]))
 779                 tg->slice_start[rw] = start;
 780
 781         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 782         throtl_log(&tg->service_queue,
 783                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 784                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 785                    tg->slice_end[rw], jiffies);
 786 }
 787
 788 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 789 {
 790         tg->bytes_disp[rw] = 0;
 791         tg->io_disp[rw] = 0;
 792         tg->slice_start[rw] = jiffies;
 793         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 794         throtl_log(&tg->service_queue,
 795                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 796                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 797                    tg->slice_end[rw], jiffies);
 798 }
 799
 800 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 801                                         unsigned long jiffy_end)
 802 {
 803         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 804 }
 805
 806 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 807                                        unsigned long jiffy_end)
 808 {
 809         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 810         throtl_log(&tg->service_queue,
 811                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 812                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 813                    tg->slice_end[rw], jiffies);
 814 }
 815
 816 /* Determine if previously allocated or extended slice is complete or not */
 817 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 818 {
 819         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 820                 return false;
 821
 822         return true;
 823 }
 824
 825 /* Trim the used slices and adjust slice start accordingly */
 826 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 827 {
 828         unsigned long nr_slices, time_elapsed, io_trim;
 829         u64 bytes_trim, tmp;
 830
 831         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 832
 833         /*
 834          * If bps are unlimited (-1), then time slice don't get
 835          * renewed. Don't try to trim the slice if slice is used. A new
 836          * slice will start when appropriate.
 837          */
 838         if (throtl_slice_used(tg, rw))
 839                 return;
 840
 841         /*
 842          * A bio has been dispatched. Also adjust slice_end. It might happen
 843          * that initially cgroup limit was very low resulting in high
 844          * slice_end, but later limit was bumped up and bio was dispached
 845          * sooner, then we need to reduce slice_end. A high bogus slice_end
 846          * is bad because it does not allow new slice to start.
 847          */
 848
 849         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 850
 851         time_elapsed = jiffies - tg->slice_start[rw];
 852
 853         nr_slices = time_elapsed / tg->td->throtl_slice;
 854
 855         if (!nr_slices)
 856                 return;
 857         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 858         do_div(tmp, HZ);
 859         bytes_trim = tmp;
 860
 861         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 862                 HZ;
 863
 864         if (!bytes_trim && !io_trim)
 865                 return;
 866
 867         if (tg->bytes_disp[rw] >= bytes_trim)
 868                 tg->bytes_disp[rw] -= bytes_trim;
 869         else
 870                 tg->bytes_disp[rw] = 0;
 871
 872         if (tg->io_disp[rw] >= io_trim)
 873                 tg->io_disp[rw] -= io_trim;
 874         else
 875                 tg->io_disp[rw] = 0;
 876
 877         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 878
 879         throtl_log(&tg->service_queue,
 880                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 881                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 882                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 883 }
 884
 885 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 886                                   unsigned long *wait)
 887 {
 888         bool rw = bio_data_dir(bio);
 889         unsigned int io_allowed;
 890         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 891         u64 tmp;
 892
 893         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 894
 895         /* Slice has just started. Consider one slice interval */
 896         if (!jiffy_elapsed)
 897                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 898
 899         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 900
 901         /*
 902          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 903          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 904          * will allow dispatch after 1 second and after that slice should
 905          * have been trimmed.
 906          */
 907
 908         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 909         do_div(tmp, HZ);
 910
 911         if (tmp > UINT_MAX)
 912                 io_allowed = UINT_MAX;
 913         else
 914                 io_allowed = tmp;
 915
 916         if (tg->io_disp[rw] + 1 <= io_allowed) {
 917                 if (wait)
 918                         *wait = 0;
 919                 return true;
 920         }
 921
 922         /* Calc approx time to dispatch */
 923         jiffy_wait = ((tg->io_disp[rw] + 1) * HZ) / tg_iops_limit(tg, rw) + 1;
 924
 925         if (jiffy_wait > jiffy_elapsed)
 926                 jiffy_wait = jiffy_wait - jiffy_elapsed;
 927         else
 928                 jiffy_wait = 1;
 929
 930         if (wait)
 931                 *wait = jiffy_wait;
 932         return false;
 933 }
 934
 935 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 936                                  unsigned long *wait)
 937 {
 938         bool rw = bio_data_dir(bio);
 939         u64 bytes_allowed, extra_bytes, tmp;
 940         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 941         unsigned int bio_size = throtl_bio_data_size(bio);
 942
 943         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 944
 945         /* Slice has just started. Consider one slice interval */
 946         if (!jiffy_elapsed)
 947                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 948
 949         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 950
 951         tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
 952         do_div(tmp, HZ);
 953         bytes_allowed = tmp;
 954
 955         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 956                 if (wait)
 957                         *wait = 0;
 958                 return true;
 959         }
 960
 961         /* Calc approx time to dispatch */
 962         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 963         jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
 964
 965         if (!jiffy_wait)
 966                 jiffy_wait = 1;
 967
 968         /*
 969          * This wait time is without taking into consideration the rounding
 970          * up we did. Add that time also.
 971          */
 972         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 973         if (wait)
 974                 *wait = jiffy_wait;
 975         return false;
 976 }
 977
 978 /*
 979  * Returns whether one can dispatch a bio or not. Also returns approx number
 980  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 981  */
 982 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 983                             unsigned long *wait)
 984 {
 985         bool rw = bio_data_dir(bio);
 986         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 987
 988         /*
 989          * Currently whole state machine of group depends on first bio
 990          * queued in the group bio list. So one should not be calling
 991          * this function with a different bio if there are other bios
 992          * queued.
 993          */
 994         BUG_ON(tg->service_queue.nr_queued[rw] &&
 995                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 996
 997         /* If tg->bps = -1, then BW is unlimited */
 998         if (tg_bps_limit(tg, rw) == U64_MAX &&
 999             tg_iops_limit(tg, rw) == UINT_MAX) {
1000                 if (wait)
1001                         *wait = 0;
1002                 return true;
1003         }
1004
1005         /*
1006          * If previous slice expired, start a new one otherwise renew/extend
1007          * existing slice to make sure it is at least throtl_slice interval
1008          * long since now. New slice is started only for empty throttle group.
1009          * If there is queued bio, that means there should be an active
1010          * slice and it should be extended instead.
