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 #include "blk-stat.h"
  17 #include "blk-throttle.h"
  18
  19 /* Max dispatch from a group in 1 round */
  20 #define THROTL_GRP_QUANTUM 8
  21
  22 /* Total max dispatch from all groups in one round */
  23 #define THROTL_QUANTUM 32
  24
  25 /* Throttling is performed over a slice and after that slice is renewed */
  26 #define DFL_THROTL_SLICE_HD (HZ / 10)
  27 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  28 #define MAX_THROTL_SLICE (HZ)
  29 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  30 #define MIN_THROTL_BPS (320 * 1024)
  31 #define MIN_THROTL_IOPS (10)
  32 #define DFL_LATENCY_TARGET (-1L)
  33 #define DFL_IDLE_THRESHOLD (0)
  34 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  35 #define LATENCY_FILTERED_SSD (0)
  36 /*
  37  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  38  * help determine if its IO is impacted by others, hence we ignore the IO
  39  */
  40 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  41
  42 /* A workqueue to queue throttle related work */
  43 static struct workqueue_struct *kthrotld_workqueue;
  44
  45 enum tg_state_flags {
  46         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  47         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  48 };
  49
  50 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  51
  52 /* We measure latency for request size from <= 4k to >= 1M */
  53 #define LATENCY_BUCKET_SIZE 9
  54
  55 struct latency_bucket {
  56         unsigned long total_latency; /* ns / 1024 */
  57         int samples;
  58 };
  59
  60 struct avg_latency_bucket {
  61         unsigned long latency; /* ns / 1024 */
  62         bool valid;
  63 };
  64
  65 struct throtl_data
  66 {
  67         /* service tree for active throtl groups */
  68         struct throtl_service_queue service_queue;
  69
  70         struct request_queue *queue;
  71
  72         /* Total Number of queued bios on READ and WRITE lists */
  73         unsigned int nr_queued[2];
  74
  75         unsigned int throtl_slice;
  76
  77         /* Work for dispatching throttled bios */
  78         struct work_struct dispatch_work;
  79         unsigned int limit_index;
  80         bool limit_valid[LIMIT_CNT];
  81
  82         unsigned long low_upgrade_time;
  83         unsigned long low_downgrade_time;
  84
  85         unsigned int scale;
  86
  87         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
  88         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
  89         struct latency_bucket __percpu *latency_buckets[2];
  90         unsigned long last_calculate_time;
  91         unsigned long filtered_latency;
  92
  93         bool track_bio_latency;
  94 };
  95
  96 static void throtl_pending_timer_fn(struct timer_list *t);
  97
  98 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
  99 {
 100         return pd_to_blkg(&tg->pd);
 101 }
 102
 103 /**
 104  * sq_to_tg - return the throl_grp the specified service queue belongs to
 105  * @sq: the throtl_service_queue of interest
 106  *
 107  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 108  * embedded in throtl_data, %NULL is returned.
 109  */
 110 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 111 {
 112         if (sq && sq->parent_sq)
 113                 return container_of(sq, struct throtl_grp, service_queue);
 114         else
 115                 return NULL;
 116 }
 117
 118 /**
 119  * sq_to_td - return throtl_data the specified service queue belongs to
 120  * @sq: the throtl_service_queue of interest
 121  *
 122  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 123  * Determine the associated throtl_data accordingly and return it.
 124  */
 125 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 126 {
 127         struct throtl_grp *tg = sq_to_tg(sq);
 128
 129         if (tg)
 130                 return tg->td;
 131         else
 132                 return container_of(sq, struct throtl_data, service_queue);
 133 }
 134
 135 /*
 136  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 137  * make the IO dispatch more smooth.
 138  * Scale up: linearly scale up according to lapsed time since upgrade. For
 139  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 140  *           limit hits .max limit
 141  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 142  */
 143 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 144 {
 145         /* arbitrary value to avoid too big scale */
 146         if (td->scale < 4096 && time_after_eq(jiffies,
 147             td->low_upgrade_time + td->scale * td->throtl_slice))
 148                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 149
 150         return low + (low >> 1) * td->scale;
 151 }
 152
 153 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 154 {
 155         struct blkcg_gq *blkg = tg_to_blkg(tg);
 156         struct throtl_data *td;
 157         uint64_t ret;
 158
 159         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 160                 return U64_MAX;
 161
 162         td = tg->td;
 163         ret = tg->bps[rw][td->limit_index];
 164         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 165                 /* intermediate node or iops isn't 0 */
 166                 if (!list_empty(&blkg->blkcg->css.children) ||
 167                     tg->iops[rw][td->limit_index])
 168                         return U64_MAX;
 169                 else
 170                         return MIN_THROTL_BPS;
 171         }
 172
 173         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 174             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 175                 uint64_t adjusted;
 176
 177                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 178                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 179         }
 180         return ret;
 181 }
 182
 183 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 184 {
 185         struct blkcg_gq *blkg = tg_to_blkg(tg);
 186         struct throtl_data *td;
 187         unsigned int ret;
 188
 189         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 190                 return UINT_MAX;
 191
 192         td = tg->td;
 193         ret = tg->iops[rw][td->limit_index];
 194         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 195                 /* intermediate node or bps isn't 0 */
 196                 if (!list_empty(&blkg->blkcg->css.children) ||
 197                     tg->bps[rw][td->limit_index])
 198                         return UINT_MAX;
 199                 else
 200                         return MIN_THROTL_IOPS;
 201         }
 202
 203         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 204             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 205                 uint64_t adjusted;
 206
 207                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 208                 if (adjusted > UINT_MAX)
 209                         adjusted = UINT_MAX;
 210                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 211         }
 212         return ret;
 213 }
 214
 215 #define request_bucket_index(sectors) \
 216         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 217
 218 /**
 219  * throtl_log - log debug message via blktrace
 220  * @sq: the service_queue being reported
 221  * @fmt: printf format string
 222  * @args: printf args
 223  *
 224  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 225  * throtl_grp; otherwise, just "throtl".
 226  */
 227 #define throtl_log(sq, fmt, args...)    do {                            \
 228         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 229         struct throtl_data *__td = sq_to_td((sq));                      \
 230                                                                         \
 231         (void)__td;                                                     \
 232         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 233                 break;                                                  \
 234         if ((__tg)) {                                                   \
 235                 blk_add_cgroup_trace_msg(__td->queue,                   \
 236                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 237         } else {                                                        \
 238                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 239         }                                                               \
 240 } while (0)
 241
 242 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 243 {
 244         /* assume it's one sector */
 245         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 246                 return 512;
 247         return bio->bi_iter.bi_size;
 248 }
 249
 250 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 251 {
 252         INIT_LIST_HEAD(&qn->node);
 253         bio_list_init(&qn->bios);
 254         qn->tg = tg;
 255 }
 256
 257 /**
 258  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 259  * @bio: bio being added
 260  * @qn: qnode to add bio to
 261  * @queued: the service_queue->queued[] list @qn belongs to
 262  *
 263  * Add @bio to @qn and put @qn on @queued if it's not already on.
 264  * @qn->tg's reference count is bumped when @qn is activated.  See the
 265  * comment on top of throtl_qnode definition for details.
 266  */
 267 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 268                                  struct list_head *queued)
 269 {
 270         bio_list_add(&qn->bios, bio);
 271         if (list_empty(&qn->node)) {
 272                 list_add_tail(&qn->node, queued);
 273                 blkg_get(tg_to_blkg(qn->tg));
 274         }
 275 }
 276
 277 /**
 278  * throtl_peek_queued - peek the first bio on a qnode list
 279  * @queued: the qnode list to peek
 280  */
 281 static struct bio *throtl_peek_queued(struct list_head *queued)
 282 {
 283         struct throtl_qnode *qn;
 284         struct bio *bio;
 285
 286         if (list_empty(queued))
 287                 return NULL;
 288
 289         qn = list_first_entry(queued, struct throtl_qnode, node);
 290         bio = bio_list_peek(&qn->bios);
 291         WARN_ON_ONCE(!bio);
 292         return bio;
 293 }
 294
 295 /**
 296  * throtl_pop_queued - pop the first bio form a qnode list
 297  * @queued: the qnode list to pop a bio from
 298  * @tg_to_put: optional out argument for throtl_grp to put
 299  *
 300  * Pop the first bio from the qnode list @queued.  After popping, the first
 301  * qnode is removed from @queued if empty or moved to the end of @queued so
 302  * that the popping order is round-robin.
 303  *
 304  * When the first qnode is removed, its associated throtl_grp should be put
 305  * too.  If @tg_to_put is NULL, this function automatically puts it;
 306  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 307  * responsible for putting it.
