block/blk-throttle.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Interface for controlling IO bandwidth on a request queue
   4  *
   5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blktrace_api.h>
  13 #include <linux/blk-cgroup.h>
  14 #include "blk.h"
  15 #include "blk-cgroup-rwstat.h"
  16
  17 /* Max dispatch from a group in 1 round */
  18 #define THROTL_GRP_QUANTUM 8
  19
  20 /* Total max dispatch from all groups in one round */
  21 #define THROTL_QUANTUM 32
  22
  23 /* Throttling is performed over a slice and after that slice is renewed */
  24 #define DFL_THROTL_SLICE_HD (HZ / 10)
  25 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  26 #define MAX_THROTL_SLICE (HZ)
  27 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  28 #define MIN_THROTL_BPS (320 * 1024)
  29 #define MIN_THROTL_IOPS (10)
  30 #define DFL_LATENCY_TARGET (-1L)
  31 #define DFL_IDLE_THRESHOLD (0)
  32 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  33 #define LATENCY_FILTERED_SSD (0)
  34 /*
  35  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  36  * help determine if its IO is impacted by others, hence we ignore the IO
  37  */
  38 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  39
  40 static struct blkcg_policy blkcg_policy_throtl;
  41
  42 /* A workqueue to queue throttle related work */
  43 static struct workqueue_struct *kthrotld_workqueue;
  44
  45 /*
  46  * To implement hierarchical throttling, throtl_grps form a tree and bios
  47  * are dispatched upwards level by level until they reach the top and get
  48  * issued.  When dispatching bios from the children and local group at each
  49  * level, if the bios are dispatched into a single bio_list, there's a risk
  50  * of a local or child group which can queue many bios at once filling up
  51  * the list starving others.
  52  *
  53  * To avoid such starvation, dispatched bios are queued separately
  54  * according to where they came from.  When they are again dispatched to
  55  * the parent, they're popped in round-robin order so that no single source
  56  * hogs the dispatch window.
  57  *
  58  * throtl_qnode is used to keep the queued bios separated by their sources.
  59  * Bios are queued to throtl_qnode which in turn is queued to
  60  * throtl_service_queue and then dispatched in round-robin order.
  61  *
  62  * It's also used to track the reference counts on blkg's.  A qnode always
  63  * belongs to a throtl_grp and gets queued on itself or the parent, so
  64  * incrementing the reference of the associated throtl_grp when a qnode is
  65  * queued and decrementing when dequeued is enough to keep the whole blkg
  66  * tree pinned while bios are in flight.
  67  */
  68 struct throtl_qnode {
  69         struct list_head        node;           /* service_queue->queued[] */
  70         struct bio_list         bios;           /* queued bios */
  71         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  72 };
  73
  74 struct throtl_service_queue {
  75         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  76
  77         /*
  78          * Bios queued directly to this service_queue or dispatched from
  79          * children throtl_grp's.
  80          */
  81         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  82         unsigned int            nr_queued[2];   /* number of queued bios */
  83
  84         /*
  85          * RB tree of active children throtl_grp's, which are sorted by
  86          * their ->disptime.
  87          */
  88         struct rb_root_cached   pending_tree;   /* RB tree of active tgs */
  89         unsigned int            nr_pending;     /* # queued in the tree */
  90         unsigned long           first_pending_disptime; /* disptime of the first tg */
  91         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  92 };
  93
  94 enum tg_state_flags {
  95         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  96         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  97 };
  98
  99 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
 100
 101 enum {
 102         LIMIT_LOW,
 103         LIMIT_MAX,
 104         LIMIT_CNT,
 105 };
 106
 107 struct throtl_grp {
 108         /* must be the first member */
 109         struct blkg_policy_data pd;
 110
 111         /* active throtl group service_queue member */
 112         struct rb_node rb_node;
 113
 114         /* throtl_data this group belongs to */
 115         struct throtl_data *td;
 116
 117         /* this group's service queue */
 118         struct throtl_service_queue service_queue;
 119
 120         /*
 121          * qnode_on_self is used when bios are directly queued to this
 122          * throtl_grp so that local bios compete fairly with bios
 123          * dispatched from children.  qnode_on_parent is used when bios are
 124          * dispatched from this throtl_grp into its parent and will compete
 125          * with the sibling qnode_on_parents and the parent's
 126          * qnode_on_self.
 127          */
 128         struct throtl_qnode qnode_on_self[2];
 129         struct throtl_qnode qnode_on_parent[2];
 130
 131         /*
 132          * Dispatch time in jiffies. This is the estimated time when group
 133          * will unthrottle and is ready to dispatch more bio. It is used as
 134          * key to sort active groups in service tree.
 135          */
 136         unsigned long disptime;
 137
 138         unsigned int flags;
 139
 140         /* are there any throtl rules between this group and td? */
 141         bool has_rules[2];
 142
 143         /* internally used bytes per second rate limits */
 144         uint64_t bps[2][LIMIT_CNT];
 145         /* user configured bps limits */
 146         uint64_t bps_conf[2][LIMIT_CNT];
 147
 148         /* internally used IOPS limits */
 149         unsigned int iops[2][LIMIT_CNT];
 150         /* user configured IOPS limits */
 151         unsigned int iops_conf[2][LIMIT_CNT];
 152
 153         /* Number of bytes dispatched 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         atomic_t io_split_cnt[2];
 182         atomic_t last_io_split_cnt[2];
 183
 184         struct blkg_rwstat stat_bytes;
 185         struct blkg_rwstat stat_ios;
 186 };
 187
 188 /* We measure latency for request size from <= 4k to >= 1M */
 189 #define LATENCY_BUCKET_SIZE 9
 190
 191 struct latency_bucket {
 192         unsigned long total_latency; /* ns / 1024 */
 193         int samples;
 194 };
 195
 196 struct avg_latency_bucket {
 197         unsigned long latency; /* ns / 1024 */
 198         bool valid;
 199 };
 200
 201 struct throtl_data
 202 {
 203         /* service tree for active throtl groups */
 204         struct throtl_service_queue service_queue;
 205
 206         struct request_queue *queue;
 207
 208         /* Total Number of queued bios on READ and WRITE lists */
 209         unsigned int nr_queued[2];
 210
 211         unsigned int throtl_slice;
 212
 213         /* Work for dispatching throttled bios */
 214         struct work_struct dispatch_work;
 215         unsigned int limit_index;
 216         bool limit_valid[LIMIT_CNT];
 217
 218         unsigned long low_upgrade_time;
 219         unsigned long low_downgrade_time;
 220
 221         unsigned int scale;
 222
 223         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
 224         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
 225         struct latency_bucket __percpu *latency_buckets[2];
 226         unsigned long last_calculate_time;
 227         unsigned long filtered_latency;
 228
 229         bool track_bio_latency;
 230 };
 231
 232 static void throtl_pending_timer_fn(struct timer_list *t);
 233
 234 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 235 {
 236         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 237 }
 238
 239 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 240 {
 241         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 242 }
 243
 244 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 245 {
 246         return pd_to_blkg(&tg->pd);
 247 }
 248
 249 /**
 250  * sq_to_tg - return the throl_grp the specified service queue belongs to
 251  * @sq: the throtl_service_queue of interest
 252  *
 253  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 254  * embedded in throtl_data, %NULL is returned.
 255  */
 256 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 257 {
 258         if (sq && sq->parent_sq)
 259                 return container_of(sq, struct throtl_grp, service_queue);
 260         else
 261                 return NULL;
 262 }
 263
 264 /**
 265  * sq_to_td - return throtl_data the specified service queue belongs to
 266  * @sq: the throtl_service_queue of interest
 267  *
 268  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 269  * Determine the associated throtl_data accordingly and return it.
 270  */
 271 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 272 {
 273         struct throtl_grp *tg = sq_to_tg(sq);
 274
 275         if (tg)
 276                 return tg->td;
 277         else
 278                 return container_of(sq, struct throtl_data, service_queue);
 279 }
 280
 281 /*
 282  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 283  * make the IO dispatch more smooth.
 284  * Scale up: linearly scale up according to lapsed time since upgrade. For
 285  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 286  *           limit hits .max limit
 287  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 288  */
 289 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 290 {
 291         /* arbitrary value to avoid too big scale */
 292         if (td->scale < 4096 && time_after_eq(jiffies,
 293             td->low_upgrade_time + td->scale * td->throtl_slice))
 294                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 295
 296         return low + (low >> 1) * td->scale;
 297 }
 298
 299 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 300 {
 301         struct blkcg_gq *blkg = tg_to_blkg(tg);
 302         struct throtl_data *td;
 303         uint64_t ret;
 304
 305         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 306                 return U64_MAX;
 307
 308         td = tg->td;
 309         ret = tg->bps[rw][td->limit_index];
 310         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 311                 /* intermediate node or iops isn't 0 */
 312                 if (!list_empty(&blkg->blkcg->css.children) ||
 313                     tg->iops[rw][td->limit_index])
 314                         return U64_MAX;
 315                 else
 316                         return MIN_THROTL_BPS;
 317         }
 318
 319         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 320             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 321                 uint64_t adjusted;
 322
 323                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 324                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 325         }
 326         return ret;
 327 }
 328
 329 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 330 {
 331         struct blkcg_gq *blkg = tg_to_blkg(tg);
 332         struct throtl_data *td;
 333         unsigned int ret;
 334
 335         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 336                 return UINT_MAX;
 337
 338         td = tg->td;
 339         ret = tg->iops[rw][td->limit_index];
 340         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 341                 /* intermediate node or bps isn't 0 */
 342                 if (!list_empty(&blkg->blkcg->css.children) ||
 343                     tg->bps[rw][td->limit_index])
 344                         return UINT_MAX;
 345                 else
 346                         return MIN_THROTL_IOPS;
 347         }
 348
 349         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 350             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 351                 uint64_t adjusted;
 352
 353                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 354                 if (adjusted > UINT_MAX)
 355                         adjusted = UINT_MAX;
 356                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 357         }
 358         return ret;
 359 }
 360
 361 #define request_bucket_index(sectors) \
 362         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 363
 364 /**
 365  * throtl_log - log debug message via blktrace
 366  * @sq: the service_queue being reported
 367  * @fmt: printf format string
 368  * @args: printf args
 369  *
 370  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 371  * throtl_grp; otherwise, just "throtl".
