root/block/blk-throttle.c

DEFINITIONS

1. pd_to_tg
2. blkg_to_tg
3. tg_to_blkg
4. sq_to_tg
5. sq_to_td
6. throtl_adjusted_limit
7. tg_bps_limit
8. tg_iops_limit
9. throtl_bio_data_size
10. throtl_qnode_init
11. throtl_qnode_add_bio
12. throtl_peek_queued
13. throtl_pop_queued
14. throtl_service_queue_init
15. throtl_pd_alloc
16. throtl_pd_init
17. tg_update_has_rules
18. throtl_pd_online
19. blk_throtl_update_limit_valid
20. throtl_pd_offline
21. throtl_pd_free
22. throtl_rb_first
23. throtl_rb_erase
24. update_min_dispatch_time
25. tg_service_queue_add
26. __throtl_enqueue_tg
27. throtl_enqueue_tg
28. __throtl_dequeue_tg
29. throtl_dequeue_tg
30. throtl_schedule_pending_timer
31. throtl_schedule_next_dispatch
32. throtl_start_new_slice_with_credit
33. throtl_start_new_slice
34. throtl_set_slice_end
35. throtl_extend_slice
36. throtl_slice_used
37. throtl_trim_slice
38. tg_with_in_iops_limit
39. tg_with_in_bps_limit
40. tg_may_dispatch
41. throtl_charge_bio
42. throtl_add_bio_tg
43. tg_update_disptime
44. start_parent_slice_with_credit
45. tg_dispatch_one_bio
46. throtl_dispatch_tg
47. throtl_select_dispatch
48. throtl_pending_timer_fn
49. blk_throtl_dispatch_work_fn
50. tg_prfill_conf_u64
51. tg_prfill_conf_uint
52. tg_print_conf_u64
53. tg_print_conf_uint
54. tg_conf_updated
55. tg_set_conf
56. tg_set_conf_u64
57. tg_set_conf_uint
58. tg_prfill_limit
59. tg_print_limit
60. tg_set_limit
61. throtl_shutdown_wq
62. __tg_last_low_overflow_time
63. tg_last_low_overflow_time
64. throtl_tg_is_idle
65. throtl_tg_can_upgrade
66. throtl_hierarchy_can_upgrade
67. throtl_can_upgrade
68. throtl_upgrade_check
69. throtl_upgrade_state
70. throtl_downgrade_state
71. throtl_tg_can_downgrade
72. throtl_hierarchy_can_downgrade
73. throtl_downgrade_check
74. blk_throtl_update_idletime
75. throtl_update_latency_buckets
76. throtl_update_latency_buckets
77. blk_throtl_bio
78. throtl_track_latency
79. blk_throtl_stat_add
80. blk_throtl_bio_endio
81. tg_drain_bios
82. blk_throtl_drain
83. blk_throtl_init
84. blk_throtl_exit
85. blk_throtl_register_queue
86. blk_throtl_sample_time_show
87. blk_throtl_sample_time_store
88. throtl_init

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Interface for controlling IO bandwidth on a request queue
   4  *
   5  * Copyright (C) 2010 Vivek Goyal <vgoyal@redhat.com>
   6  */
   7 
   8 #include <linux/module.h>
   9 #include <linux/slab.h>
  10 #include <linux/blkdev.h>
  11 #include <linux/bio.h>
  12 #include <linux/blktrace_api.h>
  13 #include <linux/blk-cgroup.h>
  14 #include "blk.h"
  15 
  16 /* Max dispatch from a group in 1 round */
  17 static int throtl_grp_quantum = 8;
  18 
  19 /* Total max dispatch from all groups in one round */
  20 static int throtl_quantum = 32;
  21 
  22 /* Throttling is performed over a slice and after that slice is renewed */
  23 #define DFL_THROTL_SLICE_HD (HZ / 10)
  24 #define DFL_THROTL_SLICE_SSD (HZ / 50)
  25 #define MAX_THROTL_SLICE (HZ)
  26 #define MAX_IDLE_TIME (5L * 1000 * 1000) /* 5 s */
  27 #define MIN_THROTL_BPS (320 * 1024)
  28 #define MIN_THROTL_IOPS (10)
  29 #define DFL_LATENCY_TARGET (-1L)
  30 #define DFL_IDLE_THRESHOLD (0)
  31 #define DFL_HD_BASELINE_LATENCY (4000L) /* 4ms */
  32 #define LATENCY_FILTERED_SSD (0)
  33 /*
  34  * For HD, very small latency comes from sequential IO. Such IO is helpless to
  35  * help determine if its IO is impacted by others, hence we ignore the IO
  36  */
  37 #define LATENCY_FILTERED_HD (1000L) /* 1ms */
  38 
  39 static struct blkcg_policy blkcg_policy_throtl;
  40 
  41 /* A workqueue to queue throttle related work */
  42 static struct workqueue_struct *kthrotld_workqueue;
  43 
  44 /*
  45  * To implement hierarchical throttling, throtl_grps form a tree and bios
  46  * are dispatched upwards level by level until they reach the top and get
  47  * issued.  When dispatching bios from the children and local group at each
  48  * level, if the bios are dispatched into a single bio_list, there's a risk
  49  * of a local or child group which can queue many bios at once filling up
  50  * the list starving others.
  51  *
  52  * To avoid such starvation, dispatched bios are queued separately
  53  * according to where they came from.  When they are again dispatched to
  54  * the parent, they're popped in round-robin order so that no single source
  55  * hogs the dispatch window.
  56  *
  57  * throtl_qnode is used to keep the queued bios separated by their sources.
  58  * Bios are queued to throtl_qnode which in turn is queued to
  59  * throtl_service_queue and then dispatched in round-robin order.
  60  *
  61  * It's also used to track the reference counts on blkg's.  A qnode always
  62  * belongs to a throtl_grp and gets queued on itself or the parent, so
  63  * incrementing the reference of the associated throtl_grp when a qnode is
  64  * queued and decrementing when dequeued is enough to keep the whole blkg
  65  * tree pinned while bios are in flight.
  66  */
  67 struct throtl_qnode {
  68         struct list_head        node;           /* service_queue->queued[] */
  69         struct bio_list         bios;           /* queued bios */
  70         struct throtl_grp       *tg;            /* tg this qnode belongs to */
  71 };
  72 
  73 struct throtl_service_queue {
  74         struct throtl_service_queue *parent_sq; /* the parent service_queue */
  75 
  76         /*
  77          * Bios queued directly to this service_queue or dispatched from
  78          * children throtl_grp's.
  79          */
  80         struct list_head        queued[2];      /* throtl_qnode [READ/WRITE] */
  81         unsigned int            nr_queued[2];   /* number of queued bios */
  82 
  83         /*
  84          * RB tree of active children throtl_grp's, which are sorted by
  85          * their ->disptime.
  86          */
  87         struct rb_root_cached   pending_tree;   /* RB tree of active tgs */
  88         unsigned int            nr_pending;     /* # queued in the tree */
  89         unsigned long           first_pending_disptime; /* disptime of the first tg */
  90         struct timer_list       pending_timer;  /* fires on first_pending_disptime */
  91 };
  92 
  93 enum tg_state_flags {
  94         THROTL_TG_PENDING       = 1 << 0,       /* on parent's pending tree */
  95         THROTL_TG_WAS_EMPTY     = 1 << 1,       /* bio_lists[] became non-empty */
  96 };
  97 
  98 #define rb_entry_tg(node)       rb_entry((node), struct throtl_grp, rb_node)
  99 
 100 enum {
 101         LIMIT_LOW,
 102         LIMIT_MAX,
 103         LIMIT_CNT,
 104 };
 105 
 106 struct throtl_grp {
 107         /* must be the first member */
 108         struct blkg_policy_data pd;
 109 
 110         /* active throtl group service_queue member */
 111         struct rb_node rb_node;
 112 
 113         /* throtl_data this group belongs to */
 114         struct throtl_data *td;
 115 
 116         /* this group's service queue */
 117         struct throtl_service_queue service_queue;
 118 
 119         /*
 120          * qnode_on_self is used when bios are directly queued to this
 121          * throtl_grp so that local bios compete fairly with bios
 122          * dispatched from children.  qnode_on_parent is used when bios are
 123          * dispatched from this throtl_grp into its parent and will compete
 124          * with the sibling qnode_on_parents and the parent's
 125          * qnode_on_self.