 308  */
 309 static struct bio *throtl_pop_queued(struct list_head *queued,
 310                                      struct throtl_grp **tg_to_put)
 311 {
 312         struct throtl_qnode *qn;
 313         struct bio *bio;
 314
 315         if (list_empty(queued))
 316                 return NULL;
 317
 318         qn = list_first_entry(queued, struct throtl_qnode, node);
 319         bio = bio_list_pop(&qn->bios);
 320         WARN_ON_ONCE(!bio);
 321
 322         if (bio_list_empty(&qn->bios)) {
 323                 list_del_init(&qn->node);
 324                 if (tg_to_put)
 325                         *tg_to_put = qn->tg;
 326                 else
 327                         blkg_put(tg_to_blkg(qn->tg));
 328         } else {
 329                 list_move_tail(&qn->node, queued);
 330         }
 331
 332         return bio;
 333 }
 334
 335 /* init a service_queue, assumes the caller zeroed it */
 336 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 337 {
 338         INIT_LIST_HEAD(&sq->queued[0]);
 339         INIT_LIST_HEAD(&sq->queued[1]);
 340         sq->pending_tree = RB_ROOT_CACHED;
 341         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 342 }
 343
 344 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
 345                                                 struct request_queue *q,
 346                                                 struct blkcg *blkcg)
 347 {
 348         struct throtl_grp *tg;
 349         int rw;
 350
 351         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
 352         if (!tg)
 353                 return NULL;
 354
 355         if (blkg_rwstat_init(&tg->stat_bytes, gfp))
 356                 goto err_free_tg;
 357
 358         if (blkg_rwstat_init(&tg->stat_ios, gfp))
 359                 goto err_exit_stat_bytes;
 360
 361         throtl_service_queue_init(&tg->service_queue);
 362
 363         for (rw = READ; rw <= WRITE; rw++) {
 364                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 365                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 366         }
 367
 368         RB_CLEAR_NODE(&tg->rb_node);
 369         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 370         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 371         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 372         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 373         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 374         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 375         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 376         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 377         /* LIMIT_LOW will have default value 0 */
 378
 379         tg->latency_target = DFL_LATENCY_TARGET;
 380         tg->latency_target_conf = DFL_LATENCY_TARGET;
 381         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 382         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 383
 384         return &tg->pd;
 385
 386 err_exit_stat_bytes:
 387         blkg_rwstat_exit(&tg->stat_bytes);
 388 err_free_tg:
 389         kfree(tg);
 390         return NULL;
 391 }
 392
 393 static void throtl_pd_init(struct blkg_policy_data *pd)
 394 {
 395         struct throtl_grp *tg = pd_to_tg(pd);
 396         struct blkcg_gq *blkg = tg_to_blkg(tg);
 397         struct throtl_data *td = blkg->q->td;
 398         struct throtl_service_queue *sq = &tg->service_queue;
 399
 400         /*
 401          * If on the default hierarchy, we switch to properly hierarchical
 402          * behavior where limits on a given throtl_grp are applied to the
 403          * whole subtree rather than just the group itself.  e.g. If 16M
 404          * read_bps limit is set on the root group, the whole system can't
 405          * exceed 16M for the device.
 406          *
 407          * If not on the default hierarchy, the broken flat hierarchy
 408          * behavior is retained where all throtl_grps are treated as if
 409          * they're all separate root groups right below throtl_data.
 410          * Limits of a group don't interact with limits of other groups
 411          * regardless of the position of the group in the hierarchy.
 412          */
 413         sq->parent_sq = &td->service_queue;
 414         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 415                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 416         tg->td = td;
 417 }
 418
 419 /*
 420  * Set has_rules[] if @tg or any of its parents have limits configured.
 421  * This doesn't require walking up to the top of the hierarchy as the
 422  * parent's has_rules[] is guaranteed to be correct.
 423  */
 424 static void tg_update_has_rules(struct throtl_grp *tg)
 425 {
 426         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 427         struct throtl_data *td = tg->td;
 428         int rw;
 429
 430         for (rw = READ; rw <= WRITE; rw++)
 431                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 432                         (td->limit_valid[td->limit_index] &&
 433                          (tg_bps_limit(tg, rw) != U64_MAX ||
 434                           tg_iops_limit(tg, rw) != UINT_MAX));
 435 }
 436
 437 static void throtl_pd_online(struct blkg_policy_data *pd)
 438 {
 439         struct throtl_grp *tg = pd_to_tg(pd);
 440         /*
 441          * We don't want new groups to escape the limits of its ancestors.
 442          * Update has_rules[] after a new group is brought online.
 443          */
 444         tg_update_has_rules(tg);
 445 }
 446
 447 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 448 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 449 {
 450         struct cgroup_subsys_state *pos_css;
 451         struct blkcg_gq *blkg;
 452         bool low_valid = false;
 453
 454         rcu_read_lock();
 455         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 456                 struct throtl_grp *tg = blkg_to_tg(blkg);
 457
 458                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 459                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 460                         low_valid = true;
 461                         break;
 462                 }
 463         }
 464         rcu_read_unlock();
 465
 466         td->limit_valid[LIMIT_LOW] = low_valid;
 467 }
 468 #else
 469 static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
 470 {
 471 }
 472 #endif
 473
 474 static void throtl_upgrade_state(struct throtl_data *td);
 475 static void throtl_pd_offline(struct blkg_policy_data *pd)
 476 {
 477         struct throtl_grp *tg = pd_to_tg(pd);
 478
 479         tg->bps[READ][LIMIT_LOW] = 0;
 480         tg->bps[WRITE][LIMIT_LOW] = 0;
 481         tg->iops[READ][LIMIT_LOW] = 0;
 482         tg->iops[WRITE][LIMIT_LOW] = 0;
 483
 484         blk_throtl_update_limit_valid(tg->td);
 485
 486         if (!tg->td->limit_valid[tg->td->limit_index])
 487                 throtl_upgrade_state(tg->td);
 488 }
 489
 490 static void throtl_pd_free(struct blkg_policy_data *pd)
 491 {
 492         struct throtl_grp *tg = pd_to_tg(pd);
 493
 494         del_timer_sync(&tg->service_queue.pending_timer);
 495         blkg_rwstat_exit(&tg->stat_bytes);
 496         blkg_rwstat_exit(&tg->stat_ios);
 497         kfree(tg);
 498 }
 499
 500 static struct throtl_grp *
 501 throtl_rb_first(struct throtl_service_queue *parent_sq)
 502 {
 503         struct rb_node *n;
 504
 505         n = rb_first_cached(&parent_sq->pending_tree);
 506         WARN_ON_ONCE(!n);
 507         if (!n)
 508                 return NULL;
 509         return rb_entry_tg(n);
 510 }
 511
 512 static void throtl_rb_erase(struct rb_node *n,
 513                             struct throtl_service_queue *parent_sq)
 514 {
 515         rb_erase_cached(n, &parent_sq->pending_tree);
 516         RB_CLEAR_NODE(n);
 517         --parent_sq->nr_pending;
 518 }
 519
 520 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 521 {
 522         struct throtl_grp *tg;
 523
 524         tg = throtl_rb_first(parent_sq);
 525         if (!tg)
 526                 return;
 527
 528         parent_sq->first_pending_disptime = tg->disptime;
 529 }
 530
 531 static void tg_service_queue_add(struct throtl_grp *tg)
 532 {
 533         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 534         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 535         struct rb_node *parent = NULL;
 536         struct throtl_grp *__tg;
 537         unsigned long key = tg->disptime;
 538         bool leftmost = true;
 539
 540         while (*node != NULL) {
 541                 parent = *node;
 542                 __tg = rb_entry_tg(parent);
 543
 544                 if (time_before(key, __tg->disptime))
 545                         node = &parent->rb_left;
 546                 else {
 547                         node = &parent->rb_right;
 548                         leftmost = false;
 549                 }
 550         }
 551
 552         rb_link_node(&tg->rb_node, parent, node);
 553         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 554                                leftmost);
 555 }
 556
 557 static void throtl_enqueue_tg(struct throtl_grp *tg)
 558 {
 559         if (!(tg->flags & THROTL_TG_PENDING)) {
 560                 tg_service_queue_add(tg);
 561                 tg->flags |= THROTL_TG_PENDING;
 562                 tg->service_queue.parent_sq->nr_pending++;
 563         }
 564 }
 565
 566 static void throtl_dequeue_tg(struct throtl_grp *tg)
 567 {
 568         if (tg->flags & THROTL_TG_PENDING) {
 569                 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 570                 tg->flags &= ~THROTL_TG_PENDING;
 571         }
 572 }
 573
 574 /* Call with queue lock held */
 575 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 576                                           unsigned long expires)
 577 {
 578         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 579
 580         /*
 581          * Since we are adjusting the throttle limit dynamically, the sleep
 582          * time calculated according to previous limit might be invalid. It's
 583          * possible the cgroup sleep time is very long and no other cgroups
 584          * have IO running so notify the limit changes. Make sure the cgroup
 585          * doesn't sleep too long to avoid the missed notification.
 586          */
 587         if (time_after(expires, max_expire))
 588                 expires = max_expire;
 589         mod_timer(&sq->pending_timer, expires);
 590         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 591                    expires - jiffies, jiffies);
 592 }
 593
 594 /**
 595  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 596  * @sq: the service_queue to schedule dispatch for
 597  * @force: force scheduling
 598  *
 599  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 600  * dispatch time of the first pending child.  Returns %true if either timer
 601  * is armed or there's no pending child left.  %false if the current
 602  * dispatch window is still open and the caller should continue
 603  * dispatching.
 604  *
 605  * If @force is %true, the dispatch timer is always scheduled and this
 606  * function is guaranteed to return %true.  This is to be used when the
 607  * caller can't dispatch itself and needs to invoke pending_timer
 608  * unconditionally.  Note that forced scheduling is likely to induce short
 609  * delay before dispatch starts even if @sq->first_pending_disptime is not
 610  * in the future and thus shouldn't be used in hot paths.