1011          */
1012         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1013                 throtl_start_new_slice(tg, rw);
1014         else {
1015                 if (time_before(tg->slice_end[rw],
1016                     jiffies + tg->td->throtl_slice))
1017                         throtl_extend_slice(tg, rw,
1018                                 jiffies + tg->td->throtl_slice);
1019         }
1020
1021         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1022             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1023                 if (wait)
1024                         *wait = 0;
1025                 return true;
1026         }
1027
1028         max_wait = max(bps_wait, iops_wait);
1029
1030         if (wait)
1031                 *wait = max_wait;
1032
1033         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1034                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1035
1036         return false;
1037 }
1038
1039 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1040 {
1041         bool rw = bio_data_dir(bio);
1042         unsigned int bio_size = throtl_bio_data_size(bio);
1043
1044         /* Charge the bio to the group */
1045         tg->bytes_disp[rw] += bio_size;
1046         tg->io_disp[rw]++;
1047         tg->last_bytes_disp[rw] += bio_size;
1048         tg->last_io_disp[rw]++;
1049
1050         /*
1051          * BIO_THROTTLED is used to prevent the same bio to be throttled
1052          * more than once as a throttled bio will go through blk-throtl the
1053          * second time when it eventually gets issued.  Set it when a bio
1054          * is being charged to a tg.
1055          */
1056         if (!bio_flagged(bio, BIO_THROTTLED))
1057                 bio_set_flag(bio, BIO_THROTTLED);
1058 }
1059
1060 /**
1061  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1062  * @bio: bio to add
1063  * @qn: qnode to use
1064  * @tg: the target throtl_grp
1065  *
1066  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1067  * tg->qnode_on_self[] is used.
1068  */
1069 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1070                               struct throtl_grp *tg)
1071 {
1072         struct throtl_service_queue *sq = &tg->service_queue;
1073         bool rw = bio_data_dir(bio);
1074
1075         if (!qn)
1076                 qn = &tg->qnode_on_self[rw];
1077
1078         /*
1079          * If @tg doesn't currently have any bios queued in the same
1080          * direction, queueing @bio can change when @tg should be
1081          * dispatched.  Mark that @tg was empty.  This is automatically
1082          * cleaered on the next tg_update_disptime().
1083          */
1084         if (!sq->nr_queued[rw])
1085                 tg->flags |= THROTL_TG_WAS_EMPTY;
1086
1087         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1088
1089         sq->nr_queued[rw]++;
1090         throtl_enqueue_tg(tg);
1091 }
1092
1093 static void tg_update_disptime(struct throtl_grp *tg)
1094 {
1095         struct throtl_service_queue *sq = &tg->service_queue;
1096         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1097         struct bio *bio;
1098
1099         bio = throtl_peek_queued(&sq->queued[READ]);
1100         if (bio)
1101                 tg_may_dispatch(tg, bio, &read_wait);
1102
1103         bio = throtl_peek_queued(&sq->queued[WRITE]);
1104         if (bio)
1105                 tg_may_dispatch(tg, bio, &write_wait);
1106
1107         min_wait = min(read_wait, write_wait);
1108         disptime = jiffies + min_wait;
1109
1110         /* Update dispatch time */
1111         throtl_dequeue_tg(tg);
1112         tg->disptime = disptime;
1113         throtl_enqueue_tg(tg);
1114
1115         /* see throtl_add_bio_tg() */
1116         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1117 }
1118
1119 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1120                                         struct throtl_grp *parent_tg, bool rw)
1121 {
1122         if (throtl_slice_used(parent_tg, rw)) {
1123                 throtl_start_new_slice_with_credit(parent_tg, rw,
1124                                 child_tg->slice_start[rw]);
1125         }
1126
1127 }
1128
1129 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1130 {
1131         struct throtl_service_queue *sq = &tg->service_queue;
1132         struct throtl_service_queue *parent_sq = sq->parent_sq;
1133         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1134         struct throtl_grp *tg_to_put = NULL;
1135         struct bio *bio;
1136
1137         /*
1138          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1139          * from @tg may put its reference and @parent_sq might end up
1140          * getting released prematurely.  Remember the tg to put and put it
1141          * after @bio is transferred to @parent_sq.
1142          */
1143         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1144         sq->nr_queued[rw]--;
1145
1146         throtl_charge_bio(tg, bio);
1147
1148         /*
1149          * If our parent is another tg, we just need to transfer @bio to
1150          * the parent using throtl_add_bio_tg().  If our parent is
1151          * @td->service_queue, @bio is ready to be issued.  Put it on its
1152          * bio_lists[] and decrease total number queued.  The caller is
1153          * responsible for issuing these bios.
1154          */
1155         if (parent_tg) {
1156                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1157                 start_parent_slice_with_credit(tg, parent_tg, rw);
1158         } else {
1159                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1160                                      &parent_sq->queued[rw]);
1161                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1162                 tg->td->nr_queued[rw]--;
1163         }
1164
1165         throtl_trim_slice(tg, rw);
1166
1167         if (tg_to_put)
1168                 blkg_put(tg_to_blkg(tg_to_put));
1169 }
1170
1171 static int throtl_dispatch_tg(struct throtl_grp *tg)
1172 {
1173         struct throtl_service_queue *sq = &tg->service_queue;
1174         unsigned int nr_reads = 0, nr_writes = 0;
1175         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1176         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1177         struct bio *bio;
1178
1179         /* Try to dispatch 75% READS and 25% WRITES */
1180
1181         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1182                tg_may_dispatch(tg, bio, NULL)) {
1183
1184                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1185                 nr_reads++;
1186
1187                 if (nr_reads >= max_nr_reads)
1188                         break;
1189         }
1190
1191         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1192                tg_may_dispatch(tg, bio, NULL)) {
1193
1194                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1195                 nr_writes++;
1196
1197                 if (nr_writes >= max_nr_writes)
1198                         break;
1199         }
1200
1201         return nr_reads + nr_writes;
1202 }
1203
1204 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1205 {
1206         unsigned int nr_disp = 0;
1207
1208         while (1) {
1209                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1210                 struct throtl_service_queue *sq;
1211
1212                 if (!tg)
1213                         break;
1214
1215                 if (time_before(jiffies, tg->disptime))
1216                         break;
1217
1218                 throtl_dequeue_tg(tg);
1219
1220                 nr_disp += throtl_dispatch_tg(tg);
1221
1222                 sq = &tg->service_queue;
1223                 if (sq->nr_queued[0] || sq->nr_queued[1])
1224                         tg_update_disptime(tg);
1225
1226                 if (nr_disp >= throtl_quantum)
1227                         break;
1228         }
1229
1230         return nr_disp;
1231 }
1232
1233 static bool throtl_can_upgrade(struct throtl_data *td,
1234         struct throtl_grp *this_tg);
1235 /**
1236  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1237  * @arg: the throtl_service_queue being serviced
1238  *
1239  * This timer is armed when a child throtl_grp with active bio's become
1240  * pending and queued on the service_queue's pending_tree and expires when
1241  * the first child throtl_grp should be dispatched.  This function
1242  * dispatches bio's from the children throtl_grps to the parent
1243  * service_queue.