 126          */
 127         struct throtl_qnode qnode_on_self[2];
 128         struct throtl_qnode qnode_on_parent[2];
 129 
 130         /*
 131          * Dispatch time in jiffies. This is the estimated time when group
 132          * will unthrottle and is ready to dispatch more bio. It is used as
 133          * key to sort active groups in service tree.
 134          */
 135         unsigned long disptime;
 136 
 137         unsigned int flags;
 138 
 139         /* are there any throtl rules between this group and td? */
 140         bool has_rules[2];
 141 
 142         /* internally used bytes per second rate limits */
 143         uint64_t bps[2][LIMIT_CNT];
 144         /* user configured bps limits */
 145         uint64_t bps_conf[2][LIMIT_CNT];
 146 
 147         /* internally used IOPS limits */
 148         unsigned int iops[2][LIMIT_CNT];
 149         /* user configured IOPS limits */
 150         unsigned int iops_conf[2][LIMIT_CNT];
 151 
 152         /* Number of bytes disptached in current slice */
 153         uint64_t bytes_disp[2];
 154         /* Number of bio's dispatched in current slice */
 155         unsigned int io_disp[2];
 156 
 157         unsigned long last_low_overflow_time[2];
 158 
 159         uint64_t last_bytes_disp[2];
 160         unsigned int last_io_disp[2];
 161 
 162         unsigned long last_check_time;
 163 
 164         unsigned long latency_target; /* us */
 165         unsigned long latency_target_conf; /* us */
 166         /* When did we start a new slice */
 167         unsigned long slice_start[2];
 168         unsigned long slice_end[2];
 169 
 170         unsigned long last_finish_time; /* ns / 1024 */
 171         unsigned long checked_last_finish_time; /* ns / 1024 */
 172         unsigned long avg_idletime; /* ns / 1024 */
 173         unsigned long idletime_threshold; /* us */
 174         unsigned long idletime_threshold_conf; /* us */
 175 
 176         unsigned int bio_cnt; /* total bios */
 177         unsigned int bad_bio_cnt; /* bios exceeding latency threshold */
 178         unsigned long bio_cnt_reset_time;
 179 };
 180 
 181 /* We measure latency for request size from <= 4k to >= 1M */
 182 #define LATENCY_BUCKET_SIZE 9
 183 
 184 struct latency_bucket {
 185         unsigned long total_latency; /* ns / 1024 */
 186         int samples;
 187 };
 188 
 189 struct avg_latency_bucket {
 190         unsigned long latency; /* ns / 1024 */
 191         bool valid;
 192 };
 193 
 194 struct throtl_data
 195 {
 196         /* service tree for active throtl groups */
 197         struct throtl_service_queue service_queue;
 198 
 199         struct request_queue *queue;
 200 
 201         /* Total Number of queued bios on READ and WRITE lists */
 202         unsigned int nr_queued[2];
 203 
 204         unsigned int throtl_slice;
 205 
 206         /* Work for dispatching throttled bios */
 207         struct work_struct dispatch_work;
 208         unsigned int limit_index;
 209         bool limit_valid[LIMIT_CNT];
 210 
 211         unsigned long low_upgrade_time;
 212         unsigned long low_downgrade_time;
 213 
 214         unsigned int scale;
 215 
 216         struct latency_bucket tmp_buckets[2][LATENCY_BUCKET_SIZE];
 217         struct avg_latency_bucket avg_buckets[2][LATENCY_BUCKET_SIZE];
 218         struct latency_bucket __percpu *latency_buckets[2];
 219         unsigned long last_calculate_time;
 220         unsigned long filtered_latency;
 221 
 222         bool track_bio_latency;
 223 };
 224 
 225 static void throtl_pending_timer_fn(struct timer_list *t);
 226 
 227 static inline struct throtl_grp *pd_to_tg(struct blkg_policy_data *pd)
 228 {
 229         return pd ? container_of(pd, struct throtl_grp, pd) : NULL;
 230 }
 231 
 232 static inline struct throtl_grp *blkg_to_tg(struct blkcg_gq *blkg)
 233 {
 234         return pd_to_tg(blkg_to_pd(blkg, &blkcg_policy_throtl));
 235 }
 236 
 237 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg)
 238 {
 239         return pd_to_blkg(&tg->pd);
 240 }
 241 
 242 /**
 243  * sq_to_tg - return the throl_grp the specified service queue belongs to
 244  * @sq: the throtl_service_queue of interest
 245  *
 246  * Return the throtl_grp @sq belongs to.  If @sq is the top-level one
 247  * embedded in throtl_data, %NULL is returned.
 248  */
 249 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq)
 250 {
 251         if (sq && sq->parent_sq)
 252                 return container_of(sq, struct throtl_grp, service_queue);
 253         else
 254                 return NULL;
 255 }
 256 
 257 /**
 258  * sq_to_td - return throtl_data the specified service queue belongs to
 259  * @sq: the throtl_service_queue of interest
 260  *
 261  * A service_queue can be embedded in either a throtl_grp or throtl_data.
 262  * Determine the associated throtl_data accordingly and return it.
 263  */
 264 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq)
 265 {
 266         struct throtl_grp *tg = sq_to_tg(sq);
 267 
 268         if (tg)
 269                 return tg->td;
 270         else
 271                 return container_of(sq, struct throtl_data, service_queue);
 272 }
 273 
 274 /*
 275  * cgroup's limit in LIMIT_MAX is scaled if low limit is set. This scale is to
 276  * make the IO dispatch more smooth.