 611  */
 612 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 613                                           bool force)
 614 {
 615         /* any pending children left? */
 616         if (!sq->nr_pending)
 617                 return true;
 618
 619         update_min_dispatch_time(sq);
 620
 621         /* is the next dispatch time in the future? */
 622         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 623                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 624                 return true;
 625         }
 626
 627         /* tell the caller to continue dispatching */
 628         return false;
 629 }
 630
 631 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 632                 bool rw, unsigned long start)
 633 {
 634         tg->bytes_disp[rw] = 0;
 635         tg->io_disp[rw] = 0;
 636
 637         atomic_set(&tg->io_split_cnt[rw], 0);
 638
 639         /*
 640          * Previous slice has expired. We must have trimmed it after last
 641          * bio dispatch. That means since start of last slice, we never used
 642          * that bandwidth. Do try to make use of that bandwidth while giving
 643          * credit.
 644          */
 645         if (time_after_eq(start, tg->slice_start[rw]))
 646                 tg->slice_start[rw] = start;
 647
 648         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 649         throtl_log(&tg->service_queue,
 650                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 651                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 652                    tg->slice_end[rw], jiffies);
 653 }
 654
 655 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 656 {
 657         tg->bytes_disp[rw] = 0;
 658         tg->io_disp[rw] = 0;
 659         tg->slice_start[rw] = jiffies;
 660         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 661
 662         atomic_set(&tg->io_split_cnt[rw], 0);
 663
 664         throtl_log(&tg->service_queue,
 665                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 666                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 667                    tg->slice_end[rw], jiffies);
 668 }
 669
 670 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 671                                         unsigned long jiffy_end)
 672 {
 673         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 674 }
 675
 676 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 677                                        unsigned long jiffy_end)
 678 {
 679         throtl_set_slice_end(tg, rw, jiffy_end);
 680         throtl_log(&tg->service_queue,
 681                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 682                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 683                    tg->slice_end[rw], jiffies);
 684 }
 685
 686 /* Determine if previously allocated or extended slice is complete or not */
 687 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 688 {
 689         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 690                 return false;
 691
 692         return true;
 693 }
 694
 695 /* Trim the used slices and adjust slice start accordingly */
 696 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 697 {
 698         unsigned long nr_slices, time_elapsed, io_trim;
 699         u64 bytes_trim, tmp;
 700
 701         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 702
 703         /*
 704          * If bps are unlimited (-1), then time slice don't get
 705          * renewed. Don't try to trim the slice if slice is used. A new
 706          * slice will start when appropriate.
 707          */
 708         if (throtl_slice_used(tg, rw))
 709                 return;
 710
 711         /*
 712          * A bio has been dispatched. Also adjust slice_end. It might happen
 713          * that initially cgroup limit was very low resulting in high
 714          * slice_end, but later limit was bumped up and bio was dispatched
 715          * sooner, then we need to reduce slice_end. A high bogus slice_end
 716          * is bad because it does not allow new slice to start.
 717          */
 718
 719         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 720
 721         time_elapsed = jiffies - tg->slice_start[rw];
 722
 723         nr_slices = time_elapsed / tg->td->throtl_slice;
 724
 725         if (!nr_slices)
 726                 return;
 727         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 728         do_div(tmp, HZ);
 729         bytes_trim = tmp;
 730
 731         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 732                 HZ;
 733
 734         if (!bytes_trim && !io_trim)
 735                 return;
 736
 737         if (tg->bytes_disp[rw] >= bytes_trim)
 738                 tg->bytes_disp[rw] -= bytes_trim;
 739         else
 740                 tg->bytes_disp[rw] = 0;
 741
 742         if (tg->io_disp[rw] >= io_trim)
 743                 tg->io_disp[rw] -= io_trim;
 744         else
 745                 tg->io_disp[rw] = 0;
 746
 747         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 748
 749         throtl_log(&tg->service_queue,
 750                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 751                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 752                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 753 }
 754
 755 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 756                                   u32 iops_limit, unsigned long *wait)
 757 {
 758         bool rw = bio_data_dir(bio);
 759         unsigned int io_allowed;
 760         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 761         u64 tmp;
 762
 763         if (iops_limit == UINT_MAX) {
 764                 if (wait)
 765                         *wait = 0;
 766                 return true;
 767         }
 768
 769         jiffy_elapsed = jiffies - tg->slice_start[rw];
 770
 771         /* Round up to the next throttle slice, wait time must be nonzero */
 772         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 773
 774         /*
 775          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 776          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 777          * will allow dispatch after 1 second and after that slice should
 778          * have been trimmed.
 779          */
 780
 781         tmp = (u64)iops_limit * jiffy_elapsed_rnd;
 782         do_div(tmp, HZ);
 783
 784         if (tmp > UINT_MAX)
 785                 io_allowed = UINT_MAX;
 786         else
 787                 io_allowed = tmp;
 788
 789         if (tg->io_disp[rw] + 1 <= io_allowed) {
 790                 if (wait)
 791                         *wait = 0;
 792                 return true;
 793         }
 794
 795         /* Calc approx time to dispatch */
 796         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 797
 798         if (wait)
 799                 *wait = jiffy_wait;
 800         return false;
 801 }
 802
 803 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 804                                  u64 bps_limit, unsigned long *wait)
 805 {
 806         bool rw = bio_data_dir(bio);
 807         u64 bytes_allowed, extra_bytes, tmp;
 808         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 809         unsigned int bio_size = throtl_bio_data_size(bio);
 810
 811         /* no need to throttle if this bio's bytes have been accounted */
 812         if (bps_limit == U64_MAX || bio_flagged(bio, BIO_THROTTLED)) {
 813                 if (wait)
 814                         *wait = 0;
 815                 return true;
 816         }
 817
 818         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 819
 820         /* Slice has just started. Consider one slice interval */
 821         if (!jiffy_elapsed)
 822                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 823
 824         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 825
 826         tmp = bps_limit * jiffy_elapsed_rnd;
 827         do_div(tmp, HZ);
 828         bytes_allowed = tmp;
 829
 830         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 831                 if (wait)
 832                         *wait = 0;
 833                 return true;
 834         }
 835
 836         /* Calc approx time to dispatch */
 837         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 838         jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit);
 839
 840         if (!jiffy_wait)
 841                 jiffy_wait = 1;
 842
 843         /*
 844          * This wait time is without taking into consideration the rounding
 845          * up we did. Add that time also.
 846          */
 847         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 848         if (wait)
 849                 *wait = jiffy_wait;
 850         return false;
 851 }
 852
 853 /*
 854  * Returns whether one can dispatch a bio or not. Also returns approx number
 855  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 856  */
 857 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 858                             unsigned long *wait)
 859 {
 860         bool rw = bio_data_dir(bio);
 861         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 862         u64 bps_limit = tg_bps_limit(tg, rw);
 863         u32 iops_limit = tg_iops_limit(tg, rw);
 864
 865         /*
 866          * Currently whole state machine of group depends on first bio
 867          * queued in the group bio list. So one should not be calling
 868          * this function with a different bio if there are other bios
 869          * queued.
 870          */
 871         BUG_ON(tg->service_queue.nr_queued[rw] &&
 872                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 873
 874         /* If tg->bps = -1, then BW is unlimited */
 875         if (bps_limit == U64_MAX && iops_limit == UINT_MAX) {
 876                 if (wait)
 877                         *wait = 0;
 878                 return true;
 879         }
 880
 881         /*
 882          * If previous slice expired, start a new one otherwise renew/extend
 883          * existing slice to make sure it is at least throtl_slice interval
 884          * long since now. New slice is started only for empty throttle group.
 885          * If there is queued bio, that means there should be an active
 886          * slice and it should be extended instead.
 887          */
 888         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 889                 throtl_start_new_slice(tg, rw);
 890         else {
 891                 if (time_before(tg->slice_end[rw],
 892                     jiffies + tg->td->throtl_slice))
 893                         throtl_extend_slice(tg, rw,
 894                                 jiffies + tg->td->throtl_slice);
 895         }
 896
 897         if (iops_limit != UINT_MAX)
 898                 tg->io_disp[rw] += atomic_xchg(&tg->io_split_cnt[rw], 0);
 899
 900         if (tg_with_in_bps_limit(tg, bio, bps_limit, &bps_wait) &&
 901             tg_with_in_iops_limit(tg, bio, iops_limit, &iops_wait)) {
 902                 if (wait)
 903                         *wait = 0;
 904                 return true;
 905         }
 906
 907         max_wait = max(bps_wait, iops_wait);
 908
 909         if (wait)
 910                 *wait = max_wait;
 911
 912         if (time_before(tg->slice_end[rw], jiffies + max_wait))
 913                 throtl_extend_slice(tg, rw, jiffies + max_wait);
 914
 915         return false;
 916 }
 917
 918 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
 919 {
 920         bool rw = bio_data_dir(bio);
 921         unsigned int bio_size = throtl_bio_data_size(bio);
 922
 923         /* Charge the bio to the group */
 924         if (!bio_flagged(bio, BIO_THROTTLED)) {
 925                 tg->bytes_disp[rw] += bio_size;
 926                 tg->last_bytes_disp[rw] += bio_size;
 927         }
 928
 929         tg->io_disp[rw]++;
 930         tg->last_io_disp[rw]++;
 931
 932         /*
 933          * BIO_THROTTLED is used to prevent the same bio to be throttled
 934          * more than once as a throttled bio will go through blk-throtl the
 935          * second time when it eventually gets issued.  Set it when a bio
 936          * is being charged to a tg.