 372  */
 373 #define throtl_log(sq, fmt, args...)    do {                            \
 374         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 375         struct throtl_data *__td = sq_to_td((sq));                      \
 376                                                                         \
 377         (void)__td;                                                     \
 378         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 379                 break;                                                  \
 380         if ((__tg)) {                                                   \
 381                 blk_add_cgroup_trace_msg(__td->queue,                   \
 382                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 383         } else {                                                        \
 384                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 385         }                                                               \
 386 } while (0)
 387
 388 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 389 {
 390         /* assume it's one sector */
 391         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 392                 return 512;
 393         return bio->bi_iter.bi_size;
 394 }
 395
 396 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 397 {
 398         INIT_LIST_HEAD(&qn->node);
 399         bio_list_init(&qn->bios);
 400         qn->tg = tg;
 401 }
 402
 403 /**
 404  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 405  * @bio: bio being added
 406  * @qn: qnode to add bio to
 407  * @queued: the service_queue->queued[] list @qn belongs to
 408  *
 409  * Add @bio to @qn and put @qn on @queued if it's not already on.
 410  * @qn->tg's reference count is bumped when @qn is activated.  See the
 411  * comment on top of throtl_qnode definition for details.
 412  */
 413 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 414                                  struct list_head *queued)
 415 {
 416         bio_list_add(&qn->bios, bio);
 417         if (list_empty(&qn->node)) {
 418                 list_add_tail(&qn->node, queued);
 419                 blkg_get(tg_to_blkg(qn->tg));
 420         }
 421 }
 422
 423 /**
 424  * throtl_peek_queued - peek the first bio on a qnode list
 425  * @queued: the qnode list to peek
 426  */
 427 static struct bio *throtl_peek_queued(struct list_head *queued)
 428 {
 429         struct throtl_qnode *qn;
 430         struct bio *bio;
 431
 432         if (list_empty(queued))
 433                 return NULL;
 434
 435         qn = list_first_entry(queued, struct throtl_qnode, node);
 436         bio = bio_list_peek(&qn->bios);
 437         WARN_ON_ONCE(!bio);
 438         return bio;
 439 }
 440
 441 /**
 442  * throtl_pop_queued - pop the first bio form a qnode list
 443  * @queued: the qnode list to pop a bio from
 444  * @tg_to_put: optional out argument for throtl_grp to put
 445  *
 446  * Pop the first bio from the qnode list @queued.  After popping, the first
 447  * qnode is removed from @queued if empty or moved to the end of @queued so
 448  * that the popping order is round-robin.
 449  *
 450  * When the first qnode is removed, its associated throtl_grp should be put
 451  * too.  If @tg_to_put is NULL, this function automatically puts it;
 452  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 453  * responsible for putting it.
 454  */
 455 static struct bio *throtl_pop_queued(struct list_head *queued,
 456                                      struct throtl_grp **tg_to_put)
 457 {
 458         struct throtl_qnode *qn;
 459         struct bio *bio;
 460
 461         if (list_empty(queued))
 462                 return NULL;
 463
 464         qn = list_first_entry(queued, struct throtl_qnode, node);
 465         bio = bio_list_pop(&qn->bios);
 466         WARN_ON_ONCE(!bio);
 467
 468         if (bio_list_empty(&qn->bios)) {
 469                 list_del_init(&qn->node);
 470                 if (tg_to_put)
 471                         *tg_to_put = qn->tg;
 472                 else
 473                         blkg_put(tg_to_blkg(qn->tg));
 474         } else {
 475                 list_move_tail(&qn->node, queued);
 476         }
 477
 478         return bio;
 479 }
 480
 481 /* init a service_queue, assumes the caller zeroed it */
 482 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 483 {
 484         INIT_LIST_HEAD(&sq->queued[0]);
 485         INIT_LIST_HEAD(&sq->queued[1]);
 486         sq->pending_tree = RB_ROOT_CACHED;
 487         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 488 }
 489
 490 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
 491                                                 struct request_queue *q,
 492                                                 struct blkcg *blkcg)
 493 {
 494         struct throtl_grp *tg;
 495         int rw;
 496
 497         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
 498         if (!tg)
 499                 return NULL;
 500
 501         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 502                 goto err_free_tg;
 503
 504         if (blkg_rwstat_init(&tg->stat_ios, gfp))
 505                 goto err_exit_stat_bytes;
 506
 507         throtl_service_queue_init(&tg->service_queue);
 508
 509         for (rw = READ; rw <= WRITE; rw++) {
 510                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 511                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 512         }
 513
 514         RB_CLEAR_NODE(&tg->rb_node);
 515         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 516         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 517         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 518         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 519         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 520         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 521         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 522         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 523         /* LIMIT_LOW will have default value 0 */
 524
 525         tg->latency_target = DFL_LATENCY_TARGET;
 526         tg->latency_target_conf = DFL_LATENCY_TARGET;
 527         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 528         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 529
 530         return &tg->pd;
 531
 532 err_exit_stat_bytes:
 533         blkg_rwstat_exit(&tg->stat_bytes);
 534 err_free_tg:
 535         kfree(tg);
 536         return NULL;
 537 }
 538
 539 static void throtl_pd_init(struct blkg_policy_data *pd)
 540 {
 541         struct throtl_grp *tg = pd_to_tg(pd);
 542         struct blkcg_gq *blkg = tg_to_blkg(tg);
 543         struct throtl_data *td = blkg->q->td;
 544         struct throtl_service_queue *sq = &tg->service_queue;
 545
 546         /*
 547          * If on the default hierarchy, we switch to properly hierarchical
 548          * behavior where limits on a given throtl_grp are applied to the
 549          * whole subtree rather than just the group itself.  e.g. If 16M
 550          * read_bps limit is set on the root group, the whole system can't
 551          * exceed 16M for the device.
 552          *
 553          * If not on the default hierarchy, the broken flat hierarchy
 554          * behavior is retained where all throtl_grps are treated as if
 555          * they're all separate root groups right below throtl_data.
 556          * Limits of a group don't interact with limits of other groups
 557          * regardless of the position of the group in the hierarchy.
 558          */
 559         sq->parent_sq = &td->service_queue;
 560         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 561                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 562         tg->td = td;
 563 }
 564
 565 /*
 566  * Set has_rules[] if @tg or any of its parents have limits configured.
 567  * This doesn't require walking up to the top of the hierarchy as the
 568  * parent's has_rules[] is guaranteed to be correct.
 569  */
 570 static void tg_update_has_rules(struct throtl_grp *tg)
 571 {
 572         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 573         struct throtl_data *td = tg->td;
 574         int rw;
 575
 576         for (rw = READ; rw <= WRITE; rw++)
 577                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 578                         (td->limit_valid[td->limit_index] &&
 579                          (tg_bps_limit(tg, rw) != U64_MAX ||
 580                           tg_iops_limit(tg, rw) != UINT_MAX));
 581 }
 582
 583 static void throtl_pd_online(struct blkg_policy_data *pd)
 584 {
 585         struct throtl_grp *tg = pd_to_tg(pd);
 586         /*
 587          * We don't want new groups to escape the limits of its ancestors.
 588          * Update has_rules[] after a new group is brought online.