 277  * Scale up: linearly scale up according to lapsed time since upgrade. For
 278  *           every throtl_slice, the limit scales up 1/2 .low limit till the
 279  *           limit hits .max limit
 280  * Scale down: exponentially scale down if a cgroup doesn't hit its .low limit
 281  */
 282 static uint64_t throtl_adjusted_limit(uint64_t low, struct throtl_data *td)
 283 {
 284         /* arbitrary value to avoid too big scale */
 285         if (td->scale < 4096 && time_after_eq(jiffies,
 286             td->low_upgrade_time + td->scale * td->throtl_slice))
 287                 td->scale = (jiffies - td->low_upgrade_time) / td->throtl_slice;
 288 
 289         return low + (low >> 1) * td->scale;
 290 }
 291 
 292 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw)
 293 {
 294         struct blkcg_gq *blkg = tg_to_blkg(tg);
 295         struct throtl_data *td;
 296         uint64_t ret;
 297 
 298         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 299                 return U64_MAX;
 300 
 301         td = tg->td;
 302         ret = tg->bps[rw][td->limit_index];
 303         if (ret == 0 && td->limit_index == LIMIT_LOW) {
 304                 /* intermediate node or iops isn't 0 */
 305                 if (!list_empty(&blkg->blkcg->css.children) ||
 306                     tg->iops[rw][td->limit_index])
 307                         return U64_MAX;
 308                 else
 309                         return MIN_THROTL_BPS;
 310         }
 311 
 312         if (td->limit_index == LIMIT_MAX && tg->bps[rw][LIMIT_LOW] &&
 313             tg->bps[rw][LIMIT_LOW] != tg->bps[rw][LIMIT_MAX]) {
 314                 uint64_t adjusted;
 315 
 316                 adjusted = throtl_adjusted_limit(tg->bps[rw][LIMIT_LOW], td);
 317                 ret = min(tg->bps[rw][LIMIT_MAX], adjusted);
 318         }
 319         return ret;
 320 }
 321 
 322 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw)
 323 {
 324         struct blkcg_gq *blkg = tg_to_blkg(tg);
 325         struct throtl_data *td;
 326         unsigned int ret;
 327 
 328         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent)
 329                 return UINT_MAX;
 330 
 331         td = tg->td;
 332         ret = tg->iops[rw][td->limit_index];
 333         if (ret == 0 && tg->td->limit_index == LIMIT_LOW) {
 334                 /* intermediate node or bps isn't 0 */
 335                 if (!list_empty(&blkg->blkcg->css.children) ||
 336                     tg->bps[rw][td->limit_index])
 337                         return UINT_MAX;
 338                 else
 339                         return MIN_THROTL_IOPS;
 340         }
 341 
 342         if (td->limit_index == LIMIT_MAX && tg->iops[rw][LIMIT_LOW] &&
 343             tg->iops[rw][LIMIT_LOW] != tg->iops[rw][LIMIT_MAX]) {
 344                 uint64_t adjusted;
 345 
 346                 adjusted = throtl_adjusted_limit(tg->iops[rw][LIMIT_LOW], td);
 347                 if (adjusted > UINT_MAX)
 348                         adjusted = UINT_MAX;
 349                 ret = min_t(unsigned int, tg->iops[rw][LIMIT_MAX], adjusted);
 350         }
 351         return ret;
 352 }
 353 
 354 #define request_bucket_index(sectors) \
 355         clamp_t(int, order_base_2(sectors) - 3, 0, LATENCY_BUCKET_SIZE - 1)
 356 
 357 /**
 358  * throtl_log - log debug message via blktrace
 359  * @sq: the service_queue being reported
 360  * @fmt: printf format string
 361  * @args: printf args
 362  *
 363  * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a
 364  * throtl_grp; otherwise, just "throtl".
 365  */
 366 #define throtl_log(sq, fmt, args...)    do {                            \
 367         struct throtl_grp *__tg = sq_to_tg((sq));                       \
 368         struct throtl_data *__td = sq_to_td((sq));                      \
 369                                                                         \
 370         (void)__td;                                                     \
 371         if (likely(!blk_trace_note_message_enabled(__td->queue)))       \
 372                 break;                                                  \
 373         if ((__tg)) {                                                   \
 374                 blk_add_cgroup_trace_msg(__td->queue,                   \
 375                         tg_to_blkg(__tg)->blkcg, "throtl " fmt, ##args);\
 376         } else {                                                        \
 377                 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args);  \
 378         }                                                               \
 379 } while (0)
 380 
 381 static inline unsigned int throtl_bio_data_size(struct bio *bio)
 382 {
 383         /* assume it's one sector */
 384         if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
 385                 return 512;
 386         return bio->bi_iter.bi_size;
 387 }
 388 
 389 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg)
 390 {
 391         INIT_LIST_HEAD(&qn->node);
 392         bio_list_init(&qn->bios);
 393         qn->tg = tg;
 394 }
 395 
 396 /**
 397  * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it
 398  * @bio: bio being added
 399  * @qn: qnode to add bio to
 400  * @queued: the service_queue->queued[] list @qn belongs to
 401  *
 402  * Add @bio to @qn and put @qn on @queued if it's not already on.
 403  * @qn->tg's reference count is bumped when @qn is activated.  See the
 404  * comment on top of throtl_qnode definition for details.
 405  */
 406 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn,
 407                                  struct list_head *queued)
 408 {
 409         bio_list_add(&qn->bios, bio);
 410         if (list_empty(&qn->node)) {
 411                 list_add_tail(&qn->node, queued);
 412                 blkg_get(tg_to_blkg(qn->tg));
 413         }
 414 }
 415 
 416 /**
 417  * throtl_peek_queued - peek the first bio on a qnode list
 418  * @queued: the qnode list to peek
 419  */
 420 static struct bio *throtl_peek_queued(struct list_head *queued)
 421 {
 422         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 423         struct bio *bio;
 424 
 425         if (list_empty(queued))
 426                 return NULL;
 427 
 428         bio = bio_list_peek(&qn->bios);
 429         WARN_ON_ONCE(!bio);
 430         return bio;
 431 }
 432 
 433 /**
 434  * throtl_pop_queued - pop the first bio form a qnode list
 435  * @queued: the qnode list to pop a bio from
 436  * @tg_to_put: optional out argument for throtl_grp to put
 437  *
 438  * Pop the first bio from the qnode list @queued.  After popping, the first
 439  * qnode is removed from @queued if empty or moved to the end of @queued so
 440  * that the popping order is round-robin.
 441  *
 442  * When the first qnode is removed, its associated throtl_grp should be put
 443  * too.  If @tg_to_put is NULL, this function automatically puts it;
 444  * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is
 445  * responsible for putting it.
 446  */
 447 static struct bio *throtl_pop_queued(struct list_head *queued,
 448                                      struct throtl_grp **tg_to_put)
 449 {
 450         struct throtl_qnode *qn = list_first_entry(queued, struct throtl_qnode, node);
 451         struct bio *bio;
 452 
 453         if (list_empty(queued))
 454                 return NULL;
 455 
 456         bio = bio_list_pop(&qn->bios);
 457         WARN_ON_ONCE(!bio);
 458 
 459         if (bio_list_empty(&qn->bios)) {
 460                 list_del_init(&qn->node);
 461                 if (tg_to_put)
 462                         *tg_to_put = qn->tg;
 463                 else
 464                         blkg_put(tg_to_blkg(qn->tg));
 465         } else {
 466                 list_move_tail(&qn->node, queued);
 467         }
 468 
 469         return bio;
 470 }
 471 
 472 /* init a service_queue, assumes the caller zeroed it */
 473 static void throtl_service_queue_init(struct throtl_service_queue *sq)
 474 {
 475         INIT_LIST_HEAD(&sq->queued[0]);
 476         INIT_LIST_HEAD(&sq->queued[1]);
 477         sq->pending_tree = RB_ROOT_CACHED;
 478         timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0);
 479 }
 480 
 481 static struct blkg_policy_data *throtl_pd_alloc(gfp_t gfp,
 482                                                 struct request_queue *q,
 483                                                 struct blkcg *blkcg)
 484 {
 485         struct throtl_grp *tg;
 486         int rw;
 487 
 488         tg = kzalloc_node(sizeof(*tg), gfp, q->node);
 489         if (!tg)
 490                 return NULL;
 491 
 492         throtl_service_queue_init(&tg->service_queue);
 493 
 494         for (rw = READ; rw <= WRITE; rw++) {
 495                 throtl_qnode_init(&tg->qnode_on_self[rw], tg);
 496                 throtl_qnode_init(&tg->qnode_on_parent[rw], tg);
 497         }
 498 
 499         RB_CLEAR_NODE(&tg->rb_node);
 500         tg->bps[READ][LIMIT_MAX] = U64_MAX;
 501         tg->bps[WRITE][LIMIT_MAX] = U64_MAX;
 502         tg->iops[READ][LIMIT_MAX] = UINT_MAX;
 503         tg->iops[WRITE][LIMIT_MAX] = UINT_MAX;
 504         tg->bps_conf[READ][LIMIT_MAX] = U64_MAX;
 505         tg->bps_conf[WRITE][LIMIT_MAX] = U64_MAX;
 506         tg->iops_conf[READ][LIMIT_MAX] = UINT_MAX;
 507         tg->iops_conf[WRITE][LIMIT_MAX] = UINT_MAX;
 508         /* LIMIT_LOW will have default value 0 */
 509 
 510         tg->latency_target = DFL_LATENCY_TARGET;
 511         tg->latency_target_conf = DFL_LATENCY_TARGET;
 512         tg->idletime_threshold = DFL_IDLE_THRESHOLD;
 513         tg->idletime_threshold_conf = DFL_IDLE_THRESHOLD;
 514 
 515         return &tg->pd;
 516 }
 517 
 518 static void throtl_pd_init(struct blkg_policy_data *pd)
 519 {
 520         struct throtl_grp *tg = pd_to_tg(pd);
 521         struct blkcg_gq *blkg = tg_to_blkg(tg);
 522         struct throtl_data *td = blkg->q->td;
 523         struct throtl_service_queue *sq = &tg->service_queue;
 524 
 525         /*
 526          * If on the default hierarchy, we switch to properly hierarchical
 527          * behavior where limits on a given throtl_grp are applied to the
 528          * whole subtree rather than just the group itself.  e.g. If 16M
 529          * read_bps limit is set on the root group, the whole system can't
 530          * exceed 16M for the device.