 937          */
 938         if (!bio_flagged(bio, BIO_THROTTLED))
 939                 bio_set_flag(bio, BIO_THROTTLED);
 940 }
 941
 942 /**
 943  * throtl_add_bio_tg - add a bio to the specified throtl_grp
 944  * @bio: bio to add
 945  * @qn: qnode to use
 946  * @tg: the target throtl_grp
 947  *
 948  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
 949  * tg->qnode_on_self[] is used.
 950  */
 951 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
 952                               struct throtl_grp *tg)
 953 {
 954         struct throtl_service_queue *sq = &tg->service_queue;
 955         bool rw = bio_data_dir(bio);
 956
 957         if (!qn)
 958                 qn = &tg->qnode_on_self[rw];
 959
 960         /*
 961          * If @tg doesn't currently have any bios queued in the same
 962          * direction, queueing @bio can change when @tg should be
 963          * dispatched.  Mark that @tg was empty.  This is automatically
 964          * cleared on the next tg_update_disptime().
 965          */
 966         if (!sq->nr_queued[rw])
 967                 tg->flags |= THROTL_TG_WAS_EMPTY;
 968
 969         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
 970
 971         sq->nr_queued[rw]++;
 972         throtl_enqueue_tg(tg);
 973 }
 974
 975 static void tg_update_disptime(struct throtl_grp *tg)
 976 {
 977         struct throtl_service_queue *sq = &tg->service_queue;
 978         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
 979         struct bio *bio;
 980
 981         bio = throtl_peek_queued(&sq->queued[READ]);
 982         if (bio)
 983                 tg_may_dispatch(tg, bio, &read_wait);
 984
 985         bio = throtl_peek_queued(&sq->queued[WRITE]);
 986         if (bio)
 987                 tg_may_dispatch(tg, bio, &write_wait);
 988
 989         min_wait = min(read_wait, write_wait);
 990         disptime = jiffies + min_wait;
 991
 992         /* Update dispatch time */
 993         throtl_dequeue_tg(tg);
 994         tg->disptime = disptime;
 995         throtl_enqueue_tg(tg);
 996
 997         /* see throtl_add_bio_tg() */
 998         tg->flags &= ~THROTL_TG_WAS_EMPTY;
 999 }
1000
1001 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1002                                         struct throtl_grp *parent_tg, bool rw)
1003 {
1004         if (throtl_slice_used(parent_tg, rw)) {
1005                 throtl_start_new_slice_with_credit(parent_tg, rw,
1006                                 child_tg->slice_start[rw]);
1007         }
1008
1009 }
1010
1011 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1012 {
1013         struct throtl_service_queue *sq = &tg->service_queue;
1014         struct throtl_service_queue *parent_sq = sq->parent_sq;
1015         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1016         struct throtl_grp *tg_to_put = NULL;
1017         struct bio *bio;
1018
1019         /*
1020          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1021          * from @tg may put its reference and @parent_sq might end up
1022          * getting released prematurely.  Remember the tg to put and put it
1023          * after @bio is transferred to @parent_sq.
1024          */
1025         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1026         sq->nr_queued[rw]--;
1027
1028         throtl_charge_bio(tg, bio);
1029
1030         /*
1031          * If our parent is another tg, we just need to transfer @bio to
1032          * the parent using throtl_add_bio_tg().  If our parent is
1033          * @td->service_queue, @bio is ready to be issued.  Put it on its
1034          * bio_lists[] and decrease total number queued.  The caller is
1035          * responsible for issuing these bios.
1036          */
1037         if (parent_tg) {
1038                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1039                 start_parent_slice_with_credit(tg, parent_tg, rw);
1040         } else {
1041                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1042                                      &parent_sq->queued[rw]);
1043                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1044                 tg->td->nr_queued[rw]--;
1045         }
1046
1047         throtl_trim_slice(tg, rw);
1048
1049         if (tg_to_put)
1050                 blkg_put(tg_to_blkg(tg_to_put));
1051 }
1052
1053 static int throtl_dispatch_tg(struct throtl_grp *tg)
1054 {
1055         struct throtl_service_queue *sq = &tg->service_queue;
1056         unsigned int nr_reads = 0, nr_writes = 0;
1057         unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4;
1058         unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads;
1059         struct bio *bio;
1060
1061         /* Try to dispatch 75% READS and 25% WRITES */
1062
1063         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1064                tg_may_dispatch(tg, bio, NULL)) {
1065
1066                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1067                 nr_reads++;
1068
1069                 if (nr_reads >= max_nr_reads)
1070                         break;
1071         }
1072
1073         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1074                tg_may_dispatch(tg, bio, NULL)) {
1075
1076                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1077                 nr_writes++;
1078
1079                 if (nr_writes >= max_nr_writes)
1080                         break;
1081         }
1082
1083         return nr_reads + nr_writes;
1084 }
1085
1086 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1087 {
1088         unsigned int nr_disp = 0;
1089
1090         while (1) {
1091                 struct throtl_grp *tg;
1092                 struct throtl_service_queue *sq;
1093
1094                 if (!parent_sq->nr_pending)
1095                         break;
1096
1097                 tg = throtl_rb_first(parent_sq);
1098                 if (!tg)
1099                         break;
1100
1101                 if (time_before(jiffies, tg->disptime))
1102                         break;
1103
1104                 throtl_dequeue_tg(tg);
1105
1106                 nr_disp += throtl_dispatch_tg(tg);
1107
1108                 sq = &tg->service_queue;
1109                 if (sq->nr_queued[0] || sq->nr_queued[1])
1110                         tg_update_disptime(tg);
1111
1112                 if (nr_disp >= THROTL_QUANTUM)
1113                         break;
1114         }
1115
1116         return nr_disp;
1117 }
1118
1119 static bool throtl_can_upgrade(struct throtl_data *td,
1120         struct throtl_grp *this_tg);
1121 /**
1122  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1123  * @t: the pending_timer member of the throtl_service_queue being serviced
1124  *
1125  * This timer is armed when a child throtl_grp with active bio's become
1126  * pending and queued on the service_queue's pending_tree and expires when
1127  * the first child throtl_grp should be dispatched.  This function
1128  * dispatches bio's from the children throtl_grps to the parent
1129  * service_queue.
1130  *
1131  * If the parent's parent is another throtl_grp, dispatching is propagated
1132  * by either arming its pending_timer or repeating dispatch directly.  If
1133  * the top-level service_tree is reached, throtl_data->dispatch_work is
1134  * kicked so that the ready bio's are issued.
1135  */
1136 static void throtl_pending_timer_fn(struct timer_list *t)
1137 {
1138         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1139         struct throtl_grp *tg = sq_to_tg(sq);
1140         struct throtl_data *td = sq_to_td(sq);
1141         struct request_queue *q = td->queue;
1142         struct throtl_service_queue *parent_sq;
1143         bool dispatched;
1144         int ret;
1145
1146         spin_lock_irq(&q->queue_lock);
1147         if (throtl_can_upgrade(td, NULL))
1148                 throtl_upgrade_state(td);
1149
1150 again:
1151         parent_sq = sq->parent_sq;
1152         dispatched = false;
1153
1154         while (true) {
1155                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1156                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1157                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1158
1159                 ret = throtl_select_dispatch(sq);
1160                 if (ret) {
1161                         throtl_log(sq, "bios disp=%u", ret);
1162                         dispatched = true;
1163                 }
1164
1165                 if (throtl_schedule_next_dispatch(sq, false))
1166                         break;
1167
1168                 /* this dispatch windows is still open, relax and repeat */
1169                 spin_unlock_irq(&q->queue_lock);
1170                 cpu_relax();
1171                 spin_lock_irq(&q->queue_lock);
1172         }
1173
1174         if (!dispatched)
1175                 goto out_unlock;
1176
1177         if (parent_sq) {
1178                 /* @parent_sq is another throl_grp, propagate dispatch */
1179                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1180                         tg_update_disptime(tg);
1181                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1182                                 /* window is already open, repeat dispatching */
1183                                 sq = parent_sq;
1184                                 tg = sq_to_tg(sq);
1185                                 goto again;
1186                         }
1187                 }
1188         } else {
1189                 /* reached the top-level, queue issuing */
1190                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1191         }
1192 out_unlock:
1193         spin_unlock_irq(&q->queue_lock);
1194 }
1195
1196 /**
1197  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1198  * @work: work item being executed
1199  *
1200  * This function is queued for execution when bios reach the bio_lists[]
1201  * of throtl_data->service_queue.  Those bios are ready and issued by this
1202  * function.