 589          */
 590         tg_update_has_rules(tg);
 591 }
 592
 593 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 594 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 595 {
 596         struct cgroup_subsys_state *pos_css;
 597         struct blkcg_gq *blkg;
 598         bool low_valid = false;
 599
 600         rcu_read_lock();
 601         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 602                 struct throtl_grp *tg = blkg_to_tg(blkg);
 603
 604                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 605                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 606                         low_valid = true;
 607                         break;
 608                 }
 609         }
 610         rcu_read_unlock();
 611
 612         td->limit_valid[LIMIT_LOW] = low_valid;
 613 }
 614 #else
 615 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
 616 {
 617 }
 618 #endif
 619
 620 static void throtl_upgrade_state(struct throtl_data *td);
 621 static void throtl_pd_offline(struct blkg_policy_data *pd)
 622 {
 623         struct throtl_grp *tg = pd_to_tg(pd);
 624
 625         tg->bps[READ][LIMIT_LOW] = 0;
 626         tg->bps[WRITE][LIMIT_LOW] = 0;
 627         tg->iops[READ][LIMIT_LOW] = 0;
 628         tg->iops[WRITE][LIMIT_LOW] = 0;
 629
 630         blk_throtl_update_limit_valid(tg->td);
 631
 632         if (!tg->td->limit_valid[tg->td->limit_index])
 633                 throtl_upgrade_state(tg->td);
 634 }
 635
 636 static void throtl_pd_free(struct blkg_policy_data *pd)
 637 {
 638         struct throtl_grp *tg = pd_to_tg(pd);
 639
 640         del_timer_sync(&tg->service_queue.pending_timer);
 641         blkg_rwstat_exit(&tg->stat_bytes);
 642         blkg_rwstat_exit(&tg->stat_ios);
 643         kfree(tg);
 644 }
 645
 646 static struct throtl_grp *
 647 throtl_rb_first(struct throtl_service_queue *parent_sq)
 648 {
 649         struct rb_node *n;
 650
 651         n = rb_first_cached(&parent_sq->pending_tree);
 652         WARN_ON_ONCE(!n);
 653         if (!n)
 654                 return NULL;
 655         return rb_entry_tg(n);
 656 }
 657
 658 static void throtl_rb_erase(struct rb_node *n,
 659                             struct throtl_service_queue *parent_sq)
 660 {
 661         rb_erase_cached(n, &parent_sq->pending_tree);
 662         RB_CLEAR_NODE(n);
 663         --parent_sq->nr_pending;
 664 }
 665
 666 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 667 {
 668         struct throtl_grp *tg;
 669
 670         tg = throtl_rb_first(parent_sq);
 671         if (!tg)
 672                 return;
 673
 674         parent_sq->first_pending_disptime = tg->disptime;
 675 }
 676
 677 static void tg_service_queue_add(struct throtl_grp *tg)
 678 {
 679         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 680         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 681         struct rb_node *parent = NULL;
 682         struct throtl_grp *__tg;
 683         unsigned long key = tg->disptime;
 684         bool leftmost = true;
 685
 686         while (*node != NULL) {
 687                 parent = *node;
 688                 __tg = rb_entry_tg(parent);
 689
 690                 if (time_before(key, __tg->disptime))
 691                         node = &parent->rb_left;
 692                 else {
 693                         node = &parent->rb_right;
 694                         leftmost = false;
 695                 }
 696         }
 697
 698         rb_link_node(&tg->rb_node, parent, node);
 699         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 700                                leftmost);
 701 }
 702
 703 static void throtl_enqueue_tg(struct throtl_grp *tg)
 704 {
 705         if (!(tg->flags & THROTL_TG_PENDING)) {
 706                 tg_service_queue_add(tg);
 707                 tg->flags |= THROTL_TG_PENDING;
 708                 tg->service_queue.parent_sq->nr_pending++;
 709         }
 710 }
 711
 712 static void throtl_dequeue_tg(struct throtl_grp *tg)
 713 {
 714         if (tg->flags & THROTL_TG_PENDING) {
 715                 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 716                 tg->flags &= ~THROTL_TG_PENDING;
 717         }
 718 }
 719
 720 /* Call with queue lock held */
 721 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 722                                           unsigned long expires)
 723 {
 724         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 725
 726         /*
 727          * Since we are adjusting the throttle limit dynamically, the sleep
 728          * time calculated according to previous limit might be invalid. It's
 729          * possible the cgroup sleep time is very long and no other cgroups
 730          * have IO running so notify the limit changes. Make sure the cgroup
 731          * doesn't sleep too long to avoid the missed notification.
 732          */
 733         if (time_after(expires, max_expire))
 734                 expires = max_expire;
 735         mod_timer(&sq->pending_timer, expires);
 736         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 737                    expires - jiffies, jiffies);
 738 }
 739
 740 /**
 741  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 742  * @sq: the service_queue to schedule dispatch for
 743  * @force: force scheduling
 744  *
 745  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 746  * dispatch time of the first pending child.  Returns %true if either timer
 747  * is armed or there's no pending child left.  %false if the current
 748  * dispatch window is still open and the caller should continue
 749  * dispatching.
 750  *
 751  * If @force is %true, the dispatch timer is always scheduled and this
 752  * function is guaranteed to return %true.  This is to be used when the
 753  * caller can't dispatch itself and needs to invoke pending_timer
 754  * unconditionally.  Note that forced scheduling is likely to induce short
 755  * delay before dispatch starts even if @sq->first_pending_disptime is not
 756  * in the future and thus shouldn't be used in hot paths.
 757  */
 758 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 759                                           bool force)
 760 {
 761         /* any pending children left? */
 762         if (!sq->nr_pending)
 763                 return true;
 764
 765         update_min_dispatch_time(sq);
 766
 767         /* is the next dispatch time in the future? */
 768         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 769                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 770                 return true;
 771         }
 772
 773         /* tell the caller to continue dispatching */
 774         return false;
 775 }
 776
 777 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 778                 bool rw, unsigned long start)
 779 {
 780         tg->bytes_disp[rw] = 0;
 781         tg->io_disp[rw] = 0;
 782
 783         atomic_set(&tg->io_split_cnt[rw], 0);
 784
 785         /*
 786          * Previous slice has expired. We must have trimmed it after last
 787          * bio dispatch. That means since start of last slice, we never used
 788          * that bandwidth. Do try to make use of that bandwidth while giving
 789          * credit.
 790          */
 791         if (time_after_eq(start, tg->slice_start[rw]))
 792                 tg->slice_start[rw] = start;
 793
 794         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 795         throtl_log(&tg->service_queue,
 796                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 797                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 798                    tg->slice_end[rw], jiffies);
 799 }
 800
 801 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 802 {
 803         tg->bytes_disp[rw] = 0;
 804         tg->io_disp[rw] = 0;
 805         tg->slice_start[rw] = jiffies;
 806         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 807
 808         atomic_set(&tg->io_split_cnt[rw], 0);
 809
 810         throtl_log(&tg->service_queue,
 811                    "[%c] new 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 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 817                                         unsigned long jiffy_end)
 818 {
 819         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 820 }
 821
 822 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 823                                        unsigned long jiffy_end)
 824 {
 825         throtl_set_slice_end(tg, rw, jiffy_end);
 826         throtl_log(&tg->service_queue,
 827                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 828                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 829                    tg->slice_end[rw], jiffies);
 830 }
 831
 832 /* Determine if previously allocated or extended slice is complete or not */
 833 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 834 {
 835         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 836                 return false;
 837
 838         return true;
 839 }
 840
 841 /* Trim the used slices and adjust slice start accordingly */
 842 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 843 {
 844         unsigned long nr_slices, time_elapsed, io_trim;
 845         u64 bytes_trim, tmp;
 846
 847         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 848
 849         /*
 850          * If bps are unlimited (-1), then time slice don't get
 851          * renewed. Don't try to trim the slice if slice is used. A new
 852          * slice will start when appropriate.
 853          */
 854         if (throtl_slice_used(tg, rw))
 855                 return;
 856
 857         /*
 858          * A bio has been dispatched. Also adjust slice_end. It might happen
 859          * that initially cgroup limit was very low resulting in high
 860          * slice_end, but later limit was bumped up and bio was dispatched
 861          * sooner, then we need to reduce slice_end. A high bogus slice_end
 862          * is bad because it does not allow new slice to start.
 863          */
 864
 865         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 866
 867         time_elapsed = jiffies - tg->slice_start[rw];
 868
 869         nr_slices = time_elapsed / tg->td->throtl_slice;
 870
 871         if (!nr_slices)
 872                 return;
 873         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 874         do_div(tmp, HZ);
 875         bytes_trim = tmp;
 876
 877         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 878                 HZ;
 879
 880         if (!bytes_trim && !io_trim)
 881                 return;
 882
 883         if (tg->bytes_disp[rw] >= bytes_trim)
 884                 tg->bytes_disp[rw] -= bytes_trim;
 885         else
 886                 tg->bytes_disp[rw] = 0;
 887
 888         if (tg->io_disp[rw] >= io_trim)
 889                 tg->io_disp[rw] -= io_trim;
 890         else
 891                 tg->io_disp[rw] = 0;
 892
 893         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 894
 895         throtl_log(&tg->service_queue,
 896                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 897                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 898                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 899 }
 900
 901 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 902                                   u32 iops_limit, unsigned long *wait)
 903 {
 904         bool rw = bio_data_dir(bio);
 905         unsigned int io_allowed;
 906         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 907         u64 tmp;
 908
 909         if (iops_limit == UINT_MAX) {
 910                 if (wait)
 911                         *wait = 0;
 912                 return true;
 913         }
 914
 915         jiffy_elapsed = jiffies - tg->slice_start[rw];
 916
 917         /* Round up to the next throttle slice, wait time must be nonzero */
 918         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 919
 920         /*
 921          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 922          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 923          * will allow dispatch after 1 second and after that slice should
 924          * have been trimmed.
 925          */
 926
 927         tmp = (u64)iops_limit * jiffy_elapsed_rnd;
 928         do_div(tmp, HZ);
 929
 930         if (tmp > UINT_MAX)
 931                 io_allowed = UINT_MAX;
 932         else
 933                 io_allowed = tmp;
 934
 935         if (tg->io_disp[rw] + 1 <= io_allowed) {
 936                 if (wait)
 937                         *wait = 0;
 938                 return true;
 939         }
 940
 941         /* Calc approx time to dispatch */
 942         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 943
 944         if (wait)
 945                 *wait = jiffy_wait;
 946         return false;
 947 }
 948
 949 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 950                                  u64 bps_limit, unsigned long *wait)
 951 {
 952         bool rw = bio_data_dir(bio);
 953         u64 bytes_allowed, extra_bytes;
 954         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 955         unsigned int bio_size = throtl_bio_data_size(bio);
 956
 957         if (bps_limit == U64_MAX) {
 958                 if (wait)
 959                         *wait = 0;
 960                 return true;
 961         }
 962
 963         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 964
 965         /* Slice has just started. Consider one slice interval */
 966         if (!jiffy_elapsed)
 967                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 968
 969         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 970         bytes_allowed = mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed_rnd,
 971                                             (u64)HZ);
 972
 973         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 974                 if (wait)
 975                         *wait = 0;
 976                 return true;
 977         }
 978
 979         /* Calc approx time to dispatch */
 980         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 981         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 982
 983         if (!jiffy_wait)
 984                 jiffy_wait = 1;
 985
 986         /*
 987          * This wait time is without taking into consideration the rounding
 988          * up we did. Add that time also.