1244  *
1245  * If the parent's parent is another throtl_grp, dispatching is propagated
1246  * by either arming its pending_timer or repeating dispatch directly.  If
1247  * the top-level service_tree is reached, throtl_data->dispatch_work is
1248  * kicked so that the ready bio's are issued.
1249  */
1250 static void throtl_pending_timer_fn(struct timer_list *t)
1251 {
1252         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1253         struct throtl_grp *tg = sq_to_tg(sq);
1254         struct throtl_data *td = sq_to_td(sq);
1255         struct request_queue *q = td->queue;
1256         struct throtl_service_queue *parent_sq;
1257         bool dispatched;
1258         int ret;
1259
1260         spin_lock_irq(q->queue_lock);
1261         if (throtl_can_upgrade(td, NULL))
1262                 throtl_upgrade_state(td);
1263
1264 again:
1265         parent_sq = sq->parent_sq;
1266         dispatched = false;
1267
1268         while (true) {
1269                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1270                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1271                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1272
1273                 ret = throtl_select_dispatch(sq);
1274                 if (ret) {
1275                         throtl_log(sq, "bios disp=%u", ret);
1276                         dispatched = true;
1277                 }
1278
1279                 if (throtl_schedule_next_dispatch(sq, false))
1280                         break;
1281
1282                 /* this dispatch windows is still open, relax and repeat */
1283                 spin_unlock_irq(q->queue_lock);
1284                 cpu_relax();
1285                 spin_lock_irq(q->queue_lock);
1286         }
1287
1288         if (!dispatched)
1289                 goto out_unlock;
1290
1291         if (parent_sq) {
1292                 /* @parent_sq is another throl_grp, propagate dispatch */
1293                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1294                         tg_update_disptime(tg);
1295                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1296                                 /* window is already open, repeat dispatching */
1297                                 sq = parent_sq;
1298                                 tg = sq_to_tg(sq);
1299                                 goto again;
1300                         }
1301                 }
1302         } else {
1303                 /* reached the top-level, queue issueing */
1304                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1305         }
1306 out_unlock:
1307         spin_unlock_irq(q->queue_lock);
1308 }
1309
1310 /**
1311  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1312  * @work: work item being executed
1313  *
1314  * This function is queued for execution when bio's reach the bio_lists[]
1315  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1316  * function.
1317  */
1318 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1319 {
1320         struct throtl_data *td = container_of(work, struct throtl_data,
1321                                               dispatch_work);
1322         struct throtl_service_queue *td_sq = &td->service_queue;
1323         struct request_queue *q = td->queue;
1324         struct bio_list bio_list_on_stack;
1325         struct bio *bio;
1326         struct blk_plug plug;
1327         int rw;
1328
1329         bio_list_init(&bio_list_on_stack);
1330
1331         spin_lock_irq(q->queue_lock);
1332         for (rw = READ; rw <= WRITE; rw++)
1333                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1334                         bio_list_add(&bio_list_on_stack, bio);
1335         spin_unlock_irq(q->queue_lock);
1336
1337         if (!bio_list_empty(&bio_list_on_stack)) {
1338                 blk_start_plug(&plug);
1339                 while((bio = bio_list_pop(&bio_list_on_stack)))
1340                         generic_make_request(bio);
1341                 blk_finish_plug(&plug);
1342         }
1343 }
1344
1345 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1346                               int off)
1347 {
1348         struct throtl_grp *tg = pd_to_tg(pd);
1349         u64 v = *(u64 *)((void *)tg + off);
1350
1351         if (v == U64_MAX)
1352                 return 0;
1353         return __blkg_prfill_u64(sf, pd, v);
1354 }
1355
1356 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1357                                int off)
1358 {
1359         struct throtl_grp *tg = pd_to_tg(pd);
1360         unsigned int v = *(unsigned int *)((void *)tg + off);
1361
1362         if (v == UINT_MAX)
1363                 return 0;
1364         return __blkg_prfill_u64(sf, pd, v);
1365 }
1366
1367 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1368 {
1369         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1370                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1371         return 0;
1372 }
1373
1374 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1375 {
1376         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1377                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1378         return 0;
1379 }
1380
1381 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1382 {
1383         struct throtl_service_queue *sq = &tg->service_queue;
1384         struct cgroup_subsys_state *pos_css;
1385         struct blkcg_gq *blkg;
1386
1387         throtl_log(&tg->service_queue,
1388                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1389                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1390                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1391
1392         /*
1393          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1394          * considered to have rules if either the tg itself or any of its
1395          * ancestors has rules.  This identifies groups without any
1396          * restrictions in the whole hierarchy and allows them to bypass
1397          * blk-throttle.
1398          */
1399         blkg_for_each_descendant_pre(blkg, pos_css,
1400                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1401                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1402                 struct throtl_grp *parent_tg;
1403
1404                 tg_update_has_rules(this_tg);
1405                 /* ignore root/second level */
1406                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1407                     !blkg->parent->parent)
1408                         continue;
1409                 parent_tg = blkg_to_tg(blkg->parent);
1410                 /*
1411                  * make sure all children has lower idle time threshold and
1412                  * higher latency target
1413                  */
1414                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1415                                 parent_tg->idletime_threshold);
1416                 this_tg->latency_target = max(this_tg->latency_target,
1417                                 parent_tg->latency_target);
1418         }
1419
1420         /*
1421          * We're already holding queue_lock and know @tg is valid.  Let's
1422          * apply the new config directly.
1423          *
1424          * Restart the slices for both READ and WRITES. It might happen
1425          * that a group's limit are dropped suddenly and we don't want to
1426          * account recently dispatched IO with new low rate.