1203  */
1204 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1205 {
1206         struct throtl_data *td = container_of(work, struct throtl_data,
1207                                               dispatch_work);
1208         struct throtl_service_queue *td_sq = &td->service_queue;
1209         struct request_queue *q = td->queue;
1210         struct bio_list bio_list_on_stack;
1211         struct bio *bio;
1212         struct blk_plug plug;
1213         int rw;
1214
1215         bio_list_init(&bio_list_on_stack);
1216
1217         spin_lock_irq(&q->queue_lock);
1218         for (rw = READ; rw <= WRITE; rw++)
1219                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1220                         bio_list_add(&bio_list_on_stack, bio);
1221         spin_unlock_irq(&q->queue_lock);
1222
1223         if (!bio_list_empty(&bio_list_on_stack)) {
1224                 blk_start_plug(&plug);
1225                 while ((bio = bio_list_pop(&bio_list_on_stack)))
1226                         submit_bio_noacct(bio);
1227                 blk_finish_plug(&plug);
1228         }
1229 }
1230
1231 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1232                               int off)
1233 {
1234         struct throtl_grp *tg = pd_to_tg(pd);
1235         u64 v = *(u64 *)((void *)tg + off);
1236
1237         if (v == U64_MAX)
1238                 return 0;
1239         return __blkg_prfill_u64(sf, pd, v);
1240 }
1241
1242 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1243                                int off)
1244 {
1245         struct throtl_grp *tg = pd_to_tg(pd);
1246         unsigned int v = *(unsigned int *)((void *)tg + off);
1247
1248         if (v == UINT_MAX)
1249                 return 0;
1250         return __blkg_prfill_u64(sf, pd, v);
1251 }
1252
1253 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1254 {
1255         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1256                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1257         return 0;
1258 }
1259
1260 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1261 {
1262         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1263                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1264         return 0;
1265 }
1266
1267 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1268 {
1269         struct throtl_service_queue *sq = &tg->service_queue;
1270         struct cgroup_subsys_state *pos_css;
1271         struct blkcg_gq *blkg;
1272
1273         throtl_log(&tg->service_queue,
1274                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1275                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1276                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1277
1278         /*
1279          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1280          * considered to have rules if either the tg itself or any of its
1281          * ancestors has rules.  This identifies groups without any
1282          * restrictions in the whole hierarchy and allows them to bypass
1283          * blk-throttle.
1284          */
1285         blkg_for_each_descendant_pre(blkg, pos_css,
1286                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1287                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1288                 struct throtl_grp *parent_tg;
1289
1290                 tg_update_has_rules(this_tg);
1291                 /* ignore root/second level */
1292                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1293                     !blkg->parent->parent)
1294                         continue;
1295                 parent_tg = blkg_to_tg(blkg->parent);
1296                 /*
1297                  * make sure all children has lower idle time threshold and
1298                  * higher latency target
1299                  */
1300                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1301                                 parent_tg->idletime_threshold);
1302                 this_tg->latency_target = max(this_tg->latency_target,
1303                                 parent_tg->latency_target);
1304         }
1305
1306         /*
1307          * We're already holding queue_lock and know @tg is valid.  Let's
1308          * apply the new config directly.
1309          *
1310          * Restart the slices for both READ and WRITES. It might happen
1311          * that a group's limit are dropped suddenly and we don't want to
1312          * account recently dispatched IO with new low rate.
1313          */
1314         throtl_start_new_slice(tg, READ);
1315         throtl_start_new_slice(tg, WRITE);
1316
1317         if (tg->flags & THROTL_TG_PENDING) {
1318                 tg_update_disptime(tg);
1319                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1320         }
1321 }
1322
1323 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1324                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1325 {
1326         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1327         struct blkg_conf_ctx ctx;
1328         struct throtl_grp *tg;
1329         int ret;
1330         u64 v;
1331
1332         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1333         if (ret)
1334                 return ret;
1335
1336         ret = -EINVAL;
1337         if (sscanf(ctx.body, "%llu", &v) != 1)
1338                 goto out_finish;
1339         if (!v)
1340                 v = U64_MAX;
1341
1342         tg = blkg_to_tg(ctx.blkg);
1343
1344         if (is_u64)
1345                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1346         else
1347                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1348
1349         tg_conf_updated(tg, false);
1350         ret = 0;
1351 out_finish:
1352         blkg_conf_finish(&ctx);
1353         return ret ?: nbytes;
1354 }
1355
1356 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1357                                char *buf, size_t nbytes, loff_t off)
1358 {
1359         return tg_set_conf(of, buf, nbytes, off, true);
1360 }
1361
1362 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1363                                 char *buf, size_t nbytes, loff_t off)
1364 {
1365         return tg_set_conf(of, buf, nbytes, off, false);
1366 }
1367
1368 static int tg_print_rwstat(struct seq_file *sf, void *v)
1369 {
1370         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1371                           blkg_prfill_rwstat, &blkcg_policy_throtl,
1372                           seq_cft(sf)->private, true);
1373         return 0;
1374 }
1375
1376 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf,
1377                                       struct blkg_policy_data *pd, int off)
1378 {
1379         struct blkg_rwstat_sample sum;
1380
1381         blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off,
1382                                   &sum);
1383         return __blkg_prfill_rwstat(sf, pd, &sum);
1384 }
1385
1386 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v)
1387 {
1388         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1389                           tg_prfill_rwstat_recursive, &blkcg_policy_throtl,
1390                           seq_cft(sf)->private, true);
1391         return 0;
1392 }
1393
1394 static struct cftype throtl_legacy_files[] = {
1395         {
1396                 .name = "throttle.read_bps_device",
1397                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1398                 .seq_show = tg_print_conf_u64,
1399                 .write = tg_set_conf_u64,
1400         },
1401         {
1402                 .name = "throttle.write_bps_device",
1403                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1404                 .seq_show = tg_print_conf_u64,
1405                 .write = tg_set_conf_u64,
1406         },
1407         {
1408                 .name = "throttle.read_iops_device",
1409                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1410                 .seq_show = tg_print_conf_uint,
1411                 .write = tg_set_conf_uint,
1412         },
1413         {
1414                 .name = "throttle.write_iops_device",
1415                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1416                 .seq_show = tg_print_conf_uint,
1417                 .write = tg_set_conf_uint,
1418         },
1419         {
1420                 .name = "throttle.io_service_bytes",
1421                 .private = offsetof(struct throtl_grp, stat_bytes),
1422                 .seq_show = tg_print_rwstat,
1423         },
1424         {
1425                 .name = "throttle.io_service_bytes_recursive",
1426                 .private = offsetof(struct throtl_grp, stat_bytes),
1427                 .seq_show = tg_print_rwstat_recursive,
1428         },
1429         {
1430                 .name = "throttle.io_serviced",
1431                 .private = offsetof(struct throtl_grp, stat_ios),
1432                 .seq_show = tg_print_rwstat,
1433         },
1434         {
1435                 .name = "throttle.io_serviced_recursive",
1436                 .private = offsetof(struct throtl_grp, stat_ios),
1437                 .seq_show = tg_print_rwstat_recursive,
1438         },
1439         { }     /* terminate */
1440 };
1441
1442 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1443                          int off)
1444 {
1445         struct throtl_grp *tg = pd_to_tg(pd);
1446         const char *dname = blkg_dev_name(pd->blkg);
1447         char bufs[4][21] = { "max", "max", "max", "max" };
1448         u64 bps_dft;
1449         unsigned int iops_dft;
1450         char idle_time[26] = "";
1451         char latency_time[26] = "";
1452
1453         if (!dname)
1454                 return 0;
1455
1456         if (off == LIMIT_LOW) {
1457                 bps_dft = 0;
1458                 iops_dft = 0;
1459         } else {
1460                 bps_dft = U64_MAX;
1461                 iops_dft = UINT_MAX;
1462         }
1463
1464         if (tg->bps_conf[READ][off] == bps_dft &&
1465             tg->bps_conf[WRITE][off] == bps_dft &&
1466             tg->iops_conf[READ][off] == iops_dft &&
1467             tg->iops_conf[WRITE][off] == iops_dft &&
1468             (off != LIMIT_LOW ||
1469              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1470               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1471                 return 0;
1472
1473         if (tg->bps_conf[READ][off] != U64_MAX)
1474                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1475                         tg->bps_conf[READ][off]);
1476         if (tg->bps_conf[WRITE][off] != U64_MAX)
1477                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1478                         tg->bps_conf[WRITE][off]);
1479         if (tg->iops_conf[READ][off] != UINT_MAX)
1480                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1481                         tg->iops_conf[READ][off]);
1482         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1483                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1484                         tg->iops_conf[WRITE][off]);
1485         if (off == LIMIT_LOW) {
1486                 if (tg->idletime_threshold_conf == ULONG_MAX)
1487                         strcpy(idle_time, " idle=max");
1488                 else
1489                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1490                                 tg->idletime_threshold_conf);
1491
1492                 if (tg->latency_target_conf == ULONG_MAX)
1493                         strcpy(latency_time, " latency=max");
1494                 else
1495                         snprintf(latency_time, sizeof(latency_time),
1496                                 " latency=%lu", tg->latency_target_conf);
1497         }
1498
1499         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1500                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1501                    latency_time);
1502         return 0;
1503 }
1504
1505 static int tg_print_limit(struct seq_file *sf, void *v)
1506 {
1507         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1508                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1509         return 0;
1510 }
1511
1512 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1513                           char *buf, size_t nbytes, loff_t off)