 531          *
 532          * If not on the default hierarchy, the broken flat hierarchy
 533          * behavior is retained where all throtl_grps are treated as if
 534          * they're all separate root groups right below throtl_data.
 535          * Limits of a group don't interact with limits of other groups
 536          * regardless of the position of the group in the hierarchy.
 537          */
 538         sq->parent_sq = &td->service_queue;
 539         if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent)
 540                 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue;
 541         tg->td = td;
 542 }
 543 
 544 /*
 545  * Set has_rules[] if @tg or any of its parents have limits configured.
 546  * This doesn't require walking up to the top of the hierarchy as the
 547  * parent's has_rules[] is guaranteed to be correct.
 548  */
 549 static void tg_update_has_rules(struct throtl_grp *tg)
 550 {
 551         struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq);
 552         struct throtl_data *td = tg->td;
 553         int rw;
 554 
 555         for (rw = READ; rw <= WRITE; rw++)
 556                 tg->has_rules[rw] = (parent_tg && parent_tg->has_rules[rw]) ||
 557                         (td->limit_valid[td->limit_index] &&
 558                          (tg_bps_limit(tg, rw) != U64_MAX ||
 559                           tg_iops_limit(tg, rw) != UINT_MAX));
 560 }
 561 
 562 static void throtl_pd_online(struct blkg_policy_data *pd)
 563 {
 564         struct throtl_grp *tg = pd_to_tg(pd);
 565         /*
 566          * We don't want new groups to escape the limits of its ancestors.
 567          * Update has_rules[] after a new group is brought online.
 568          */
 569         tg_update_has_rules(tg);
 570 }
 571 
 572 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 573 {
 574         struct cgroup_subsys_state *pos_css;
 575         struct blkcg_gq *blkg;
 576         bool low_valid = false;
 577 
 578         rcu_read_lock();
 579         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
 580                 struct throtl_grp *tg = blkg_to_tg(blkg);
 581 
 582                 if (tg->bps[READ][LIMIT_LOW] || tg->bps[WRITE][LIMIT_LOW] ||
 583                     tg->iops[READ][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) {
 584                         low_valid = true;
 585                         break;
 586                 }
 587         }
 588         rcu_read_unlock();
 589 
 590         td->limit_valid[LIMIT_LOW] = low_valid;
 591 }
 592 
 593 static void throtl_upgrade_state(struct throtl_data *td);
 594 static void throtl_pd_offline(struct blkg_policy_data *pd)
 595 {
 596         struct throtl_grp *tg = pd_to_tg(pd);
 597 
 598         tg->bps[READ][LIMIT_LOW] = 0;
 599         tg->bps[WRITE][LIMIT_LOW] = 0;
 600         tg->iops[READ][LIMIT_LOW] = 0;
 601         tg->iops[WRITE][LIMIT_LOW] = 0;
 602 
 603         blk_throtl_update_limit_valid(tg->td);
 604 
 605         if (!tg->td->limit_valid[tg->td->limit_index])
 606                 throtl_upgrade_state(tg->td);
 607 }
 608 
 609 static void throtl_pd_free(struct blkg_policy_data *pd)
 610 {
 611         struct throtl_grp *tg = pd_to_tg(pd);
 612 
 613         del_timer_sync(&tg->service_queue.pending_timer);
 614         kfree(tg);
 615 }
 616 
 617 static struct throtl_grp *
 618 throtl_rb_first(struct throtl_service_queue *parent_sq)
 619 {
 620         struct rb_node *n;
 621         /* Service tree is empty */
 622         if (!parent_sq->nr_pending)
 623                 return NULL;
 624 
 625         n = rb_first_cached(&parent_sq->pending_tree);
 626         WARN_ON_ONCE(!n);
 627         if (!n)
 628                 return NULL;
 629         return rb_entry_tg(n);
 630 }
 631 
 632 static void throtl_rb_erase(struct rb_node *n,
 633                             struct throtl_service_queue *parent_sq)
 634 {
 635         rb_erase_cached(n, &parent_sq->pending_tree);
 636         RB_CLEAR_NODE(n);
 637         --parent_sq->nr_pending;
 638 }
 639 
 640 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq)
 641 {
 642         struct throtl_grp *tg;
 643 
 644         tg = throtl_rb_first(parent_sq);
 645         if (!tg)
 646                 return;
 647 
 648         parent_sq->first_pending_disptime = tg->disptime;
 649 }
 650 
 651 static void tg_service_queue_add(struct throtl_grp *tg)
 652 {
 653         struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq;
 654         struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node;
 655         struct rb_node *parent = NULL;
 656         struct throtl_grp *__tg;
 657         unsigned long key = tg->disptime;
 658         bool leftmost = true;
 659 
 660         while (*node != NULL) {
 661                 parent = *node;
 662                 __tg = rb_entry_tg(parent);
 663 
 664                 if (time_before(key, __tg->disptime))
 665                         node = &parent->rb_left;
 666                 else {
 667                         node = &parent->rb_right;
 668                         leftmost = false;
 669                 }
 670         }
 671 
 672         rb_link_node(&tg->rb_node, parent, node);
 673         rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree,
 674                                leftmost);
 675 }
 676 
 677 static void __throtl_enqueue_tg(struct throtl_grp *tg)
 678 {
 679         tg_service_queue_add(tg);
 680         tg->flags |= THROTL_TG_PENDING;
 681         tg->service_queue.parent_sq->nr_pending++;
 682 }
 683 
 684 static void throtl_enqueue_tg(struct throtl_grp *tg)
 685 {
 686         if (!(tg->flags & THROTL_TG_PENDING))
 687                 __throtl_enqueue_tg(tg);
 688 }
 689 
 690 static void __throtl_dequeue_tg(struct throtl_grp *tg)
 691 {
 692         throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq);
 693         tg->flags &= ~THROTL_TG_PENDING;
 694 }
 695 
 696 static void throtl_dequeue_tg(struct throtl_grp *tg)
 697 {
 698         if (tg->flags & THROTL_TG_PENDING)
 699                 __throtl_dequeue_tg(tg);
 700 }
 701 
 702 /* Call with queue lock held */
 703 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq,
 704                                           unsigned long expires)
 705 {
 706         unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice;
 707 
 708         /*
 709          * Since we are adjusting the throttle limit dynamically, the sleep
 710          * time calculated according to previous limit might be invalid. It's
 711          * possible the cgroup sleep time is very long and no other cgroups
 712          * have IO running so notify the limit changes. Make sure the cgroup
 713          * doesn't sleep too long to avoid the missed notification.
 714          */
 715         if (time_after(expires, max_expire))
 716                 expires = max_expire;
 717         mod_timer(&sq->pending_timer, expires);
 718         throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu",
 719                    expires - jiffies, jiffies);
 720 }
 721 
 722 /**
 723  * throtl_schedule_next_dispatch - schedule the next dispatch cycle
 724  * @sq: the service_queue to schedule dispatch for
 725  * @force: force scheduling
 726  *
 727  * Arm @sq->pending_timer so that the next dispatch cycle starts on the
 728  * dispatch time of the first pending child.  Returns %true if either timer
 729  * is armed or there's no pending child left.  %false if the current
 730  * dispatch window is still open and the caller should continue
 731  * dispatching.