1427          */
1428         throtl_start_new_slice(tg, 0);
1429         throtl_start_new_slice(tg, 1);
1430
1431         if (tg->flags & THROTL_TG_PENDING) {
1432                 tg_update_disptime(tg);
1433                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1434         }
1435 }
1436
1437 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1438                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1439 {
1440         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1441         struct blkg_conf_ctx ctx;
1442         struct throtl_grp *tg;
1443         int ret;
1444         u64 v;
1445
1446         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1447         if (ret)
1448                 return ret;
1449
1450         ret = -EINVAL;
1451         if (sscanf(ctx.body, "%llu", &v) != 1)
1452                 goto out_finish;
1453         if (!v)
1454                 v = U64_MAX;
1455
1456         tg = blkg_to_tg(ctx.blkg);
1457
1458         if (is_u64)
1459                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1460         else
1461                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1462
1463         tg_conf_updated(tg, false);
1464         ret = 0;
1465 out_finish:
1466         blkg_conf_finish(&ctx);
1467         return ret ?: nbytes;
1468 }
1469
1470 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1471                                char *buf, size_t nbytes, loff_t off)
1472 {
1473         return tg_set_conf(of, buf, nbytes, off, true);
1474 }
1475
1476 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1477                                 char *buf, size_t nbytes, loff_t off)
1478 {
1479         return tg_set_conf(of, buf, nbytes, off, false);
1480 }
1481
1482 static struct cftype throtl_legacy_files[] = {
1483         {
1484                 .name = "throttle.read_bps_device",
1485                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1486                 .seq_show = tg_print_conf_u64,
1487                 .write = tg_set_conf_u64,
1488         },
1489         {
1490                 .name = "throttle.write_bps_device",
1491                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1492                 .seq_show = tg_print_conf_u64,
1493                 .write = tg_set_conf_u64,
1494         },
1495         {
1496                 .name = "throttle.read_iops_device",
1497                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1498                 .seq_show = tg_print_conf_uint,
1499                 .write = tg_set_conf_uint,
1500         },
1501         {
1502                 .name = "throttle.write_iops_device",
1503                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1504                 .seq_show = tg_print_conf_uint,
1505                 .write = tg_set_conf_uint,
1506         },
1507         {
1508                 .name = "throttle.io_service_bytes",
1509                 .private = (unsigned long)&blkcg_policy_throtl,
1510                 .seq_show = blkg_print_stat_bytes,
1511         },
1512         {
1513                 .name = "throttle.io_service_bytes_recursive",
1514                 .private = (unsigned long)&blkcg_policy_throtl,
1515                 .seq_show = blkg_print_stat_bytes_recursive,
1516         },
1517         {
1518                 .name = "throttle.io_serviced",
1519                 .private = (unsigned long)&blkcg_policy_throtl,
1520                 .seq_show = blkg_print_stat_ios,
1521         },
1522         {
1523                 .name = "throttle.io_serviced_recursive",
1524                 .private = (unsigned long)&blkcg_policy_throtl,
1525                 .seq_show = blkg_print_stat_ios_recursive,
1526         },
1527         { }     /* terminate */
1528 };
1529
1530 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1531                          int off)
1532 {
1533         struct throtl_grp *tg = pd_to_tg(pd);
1534         const char *dname = blkg_dev_name(pd->blkg);
1535         char bufs[4][21] = { "max", "max", "max", "max" };
1536         u64 bps_dft;
1537         unsigned int iops_dft;
1538         char idle_time[26] = "";
1539         char latency_time[26] = "";
1540
1541         if (!dname)
1542                 return 0;
1543
1544         if (off == LIMIT_LOW) {
1545                 bps_dft = 0;
1546                 iops_dft = 0;
1547         } else {
1548                 bps_dft = U64_MAX;
1549                 iops_dft = UINT_MAX;
1550         }
1551
1552         if (tg->bps_conf[READ][off] == bps_dft &&
1553             tg->bps_conf[WRITE][off] == bps_dft &&
1554             tg->iops_conf[READ][off] == iops_dft &&
1555             tg->iops_conf[WRITE][off] == iops_dft &&
1556             (off != LIMIT_LOW ||
1557              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1558               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1559                 return 0;
1560
1561         if (tg->bps_conf[READ][off] != U64_MAX)
1562                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1563                         tg->bps_conf[READ][off]);
1564         if (tg->bps_conf[WRITE][off] != U64_MAX)
1565                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1566                         tg->bps_conf[WRITE][off]);
1567         if (tg->iops_conf[READ][off] != UINT_MAX)
1568                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1569                         tg->iops_conf[READ][off]);
1570         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1571                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1572                         tg->iops_conf[WRITE][off]);
1573         if (off == LIMIT_LOW) {
1574                 if (tg->idletime_threshold_conf == ULONG_MAX)
1575                         strcpy(idle_time, " idle=max");
1576                 else
1577                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1578                                 tg->idletime_threshold_conf);
1579
1580                 if (tg->latency_target_conf == ULONG_MAX)
1581                         strcpy(latency_time, " latency=max");
1582                 else
1583                         snprintf(latency_time, sizeof(latency_time),
1584                                 " latency=%lu", tg->latency_target_conf);
1585         }
1586
1587         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1588                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1589                    latency_time);
1590         return 0;
1591 }
1592
1593 static int tg_print_limit(struct seq_file *sf, void *v)
1594 {
1595         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1596                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1597         return 0;
1598 }
1599
1600 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1601                           char *buf, size_t nbytes, loff_t off)
1602 {
1603         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1604         struct blkg_conf_ctx ctx;
1605         struct throtl_grp *tg;
1606         u64 v[4];
1607         unsigned long idle_time;
1608         unsigned long latency_time;
1609         int ret;
1610         int index = of_cft(of)->private;
1611
1612         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1613         if (ret)
1614                 return ret;
1615
1616         tg = blkg_to_tg(ctx.blkg);
1617
1618         v[0] = tg->bps_conf[READ][index];
1619         v[1] = tg->bps_conf[WRITE][index];
1620         v[2] = tg->iops_conf[READ][index];
1621         v[3] = tg->iops_conf[WRITE][index];
1622
1623         idle_time = tg->idletime_threshold_conf;
1624         latency_time = tg->latency_target_conf;
1625         while (true) {
1626                 char tok[27];   /* wiops=18446744073709551616 */
1627                 char *p;
1628                 u64 val = U64_MAX;
1629                 int len;
1630
1631                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1632                         break;
1633                 if (tok[0] == '\0')
1634                         break;
1635                 ctx.body += len;
1636
1637                 ret = -EINVAL;