1514 {
1515         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1516         struct blkg_conf_ctx ctx;
1517         struct throtl_grp *tg;
1518         u64 v[4];
1519         unsigned long idle_time;
1520         unsigned long latency_time;
1521         int ret;
1522         int index = of_cft(of)->private;
1523
1524         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1525         if (ret)
1526                 return ret;
1527
1528         tg = blkg_to_tg(ctx.blkg);
1529
1530         v[0] = tg->bps_conf[READ][index];
1531         v[1] = tg->bps_conf[WRITE][index];
1532         v[2] = tg->iops_conf[READ][index];
1533         v[3] = tg->iops_conf[WRITE][index];
1534
1535         idle_time = tg->idletime_threshold_conf;
1536         latency_time = tg->latency_target_conf;
1537         while (true) {
1538                 char tok[27];   /* wiops=18446744073709551616 */
1539                 char *p;
1540                 u64 val = U64_MAX;
1541                 int len;
1542
1543                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1544                         break;
1545                 if (tok[0] == '\0')
1546                         break;
1547                 ctx.body += len;
1548
1549                 ret = -EINVAL;
1550                 p = tok;
1551                 strsep(&p, "=");
1552                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1553                         goto out_finish;
1554
1555                 ret = -ERANGE;
1556                 if (!val)
1557                         goto out_finish;
1558
1559                 ret = -EINVAL;
1560                 if (!strcmp(tok, "rbps") && val > 1)
1561                         v[0] = val;
1562                 else if (!strcmp(tok, "wbps") && val > 1)
1563                         v[1] = val;
1564                 else if (!strcmp(tok, "riops") && val > 1)
1565                         v[2] = min_t(u64, val, UINT_MAX);
1566                 else if (!strcmp(tok, "wiops") && val > 1)
1567                         v[3] = min_t(u64, val, UINT_MAX);
1568                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1569                         idle_time = val;
1570                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1571                         latency_time = val;
1572                 else
1573                         goto out_finish;
1574         }
1575
1576         tg->bps_conf[READ][index] = v[0];
1577         tg->bps_conf[WRITE][index] = v[1];
1578         tg->iops_conf[READ][index] = v[2];
1579         tg->iops_conf[WRITE][index] = v[3];
1580
1581         if (index == LIMIT_MAX) {
1582                 tg->bps[READ][index] = v[0];
1583                 tg->bps[WRITE][index] = v[1];
1584                 tg->iops[READ][index] = v[2];
1585                 tg->iops[WRITE][index] = v[3];
1586         }
1587         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1588                 tg->bps_conf[READ][LIMIT_MAX]);
1589         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1590                 tg->bps_conf[WRITE][LIMIT_MAX]);
1591         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1592                 tg->iops_conf[READ][LIMIT_MAX]);
1593         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1594                 tg->iops_conf[WRITE][LIMIT_MAX]);
1595         tg->idletime_threshold_conf = idle_time;
1596         tg->latency_target_conf = latency_time;
1597
1598         /* force user to configure all settings for low limit  */
1599         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1600               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1601             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1602             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1603                 tg->bps[READ][LIMIT_LOW] = 0;
1604                 tg->bps[WRITE][LIMIT_LOW] = 0;
1605                 tg->iops[READ][LIMIT_LOW] = 0;
1606                 tg->iops[WRITE][LIMIT_LOW] = 0;
1607                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1608                 tg->latency_target = DFL_LATENCY_TARGET;
1609         } else if (index == LIMIT_LOW) {
1610                 tg->idletime_threshold = tg->idletime_threshold_conf;
1611                 tg->latency_target = tg->latency_target_conf;
1612         }
1613
1614         blk_throtl_update_limit_valid(tg->td);
1615         if (tg->td->limit_valid[LIMIT_LOW]) {
1616                 if (index == LIMIT_LOW)
1617                         tg->td->limit_index = LIMIT_LOW;
1618         } else
1619                 tg->td->limit_index = LIMIT_MAX;
1620         tg_conf_updated(tg, index == LIMIT_LOW &&
1621                 tg->td->limit_valid[LIMIT_LOW]);
1622         ret = 0;
1623 out_finish:
1624         blkg_conf_finish(&ctx);
1625         return ret ?: nbytes;
1626 }
1627
1628 static struct cftype throtl_files[] = {
1629 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1630         {
1631                 .name = "low",
1632                 .flags = CFTYPE_NOT_ON_ROOT,
1633                 .seq_show = tg_print_limit,
1634                 .write = tg_set_limit,
1635                 .private = LIMIT_LOW,
1636         },
1637 #endif
1638         {
1639                 .name = "max",
1640                 .flags = CFTYPE_NOT_ON_ROOT,
1641                 .seq_show = tg_print_limit,
1642                 .write = tg_set_limit,
1643                 .private = LIMIT_MAX,
1644         },
1645         { }     /* terminate */
1646 };
1647
1648 static void throtl_shutdown_wq(struct request_queue *q)
1649 {
1650         struct throtl_data *td = q->td;
1651
1652         cancel_work_sync(&td->dispatch_work);
1653 }
1654
1655 struct blkcg_policy blkcg_policy_throtl = {
1656         .dfl_cftypes            = throtl_files,
1657         .legacy_cftypes         = throtl_legacy_files,
1658
1659         .pd_alloc_fn            = throtl_pd_alloc,
1660         .pd_init_fn             = throtl_pd_init,
1661         .pd_online_fn           = throtl_pd_online,
1662         .pd_offline_fn          = throtl_pd_offline,
1663         .pd_free_fn             = throtl_pd_free,
1664 };
1665
1666 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1667 {
1668         unsigned long rtime = jiffies, wtime = jiffies;
1669
1670         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1671                 rtime = tg->last_low_overflow_time[READ];
1672         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1673                 wtime = tg->last_low_overflow_time[WRITE];
1674         return min(rtime, wtime);
1675 }
1676
1677 /* tg should not be an intermediate node */
1678 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1679 {
1680         struct throtl_service_queue *parent_sq;
1681         struct throtl_grp *parent = tg;
1682         unsigned long ret = __tg_last_low_overflow_time(tg);
1683
1684         while (true) {
1685                 parent_sq = parent->service_queue.parent_sq;
1686                 parent = sq_to_tg(parent_sq);
1687                 if (!parent)
1688                         break;
1689
1690                 /*
1691                  * The parent doesn't have low limit, it always reaches low
1692                  * limit. Its overflow time is useless for children
1693                  */
1694                 if (!parent->bps[READ][LIMIT_LOW] &&
1695                     !parent->iops[READ][LIMIT_LOW] &&
1696                     !parent->bps[WRITE][LIMIT_LOW] &&
1697                     !parent->iops[WRITE][LIMIT_LOW])
1698                         continue;
1699                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1700                         ret = __tg_last_low_overflow_time(parent);
1701         }
1702         return ret;
1703 }
1704
1705 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1706 {
1707         /*
1708          * cgroup is idle if:
1709          * - single idle is too long, longer than a fixed value (in case user
1710          *   configure a too big threshold) or 4 times of idletime threshold
1711          * - average think time is more than threshold
1712          * - IO latency is largely below threshold
1713          */
1714         unsigned long time;
1715         bool ret;
1716
1717         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1718         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1719               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1720               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1721               tg->avg_idletime > tg->idletime_threshold ||
1722               (tg->latency_target && tg->bio_cnt &&
1723                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1724         throtl_log(&tg->service_queue,
1725                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1726                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1727                 tg->bio_cnt, ret, tg->td->scale);
1728         return ret;
1729 }
1730
1731 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1732 {
1733         struct throtl_service_queue *sq = &tg->service_queue;
1734         bool read_limit, write_limit;
1735
1736         /*
1737          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1738          * reaches), it's ok to upgrade to next limit
1739          */
1740         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1741         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1742         if (!read_limit && !write_limit)
1743                 return true;
1744         if (read_limit && sq->nr_queued[READ] &&
1745             (!write_limit || sq->nr_queued[WRITE]))
1746                 return true;
1747         if (write_limit && sq->nr_queued[WRITE] &&
1748             (!read_limit || sq->nr_queued[READ]))
1749                 return true;
1750
1751         if (time_after_eq(jiffies,
1752                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1753             throtl_tg_is_idle(tg))
1754                 return true;
1755         return false;
1756 }
1757
1758 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1759 {
1760         while (true) {
1761                 if (throtl_tg_can_upgrade(tg))
1762                         return true;
1763                 tg = sq_to_tg(tg->service_queue.parent_sq);
1764                 if (!tg || !tg_to_blkg(tg)->parent)
1765                         return false;
1766         }
1767         return false;
1768 }
1769
1770 static bool throtl_can_upgrade(struct throtl_data *td,
1771         struct throtl_grp *this_tg)
1772 {
1773         struct cgroup_subsys_state *pos_css;
1774         struct blkcg_gq *blkg;
1775
1776         if (td->limit_index != LIMIT_LOW)
1777                 return false;
1778
1779         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1780                 return false;
1781
1782         rcu_read_lock();
1783         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1784                 struct throtl_grp *tg = blkg_to_tg(blkg);
1785
1786                 if (tg == this_tg)
1787                         continue;
1788                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1789                         continue;
1790                 if (!throtl_hierarchy_can_upgrade(tg)) {
1791                         rcu_read_unlock();
1792                         return false;
1793                 }
1794         }
1795         rcu_read_unlock();
1796         return true;
1797 }
1798
1799 static void throtl_upgrade_check(struct throtl_grp *tg)
1800 {
1801         unsigned long now = jiffies;
1802
1803         if (tg->td->limit_index != LIMIT_LOW)
1804                 return;
1805
1806         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1807                 return;
1808
1809         tg->last_check_time = now;
1810
1811         if (!time_after_eq(now,
1812              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1813                 return;
1814
1815         if (throtl_can_upgrade(tg->td, NULL))
1816                 throtl_upgrade_state(tg->td);
1817 }
1818
1819 static void throtl_upgrade_state(struct throtl_data *td)
1820 {
1821         struct cgroup_subsys_state *pos_css;
1822         struct blkcg_gq *blkg;
1823
1824         throtl_log(&td->service_queue, "upgrade to max");
1825         td->limit_index = LIMIT_MAX;
1826         td->low_upgrade_time = jiffies;