 732  *
 733  * If @force is %true, the dispatch timer is always scheduled and this
 734  * function is guaranteed to return %true.  This is to be used when the
 735  * caller can't dispatch itself and needs to invoke pending_timer
 736  * unconditionally.  Note that forced scheduling is likely to induce short
 737  * delay before dispatch starts even if @sq->first_pending_disptime is not
 738  * in the future and thus shouldn't be used in hot paths.
 739  */
 740 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq,
 741                                           bool force)
 742 {
 743         /* any pending children left? */
 744         if (!sq->nr_pending)
 745                 return true;
 746 
 747         update_min_dispatch_time(sq);
 748 
 749         /* is the next dispatch time in the future? */
 750         if (force || time_after(sq->first_pending_disptime, jiffies)) {
 751                 throtl_schedule_pending_timer(sq, sq->first_pending_disptime);
 752                 return true;
 753         }
 754 
 755         /* tell the caller to continue dispatching */
 756         return false;
 757 }
 758 
 759 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg,
 760                 bool rw, unsigned long start)
 761 {
 762         tg->bytes_disp[rw] = 0;
 763         tg->io_disp[rw] = 0;
 764 
 765         /*
 766          * Previous slice has expired. We must have trimmed it after last
 767          * bio dispatch. That means since start of last slice, we never used
 768          * that bandwidth. Do try to make use of that bandwidth while giving
 769          * credit.
 770          */
 771         if (time_after_eq(start, tg->slice_start[rw]))
 772                 tg->slice_start[rw] = start;
 773 
 774         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 775         throtl_log(&tg->service_queue,
 776                    "[%c] new slice with credit start=%lu end=%lu jiffies=%lu",
 777                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 778                    tg->slice_end[rw], jiffies);
 779 }
 780 
 781 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw)
 782 {
 783         tg->bytes_disp[rw] = 0;
 784         tg->io_disp[rw] = 0;
 785         tg->slice_start[rw] = jiffies;
 786         tg->slice_end[rw] = jiffies + tg->td->throtl_slice;
 787         throtl_log(&tg->service_queue,
 788                    "[%c] new slice start=%lu end=%lu jiffies=%lu",
 789                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 790                    tg->slice_end[rw], jiffies);
 791 }
 792 
 793 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw,
 794                                         unsigned long jiffy_end)
 795 {
 796         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 797 }
 798 
 799 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw,
 800                                        unsigned long jiffy_end)
 801 {
 802         tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice);
 803         throtl_log(&tg->service_queue,
 804                    "[%c] extend slice start=%lu end=%lu jiffies=%lu",
 805                    rw == READ ? 'R' : 'W', tg->slice_start[rw],
 806                    tg->slice_end[rw], jiffies);
 807 }
 808 
 809 /* Determine if previously allocated or extended slice is complete or not */
 810 static bool throtl_slice_used(struct throtl_grp *tg, bool rw)
 811 {
 812         if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw]))
 813                 return false;
 814 
 815         return true;
 816 }
 817 
 818 /* Trim the used slices and adjust slice start accordingly */
 819 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw)
 820 {
 821         unsigned long nr_slices, time_elapsed, io_trim;
 822         u64 bytes_trim, tmp;
 823 
 824         BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw]));
 825 
 826         /*
 827          * If bps are unlimited (-1), then time slice don't get
 828          * renewed. Don't try to trim the slice if slice is used. A new
 829          * slice will start when appropriate.
 830          */
 831         if (throtl_slice_used(tg, rw))
 832                 return;
 833 
 834         /*
 835          * A bio has been dispatched. Also adjust slice_end. It might happen
 836          * that initially cgroup limit was very low resulting in high
 837          * slice_end, but later limit was bumped up and bio was dispached
 838          * sooner, then we need to reduce slice_end. A high bogus slice_end
 839          * is bad because it does not allow new slice to start.
 840          */
 841 
 842         throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice);
 843 
 844         time_elapsed = jiffies - tg->slice_start[rw];
 845 
 846         nr_slices = time_elapsed / tg->td->throtl_slice;
 847 
 848         if (!nr_slices)
 849                 return;
 850         tmp = tg_bps_limit(tg, rw) * tg->td->throtl_slice * nr_slices;
 851         do_div(tmp, HZ);
 852         bytes_trim = tmp;
 853 
 854         io_trim = (tg_iops_limit(tg, rw) * tg->td->throtl_slice * nr_slices) /
 855                 HZ;
 856 
 857         if (!bytes_trim && !io_trim)
 858                 return;
 859 
 860         if (tg->bytes_disp[rw] >= bytes_trim)
 861                 tg->bytes_disp[rw] -= bytes_trim;
 862         else
 863                 tg->bytes_disp[rw] = 0;
 864 
 865         if (tg->io_disp[rw] >= io_trim)
 866                 tg->io_disp[rw] -= io_trim;
 867         else
 868                 tg->io_disp[rw] = 0;
 869 
 870         tg->slice_start[rw] += nr_slices * tg->td->throtl_slice;
 871 
 872         throtl_log(&tg->service_queue,
 873                    "[%c] trim slice nr=%lu bytes=%llu io=%lu start=%lu end=%lu jiffies=%lu",
 874                    rw == READ ? 'R' : 'W', nr_slices, bytes_trim, io_trim,
 875                    tg->slice_start[rw], tg->slice_end[rw], jiffies);
 876 }
 877 
 878 static bool tg_with_in_iops_limit(struct throtl_grp *tg, struct bio *bio,
 879                                   unsigned long *wait)
 880 {
 881         bool rw = bio_data_dir(bio);
 882         unsigned int io_allowed;
 883         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 884         u64 tmp;
 885 
 886         jiffy_elapsed = jiffies - tg->slice_start[rw];
 887 
 888         /* Round up to the next throttle slice, wait time must be nonzero */
 889         jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice);
 890 
 891         /*
 892          * jiffy_elapsed_rnd should not be a big value as minimum iops can be
 893          * 1 then at max jiffy elapsed should be equivalent of 1 second as we
 894          * will allow dispatch after 1 second and after that slice should
 895          * have been trimmed.
 896          */
 897 
 898         tmp = (u64)tg_iops_limit(tg, rw) * jiffy_elapsed_rnd;
 899         do_div(tmp, HZ);
 900 
 901         if (tmp > UINT_MAX)
 902                 io_allowed = UINT_MAX;
 903         else
 904                 io_allowed = tmp;
 905 
 906         if (tg->io_disp[rw] + 1 <= io_allowed) {
 907                 if (wait)
 908                         *wait = 0;
 909                 return true;
 910         }
 911 
 912         /* Calc approx time to dispatch */
 913         jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed;
 914 
 915         if (wait)
 916                 *wait = jiffy_wait;
 917         return false;
 918 }
 919 
 920 static bool tg_with_in_bps_limit(struct throtl_grp *tg, struct bio *bio,
 921                                  unsigned long *wait)
 922 {
 923         bool rw = bio_data_dir(bio);
 924         u64 bytes_allowed, extra_bytes, tmp;
 925         unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd;
 926         unsigned int bio_size = throtl_bio_data_size(bio);
 927 
 928         jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw];
 929 
 930         /* Slice has just started. Consider one slice interval */
 931         if (!jiffy_elapsed)
 932                 jiffy_elapsed_rnd = tg->td->throtl_slice;
 933 
 934         jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice);
 935 
 936         tmp = tg_bps_limit(tg, rw) * jiffy_elapsed_rnd;
 937         do_div(tmp, HZ);
 938         bytes_allowed = tmp;
 939 
 940         if (tg->bytes_disp[rw] + bio_size <= bytes_allowed) {
 941                 if (wait)
 942                         *wait = 0;
 943                 return true;
 944         }
 945 
 946         /* Calc approx time to dispatch */
 947         extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed;
 948         jiffy_wait = div64_u64(extra_bytes * HZ, tg_bps_limit(tg, rw));
 949 
 950         if (!jiffy_wait)
 951                 jiffy_wait = 1;
 952 
 953         /*
 954          * This wait time is without taking into consideration the rounding
 955          * up we did. Add that time also.