 989          */
 990         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 991         if (wait)
 992                 *wait = jiffy_wait;
 993         return false;
 994 }
 995
 996 /*
 997  * Returns whether one can dispatch a bio or not. Also returns approx number
 998  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 999  */
1000 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
1001                             unsigned long *wait)
1002 {
1003         bool rw = bio_data_dir(bio);
1004         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
1005         u64 bps_limit = tg_bps_limit(tg, rw);
1006         u32 iops_limit = tg_iops_limit(tg, rw);
1007
1008         /*
1009          * Currently whole state machine of group depends on first bio
1010          * queued in the group bio list. So one should not be calling
1011          * this function with a different bio if there are other bios
1012          * queued.
1013          */
1014         BUG_ON(tg->service_queue.nr_queued[rw] &&
1015                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
1016
1017         /* If tg->bps = -1, then BW is unlimited */
1018         if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
1019                 if (wait)
1020                         *wait = 0;
1021                 return true;
1022         }
1023
1024         /*
1025          * If previous slice expired, start a new one otherwise renew/extend
1026          * existing slice to make sure it is at least throtl_slice interval
1027          * long since now. New slice is started only for empty throttle group.
1028          * If there is queued bio, that means there should be an active
1029          * slice and it should be extended instead.
1030          */
1031         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
1032                 throtl_start_new_slice(tg, rw);
1033         else {
1034                 if (time_before(tg->slice_end[rw],
1035                     jiffies + tg->td->throtl_slice))
1036                         throtl_extend_slice(tg, rw,
1037                                 jiffies + tg->td->throtl_slice);
1038         }
1039
1040         if (iops_limit != UINT_MAX)
1041                 tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
1042
1043         if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
1044             tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
1045                 if (wait)
1046                         *wait = 0;
1047                 return true;
1048         }
1049
1050         max_wait = max(bps_wait, iops_wait);
1051
1052         if (wait)
1053                 *wait = max_wait;
1054
1055         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1056                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1057
1058         return false;
1059 }
1060
1061 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1062 {
1063         bool rw = bio_data_dir(bio);
1064         unsigned int bio_size = throtl_bio_data_size(bio);
1065
1066         /* Charge the bio to the group */
1067         tg->bytes_disp[rw] += bio_size;
1068         tg->io_disp[rw]++;
1069         tg->last_bytes_disp[rw] += bio_size;
1070         tg->last_io_disp[rw]++;
1071
1072         /*
1073          * BIO_THROTTLED is used to prevent the same bio to be throttled
1074          * more than once as a throttled bio will go through blk-throtl the
1075          * second time when it eventually gets issued.  Set it when a bio
1076          * is being charged to a tg.
1077          */
1078         if (!bio_flagged(bio, BIO_THROTTLED))
1079                 bio_set_flag(bio, BIO_THROTTLED);
1080 }
1081
1082 /**
1083  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1084  * @bio: bio to add
1085  * @qn: qnode to use
1086  * @tg: the target throtl_grp
1087  *
1088  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1089  * tg->qnode_on_self[] is used.
1090  */
1091 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1092                               struct throtl_grp *tg)
1093 {
1094         struct throtl_service_queue *sq = &tg->service_queue;
1095         bool rw = bio_data_dir(bio);
1096
1097         if (!qn)
1098                 qn = &tg->qnode_on_self[rw];
1099
1100         /*
1101          * If @tg doesn't currently have any bios queued in the same
1102          * direction, queueing @bio can change when @tg should be
1103          * dispatched.  Mark that @tg was empty.  This is automatically
1104          * cleared on the next tg_update_disptime().
1105          */
1106         if (!sq->nr_queued[rw])
1107                 tg->flags |= THROTL_TG_WAS_EMPTY;
1108
1109         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1110
1111         sq->nr_queued[rw]++;
1112         throtl_enqueue_tg(tg);
1113 }
1114
1115 static void tg_update_disptime(struct throtl_grp *tg)
1116 {
1117         struct throtl_service_queue *sq = &tg->service_queue;
1118         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1119         struct bio *bio;
1120
1121         bio = throtl_peek_queued(&sq->queued[READ]);
1122         if (bio)
1123                 tg_may_dispatch(tg, bio, &read_wait);
1124
1125         bio = throtl_peek_queued(&sq->queued[WRITE]);
1126         if (bio)
1127                 tg_may_dispatch(tg, bio, &write_wait);
1128
1129         min_wait = min(read_wait, write_wait);
1130         disptime = jiffies + min_wait;
1131
1132         /* Update dispatch time */
1133         throtl_dequeue_tg(tg);
1134         tg->disptime = disptime;
1135         throtl_enqueue_tg(tg);
1136
1137         /* see throtl_add_bio_tg() */
1138         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1139 }
1140
1141 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1142                                         struct throtl_grp *parent_tg, bool rw)
1143 {
1144         if (throtl_slice_used(parent_tg, rw)) {
1145                 throtl_start_new_slice_with_credit(parent_tg, rw,
1146                                 child_tg->slice_start[rw]);
1147         }
1148
1149 }
1150
1151 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1152 {
1153         struct throtl_service_queue *sq = &tg->service_queue;
1154         struct throtl_service_queue *parent_sq = sq->parent_sq;
1155         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1156         struct throtl_grp *tg_to_put = NULL;
1157         struct bio *bio;
1158
1159         /*
1160          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1161          * from @tg may put its reference and @parent_sq might end up
1162          * getting released prematurely.  Remember the tg to put and put it
1163          * after @bio is transferred to @parent_sq.
1164          */
1165         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1166         sq->nr_queued[rw]--;
1167
1168         throtl_charge_bio(tg, bio);
1169
1170         /*
1171          * If our parent is another tg, we just need to transfer @bio to
1172          * the parent using throtl_add_bio_tg().  If our parent is
1173          * @td->service_queue, @bio is ready to be issued.  Put it on its
1174          * bio_lists[] and decrease total number queued.  The caller is
1175          * responsible for issuing these bios.
1176          */
1177         if (parent_tg) {
1178                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1179                 start_parent_slice_with_credit(tg, parent_tg, rw);
1180         } else {
1181                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1182                                      &parent_sq->queued[rw]);
1183                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1184                 tg->td->nr_queued[rw]--;
1185         }
1186
1187         throtl_trim_slice(tg, rw);
1188
1189         if (tg_to_put)
1190                 blkg_put(tg_to_blkg(tg_to_put));
1191 }
1192
1193 static int throtl_dispatch_tg(struct throtl_grp *tg)
1194 {
1195         struct throtl_service_queue *sq = &tg->service_queue;
1196         unsigned int nr_reads = 0, nr_writes = 0;
1197         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1198         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1199         struct bio *bio;
1200
1201         /* Try to dispatch 75% READS and 25% WRITES */
1202
1203         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1204                tg_may_dispatch(tg, bio, NULL)) {
1205
1206                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1207                 nr_reads++;
1208
1209                 if (nr_reads >= max_nr_reads)
1210                         break;
1211         }
1212
1213         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1214                tg_may_dispatch(tg, bio, NULL)) {
1215
1216                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1217                 nr_writes++;
1218
1219                 if (nr_writes >= max_nr_writes)
1220                         break;
1221         }
1222
1223         return nr_reads + nr_writes;
1224 }
1225
1226 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1227 {
1228         unsigned int nr_disp = 0;
1229
1230         while (1) {
1231                 struct throtl_grp *tg;
1232                 struct throtl_service_queue *sq;
1233
1234                 if (!parent_sq->nr_pending)
1235                         break;
1236
1237                 tg = throtl_rb_first(parent_sq);
1238                 if (!tg)
1239                         break;
1240
1241                 if (time_before(jiffies, tg->disptime))
1242                         break;
1243
1244                 throtl_dequeue_tg(tg);
1245
1246                 nr_disp += throtl_dispatch_tg(tg);
1247
1248                 sq = &tg->service_queue;
1249                 if (sq->nr_queued[0] || sq->nr_queued[1])
1250                         tg_update_disptime(tg);
1251
1252                 if (nr_disp >= THROTL_QUANTUM)
1253                         break;
1254         }
1255
1256         return nr_disp;
1257 }
1258
1259 static bool throtl_can_upgrade(struct throtl_data *td,
1260         struct throtl_grp *this_tg);
1261 /**
1262  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1263  * @t: the pending_timer member of the throtl_service_queue being serviced
1264  *
1265  * This timer is armed when a child throtl_grp with active bio's become
1266  * pending and queued on the service_queue's pending_tree and expires when
1267  * the first child throtl_grp should be dispatched.  This function
1268  * dispatches bio's from the children throtl_grps to the parent
1269  * service_queue.
1270  *
1271  * If the parent's parent is another throtl_grp, dispatching is propagated
1272  * by either arming its pending_timer or repeating dispatch directly.  If
1273  * the top-level service_tree is reached, throtl_data->dispatch_work is
1274  * kicked so that the ready bio's are issued.