1638                 p = tok;
1639                 strsep(&p, "=");
1640                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1641                         goto out_finish;
1642
1643                 ret = -ERANGE;
1644                 if (!val)
1645                         goto out_finish;
1646
1647                 ret = -EINVAL;
1648                 if (!strcmp(tok, "rbps"))
1649                         v[0] = val;
1650                 else if (!strcmp(tok, "wbps"))
1651                         v[1] = val;
1652                 else if (!strcmp(tok, "riops"))
1653                         v[2] = min_t(u64, val, UINT_MAX);
1654                 else if (!strcmp(tok, "wiops"))
1655                         v[3] = min_t(u64, val, UINT_MAX);
1656                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1657                         idle_time = val;
1658                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1659                         latency_time = val;
1660                 else
1661                         goto out_finish;
1662         }
1663
1664         tg->bps_conf[READ][index] = v[0];
1665         tg->bps_conf[WRITE][index] = v[1];
1666         tg->iops_conf[READ][index] = v[2];
1667         tg->iops_conf[WRITE][index] = v[3];
1668
1669         if (index == LIMIT_MAX) {
1670                 tg->bps[READ][index] = v[0];
1671                 tg->bps[WRITE][index] = v[1];
1672                 tg->iops[READ][index] = v[2];
1673                 tg->iops[WRITE][index] = v[3];
1674         }
1675         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1676                 tg->bps_conf[READ][LIMIT_MAX]);
1677         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1678                 tg->bps_conf[WRITE][LIMIT_MAX]);
1679         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1680                 tg->iops_conf[READ][LIMIT_MAX]);
1681         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1682                 tg->iops_conf[WRITE][LIMIT_MAX]);
1683         tg->idletime_threshold_conf = idle_time;
1684         tg->latency_target_conf = latency_time;
1685
1686         /* force user to configure all settings for low limit  */
1687         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1688               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1689             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1690             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1691                 tg->bps[READ][LIMIT_LOW] = 0;
1692                 tg->bps[WRITE][LIMIT_LOW] = 0;
1693                 tg->iops[READ][LIMIT_LOW] = 0;
1694                 tg->iops[WRITE][LIMIT_LOW] = 0;
1695                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1696                 tg->latency_target = DFL_LATENCY_TARGET;
1697         } else if (index == LIMIT_LOW) {
1698                 tg->idletime_threshold = tg->idletime_threshold_conf;
1699                 tg->latency_target = tg->latency_target_conf;
1700         }
1701
1702         blk_throtl_update_limit_valid(tg->td);
1703         if (tg->td->limit_valid[LIMIT_LOW]) {
1704                 if (index == LIMIT_LOW)
1705                         tg->td->limit_index = LIMIT_LOW;
1706         } else
1707                 tg->td->limit_index = LIMIT_MAX;
1708         tg_conf_updated(tg, index == LIMIT_LOW &&
1709                 tg->td->limit_valid[LIMIT_LOW]);
1710         ret = 0;
1711 out_finish:
1712         blkg_conf_finish(&ctx);
1713         return ret ?: nbytes;
1714 }
1715
1716 static struct cftype throtl_files[] = {
1717 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1718         {
1719                 .name = "low",
1720                 .flags = CFTYPE_NOT_ON_ROOT,
1721                 .seq_show = tg_print_limit,
1722                 .write = tg_set_limit,
1723                 .private = LIMIT_LOW,
1724         },
1725 #endif
1726         {
1727                 .name = "max",
1728                 .flags = CFTYPE_NOT_ON_ROOT,
1729                 .seq_show = tg_print_limit,
1730                 .write = tg_set_limit,
1731                 .private = LIMIT_MAX,
1732         },
1733         { }     /* terminate */
1734 };
1735
1736 static void throtl_shutdown_wq(struct request_queue *q)
1737 {
1738         struct throtl_data *td = q->td;
1739
1740         cancel_work_sync(&td->dispatch_work);
1741 }
1742
1743 static struct blkcg_policy blkcg_policy_throtl = {
1744         .dfl_cftypes            = throtl_files,
1745         .legacy_cftypes         = throtl_legacy_files,
1746
1747         .pd_alloc_fn            = throtl_pd_alloc,
1748         .pd_init_fn             = throtl_pd_init,
1749         .pd_online_fn           = throtl_pd_online,
1750         .pd_offline_fn          = throtl_pd_offline,
1751         .pd_free_fn             = throtl_pd_free,
1752 };
1753
1754 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1755 {
1756         unsigned long rtime = jiffies, wtime = jiffies;
1757
1758         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1759                 rtime = tg->last_low_overflow_time[READ];
1760         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1761                 wtime = tg->last_low_overflow_time[WRITE];
1762         return min(rtime, wtime);
1763 }
1764
1765 /* tg should not be an intermediate node */
1766 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1767 {
1768         struct throtl_service_queue *parent_sq;
1769         struct throtl_grp *parent = tg;
1770         unsigned long ret = __tg_last_low_overflow_time(tg);
1771
1772         while (true) {
1773                 parent_sq = parent->service_queue.parent_sq;
1774                 parent = sq_to_tg(parent_sq);
1775                 if (!parent)
1776                         break;
1777
1778                 /*
1779                  * The parent doesn't have low limit, it always reaches low
1780                  * limit. Its overflow time is useless for children
1781                  */
1782                 if (!parent->bps[READ][LIMIT_LOW] &&
1783                     !parent->iops[READ][LIMIT_LOW] &&
1784                     !parent->bps[WRITE][LIMIT_LOW] &&
1785                     !parent->iops[WRITE][LIMIT_LOW])
1786                         continue;
1787                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1788                         ret = __tg_last_low_overflow_time(parent);
1789         }
1790         return ret;
1791 }
1792
1793 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1794 {
1795         /*
1796          * cgroup is idle if:
1797          * - single idle is too long, longer than a fixed value (in case user
1798          *   configure a too big threshold) or 4 times of idletime threshold
1799          * - average think time is more than threshold
1800          * - IO latency is largely below threshold
1801          */
1802         unsigned long time;
1803         bool ret;
1804
1805         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1806         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1807               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1808               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1809               tg->avg_idletime > tg->idletime_threshold ||
1810               (tg->latency_target && tg->bio_cnt &&
1811                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1812         throtl_log(&tg->service_queue,
1813                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1814                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1815                 tg->bio_cnt, ret, tg->td->scale);
1816         return ret;
1817 }
1818
1819 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1820 {
1821         struct throtl_service_queue *sq = &tg->service_queue;
1822         bool read_limit, write_limit;
1823
1824         /*
1825          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1826          * reaches), it's ok to upgrade to next limit
1827          */
1828         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1829         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1830         if (!read_limit && !write_limit)