 956          */
 957         jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed);
 958         if (wait)
 959                 *wait = jiffy_wait;
 960         return false;
 961 }
 962 
 963 /*
 964  * Returns whether one can dispatch a bio or not. Also returns approx number
 965  * of jiffies to wait before this bio is with-in IO rate and can be dispatched
 966  */
 967 static bool tg_may_dispatch(struct throtl_grp *tg, struct bio *bio,
 968                             unsigned long *wait)
 969 {
 970         bool rw = bio_data_dir(bio);
 971         unsigned long bps_wait = 0, iops_wait = 0, max_wait = 0;
 972 
 973         /*
 974          * Currently whole state machine of group depends on first bio
 975          * queued in the group bio list. So one should not be calling
 976          * this function with a different bio if there are other bios
 977          * queued.
 978          */
 979         BUG_ON(tg->service_queue.nr_queued[rw] &&
 980                bio != throtl_peek_queued(&tg->service_queue.queued[rw]));
 981 
 982         /* If tg->bps = -1, then BW is unlimited */
 983         if (tg_bps_limit(tg, rw) == U64_MAX &&
 984             tg_iops_limit(tg, rw) == UINT_MAX) {
 985                 if (wait)
 986                         *wait = 0;
 987                 return true;
 988         }
 989 
 990         /*
 991          * If previous slice expired, start a new one otherwise renew/extend
 992          * existing slice to make sure it is at least throtl_slice interval
 993          * long since now. New slice is started only for empty throttle group.
 994          * If there is queued bio, that means there should be an active
 995          * slice and it should be extended instead.
 996          */
 997         if (throtl_slice_used(tg, rw) && !(tg->service_queue.nr_queued[rw]))
 998                 throtl_start_new_slice(tg, rw);
 999         else {
1000                 if (time_before(tg->slice_end[rw],
1001                     jiffies + tg->td->throtl_slice))
1002                         throtl_extend_slice(tg, rw,
1003                                 jiffies + tg->td->throtl_slice);
1004         }
1005 
1006         if (tg_with_in_bps_limit(tg, bio, &bps_wait) &&
1007             tg_with_in_iops_limit(tg, bio, &iops_wait)) {
1008                 if (wait)
1009                         *wait = 0;
1010                 return true;
1011         }
1012 
1013         max_wait = max(bps_wait, iops_wait);
1014 
1015         if (wait)
1016                 *wait = max_wait;
1017 
1018         if (time_before(tg->slice_end[rw], jiffies + max_wait))
1019                 throtl_extend_slice(tg, rw, jiffies + max_wait);
1020 
1021         return false;
1022 }
1023 
1024 static void throtl_charge_bio(struct throtl_grp *tg, struct bio *bio)
1025 {
1026         bool rw = bio_data_dir(bio);
1027         unsigned int bio_size = throtl_bio_data_size(bio);
1028 
1029         /* Charge the bio to the group */
1030         tg->bytes_disp[rw] += bio_size;
1031         tg->io_disp[rw]++;
1032         tg->last_bytes_disp[rw] += bio_size;
1033         tg->last_io_disp[rw]++;
1034 
1035         /*
1036          * BIO_THROTTLED is used to prevent the same bio to be throttled
1037          * more than once as a throttled bio will go through blk-throtl the
1038          * second time when it eventually gets issued.  Set it when a bio
1039          * is being charged to a tg.
1040          */
1041         if (!bio_flagged(bio, BIO_THROTTLED))
1042                 bio_set_flag(bio, BIO_THROTTLED);
1043 }
1044 
1045 /**
1046  * throtl_add_bio_tg - add a bio to the specified throtl_grp
1047  * @bio: bio to add
1048  * @qn: qnode to use
1049  * @tg: the target throtl_grp
1050  *
1051  * Add @bio to @tg's service_queue using @qn.  If @qn is not specified,
1052  * tg->qnode_on_self[] is used.
1053  */
1054 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn,
1055                               struct throtl_grp *tg)
1056 {
1057         struct throtl_service_queue *sq = &tg->service_queue;
1058         bool rw = bio_data_dir(bio);
1059 
1060         if (!qn)
1061                 qn = &tg->qnode_on_self[rw];
1062 
1063         /*
1064          * If @tg doesn't currently have any bios queued in the same
1065          * direction, queueing @bio can change when @tg should be
1066          * dispatched.  Mark that @tg was empty.  This is automatically
1067          * cleaered on the next tg_update_disptime().
1068          */
1069         if (!sq->nr_queued[rw])
1070                 tg->flags |= THROTL_TG_WAS_EMPTY;
1071 
1072         throtl_qnode_add_bio(bio, qn, &sq->queued[rw]);
1073 
1074         sq->nr_queued[rw]++;
1075         throtl_enqueue_tg(tg);
1076 }
1077 
1078 static void tg_update_disptime(struct throtl_grp *tg)
1079 {
1080         struct throtl_service_queue *sq = &tg->service_queue;
1081         unsigned long read_wait = -1, write_wait = -1, min_wait = -1, disptime;
1082         struct bio *bio;
1083 
1084         bio = throtl_peek_queued(&sq->queued[READ]);
1085         if (bio)
1086                 tg_may_dispatch(tg, bio, &read_wait);
1087 
1088         bio = throtl_peek_queued(&sq->queued[WRITE]);
1089         if (bio)
1090                 tg_may_dispatch(tg, bio, &write_wait);
1091 
1092         min_wait = min(read_wait, write_wait);
1093         disptime = jiffies + min_wait;
1094 
1095         /* Update dispatch time */
1096         throtl_dequeue_tg(tg);
1097         tg->disptime = disptime;
1098         throtl_enqueue_tg(tg);
1099 
1100         /* see throtl_add_bio_tg() */
1101         tg->flags &= ~THROTL_TG_WAS_EMPTY;
1102 }
1103 
1104 static void start_parent_slice_with_credit(struct throtl_grp *child_tg,
1105                                         struct throtl_grp *parent_tg, bool rw)
1106 {
1107         if (throtl_slice_used(parent_tg, rw)) {
1108                 throtl_start_new_slice_with_credit(parent_tg, rw,
1109                                 child_tg->slice_start[rw]);
1110         }
1111 
1112 }
1113 
1114 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw)
1115 {
1116         struct throtl_service_queue *sq = &tg->service_queue;
1117         struct throtl_service_queue *parent_sq = sq->parent_sq;
1118         struct throtl_grp *parent_tg = sq_to_tg(parent_sq);
1119         struct throtl_grp *tg_to_put = NULL;
1120         struct bio *bio;
1121 
1122         /*
1123          * @bio is being transferred from @tg to @parent_sq.  Popping a bio
1124          * from @tg may put its reference and @parent_sq might end up
1125          * getting released prematurely.  Remember the tg to put and put it
1126          * after @bio is transferred to @parent_sq.
1127          */
1128         bio = throtl_pop_queued(&sq->queued[rw], &tg_to_put);
1129         sq->nr_queued[rw]--;
1130 
1131         throtl_charge_bio(tg, bio);
1132 
1133         /*
1134          * If our parent is another tg, we just need to transfer @bio to
1135          * the parent using throtl_add_bio_tg().  If our parent is
1136          * @td->service_queue, @bio is ready to be issued.  Put it on its
1137          * bio_lists[] and decrease total number queued.  The caller is
1138          * responsible for issuing these bios.