1275  */
1276 static void throtl_pending_timer_fn(struct timer_list *t)
1277 {
1278         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1279         struct throtl_grp *tg = sq_to_tg(sq);
1280         struct throtl_data *td = sq_to_td(sq);
1281         struct request_queue *q = td->queue;
1282         struct throtl_service_queue *parent_sq;
1283         bool dispatched;
1284         int ret;
1285
1286         spin_lock_irq(&q->queue_lock);
1287         if (throtl_can_upgrade(td, NULL))
1288                 throtl_upgrade_state(td);
1289
1290 again:
1291         parent_sq = sq->parent_sq;
1292         dispatched = false;
1293
1294         while (true) {
1295                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1296                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1297                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1298
1299                 ret = throtl_select_dispatch(sq);
1300                 if (ret) {
1301                         throtl_log(sq, "bios disp=%u", ret);
1302                         dispatched = true;
1303                 }
1304
1305                 if (throtl_schedule_next_dispatch(sq, false))
1306                         break;
1307
1308                 /* this dispatch windows is still open, relax and repeat */
1309                 spin_unlock_irq(&q->queue_lock);
1310                 cpu_relax();
1311                 spin_lock_irq(&q->queue_lock);
1312         }
1313
1314         if (!dispatched)
1315                 goto out_unlock;
1316
1317         if (parent_sq) {
1318                 /* @parent_sq is another throl_grp, propagate dispatch */
1319                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1320                         tg_update_disptime(tg);
1321                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1322                                 /* window is already open, repeat dispatching */
1323                                 sq = parent_sq;
1324                                 tg = sq_to_tg(sq);
1325                                 goto again;
1326                         }
1327                 }
1328         } else {
1329                 /* reached the top-level, queue issuing */
1330                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1331         }
1332 out_unlock:
1333         spin_unlock_irq(&q->queue_lock);
1334 }
1335
1336 /**
1337  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1338  * @work: work item being executed
1339  *
1340  * This function is queued for execution when bios reach the bio_lists[]
1341  * of throtl_data->service_queue.  Those bios are ready and issued by this
1342  * function.
1343  */
1344 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1345 {
1346         struct throtl_data *td = container_of(work, struct throtl_data,
1347                                               dispatch_work);
1348         struct throtl_service_queue *td_sq = &td->service_queue;
1349         struct request_queue *q = td->queue;
1350         struct bio_list bio_list_on_stack;
1351         struct bio *bio;
1352         struct blk_plug plug;
1353         int rw;
1354
1355         bio_list_init(&bio_list_on_stack);
1356
1357         spin_lock_irq(&q->queue_lock);
1358         for (rw = READ; rw <= WRITE; rw++)
1359                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1360                         bio_list_add(&bio_list_on_stack, bio);
1361         spin_unlock_irq(&q->queue_lock);
1362
1363         if (!bio_list_empty(&bio_list_on_stack)) {
1364                 blk_start_plug(&plug);
1365                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1366                         submit_bio_noacct(bio);
1367                 blk_finish_plug(&plug);
1368         }
1369 }
1370
1371 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1372                               int off)
1373 {
1374         struct throtl_grp *tg = pd_to_tg(pd);
1375         u64 v = *(u64 *)((void *)tg + off);
1376
1377         if (v == U64_MAX)
1378                 return 0;
1379         return __blkg_prfill_u64(sf, pd, v);
1380 }
1381
1382 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1383                                int off)
1384 {
1385         struct throtl_grp *tg = pd_to_tg(pd);
1386         unsigned int v = *(unsigned int *)((void *)tg + off);
1387
1388         if (v == UINT_MAX)
1389                 return 0;
1390         return __blkg_prfill_u64(sf, pd, v);
1391 }
1392
1393 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1394 {
1395         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1396                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1397         return 0;
1398 }
1399
1400 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1401 {
1402         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1403                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1404         return 0;
1405 }
1406
1407 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1408 {
1409         struct throtl_service_queue *sq = &tg->service_queue;
1410         struct cgroup_subsys_state *pos_css;
1411         struct blkcg_gq *blkg;
1412
1413         throtl_log(&tg->service_queue,
1414                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1415                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1416                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1417
1418         /*
1419          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1420          * considered to have rules if either the tg itself or any of its
1421          * ancestors has rules.  This identifies groups without any
1422          * restrictions in the whole hierarchy and allows them to bypass
1423          * blk-throttle.
1424          */
1425         blkg_for_each_descendant_pre(blkg, pos_css,
1426                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1427                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1428                 struct throtl_grp *parent_tg;
1429
1430                 tg_update_has_rules(this_tg);
1431                 /* ignore root/second level */
1432                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1433                     !blkg->parent->parent)
1434                         continue;
1435                 parent_tg = blkg_to_tg(blkg->parent);
1436                 /*
1437                  * make sure all children has lower idle time threshold and
1438                  * higher latency target
1439                  */
1440                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1441                                 parent_tg->idletime_threshold);
1442                 this_tg->latency_target = max(this_tg->latency_target,
1443                                 parent_tg->latency_target);
1444         }
1445
1446         /*
1447          * We're already holding queue_lock and know @tg is valid.  Let's
1448          * apply the new config directly.
1449          *
1450          * Restart the slices for both READ and WRITES. It might happen
1451          * that a group's limit are dropped suddenly and we don't want to
1452          * account recently dispatched IO with new low rate.
1453          */
1454         throtl_start_new_slice(tg, READ);
1455         throtl_start_new_slice(tg, WRITE);
1456
1457         if (tg->flags & THROTL_TG_PENDING) {
1458                 tg_update_disptime(tg);
1459                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1460         }
1461 }
1462
1463 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1464                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1465 {
1466         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1467         struct blkg_conf_ctx ctx;
1468         struct throtl_grp *tg;
1469         int ret;
1470         u64 v;
1471
1472         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1473         if (ret)
1474                 return ret;
1475
1476         ret = -EINVAL;
1477         if (sscanf(ctx.body, "%llu", &v) != 1)
1478                 goto out_finish;
1479         if (!v)
1480                 v = U64_MAX;
1481
1482         tg = blkg_to_tg(ctx.blkg);
1483
1484         if (is_u64)
1485                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1486         else
1487                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1488
1489         tg_conf_updated(tg, false);
1490         ret = 0;
1491 out_finish:
1492         blkg_conf_finish(&ctx);
1493         return ret ?: nbytes;
1494 }
1495
1496 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1497                                char *buf, size_t nbytes, loff_t off)
1498 {
1499         return tg_set_conf(of, buf, nbytes, off, true);
1500 }
1501
1502 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1503                                 char *buf, size_t nbytes, loff_t off)
1504 {
1505         return tg_set_conf(of, buf, nbytes, off, false);
1506 }
1507
1508 static int tg_print_rwstat(struct seq_file *sf, void *v)
1509 {
1510         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1511                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1512                           seq_cft(sf)->private, true);
1513         return 0;
1514 }
1515
1516 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1517                                       struct blkg_policy_data *pd, int off)
1518 {
1519         struct blkg_rwstat_sample sum;
1520
1521         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1522                                   &sum);
1523         return __blkg_prfill_rwstat(sf, pd, &sum);
1524 }
1525
1526 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1527 {
1528         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1529                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1530                           seq_cft(sf)->private, true);
1531         return 0;
1532 }
1533
1534 static struct cftype throtl_legacy_files[] = {
1535         {
1536                 .name = "throttle.read_bps_device",
1537                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1538                 .seq_show = tg_print_conf_u64,
1539                 .write = tg_set_conf_u64,
1540         },
1541         {
1542                 .name = "throttle.write_bps_device",
1543                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1544                 .seq_show = tg_print_conf_u64,
1545                 .write = tg_set_conf_u64,
1546         },
1547         {
1548                 .name = "throttle.read_iops_device",
1549                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1550                 .seq_show = tg_print_conf_uint,
1551                 .write = tg_set_conf_uint,
1552         },
1553         {
1554                 .name = "throttle.write_iops_device",
1555                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1556                 .seq_show = tg_print_conf_uint,
1557                 .write = tg_set_conf_uint,
1558         },
1559         {
1560                 .name = "throttle.io_service_bytes",
1561                 .private = offsetof(struct throtl_grp, stat_bytes),
1562                 .seq_show = tg_print_rwstat,
1563         },
1564         {
1565                 .name = "throttle.io_service_bytes_recursive",
1566                 .private = offsetof(struct throtl_grp, stat_bytes),
1567                 .seq_show = tg_print_rwstat_recursive,
1568         },
1569         {
1570                 .name = "throttle.io_serviced",
1571                 .private = offsetof(struct throtl_grp, stat_ios),
1572                 .seq_show = tg_print_rwstat,
1573         },
1574         {
1575                 .name = "throttle.io_serviced_recursive",
1576                 .private = offsetof(struct throtl_grp, stat_ios),
1577                 .seq_show = tg_print_rwstat_recursive,
1578         },
1579         { }     /* terminate */
1580 };
1581
1582 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1583                          int off)
1584 {
1585         struct throtl_grp *tg = pd_to_tg(pd);
1586         const char *dname = blkg_dev_name(pd->blkg);
1587         char bufs[4][21] = { "max", "max", "max", "max" };
1588         u64 bps_dft;
1589         unsigned int iops_dft;
1590         char idle_time[26] = "";
1591         char latency_time[26] = "";
1592
1593         if (!dname)
1594                 return 0;
1595
1596         if (off == LIMIT_LOW) {
1597                 bps_dft = 0;
1598                 iops_dft = 0;
1599         } else {
1600                 bps_dft = U64_MAX;
1601                 iops_dft = UINT_MAX;
1602         }
1603
1604         if (tg->bps_conf[READ][off] == bps_dft &&
1605             tg->bps_conf[WRITE][off] == bps_dft &&
1606             tg->iops_conf[READ][off] == iops_dft &&
1607             tg->iops_conf[WRITE][off] == iops_dft &&
1608             (off != LIMIT_LOW ||
1609              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1610               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1611                 return 0;
1612
1613         if (tg->bps_conf[READ][off] != U64_MAX)
1614                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1615                         tg->bps_conf[READ][off]);
1616         if (tg->bps_conf[WRITE][off] != U64_MAX)