1831                 return true;
1832         if (read_limit && sq->nr_queued[READ] &&
1833             (!write_limit || sq->nr_queued[WRITE]))
1834                 return true;
1835         if (write_limit && sq->nr_queued[WRITE] &&
1836             (!read_limit || sq->nr_queued[READ]))
1837                 return true;
1838
1839         if (time_after_eq(jiffies,
1840                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1841             throtl_tg_is_idle(tg))
1842                 return true;
1843         return false;
1844 }
1845
1846 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1847 {
1848         while (true) {
1849                 if (throtl_tg_can_upgrade(tg))
1850                         return true;
1851                 tg = sq_to_tg(tg->service_queue.parent_sq);
1852                 if (!tg || !tg_to_blkg(tg)->parent)
1853                         return false;
1854         }
1855         return false;
1856 }
1857
1858 static bool throtl_can_upgrade(struct throtl_data *td,
1859         struct throtl_grp *this_tg)
1860 {
1861         struct cgroup_subsys_state *pos_css;
1862         struct blkcg_gq *blkg;
1863
1864         if (td->limit_index != LIMIT_LOW)
1865                 return false;
1866
1867         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1868                 return false;
1869
1870         rcu_read_lock();
1871         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1872                 struct throtl_grp *tg = blkg_to_tg(blkg);
1873
1874                 if (tg == this_tg)
1875                         continue;
1876                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1877                         continue;
1878                 if (!throtl_hierarchy_can_upgrade(tg)) {
1879                         rcu_read_unlock();
1880                         return false;
1881                 }
1882         }
1883         rcu_read_unlock();
1884         return true;
1885 }
1886
1887 static void throtl_upgrade_check(struct throtl_grp *tg)
1888 {
1889         unsigned long now = jiffies;
1890
1891         if (tg->td->limit_index != LIMIT_LOW)
1892                 return;
1893
1894         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1895                 return;
1896
1897         tg->last_check_time = now;
1898
1899         if (!time_after_eq(now,
1900              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1901                 return;
1902
1903         if (throtl_can_upgrade(tg->td, NULL))
1904                 throtl_upgrade_state(tg->td);
1905 }
1906
1907 static void throtl_upgrade_state(struct throtl_data *td)
1908 {
1909         struct cgroup_subsys_state *pos_css;
1910         struct blkcg_gq *blkg;
1911
1912         throtl_log(&td->service_queue, "upgrade to max");
1913         td->limit_index = LIMIT_MAX;
1914         td->low_upgrade_time = jiffies;
1915         td->scale = 0;
1916         rcu_read_lock();
1917         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1918                 struct throtl_grp *tg = blkg_to_tg(blkg);
1919                 struct throtl_service_queue *sq = &tg->service_queue;
1920
1921                 tg->disptime = jiffies - 1;
1922                 throtl_select_dispatch(sq);
1923                 throtl_schedule_next_dispatch(sq, true);
1924         }
1925         rcu_read_unlock();
1926         throtl_select_dispatch(&td->service_queue);
1927         throtl_schedule_next_dispatch(&td->service_queue, true);
1928         queue_work(kthrotld_workqueue, &td->dispatch_work);
1929 }
1930
1931 static void throtl_downgrade_state(struct throtl_data *td, int new)
1932 {
1933         td->scale /= 2;
1934
1935         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1936         if (td->scale) {
1937                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1938                 return;
1939         }
1940
1941         td->limit_index = new;
1942         td->low_downgrade_time = jiffies;
1943 }
1944
1945 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1946 {
1947         struct throtl_data *td = tg->td;
1948         unsigned long now = jiffies;
1949
1950         /*
1951          * If cgroup is below low limit, consider downgrade and throttle other
1952          * cgroups
1953          */
1954         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1955             time_after_eq(now, tg_last_low_overflow_time(tg) +
1956                                         td->throtl_slice) &&
1957             (!throtl_tg_is_idle(tg) ||
1958              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1959                 return true;
1960         return false;
1961 }
1962
1963 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1964 {
1965         while (true) {
1966                 if (!throtl_tg_can_downgrade(tg))
1967                         return false;
1968                 tg = sq_to_tg(tg->service_queue.parent_sq);
1969                 if (!tg || !tg_to_blkg(tg)->parent)
1970                         break;
1971         }
1972         return true;
1973 }
1974
1975 static void throtl_downgrade_check(struct throtl_grp *tg)
1976 {
1977         uint64_t bps;
1978         unsigned int iops;
1979         unsigned long elapsed_time;
1980         unsigned long now = jiffies;
1981
1982         if (tg->td->limit_index != LIMIT_MAX ||
1983             !tg->td->limit_valid[LIMIT_LOW])
1984                 return;
1985         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1986                 return;
1987         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1988                 return;
1989
1990         elapsed_time = now - tg->last_check_time;
1991         tg->last_check_time = now;
1992
1993         if (time_before(now, tg_last_low_overflow_time(tg) +
1994                         tg->td->throtl_slice))
1995                 return;
1996
1997         if (tg->bps[READ][LIMIT_LOW]) {
1998                 bps = tg->last_bytes_disp[READ] * HZ;
1999                 do_div(bps, elapsed_time);
2000                 if (bps >= tg->bps[READ][LIMIT_LOW])
2001                         tg->last_low_overflow_time[READ] = now;
2002         }
2003
2004         if (tg->bps[WRITE][LIMIT_LOW]) {
2005                 bps = tg->last_bytes_disp[WRITE] * HZ;
2006                 do_div(bps, elapsed_time);
2007                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2008                         tg->last_low_overflow_time[WRITE] = now;
2009         }
2010
2011         if (tg->iops[READ][LIMIT_LOW]) {
2012                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2013                 if (iops >= tg->iops[READ][LIMIT_LOW])
2014                         tg->last_low_overflow_time[READ] = now;
2015         }
2016
2017         if (tg->iops[WRITE][LIMIT_LOW]) {
2018                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2019                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2020                         tg->last_low_overflow_time[WRITE] = now;
2021         }
2022
2023         /*
2024          * If cgroup is below low limit, consider downgrade and throttle other
2025          * cgroups
2026          */
2027         if (throtl_hierarchy_can_downgrade(tg))
2028                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2029
2030         tg->last_bytes_disp[READ] = 0;
2031         tg->last_bytes_disp[WRITE] = 0;
2032         tg->last_io_disp[READ] = 0;
2033         tg->last_io_disp[WRITE] = 0;
2034 }
2035
2036 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2037 {
2038         unsigned long now = ktime_get_ns() >> 10;
2039         unsigned long last_finish_time = tg->last_finish_time;
2040
2041         if (now <= last_finish_time || last_finish_time == 0 ||
2042             last_finish_time == tg->checked_last_finish_time)
2043                 return;
2044
2045         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2046         tg->checked_last_finish_time = last_finish_time;
2047 }
2048
2049 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2050 static void throtl_update_latency_buckets(struct throtl_data *td)
2051 {
2052         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2053         int i, cpu, rw;