1139          */
1140         if (parent_tg) {
1141                 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg);
1142                 start_parent_slice_with_credit(tg, parent_tg, rw);
1143         } else {
1144                 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw],
1145                                      &parent_sq->queued[rw]);
1146                 BUG_ON(tg->td->nr_queued[rw] <= 0);
1147                 tg->td->nr_queued[rw]--;
1148         }
1149 
1150         throtl_trim_slice(tg, rw);
1151 
1152         if (tg_to_put)
1153                 blkg_put(tg_to_blkg(tg_to_put));
1154 }
1155 
1156 static int throtl_dispatch_tg(struct throtl_grp *tg)
1157 {
1158         struct throtl_service_queue *sq = &tg->service_queue;
1159         unsigned int nr_reads = 0, nr_writes = 0;
1160         unsigned int max_nr_reads = throtl_grp_quantum*3/4;
1161         unsigned int max_nr_writes = throtl_grp_quantum - max_nr_reads;
1162         struct bio *bio;
1163 
1164         /* Try to dispatch 75% READS and 25% WRITES */
1165 
1166         while ((bio = throtl_peek_queued(&sq->queued[READ])) &&
1167                tg_may_dispatch(tg, bio, NULL)) {
1168 
1169                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1170                 nr_reads++;
1171 
1172                 if (nr_reads >= max_nr_reads)
1173                         break;
1174         }
1175 
1176         while ((bio = throtl_peek_queued(&sq->queued[WRITE])) &&
1177                tg_may_dispatch(tg, bio, NULL)) {
1178 
1179                 tg_dispatch_one_bio(tg, bio_data_dir(bio));
1180                 nr_writes++;
1181 
1182                 if (nr_writes >= max_nr_writes)
1183                         break;
1184         }
1185 
1186         return nr_reads + nr_writes;
1187 }
1188 
1189 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq)
1190 {
1191         unsigned int nr_disp = 0;
1192 
1193         while (1) {
1194                 struct throtl_grp *tg = throtl_rb_first(parent_sq);
1195                 struct throtl_service_queue *sq;
1196 
1197                 if (!tg)
1198                         break;
1199 
1200                 if (time_before(jiffies, tg->disptime))
1201                         break;
1202 
1203                 throtl_dequeue_tg(tg);
1204 
1205                 nr_disp += throtl_dispatch_tg(tg);
1206 
1207                 sq = &tg->service_queue;
1208                 if (sq->nr_queued[0] || sq->nr_queued[1])
1209                         tg_update_disptime(tg);
1210 
1211                 if (nr_disp >= throtl_quantum)
1212                         break;
1213         }
1214 
1215         return nr_disp;
1216 }
1217 
1218 static bool throtl_can_upgrade(struct throtl_data *td,
1219         struct throtl_grp *this_tg);
1220 /**
1221  * throtl_pending_timer_fn - timer function for service_queue->pending_timer
1222  * @t: the pending_timer member of the throtl_service_queue being serviced
1223  *
1224  * This timer is armed when a child throtl_grp with active bio's become
1225  * pending and queued on the service_queue's pending_tree and expires when
1226  * the first child throtl_grp should be dispatched.  This function
1227  * dispatches bio's from the children throtl_grps to the parent
1228  * service_queue.
1229  *
1230  * If the parent's parent is another throtl_grp, dispatching is propagated
1231  * by either arming its pending_timer or repeating dispatch directly.  If
1232  * the top-level service_tree is reached, throtl_data->dispatch_work is
1233  * kicked so that the ready bio's are issued.
1234  */
1235 static void throtl_pending_timer_fn(struct timer_list *t)
1236 {
1237         struct throtl_service_queue *sq = from_timer(sq, t, pending_timer);
1238         struct throtl_grp *tg = sq_to_tg(sq);
1239         struct throtl_data *td = sq_to_td(sq);
1240         struct request_queue *q = td->queue;
1241         struct throtl_service_queue *parent_sq;
1242         bool dispatched;
1243         int ret;
1244 
1245         spin_lock_irq(&q->queue_lock);
1246         if (throtl_can_upgrade(td, NULL))
1247                 throtl_upgrade_state(td);
1248 
1249 again:
1250         parent_sq = sq->parent_sq;
1251         dispatched = false;
1252 
1253         while (true) {
1254                 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u",
1255                            sq->nr_queued[READ] + sq->nr_queued[WRITE],
1256                            sq->nr_queued[READ], sq->nr_queued[WRITE]);
1257 
1258                 ret = throtl_select_dispatch(sq);
1259                 if (ret) {
1260                         throtl_log(sq, "bios disp=%u", ret);
1261                         dispatched = true;
1262                 }
1263 
1264                 if (throtl_schedule_next_dispatch(sq, false))
1265                         break;
1266 
1267                 /* this dispatch windows is still open, relax and repeat */
1268                 spin_unlock_irq(&q->queue_lock);
1269                 cpu_relax();
1270                 spin_lock_irq(&q->queue_lock);
1271         }
1272 
1273         if (!dispatched)
1274                 goto out_unlock;
1275 
1276         if (parent_sq) {
1277                 /* @parent_sq is another throl_grp, propagate dispatch */
1278                 if (tg->flags & THROTL_TG_WAS_EMPTY) {
1279                         tg_update_disptime(tg);
1280                         if (!throtl_schedule_next_dispatch(parent_sq, false)) {
1281                                 /* window is already open, repeat dispatching */
1282                                 sq = parent_sq;
1283                                 tg = sq_to_tg(sq);
1284                                 goto again;
1285                         }
1286                 }
1287         } else {
1288                 /* reached the top-level, queue issueing */
1289                 queue_work(kthrotld_workqueue, &td->dispatch_work);
1290         }
1291 out_unlock:
1292         spin_unlock_irq(&q->queue_lock);
1293 }
1294 
1295 /**
1296  * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work
1297  * @work: work item being executed
1298  *
1299  * This function is queued for execution when bio's reach the bio_lists[]
1300  * of throtl_data->service_queue.  Those bio's are ready and issued by this
1301  * function.
1302  */
1303 static void blk_throtl_dispatch_work_fn(struct work_struct *work)
1304 {
1305         struct throtl_data *td = container_of(work, struct throtl_data,
1306                                               dispatch_work);
1307         struct throtl_service_queue *td_sq = &td->service_queue;
1308         struct request_queue *q = td->queue;
1309         struct bio_list bio_list_on_stack;
1310         struct bio *bio;
1311         struct blk_plug plug;
1312         int rw;
1313 
1314         bio_list_init(&bio_list_on_stack);
1315 
1316         spin_lock_irq(&q->queue_lock);
1317         for (rw = READ; rw <= WRITE; rw++)
1318                 while ((bio = throtl_pop_queued(&td_sq->queued[rw], NULL)))
1319                         bio_list_add(&bio_list_on_stack, bio);
1320         spin_unlock_irq(&q->queue_lock);
1321 
1322         if (!bio_list_empty(&bio_list_on_stack)) {
1323                 blk_start_plug(&plug);
1324                 while((bio = bio_list_pop(&bio_list_on_stack)))
1325                         generic_make_request(bio);
1326                 blk_finish_plug(&plug);
1327         }
1328 }
1329 
1330 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd,
1331                               int off)
1332 {
1333         struct throtl_grp *tg = pd_to_tg(pd);
1334         u64 v = *(u64 *)((void *)tg + off);
1335 
1336         if (v == U64_MAX)
1337                 return 0;
1338         return __blkg_prfill_u64(sf, pd, v);
1339 }
1340 
1341 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd,
1342                                int off)
1343 {
1344         struct throtl_grp *tg = pd_to_tg(pd);
1345         unsigned int v = *(unsigned int *)((void *)tg + off);
1346 
1347         if (v == UINT_MAX)
1348                 return 0;
1349         return __blkg_prfill_u64(sf, pd, v);
1350 }
1351 
1352 static int tg_print_conf_u64(struct seq_file *sf, void *v)
1353 {
1354         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64,
1355                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1356         return 0;
1357 }
1358 
1359 static int tg_print_conf_uint(struct seq_file *sf, void *v)
1360 {
1361         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint,
1362                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1363         return 0;
1364 }
1365 
1366 static void tg_conf_updated(struct throtl_grp *tg, bool global)
1367 {
1368         struct throtl_service_queue *sq = &tg->service_queue;
1369         struct cgroup_subsys_state *pos_css;
1370         struct blkcg_gq *blkg;
1371 
1372         throtl_log(&tg->service_queue,
1373                    "limit change rbps=%llu wbps=%llu riops=%u wiops=%u",
1374                    tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE),
1375                    tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE));
1376 
1377         /*
1378          * Update has_rules[] flags for the updated tg's subtree.  A tg is
1379          * considered to have rules if either the tg itself or any of its
1380          * ancestors has rules.  This identifies groups without any
1381          * restrictions in the whole hierarchy and allows them to bypass
1382          * blk-throttle.