1617                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1618                         tg->bps_conf[WRITE][off]);
1619         if (tg->iops_conf[READ][off] != UINT_MAX)
1620                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1621                         tg->iops_conf[READ][off]);
1622         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1623                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1624                         tg->iops_conf[WRITE][off]);
1625         if (off == LIMIT_LOW) {
1626                 if (tg->idletime_threshold_conf == ULONG_MAX)
1627                         strcpy(idle_time, " idle=max");
1628                 else
1629                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1630                                 tg->idletime_threshold_conf);
1631
1632                 if (tg->latency_target_conf == ULONG_MAX)
1633                         strcpy(latency_time, " latency=max");
1634                 else
1635                         snprintf(latency_time, sizeof(latency_time),
1636                                 " latency=%lu", tg->latency_target_conf);
1637         }
1638
1639         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1640                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1641                    latency_time);
1642         return 0;
1643 }
1644
1645 static int tg_print_limit(struct seq_file *sf, void *v)
1646 {
1647         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1648                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1649         return 0;
1650 }
1651
1652 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1653                           char *buf, size_t nbytes, loff_t off)
1654 {
1655         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1656         struct blkg_conf_ctx ctx;
1657         struct throtl_grp *tg;
1658         u64 v[4];
1659         unsigned long idle_time;
1660         unsigned long latency_time;
1661         int ret;
1662         int index = of_cft(of)->private;
1663
1664         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1665         if (ret)
1666                 return ret;
1667
1668         tg = blkg_to_tg(ctx.blkg);
1669
1670         v[0] = tg->bps_conf[READ][index];
1671         v[1] = tg->bps_conf[WRITE][index];
1672         v[2] = tg->iops_conf[READ][index];
1673         v[3] = tg->iops_conf[WRITE][index];
1674
1675         idle_time = tg->idletime_threshold_conf;
1676         latency_time = tg->latency_target_conf;
1677         while (true) {
1678                 char tok[27];   /* wiops=18446744073709551616 */
1679                 char *p;
1680                 u64 val = U64_MAX;
1681                 int len;
1682
1683                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1684                         break;
1685                 if (tok[0] == '\0')
1686                         break;
1687                 ctx.body += len;
1688
1689                 ret = -EINVAL;
1690                 p = tok;
1691                 strsep(&p, "=");
1692                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1693                         goto out_finish;
1694
1695                 ret = -ERANGE;
1696                 if (!val)
1697                         goto out_finish;
1698
1699                 ret = -EINVAL;
1700                 if (!strcmp(tok, "rbps") && val > 1)
1701                         v[0] = val;
1702                 else if (!strcmp(tok, "wbps") && val > 1)
1703                         v[1] = val;
1704                 else if (!strcmp(tok, "riops") && val > 1)
1705                         v[2] = min_t(u64, val, UINT_MAX);
1706                 else if (!strcmp(tok, "wiops") && val > 1)
1707                         v[3] = min_t(u64, val, UINT_MAX);
1708                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1709                         idle_time = val;
1710                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1711                         latency_time = val;
1712                 else
1713                         goto out_finish;
1714         }
1715
1716         tg->bps_conf[READ][index] = v[0];
1717         tg->bps_conf[WRITE][index] = v[1];
1718         tg->iops_conf[READ][index] = v[2];
1719         tg->iops_conf[WRITE][index] = v[3];
1720
1721         if (index == LIMIT_MAX) {
1722                 tg->bps[READ][index] = v[0];
1723                 tg->bps[WRITE][index] = v[1];
1724                 tg->iops[READ][index] = v[2];
1725                 tg->iops[WRITE][index] = v[3];
1726         }
1727         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1728                 tg->bps_conf[READ][LIMIT_MAX]);
1729         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1730                 tg->bps_conf[WRITE][LIMIT_MAX]);
1731         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1732                 tg->iops_conf[READ][LIMIT_MAX]);
1733         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1734                 tg->iops_conf[WRITE][LIMIT_MAX]);
1735         tg->idletime_threshold_conf = idle_time;
1736         tg->latency_target_conf = latency_time;
1737
1738         /* force user to configure all settings for low limit  */
1739         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1740               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1741             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1742             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1743                 tg->bps[READ][LIMIT_LOW] = 0;
1744                 tg->bps[WRITE][LIMIT_LOW] = 0;
1745                 tg->iops[READ][LIMIT_LOW] = 0;
1746                 tg->iops[WRITE][LIMIT_LOW] = 0;
1747                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1748                 tg->latency_target = DFL_LATENCY_TARGET;
1749         } else if (index == LIMIT_LOW) {
1750                 tg->idletime_threshold = tg->idletime_threshold_conf;
1751                 tg->latency_target = tg->latency_target_conf;
1752         }
1753
1754         blk_throtl_update_limit_valid(tg->td);
1755         if (tg->td->limit_valid[LIMIT_LOW]) {
1756                 if (index == LIMIT_LOW)
1757                         tg->td->limit_index = LIMIT_LOW;
1758         } else
1759                 tg->td->limit_index = LIMIT_MAX;
1760         tg_conf_updated(tg, index == LIMIT_LOW &&
1761                 tg->td->limit_valid[LIMIT_LOW]);
1762         ret = 0;
1763 out_finish:
1764         blkg_conf_finish(&ctx);
1765         return ret ?: nbytes;
1766 }
1767
1768 static struct cftype throtl_files[] = {
1769 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1770         {
1771                 .name = "low",
1772                 .flags = CFTYPE_NOT_ON_ROOT,
1773                 .seq_show = tg_print_limit,
1774                 .write = tg_set_limit,
1775                 .private = LIMIT_LOW,
1776         },
1777 #endif
1778         {
1779                 .name = "max",
1780                 .flags = CFTYPE_NOT_ON_ROOT,
1781                 .seq_show = tg_print_limit,
1782                 .write = tg_set_limit,
1783                 .private = LIMIT_MAX,
1784         },
1785         { }     /* terminate */
1786 };
1787
1788 static void throtl_shutdown_wq(struct request_queue *q)
1789 {
1790         struct throtl_data *td = q->td;
1791
1792         cancel_work_sync(&td->dispatch_work);
1793 }
1794
1795 static struct blkcg_policy blkcg_policy_throtl = {
1796         .dfl_cftypes            = throtl_files,
1797         .legacy_cftypes         = throtl_legacy_files,
1798
1799         .pd_alloc_fn            = throtl_pd_alloc,
1800         .pd_init_fn             = throtl_pd_init,
1801         .pd_online_fn           = throtl_pd_online,
1802         .pd_offline_fn          = throtl_pd_offline,
1803         .pd_free_fn             = throtl_pd_free,
1804 };
1805
1806 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1807 {
1808         unsigned long rtime = jiffies, wtime = jiffies;
1809
1810         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1811                 rtime = tg->last_low_overflow_time[READ];
1812         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1813                 wtime = tg->last_low_overflow_time[WRITE];
1814         return min(rtime, wtime);
1815 }
1816
1817 /* tg should not be an intermediate node */
1818 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1819 {
1820         struct throtl_service_queue *parent_sq;
1821         struct throtl_grp *parent = tg;
1822         unsigned long ret = __tg_last_low_overflow_time(tg);
1823
1824         while (true) {
1825                 parent_sq = parent->service_queue.parent_sq;
1826                 parent = sq_to_tg(parent_sq);
1827                 if (!parent)
1828                         break;
1829
1830                 /*
1831                  * The parent doesn't have low limit, it always reaches low
1832                  * limit. Its overflow time is useless for children
1833                  */
1834                 if (!parent->bps[READ][LIMIT_LOW] &&
1835                     !parent->iops[READ][LIMIT_LOW] &&
1836                     !parent->bps[WRITE][LIMIT_LOW] &&
1837                     !parent->iops[WRITE][LIMIT_LOW])
1838                         continue;
1839                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1840                         ret = __tg_last_low_overflow_time(parent);
1841         }
1842         return ret;
1843 }
1844
1845 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1846 {
1847         /*
1848          * cgroup is idle if:
1849          * - single idle is too long, longer than a fixed value (in case user
1850          *   configure a too big threshold) or 4 times of idletime threshold
1851          * - average think time is more than threshold
1852          * - IO latency is largely below threshold
1853          */
1854         unsigned long time;
1855         bool ret;
1856
1857         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1858         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1859               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1860               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1861               tg->avg_idletime > tg->idletime_threshold ||
1862               (tg->latency_target && tg->bio_cnt &&
1863                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1864         throtl_log(&tg->service_queue,
1865                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1866                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1867                 tg->bio_cnt, ret, tg->td->scale);
1868         return ret;
1869 }
1870
1871 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1872 {
1873         struct throtl_service_queue *sq = &tg->service_queue;
1874         bool read_limit, write_limit;
1875
1876         /*
1877          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1878          * reaches), it's ok to upgrade to next limit
1879          */
1880         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1881         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1882         if (!read_limit && !write_limit)
1883                 return true;
1884         if (read_limit && sq->nr_queued[READ] &&
1885             (!write_limit || sq->nr_queued[WRITE]))
1886                 return true;
1887         if (write_limit && sq->nr_queued[WRITE] &&
1888             (!read_limit || sq->nr_queued[READ]))
1889                 return true;
1890
1891         if (time_after_eq(jiffies,
1892                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1893             throtl_tg_is_idle(tg))
1894                 return true;
1895         return false;
1896 }
1897
1898 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1899 {
1900         while (true) {
1901                 if (throtl_tg_can_upgrade(tg))
1902                         return true;
1903                 tg = sq_to_tg(tg->service_queue.parent_sq);
1904                 if (!tg || !tg_to_blkg(tg)->parent)
1905                         return false;
1906         }
1907         return false;
1908 }
1909
1910 static bool throtl_can_upgrade(struct throtl_data *td,
1911         struct throtl_grp *this_tg)
1912 {
1913         struct cgroup_subsys_state *pos_css;
1914         struct blkcg_gq *blkg;
1915
1916         if (td->limit_index != LIMIT_LOW)
1917                 return false;
1918
1919         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1920                 return false;
1921
1922         rcu_read_lock();
1923         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1924                 struct throtl_grp *tg = blkg_to_tg(blkg);
1925
1926                 if (tg == this_tg)
1927                         continue;
1928                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1929                         continue;
1930                 if (!throtl_hierarchy_can_upgrade(tg)) {
1931                         rcu_read_unlock();
1932                         return false;
1933                 }
1934         }
1935         rcu_read_unlock();
1936         return true;
1937 }
1938
1939 static void throtl_upgrade_check(struct throtl_grp *tg)
1940 {
1941         unsigned long now = jiffies;
1942