1827         td->scale = 0;
1828         rcu_read_lock();
1829         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1830                 struct throtl_grp *tg = blkg_to_tg(blkg);
1831                 struct throtl_service_queue *sq = &tg->service_queue;
1832
1833                 tg->disptime = jiffies - 1;
1834                 throtl_select_dispatch(sq);
1835                 throtl_schedule_next_dispatch(sq, true);
1836         }
1837         rcu_read_unlock();
1838         throtl_select_dispatch(&td->service_queue);
1839         throtl_schedule_next_dispatch(&td->service_queue, true);
1840         queue_work(kthrotld_workqueue, &td->dispatch_work);
1841 }
1842
1843 static void throtl_downgrade_state(struct throtl_data *td)
1844 {
1845         td->scale /= 2;
1846
1847         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1848         if (td->scale) {
1849                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1850                 return;
1851         }
1852
1853         td->limit_index = LIMIT_LOW;
1854         td->low_downgrade_time = jiffies;
1855 }
1856
1857 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1858 {
1859         struct throtl_data *td = tg->td;
1860         unsigned long now = jiffies;
1861
1862         /*
1863          * If cgroup is below low limit, consider downgrade and throttle other
1864          * cgroups
1865          */
1866         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1867             time_after_eq(now, tg_last_low_overflow_time(tg) +
1868                                         td->throtl_slice) &&
1869             (!throtl_tg_is_idle(tg) ||
1870              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1871                 return true;
1872         return false;
1873 }
1874
1875 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1876 {
1877         while (true) {
1878                 if (!throtl_tg_can_downgrade(tg))
1879                         return false;
1880                 tg = sq_to_tg(tg->service_queue.parent_sq);
1881                 if (!tg || !tg_to_blkg(tg)->parent)
1882                         break;
1883         }
1884         return true;
1885 }
1886
1887 static void throtl_downgrade_check(struct throtl_grp *tg)
1888 {
1889         uint64_t bps;
1890         unsigned int iops;
1891         unsigned long elapsed_time;
1892         unsigned long now = jiffies;
1893
1894         if (tg->td->limit_index != LIMIT_MAX ||
1895             !tg->td->limit_valid[LIMIT_LOW])
1896                 return;
1897         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1898                 return;
1899         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1900                 return;
1901
1902         elapsed_time = now - tg->last_check_time;
1903         tg->last_check_time = now;
1904
1905         if (time_before(now, tg_last_low_overflow_time(tg) +
1906                         tg->td->throtl_slice))
1907                 return;
1908
1909         if (tg->bps[READ][LIMIT_LOW]) {
1910                 bps = tg->last_bytes_disp[READ] * HZ;
1911                 do_div(bps, elapsed_time);
1912                 if (bps >= tg->bps[READ][LIMIT_LOW])
1913                         tg->last_low_overflow_time[READ] = now;
1914         }
1915
1916         if (tg->bps[WRITE][LIMIT_LOW]) {
1917                 bps = tg->last_bytes_disp[WRITE] * HZ;
1918                 do_div(bps, elapsed_time);
1919                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1920                         tg->last_low_overflow_time[WRITE] = now;
1921         }
1922
1923         if (tg->iops[READ][LIMIT_LOW]) {
1924                 tg->last_io_disp[READ] += atomic_xchg(&tg->last_io_split_cnt[READ], 0);
1925                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1926                 if (iops >= tg->iops[READ][LIMIT_LOW])
1927                         tg->last_low_overflow_time[READ] = now;
1928         }
1929
1930         if (tg->iops[WRITE][LIMIT_LOW]) {
1931                 tg->last_io_disp[WRITE] += atomic_xchg(&tg->last_io_split_cnt[WRITE], 0);
1932                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
1933                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
1934                         tg->last_low_overflow_time[WRITE] = now;
1935         }
1936
1937         /*
1938          * If cgroup is below low limit, consider downgrade and throttle other
1939          * cgroups
1940          */
1941         if (throtl_hierarchy_can_downgrade(tg))
1942                 throtl_downgrade_state(tg->td);
1943
1944         tg->last_bytes_disp[READ] = 0;
1945         tg->last_bytes_disp[WRITE] = 0;
1946         tg->last_io_disp[READ] = 0;
1947         tg->last_io_disp[WRITE] = 0;
1948 }
1949
1950 static void blk_throtl_update_idletime(struct throtl_grp *tg)
1951 {
1952         unsigned long now;
1953         unsigned long last_finish_time = tg->last_finish_time;
1954
1955         if (last_finish_time == 0)
1956                 return;
1957
1958         now = ktime_get_ns() >> 10;
1959         if (now <= last_finish_time ||
1960             last_finish_time == tg->checked_last_finish_time)
1961                 return;
1962
1963         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
1964         tg->checked_last_finish_time = last_finish_time;
1965 }
1966
1967 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1968 static void throtl_update_latency_buckets(struct throtl_data *td)
1969 {
1970         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
1971         int i, cpu, rw;
1972         unsigned long last_latency[2] = { 0 };
1973         unsigned long latency[2];
1974
1975         if (!blk_queue_nonrot(td->queue) || !td->limit_valid[LIMIT_LOW])
1976                 return;
1977         if (time_before(jiffies, td->last_calculate_time + HZ))
1978                 return;
1979         td->last_calculate_time = jiffies;
1980
1981         memset(avg_latency, 0, sizeof(avg_latency));
1982         for (rw = READ; rw <= WRITE; rw++) {
1983                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
1984                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
1985
1986                         for_each_possible_cpu(cpu) {
1987                                 struct latency_bucket *bucket;
1988
1989                                 /* this isn't race free, but ok in practice */
1990                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
1991                                         cpu);
1992                                 tmp->total_latency += bucket[i].total_latency;
1993                                 tmp->samples += bucket[i].samples;
1994                                 bucket[i].total_latency = 0;
1995                                 bucket[i].samples = 0;
1996                         }
1997
1998                         if (tmp->samples >= 32) {
1999                                 int samples = tmp->samples;
2000
2001                                 latency[rw] = tmp->total_latency;
2002
2003                                 tmp->total_latency = 0;
2004                                 tmp->samples = 0;
2005                                 latency[rw] /= samples;
2006                                 if (latency[rw] == 0)
2007                                         continue;
2008                                 avg_latency[rw][i].latency = latency[rw];
2009                         }
2010                 }
2011         }
2012
2013         for (rw = READ; rw <= WRITE; rw++) {
2014                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2015                         if (!avg_latency[rw][i].latency) {
2016                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2017                                         td->avg_buckets[rw][i].latency =
2018                                                 last_latency[rw];
2019                                 continue;
2020                         }
2021
2022                         if (!td->avg_buckets[rw][i].valid)
2023                                 latency[rw] = avg_latency[rw][i].latency;
2024                         else
2025                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2026                                         avg_latency[rw][i].latency) >> 3;
2027
2028                         td->avg_buckets[rw][i].latency = max(latency[rw],
2029                                 last_latency[rw]);
2030                         td->avg_buckets[rw][i].valid = true;
2031                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2032                 }
2033         }
2034
2035         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2036                 throtl_log(&td->service_queue,
2037                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2038                         "write latency=%ld, write valid=%d", i,
2039                         td->avg_buckets[READ][i].latency,
2040                         td->avg_buckets[READ][i].valid,
2041                         td->avg_buckets[WRITE][i].latency,
2042                         td->avg_buckets[WRITE][i].valid);
2043 }
2044 #else
2045 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2046 {
2047 }
2048 #endif
2049
2050 void blk_throtl_charge_bio_split(struct bio *bio)
2051 {
2052         struct blkcg_gq *blkg = bio->bi_blkg;
2053         struct throtl_grp *parent = blkg_to_tg(blkg);
2054         struct throtl_service_queue *parent_sq;
2055         bool rw = bio_data_dir(bio);
2056
2057         do {
2058                 if (!parent->has_rules[rw])
2059                         break;
2060
2061                 atomic_inc(&parent->io_split_cnt[rw]);
2062                 atomic_inc(&parent->last_io_split_cnt[rw]);
2063
2064                 parent_sq = parent->service_queue.parent_sq;
2065                 parent = sq_to_tg(parent_sq);
2066         } while (parent);
2067 }
2068
2069 bool __blk_throtl_bio(struct bio *bio)
2070 {
2071         struct request_queue *q = bdev_get_queue(bio->bi_bdev);
2072         struct blkcg_gq *blkg = bio->bi_blkg;
2073         struct throtl_qnode *qn = NULL;
2074         struct throtl_grp *tg = blkg_to_tg(blkg);
2075         struct throtl_service_queue *sq;
2076         bool rw = bio_data_dir(bio);
2077         bool throttled = false;
2078         struct throtl_data *td = tg->td;
2079
2080         rcu_read_lock();
2081
2082         if (!cgroup_subsys_on_dfl(io_cgrp_subsys)) {
2083                 blkg_rwstat_add(&tg->stat_bytes, bio->bi_opf,
2084                                 bio->bi_iter.bi_size);
2085                 blkg_rwstat_add(&tg->stat_ios, bio->bi_opf, 1);
2086         }
2087
2088         spin_lock_irq(&q->queue_lock);
2089
2090         throtl_update_latency_buckets(td);
2091
2092         blk_throtl_update_idletime(tg);
2093
2094         sq = &tg->service_queue;
2095
2096 again:
2097         while (true) {
2098                 if (tg->last_low_overflow_time[rw] == 0)
2099                         tg->last_low_overflow_time[rw] = jiffies;
2100                 throtl_downgrade_check(tg);
2101                 throtl_upgrade_check(tg);
2102                 /* throtl is FIFO - if bios are already queued, should queue */
2103                 if (sq->nr_queued[rw])
2104                         break;
2105
2106                 /* if above limits, break to queue */
2107                 if (!tg_may_dispatch(tg, bio, NULL)) {
2108                         tg->last_low_overflow_time[rw] = jiffies;
2109                         if (throtl_can_upgrade(td, tg)) {
2110                                 throtl_upgrade_state(td);
2111                                 goto again;
2112                         }
2113                         break;
2114                 }
2115
2116                 /* within limits, let's charge and dispatch directly */
2117                 throtl_charge_bio(tg, bio);
2118
2119                 /*
2120                  * We need to trim slice even when bios are not being queued
2121                  * otherwise it might happen that a bio is not queued for
2122                  * a long time and slice keeps on extending and trim is not
2123                  * called for a long time. Now if limits are reduced suddenly
2124                  * we take into account all the IO dispatched so far at new
2125                  * low rate and * newly queued IO gets a really long dispatch
2126                  * time.