2054         unsigned long last_latency[2] = { 0 };
2055         unsigned long latency[2];
2056
2057         if (!blk_queue_nonrot(td->queue))
2058                 return;
2059         if (time_before(jiffies, td->last_calculate_time + HZ))
2060                 return;
2061         td->last_calculate_time = jiffies;
2062
2063         memset(avg_latency, 0, sizeof(avg_latency));
2064         for (rw = READ; rw <= WRITE; rw++) {
2065                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2066                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2067
2068                         for_each_possible_cpu(cpu) {
2069                                 struct latency_bucket *bucket;
2070
2071                                 /* this isn't race free, but ok in practice */
2072                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2073                                         cpu);
2074                                 tmp->total_latency += bucket[i].total_latency;
2075                                 tmp->samples += bucket[i].samples;
2076                                 bucket[i].total_latency = 0;
2077                                 bucket[i].samples = 0;
2078                         }
2079
2080                         if (tmp->samples >= 32) {
2081                                 int samples = tmp->samples;
2082
2083                                 latency[rw] = tmp->total_latency;
2084
2085                                 tmp->total_latency = 0;
2086                                 tmp->samples = 0;
2087                                 latency[rw] /= samples;
2088                                 if (latency[rw] == 0)
2089                                         continue;
2090                                 avg_latency[rw][i].latency = latency[rw];
2091                         }
2092                 }
2093         }
2094
2095         for (rw = READ; rw <= WRITE; rw++) {
2096                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2097                         if (!avg_latency[rw][i].latency) {
2098                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2099                                         td->avg_buckets[rw][i].latency =
2100                                                 last_latency[rw];
2101                                 continue;
2102                         }
2103
2104                         if (!td->avg_buckets[rw][i].valid)
2105                                 latency[rw] = avg_latency[rw][i].latency;
2106                         else
2107                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2108                                         avg_latency[rw][i].latency) >> 3;
2109
2110                         td->avg_buckets[rw][i].latency = max(latency[rw],
2111                                 last_latency[rw]);
2112                         td->avg_buckets[rw][i].valid = true;
2113                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2114                 }
2115         }
2116
2117         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2118                 throtl_log(&td->service_queue,
2119                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2120                         "write latency=%ld, write valid=%d", i,
2121                         td->avg_buckets[READ][i].latency,
2122                         td->avg_buckets[READ][i].valid,
2123                         td->avg_buckets[WRITE][i].latency,
2124                         td->avg_buckets[WRITE][i].valid);
2125 }
2126 #else
2127 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2128 {
2129 }
2130 #endif
2131
2132 static void blk_throtl_assoc_bio(struct throtl_grp *tg, struct bio *bio)
2133 {
2134 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2135         if (bio->bi_css) {
2136                 if (bio->bi_cg_private)
2137                         blkg_put(tg_to_blkg(bio->bi_cg_private));
2138                 bio->bi_cg_private = tg;
2139                 blkg_get(tg_to_blkg(tg));
2140         }
2141         bio_issue_init(&bio->bi_issue, bio_sectors(bio));
2142 #endif
2143 }
2144
2145 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2146                     struct bio *bio)
2147 {
2148         struct throtl_qnode *qn = NULL;
2149         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2150         struct throtl_service_queue *sq;
2151         bool rw = bio_data_dir(bio);
2152         bool throttled = false;
2153         struct throtl_data *td = tg->td;
2154
2155         WARN_ON_ONCE(!rcu_read_lock_held());
2156
2157         /* see throtl_charge_bio() */
2158         if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2159                 goto out;
2160
2161         spin_lock_irq(q->queue_lock);
2162
2163         throtl_update_latency_buckets(td);
2164
2165         if (unlikely(blk_queue_bypass(q)))
2166                 goto out_unlock;
2167
2168         blk_throtl_assoc_bio(tg, bio);
2169         blk_throtl_update_idletime(tg);
2170
2171         sq = &tg->service_queue;
2172
2173 again:
2174         while (true) {
2175                 if (tg->last_low_overflow_time[rw] == 0)
2176                         tg->last_low_overflow_time[rw] = jiffies;
2177                 throtl_downgrade_check(tg);
2178                 throtl_upgrade_check(tg);
2179                 /* throtl is FIFO - if bios are already queued, should queue */
2180                 if (sq->nr_queued[rw])
2181                         break;
2182
2183                 /* if above limits, break to queue */
2184                 if (!tg_may_dispatch(tg, bio, NULL)) {
2185                         tg->last_low_overflow_time[rw] = jiffies;
2186                         if (throtl_can_upgrade(td, tg)) {
2187                                 throtl_upgrade_state(td);
2188                                 goto again;
2189                         }
2190                         break;
2191                 }
2192
2193                 /* within limits, let's charge and dispatch directly */
2194                 throtl_charge_bio(tg, bio);
2195
2196                 /*
2197                  * We need to trim slice even when bios are not being queued
2198                  * otherwise it might happen that a bio is not queued for
2199                  * a long time and slice keeps on extending and trim is not
2200                  * called for a long time. Now if limits are reduced suddenly
2201                  * we take into account all the IO dispatched so far at new
2202                  * low rate and * newly queued IO gets a really long dispatch
2203                  * time.
2204                  *
2205                  * So keep on trimming slice even if bio is not queued.
2206                  */
2207                 throtl_trim_slice(tg, rw);
2208
2209                 /*
2210                  * @bio passed through this layer without being throttled.
2211                  * Climb up the ladder.  If we''re already at the top, it
2212                  * can be executed directly.
2213                  */
2214                 qn = &tg->qnode_on_parent[rw];
2215                 sq = sq->parent_sq;
2216                 tg = sq_to_tg(sq);
2217                 if (!tg)
2218                         goto out_unlock;
2219         }
2220
2221         /* out-of-limit, queue to @tg */
2222         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2223                    rw == READ ? 'R' : 'W',
2224                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2225                    tg_bps_limit(tg, rw),
2226                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2227                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2228
2229         tg->last_low_overflow_time[rw] = jiffies;
2230
2231         td->nr_queued[rw]++;
2232         throtl_add_bio_tg(bio, qn, tg);
2233         throttled = true;
2234
2235         /*
2236          * Update @tg's dispatch time and force schedule dispatch if @tg
2237          * was empty before @bio.  The forced scheduling isn't likely to
2238          * cause undue delay as @bio is likely to be dispatched directly if
2239          * its @tg's disptime is not in the future.