1943         if (tg->td->limit_index != LIMIT_LOW)
1944                 return;
1945
1946         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1947                 return;
1948
1949         tg->last_check_time = now;
1950
1951         if (!time_after_eq(now,
1952              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1953                 return;
1954
1955         if (throtl_can_upgrade(tg->td, NULL))
1956                 throtl_upgrade_state(tg->td);
1957 }
1958
1959 static void throtl_upgrade_state(struct throtl_data *td)
1960 {
1961         struct cgroup_subsys_state *pos_css;
1962         struct blkcg_gq *blkg;
1963
1964         throtl_log(&td->service_queue, "upgrade to max");
1965         td->limit_index = LIMIT_MAX;
1966         td->low_upgrade_time = jiffies;
1967         td->scale = 0;
1968         rcu_read_lock();
1969         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1970                 struct throtl_grp *tg = blkg_to_tg(blkg);
1971                 struct throtl_service_queue *sq = &tg->service_queue;
1972
1973                 tg->disptime = jiffies - 1;
1974                 throtl_select_dispatch(sq);
1975                 throtl_schedule_next_dispatch(sq, true);
1976         }
1977         rcu_read_unlock();
1978         throtl_select_dispatch(&td->service_queue);
1979         throtl_schedule_next_dispatch(&td->service_queue, true);
1980         queue_work(kthrotld_workqueue, &td->dispatch_work);
1981 }
1982
1983 static void throtl_downgrade_state(struct throtl_data *td)
1984 {
1985         td->scale /= 2;
1986
1987         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1988         if (td->scale) {
1989                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1990                 return;
1991         }
1992
1993         td->limit_index = LIMIT_LOW;
1994         td->low_downgrade_time = jiffies;
1995 }
1996
1997 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1998 {
1999         struct throtl_data *td = tg->td;
2000         unsigned long now = jiffies;
2001
2002         /*
2003          * If cgroup is below low limit, consider downgrade and throttle other
2004          * cgroups
2005          */
2006         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
2007             time_after_eq(now, tg_last_low_overflow_time(tg) +
2008                                         td->throtl_slice) &&
2009             (!throtl_tg_is_idle(tg) ||
2010              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
2011                 return true;
2012         return false;
2013 }
2014
2015 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
2016 {
2017         while (true) {
2018                 if (!throtl_tg_can_downgrade(tg))
2019                         return false;
2020                 tg = sq_to_tg(tg->service_queue.parent_sq);
2021                 if (!tg || !tg_to_blkg(tg)->parent)
2022                         break;
2023         }
2024         return true;
2025 }
2026
2027 static void throtl_downgrade_check(struct throtl_grp *tg)
2028 {
2029         uint64_t bps;
2030         unsigned int iops;
2031         unsigned long elapsed_time;
2032         unsigned long now = jiffies;
2033
2034         if (tg->td->limit_index != LIMIT_MAX ||
2035             !tg->td->limit_valid[LIMIT_LOW])
2036                 return;
2037         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
2038                 return;
2039         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
2040                 return;
2041
2042         elapsed_time = now - tg->last_check_time;
2043         tg->last_check_time = now;
2044
2045         if (time_before(now, tg_last_low_overflow_time(tg) +
2046                         tg->td->throtl_slice))
2047                 return;
2048
2049         if (tg->bps[READ][LIMIT_LOW]) {
2050                 bps = tg->last_bytes_disp[READ] * HZ;
2051                 do_div(bps, elapsed_time);
2052                 if (bps >= tg->bps[READ][LIMIT_LOW])
2053                         tg->last_low_overflow_time[READ] = now;
2054         }
2055
2056         if (tg->bps[WRITE][LIMIT_LOW]) {
2057                 bps = tg->last_bytes_disp[WRITE] * HZ;
2058                 do_div(bps, elapsed_time);
2059                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
2060                         tg->last_low_overflow_time[WRITE] = now;
2061         }
2062
2063         if (tg->iops[READ][LIMIT_LOW]) {
2064                 tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
2065                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
2066                 if (iops >= tg->iops[READ][LIMIT_LOW])
2067                         tg->last_low_overflow_time[READ] = now;
2068         }
2069
2070         if (tg->iops[WRITE][LIMIT_LOW]) {
2071                 tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
2072                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2073                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2074                         tg->last_low_overflow_time[WRITE] = now;
2075         }
2076
2077         /*
2078          * If cgroup is below low limit, consider downgrade and throttle other
2079          * cgroups
2080          */
2081         if (throtl_hierarchy_can_downgrade(tg))
2082                 throtl_downgrade_state(tg->td);
2083
2084         tg->last_bytes_disp[READ] = 0;
2085         tg->last_bytes_disp[WRITE] = 0;
2086         tg->last_io_disp[READ] = 0;
2087         tg->last_io_disp[WRITE] = 0;
2088 }
2089
2090 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2091 {
2092         unsigned long now;
2093         unsigned long last_finish_time = tg->last_finish_time;
2094
2095         if (last_finish_time == 0)
2096                 return;
2097
2098         now = ktime_get_ns() >> 10;
2099         if (now <= last_finish_time ||
2100             last_finish_time == tg->checked_last_finish_time)
2101                 return;
2102
2103         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2104         tg->checked_last_finish_time = last_finish_time;
2105 }
2106
2107 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2108 static void throtl_update_latency_buckets(struct throtl_data *td)
2109 {
2110         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2111         int i, cpu, rw;
2112         unsigned long last_latency[2] = { 0 };
2113         unsigned long latency[2];
2114
2115         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
2116                 return;
2117         if (time_before(jiffies, td->last_calculate_time + HZ))
2118                 return;
2119         td->last_calculate_time = jiffies;
2120
2121         memset(avg_latency, 0, sizeof(avg_latency));
2122         for (rw = READ; rw <= WRITE; rw++) {
2123                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2124                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2125
2126                         for_each_possible_cpu(cpu) {
2127                                 struct latency_bucket *bucket;
2128
2129                                 /* this isn't race free, but ok in practice */
2130                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2131                                         cpu);
2132                                 tmp->total_latency += bucket[i].total_latency;
2133                                 tmp->samples += bucket[i].samples;
2134                                 bucket[i].total_latency = 0;
2135                                 bucket[i].samples = 0;
2136                         }
2137
2138                         if (tmp->samples >= 32) {
2139                                 int samples = tmp->samples;
2140
2141                                 latency[rw] = tmp->total_latency;
2142
2143                                 tmp->total_latency = 0;
2144                                 tmp->samples = 0;
2145                                 latency[rw] /= samples;
2146                                 if (latency[rw] == 0)
2147                                         continue;
2148                                 avg_latency[rw][i].latency = latency[rw];
2149                         }
2150                 }
2151         }
2152
2153         for (rw = READ; rw <= WRITE; rw++) {
2154                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2155                         if (!avg_latency[rw][i].latency) {
2156                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2157                                         td->avg_buckets[rw][i].latency =
2158                                                 last_latency[rw];
2159                                 continue;
2160                         }
2161
2162                         if (!td->avg_buckets[rw][i].valid)
2163                                 latency[rw] = avg_latency[rw][i].latency;
2164                         else
2165                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2166                                         avg_latency[rw][i].latency) >> 3;
2167
2168                         td->avg_buckets[rw][i].latency = max(latency[rw],
2169                                 last_latency[rw]);
2170                         td->avg_buckets[rw][i].valid = true;
2171                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2172                 }
2173         }
2174
2175         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2176                 throtl_log(&td->service_queue,
2177                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2178                         "write latency=%ld, write valid=%d", i,
2179                         td->avg_buckets[READ][i].latency,
2180                         td->avg_buckets[READ][i].valid,
2181                         td->avg_buckets[WRITE][i].latency,
2182                         td->avg_buckets[WRITE][i].valid);
2183 }
2184 #else
2185 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2186 {
2187 }
2188 #endif
2189
2190 void blk_throtl_charge_bio_split(struct bio *bio)
2191 {
2192         struct blkcg_gq *blkg = bio->bi_blkg;
2193         struct throtl_grp *parent = blkg_to_tg(blkg);
2194         struct throtl_service_queue *parent_sq;
2195         bool rw = bio_data_dir(bio);
2196
2197         do {
2198                 if (!parent->has_rules[rw])
2199                         break;
2200
2201                 atomic_inc(&parent->io_split_cnt[rw]);
2202                 atomic_inc(&parent->last_io_split_cnt[rw]);
2203
2204                 parent_sq = parent->service_queue.parent_sq;
2205                 parent = sq_to_tg(parent_sq);
2206         } while (parent);
2207 }
2208
2209 bool blk_throtl_bio(struct bio *bio)
2210 {
2211         struct request_queue *q = bio->bi_bdev->bd_disk->queue;
2212         struct blkcg_gq *blkg = bio->bi_blkg;
2213         struct throtl_qnode *qn = NULL;
2214         struct throtl_grp *tg = blkg_to_tg(blkg);
2215         struct throtl_service_queue *sq;
2216         bool rw = bio_data_dir(bio);
2217         bool throttled = false;
2218         struct throtl_data *td = tg->td;
2219
2220         rcu_read_lock();
2221
2222         /* see throtl_charge_bio() */
2223         if (bio_flagged(bio, BIO_THROTTLED))
2224                 goto out;
2225
2226         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2227                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2228                                 bio->bi_iter.bi_size);
2229                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2230         }
2231
2232         if (!tg->has_rules[rw])
2233                 goto out;
2234
2235         spin_lock_irq(&q->queue_lock);
2236
2237         throtl_update_latency_buckets(td);
2238
2239         blk_throtl_update_idletime(tg);
2240
2241         sq = &tg->service_queue;
2242
2243 again:
2244         while (true) {
2245                 if (tg->last_low_overflow_time[rw] == 0)
2246                         tg->last_low_overflow_time[rw] = jiffies;
2247                 throtl_downgrade_check(tg);
2248                 throtl_upgrade_check(tg);
2249                 /* throtl is FIFO - if bios are already queued, should queue */
2250                 if (sq->nr_queued[rw])
2251                         break;
2252
2253                 /* if above limits, break to queue */
2254                 if (!tg_may_dispatch(tg, bio, NULL)) {
2255                         tg->last_low_overflow_time[rw] = jiffies;
2256                         if (throtl_can_upgrade(td, tg)) {
2257                                 throtl_upgrade_state(td);
2258                                 goto again;
2259                         }
2260                         break;
2261                 }
2262
2263                 /* within limits, let's charge and dispatch directly */
2264                 throtl_charge_bio(tg, bio);
2265
2266                 /*
2267                  * We need to trim slice even when bios are not being queued
2268                  * otherwise it might happen that a bio is not queued for
2269                  * a long time and slice keeps on extending and trim is not
2270                  * called for a long time. Now if limits are reduced suddenly
2271                  * we take into account all the IO dispatched so far at new
2272                  * low rate and * newly queued IO gets a really long dispatch
2273                  * time.