2127                  *
2128                  * So keep on trimming slice even if bio is not queued.
2129                  */
2130                 throtl_trim_slice(tg, rw);
2131
2132                 /*
2133                  * @bio passed through this layer without being throttled.
2134                  * Climb up the ladder.  If we're already at the top, it
2135                  * can be executed directly.
2136                  */
2137                 qn = &tg->qnode_on_parent[rw];
2138                 sq = sq->parent_sq;
2139                 tg = sq_to_tg(sq);
2140                 if (!tg)
2141                         goto out_unlock;
2142         }
2143
2144         /* out-of-limit, queue to @tg */
2145         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2146                    rw == READ ? 'R' : 'W',
2147                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2148                    tg_bps_limit(tg, rw),
2149                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2150                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2151
2152         tg->last_low_overflow_time[rw] = jiffies;
2153
2154         td->nr_queued[rw]++;
2155         throtl_add_bio_tg(bio, qn, tg);
2156         throttled = true;
2157
2158         /*
2159          * Update @tg's dispatch time and force schedule dispatch if @tg
2160          * was empty before @bio.  The forced scheduling isn't likely to
2161          * cause undue delay as @bio is likely to be dispatched directly if
2162          * its @tg's disptime is not in the future.
2163          */
2164         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2165                 tg_update_disptime(tg);
2166                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2167         }
2168
2169 out_unlock:
2170         spin_unlock_irq(&q->queue_lock);
2171         bio_set_flag(bio, BIO_THROTTLED);
2172
2173 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2174         if (throttled || !td->track_bio_latency)
2175                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2176 #endif
2177         rcu_read_unlock();
2178         return throttled;
2179 }
2180
2181 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2182 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2183         int op, unsigned long time)
2184 {
2185         struct latency_bucket *latency;
2186         int index;
2187
2188         if (!td || td->limit_index != LIMIT_LOW ||
2189             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2190             !blk_queue_nonrot(td->queue))
2191                 return;
2192
2193         index = request_bucket_index(size);
2194
2195         latency = get_cpu_ptr(td->latency_buckets[op]);
2196         latency[index].total_latency += time;
2197         latency[index].samples++;
2198         put_cpu_ptr(td->latency_buckets[op]);
2199 }
2200
2201 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2202 {
2203         struct request_queue *q = rq->q;
2204         struct throtl_data *td = q->td;
2205
2206         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2207                              time_ns >> 10);
2208 }
2209
2210 void blk_throtl_bio_endio(struct bio *bio)
2211 {
2212         struct blkcg_gq *blkg;
2213         struct throtl_grp *tg;
2214         u64 finish_time_ns;
2215         unsigned long finish_time;
2216         unsigned long start_time;
2217         unsigned long lat;
2218         int rw = bio_data_dir(bio);
2219
2220         blkg = bio->bi_blkg;
2221         if (!blkg)
2222                 return;
2223         tg = blkg_to_tg(blkg);
2224         if (!tg->td->limit_valid[LIMIT_LOW])
2225                 return;
2226
2227         finish_time_ns = ktime_get_ns();
2228         tg->last_finish_time = finish_time_ns >> 10;
2229
2230         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2231         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2232         if (!start_time || finish_time <= start_time)
2233                 return;
2234
2235         lat = finish_time - start_time;
2236         /* this is only for bio based driver */
2237         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2238                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2239                                      bio_op(bio), lat);
2240
2241         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2242                 int bucket;
2243                 unsigned int threshold;
2244
2245                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2246                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2247                         tg->latency_target;
2248                 if (lat > threshold)
2249                         tg->bad_bio_cnt++;
2250                 /*
2251                  * Not race free, could get wrong count, which means cgroups
2252                  * will be throttled
2253                  */
2254                 tg->bio_cnt++;
2255         }
2256
2257         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2258                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2259                 tg->bio_cnt /= 2;
2260                 tg->bad_bio_cnt /= 2;
2261         }
2262 }
2263 #endif
2264
2265 int blk_throtl_init(struct request_queue *q)
2266 {
2267         struct throtl_data *td;
2268         int ret;
2269
2270         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2271         if (!td)
2272                 return -ENOMEM;
2273         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2274                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2275         if (!td->latency_buckets[READ]) {
2276                 kfree(td);
2277                 return -ENOMEM;
2278         }
2279         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2280                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2281         if (!td->latency_buckets[WRITE]) {
2282                 free_percpu(td->latency_buckets[READ]);
2283                 kfree(td);
2284                 return -ENOMEM;
2285         }
2286
2287         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2288         throtl_service_queue_init(&td->service_queue);
2289
2290         q->td = td;
2291         td->queue = q;
2292
2293         td->limit_valid[LIMIT_MAX] = true;
2294         td->limit_index = LIMIT_MAX;
2295         td->low_upgrade_time = jiffies;
2296         td->low_downgrade_time = jiffies;
2297
2298         /* activate policy */
2299         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2300         if (ret) {
2301                 free_percpu(td->latency_buckets[READ]);
2302                 free_percpu(td->latency_buckets[WRITE]);
2303                 kfree(td);
2304         }
2305         return ret;
2306 }
2307
2308 void blk_throtl_exit(struct request_queue *q)
2309 {
2310         BUG_ON(!q->td);
2311         del_timer_sync(&q->td->service_queue.pending_timer);
2312         throtl_shutdown_wq(q);
2313         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2314         free_percpu(q->td->latency_buckets[READ]);
2315         free_percpu(q->td->latency_buckets[WRITE]);
2316         kfree(q->td);
2317 }
2318
2319 void blk_throtl_register_queue(struct request_queue *q)
2320 {
2321         struct throtl_data *td;
2322         int i;
2323
2324         td = q->td;
2325         BUG_ON(!td);
2326
2327         if (blk_queue_nonrot(q)) {
2328                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2329                 td->filtered_latency = LATENCY_FILTERED_SSD;
2330         } else {
2331                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2332                 td->filtered_latency = LATENCY_FILTERED_HD;
2333                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2334                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2335                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2336                 }
2337         }
2338 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2339         /* if no low limit, use previous default */
2340         td->throtl_slice = DFL_THROTL_SLICE_HD;
2341 #endif
2342
2343         td->track_bio_latency = !queue_is_mq(q);
2344         if (!td->track_bio_latency)
2345                 blk_stat_enable_accounting(q);
2346 }
2347
2348 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2349 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2350 {
2351         if (!q->td)
2352                 return -EINVAL;
2353         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2354 }
2355
2356 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2357         const char *page, size_t count)
2358 {
2359         unsigned long v;
2360         unsigned long t;
2361
2362         if (!q->td)
2363                 return -EINVAL;
2364         if (kstrtoul(page, 10, &v))
2365                 return -EINVAL;
2366         t = msecs_to_jiffies(v);
2367         if (t == 0 || t > MAX_THROTL_SLICE)
2368                 return -EINVAL;
2369         q->td->throtl_slice = t;
2370         return count;
2371 }
2372 #endif
2373
2374 static int __init throtl_init(void)
2375 {
2376         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2377         if (!kthrotld_workqueue)
2378                 panic("Failed to create kthrotld\n");
2379
2380         return blkcg_policy_register(&blkcg_policy_throtl);
2381 }
2382
2383 module_init(throtl_init);