2240          */
2241         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2242                 tg_update_disptime(tg);
2243                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2244         }
2245
2246 out_unlock:
2247         spin_unlock_irq(q->queue_lock);
2248 out:
2249         bio_set_flag(bio, BIO_THROTTLED);
2250
2251 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2252         if (throttled || !td->track_bio_latency)
2253                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2254 #endif
2255         return throttled;
2256 }
2257
2258 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2259 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2260         int op, unsigned long time)
2261 {
2262         struct latency_bucket *latency;
2263         int index;
2264
2265         if (!td || td->limit_index != LIMIT_LOW ||
2266             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2267             !blk_queue_nonrot(td->queue))
2268                 return;
2269
2270         index = request_bucket_index(size);
2271
2272         latency = get_cpu_ptr(td->latency_buckets[op]);
2273         latency[index].total_latency += time;
2274         latency[index].samples++;
2275         put_cpu_ptr(td->latency_buckets[op]);
2276 }
2277
2278 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2279 {
2280         struct request_queue *q = rq->q;
2281         struct throtl_data *td = q->td;
2282
2283         throtl_track_latency(td, rq->throtl_size, req_op(rq), time_ns >> 10);
2284 }
2285
2286 void blk_throtl_bio_endio(struct bio *bio)
2287 {
2288         struct throtl_grp *tg;
2289         u64 finish_time_ns;
2290         unsigned long finish_time;
2291         unsigned long start_time;
2292         unsigned long lat;
2293         int rw = bio_data_dir(bio);
2294
2295         tg = bio->bi_cg_private;
2296         if (!tg)
2297                 return;
2298         bio->bi_cg_private = NULL;
2299
2300         finish_time_ns = ktime_get_ns();
2301         tg->last_finish_time = finish_time_ns >> 10;
2302
2303         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2304         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2305         if (!start_time || finish_time <= start_time) {
2306                 blkg_put(tg_to_blkg(tg));
2307                 return;
2308         }
2309
2310         lat = finish_time - start_time;
2311         /* this is only for bio based driver */
2312         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2313                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2314                                      bio_op(bio), lat);
2315
2316         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2317                 int bucket;
2318                 unsigned int threshold;
2319
2320                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2321                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2322                         tg->latency_target;
2323                 if (lat > threshold)
2324                         tg->bad_bio_cnt++;
2325                 /*
2326                  * Not race free, could get wrong count, which means cgroups
2327                  * will be throttled
2328                  */
2329                 tg->bio_cnt++;
2330         }
2331
2332         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2333                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2334                 tg->bio_cnt /= 2;
2335                 tg->bad_bio_cnt /= 2;
2336         }
2337
2338         blkg_put(tg_to_blkg(tg));
2339 }
2340 #endif
2341
2342 /*
2343  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2344  * return, @parent_sq is guaranteed to not have any active children tg's
2345  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2346  */
2347 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2348 {
2349         struct throtl_grp *tg;
2350
2351         while ((tg = throtl_rb_first(parent_sq))) {
2352                 struct throtl_service_queue *sq = &tg->service_queue;
2353                 struct bio *bio;
2354
2355                 throtl_dequeue_tg(tg);
2356
2357                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2358                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2359                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2360                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2361         }
2362 }
2363
2364 /**
2365  * blk_throtl_drain - drain throttled bios
2366  * @q: request_queue to drain throttled bios for
2367  *
2368  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2369  */
2370 void blk_throtl_drain(struct request_queue *q)
2371         __releases(q->queue_lock) __acquires(q->queue_lock)
2372 {
2373         struct throtl_data *td = q->td;
2374         struct blkcg_gq *blkg;
2375         struct cgroup_subsys_state *pos_css;
2376         struct bio *bio;
2377         int rw;
2378
2379         queue_lockdep_assert_held(q);
2380         rcu_read_lock();
2381
2382         /*
2383          * Drain each tg while doing post-order walk on the blkg tree, so
2384          * that all bios are propagated to td->service_queue.  It'd be
2385          * better to walk service_queue tree directly but blkg walk is
2386          * easier.
2387          */
2388         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2389                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2390
2391         /* finally, transfer bios from top-level tg's into the td */
2392         tg_drain_bios(&td->service_queue);
2393
2394         rcu_read_unlock();
2395         spin_unlock_irq(q->queue_lock);
2396
2397         /* all bios now should be in td->service_queue, issue them */
2398         for (rw = READ; rw <= WRITE; rw++)
2399                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2400                                                 NULL)))
2401                         generic_make_request(bio);
2402
2403         spin_lock_irq(q->queue_lock);
2404 }
2405
2406 int blk_throtl_init(struct request_queue *q)
2407 {
2408         struct throtl_data *td;
2409         int ret;
2410
2411         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2412         if (!td)
2413                 return -ENOMEM;
2414         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2415                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2416         if (!td->latency_buckets[READ]) {
2417                 kfree(td);
2418                 return -ENOMEM;
2419         }
2420         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2421                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2422         if (!td->latency_buckets[WRITE]) {
2423                 free_percpu(td->latency_buckets[READ]);
2424                 kfree(td);
2425                 return -ENOMEM;
2426         }
2427
2428         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2429         throtl_service_queue_init(&td->service_queue);
2430
2431         q->td = td;
2432         td->queue = q;
2433
2434         td->limit_valid[LIMIT_MAX] = true;
2435         td->limit_index = LIMIT_MAX;
2436         td->low_upgrade_time = jiffies;
2437         td->low_downgrade_time = jiffies;
2438
2439         /* activate policy */
2440         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2441         if (ret) {
2442                 free_percpu(td->latency_buckets[READ]);
2443                 free_percpu(td->latency_buckets[WRITE]);
2444                 kfree(td);
2445         }
2446         return ret;
2447 }
2448
2449 void blk_throtl_exit(struct request_queue *q)
2450 {
2451         BUG_ON(!q->td);
2452         throtl_shutdown_wq(q);
2453         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2454         free_percpu(q->td->latency_buckets[READ]);
2455         free_percpu(q->td->latency_buckets[WRITE]);
2456         kfree(q->td);
2457 }
2458
2459 void blk_throtl_register_queue(struct request_queue *q)
2460 {
2461         struct throtl_data *td;
2462         int i;
2463
2464         td = q->td;
2465         BUG_ON(!td);
2466
2467         if (blk_queue_nonrot(q)) {
2468                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2469                 td->filtered_latency = LATENCY_FILTERED_SSD;
2470         } else {
2471                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2472                 td->filtered_latency = LATENCY_FILTERED_HD;
2473                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2474                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2475                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2476                 }
2477         }
2478 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2479         /* if no low limit, use previous default */
2480         td->throtl_slice = DFL_THROTL_SLICE_HD;
2481 #endif
2482
2483         td->track_bio_latency = !queue_is_rq_based(q);
2484         if (!td->track_bio_latency)
2485                 blk_stat_enable_accounting(q);
2486 }
2487
2488 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2489 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2490 {
2491         if (!q->td)
2492                 return -EINVAL;
2493         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2494 }
2495
2496 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2497         const char *page, size_t count)
2498 {
2499         unsigned long v;
2500         unsigned long t;
2501
2502         if (!q->td)
2503                 return -EINVAL;
2504         if (kstrtoul(page, 10, &v))
2505                 return -EINVAL;
2506         t = msecs_to_jiffies(v);
2507         if (t == 0 || t > MAX_THROTL_SLICE)
2508                 return -EINVAL;
2509         q->td->throtl_slice = t;
2510         return count;
2511 }
2512 #endif
2513
2514 static int __init throtl_init(void)
2515 {
2516         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2517         if (!kthrotld_workqueue)
2518                 panic("Failed to create kthrotld\n");
2519
2520         return blkcg_policy_register(&blkcg_policy_throtl);
2521 }
2522
2523 module_init(throtl_init);