2274                  *
2275                  * So keep on trimming slice even if bio is not queued.
2276                  */
2277                 throtl_trim_slice(tg, rw);
2278
2279                 /*
2280                  * @bio passed through this layer without being throttled.
2281                  * Climb up the ladder.  If we're already at the top, it
2282                  * can be executed directly.
2283                  */
2284                 qn = &tg->qnode_on_parent[rw];
2285                 sq = sq->parent_sq;
2286                 tg = sq_to_tg(sq);
2287                 if (!tg)
2288                         goto out_unlock;
2289         }
2290
2291         /* out-of-limit, queue to @tg */
2292         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2293                    rw == READ ? 'R' : 'W',
2294                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2295                    tg_bps_limit(tg, rw),
2296                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2297                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2298
2299         tg->last_low_overflow_time[rw] = jiffies;
2300
2301         td->nr_queued[rw]++;
2302         throtl_add_bio_tg(bio, qn, tg);
2303         throttled = true;
2304
2305         /*
2306          * Update @tg's dispatch time and force schedule dispatch if @tg
2307          * was empty before @bio.  The forced scheduling isn't likely to
2308          * cause undue delay as @bio is likely to be dispatched directly if
2309          * its @tg's disptime is not in the future.
2310          */
2311         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2312                 tg_update_disptime(tg);
2313                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2314         }
2315
2316 out_unlock:
2317         spin_unlock_irq(&q->queue_lock);
2318 out:
2319         bio_set_flag(bio, BIO_THROTTLED);
2320
2321 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2322         if (throttled || !td->track_bio_latency)
2323                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2324 #endif
2325         rcu_read_unlock();
2326         return throttled;
2327 }
2328
2329 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2330 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2331         int op, unsigned long time)
2332 {
2333         struct latency_bucket *latency;
2334         int index;
2335
2336         if (!td || td->limit_index != LIMIT_LOW ||
2337             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2338             !blk_queue_nonrot(td->queue))
2339                 return;
2340
2341         index = request_bucket_index(size);
2342
2343         latency = get_cpu_ptr(td->latency_buckets[op]);
2344         latency[index].total_latency += time;
2345         latency[index].samples++;
2346         put_cpu_ptr(td->latency_buckets[op]);
2347 }
2348
2349 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2350 {
2351         struct request_queue *q = rq->q;
2352         struct throtl_data *td = q->td;
2353
2354         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2355                              time_ns >> 10);
2356 }
2357
2358 void blk_throtl_bio_endio(struct bio *bio)
2359 {
2360         struct blkcg_gq *blkg;
2361         struct throtl_grp *tg;
2362         u64 finish_time_ns;
2363         unsigned long finish_time;
2364         unsigned long start_time;
2365         unsigned long lat;
2366         int rw = bio_data_dir(bio);
2367
2368         blkg = bio->bi_blkg;
2369         if (!blkg)
2370                 return;
2371         tg = blkg_to_tg(blkg);
2372         if (!tg->td->limit_valid[LIMIT_LOW])
2373                 return;
2374
2375         finish_time_ns = ktime_get_ns();
2376         tg->last_finish_time = finish_time_ns >> 10;
2377
2378         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2379         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2380         if (!start_time || finish_time <= start_time)
2381                 return;
2382
2383         lat = finish_time - start_time;
2384         /* this is only for bio based driver */
2385         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2386                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2387                                      bio_op(bio), lat);
2388
2389         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2390                 int bucket;
2391                 unsigned int threshold;
2392
2393                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2394                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2395                         tg->latency_target;
2396                 if (lat > threshold)
2397                         tg->bad_bio_cnt++;
2398                 /*
2399                  * Not race free, could get wrong count, which means cgroups
2400                  * will be throttled
2401                  */
2402                 tg->bio_cnt++;
2403         }
2404
2405         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2406                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2407                 tg->bio_cnt /= 2;
2408                 tg->bad_bio_cnt /= 2;
2409         }
2410 }
2411 #endif
2412
2413 int blk_throtl_init(struct request_queue *q)
2414 {
2415         struct throtl_data *td;
2416         int ret;
2417
2418         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2419         if (!td)
2420                 return -ENOMEM;
2421         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2422                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2423         if (!td->latency_buckets[READ]) {
2424                 kfree(td);
2425                 return -ENOMEM;
2426         }
2427         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2428                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2429         if (!td->latency_buckets[WRITE]) {
2430                 free_percpu(td->latency_buckets[READ]);
2431                 kfree(td);
2432                 return -ENOMEM;
2433         }
2434
2435         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2436         throtl_service_queue_init(&td->service_queue);
2437
2438         q->td = td;
2439         td->queue = q;
2440
2441         td->limit_valid[LIMIT_MAX] = true;
2442         td->limit_index = LIMIT_MAX;
2443         td->low_upgrade_time = jiffies;
2444         td->low_downgrade_time = jiffies;
2445
2446         /* activate policy */
2447         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2448         if (ret) {
2449                 free_percpu(td->latency_buckets[READ]);
2450                 free_percpu(td->latency_buckets[WRITE]);
2451                 kfree(td);
2452         }
2453         return ret;
2454 }
2455
2456 void blk_throtl_exit(struct request_queue *q)
2457 {
2458         BUG_ON(!q->td);
2459         del_timer_sync(&q->td->service_queue.pending_timer);
2460         throtl_shutdown_wq(q);
2461         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2462         free_percpu(q->td->latency_buckets[READ]);
2463         free_percpu(q->td->latency_buckets[WRITE]);
2464         kfree(q->td);
2465 }
2466
2467 void blk_throtl_register_queue(struct request_queue *q)
2468 {
2469         struct throtl_data *td;
2470         int i;
2471
2472         td = q->td;
2473         BUG_ON(!td);
2474
2475         if (blk_queue_nonrot(q)) {
2476                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2477                 td->filtered_latency = LATENCY_FILTERED_SSD;
2478         } else {
2479                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2480                 td->filtered_latency = LATENCY_FILTERED_HD;
2481                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2482                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2483                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2484                 }
2485         }
2486 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2487         /* if no low limit, use previous default */
2488         td->throtl_slice = DFL_THROTL_SLICE_HD;
2489 #endif
2490
2491         td->track_bio_latency = !queue_is_mq(q);
2492         if (!td->track_bio_latency)
2493                 blk_stat_enable_accounting(q);
2494 }
2495
2496 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2497 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2498 {
2499         if (!q->td)
2500                 return -EINVAL;
2501         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2502 }
2503
2504 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2505         const char *page, size_t count)
2506 {
2507         unsigned long v;
2508         unsigned long t;
2509
2510         if (!q->td)
2511                 return -EINVAL;
2512         if (kstrtoul(page, 10, &v))
2513                 return -EINVAL;
2514         t = msecs_to_jiffies(v);
2515         if (t == 0 || t > MAX_THROTL_SLICE)
2516                 return -EINVAL;
2517         q->td->throtl_slice = t;
2518         return count;
2519 }
2520 #endif
2521
2522 static int __init throtl_init(void)
2523 {
2524         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2525         if (!kthrotld_workqueue)
2526                 panic("Failed to create kthrotld\n");
2527
2528         return blkcg_policy_register(&blkcg_policy_throtl);
2529 }
2530
2531 module_init(throtl_init);