1383          */
1384         blkg_for_each_descendant_pre(blkg, pos_css,
1385                         global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) {
1386                 struct throtl_grp *this_tg = blkg_to_tg(blkg);
1387                 struct throtl_grp *parent_tg;
1388 
1389                 tg_update_has_rules(this_tg);
1390                 /* ignore root/second level */
1391                 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent ||
1392                     !blkg->parent->parent)
1393                         continue;
1394                 parent_tg = blkg_to_tg(blkg->parent);
1395                 /*
1396                  * make sure all children has lower idle time threshold and
1397                  * higher latency target
1398                  */
1399                 this_tg->idletime_threshold = min(this_tg->idletime_threshold,
1400                                 parent_tg->idletime_threshold);
1401                 this_tg->latency_target = max(this_tg->latency_target,
1402                                 parent_tg->latency_target);
1403         }
1404 
1405         /*
1406          * We're already holding queue_lock and know @tg is valid.  Let's
1407          * apply the new config directly.
1408          *
1409          * Restart the slices for both READ and WRITES. It might happen
1410          * that a group's limit are dropped suddenly and we don't want to
1411          * account recently dispatched IO with new low rate.
1412          */
1413         throtl_start_new_slice(tg, 0);
1414         throtl_start_new_slice(tg, 1);
1415 
1416         if (tg->flags & THROTL_TG_PENDING) {
1417                 tg_update_disptime(tg);
1418                 throtl_schedule_next_dispatch(sq->parent_sq, true);
1419         }
1420 }
1421 
1422 static ssize_t tg_set_conf(struct kernfs_open_file *of,
1423                            char *buf, size_t nbytes, loff_t off, bool is_u64)
1424 {
1425         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1426         struct blkg_conf_ctx ctx;
1427         struct throtl_grp *tg;
1428         int ret;
1429         u64 v;
1430 
1431         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1432         if (ret)
1433                 return ret;
1434 
1435         ret = -EINVAL;
1436         if (sscanf(ctx.body, "%llu", &v) != 1)
1437                 goto out_finish;
1438         if (!v)
1439                 v = U64_MAX;
1440 
1441         tg = blkg_to_tg(ctx.blkg);
1442 
1443         if (is_u64)
1444                 *(u64 *)((void *)tg + of_cft(of)->private) = v;
1445         else
1446                 *(unsigned int *)((void *)tg + of_cft(of)->private) = v;
1447 
1448         tg_conf_updated(tg, false);
1449         ret = 0;
1450 out_finish:
1451         blkg_conf_finish(&ctx);
1452         return ret ?: nbytes;
1453 }
1454 
1455 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of,
1456                                char *buf, size_t nbytes, loff_t off)
1457 {
1458         return tg_set_conf(of, buf, nbytes, off, true);
1459 }
1460 
1461 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of,
1462                                 char *buf, size_t nbytes, loff_t off)
1463 {
1464         return tg_set_conf(of, buf, nbytes, off, false);
1465 }
1466 
1467 static struct cftype throtl_legacy_files[] = {
1468         {
1469                 .name = "throttle.read_bps_device",
1470                 .private = offsetof(struct throtl_grp, bps[READ][LIMIT_MAX]),
1471                 .seq_show = tg_print_conf_u64,
1472                 .write = tg_set_conf_u64,
1473         },
1474         {
1475                 .name = "throttle.write_bps_device",
1476                 .private = offsetof(struct throtl_grp, bps[WRITE][LIMIT_MAX]),
1477                 .seq_show = tg_print_conf_u64,
1478                 .write = tg_set_conf_u64,
1479         },
1480         {
1481                 .name = "throttle.read_iops_device",
1482                 .private = offsetof(struct throtl_grp, iops[READ][LIMIT_MAX]),
1483                 .seq_show = tg_print_conf_uint,
1484                 .write = tg_set_conf_uint,
1485         },
1486         {
1487                 .name = "throttle.write_iops_device",
1488                 .private = offsetof(struct throtl_grp, iops[WRITE][LIMIT_MAX]),
1489                 .seq_show = tg_print_conf_uint,
1490                 .write = tg_set_conf_uint,
1491         },
1492         {
1493                 .name = "throttle.io_service_bytes",
1494                 .private = (unsigned long)&blkcg_policy_throtl,
1495                 .seq_show = blkg_print_stat_bytes,
1496         },
1497         {
1498                 .name = "throttle.io_service_bytes_recursive",
1499                 .private = (unsigned long)&blkcg_policy_throtl,
1500                 .seq_show = blkg_print_stat_bytes_recursive,
1501         },
1502         {
1503                 .name = "throttle.io_serviced",
1504                 .private = (unsigned long)&blkcg_policy_throtl,
1505                 .seq_show = blkg_print_stat_ios,
1506         },
1507         {
1508                 .name = "throttle.io_serviced_recursive",
1509                 .private = (unsigned long)&blkcg_policy_throtl,
1510                 .seq_show = blkg_print_stat_ios_recursive,
1511         },
1512         { }     /* terminate */
1513 };
1514 
1515 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd,
1516                          int off)
1517 {
1518         struct throtl_grp *tg = pd_to_tg(pd);
1519         const char *dname = blkg_dev_name(pd->blkg);
1520         char bufs[4][21] = { "max", "max", "max", "max" };
1521         u64 bps_dft;
1522         unsigned int iops_dft;
1523         char idle_time[26] = "";
1524         char latency_time[26] = "";
1525 
1526         if (!dname)
1527                 return 0;
1528 
1529         if (off == LIMIT_LOW) {
1530                 bps_dft = 0;
1531                 iops_dft = 0;
1532         } else {
1533                 bps_dft = U64_MAX;
1534                 iops_dft = UINT_MAX;
1535         }
1536 
1537         if (tg->bps_conf[READ][off] == bps_dft &&
1538             tg->bps_conf[WRITE][off] == bps_dft &&
1539             tg->iops_conf[READ][off] == iops_dft &&
1540             tg->iops_conf[WRITE][off] == iops_dft &&
1541             (off != LIMIT_LOW ||
1542              (tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD &&
1543               tg->latency_target_conf == DFL_LATENCY_TARGET)))
1544                 return 0;
1545 
1546         if (tg->bps_conf[READ][off] != U64_MAX)
1547                 snprintf(bufs[0], sizeof(bufs[0]), "%llu",
1548                         tg->bps_conf[READ][off]);
1549         if (tg->bps_conf[WRITE][off] != U64_MAX)
1550                 snprintf(bufs[1], sizeof(bufs[1]), "%llu",
1551                         tg->bps_conf[WRITE][off]);
1552         if (tg->iops_conf[READ][off] != UINT_MAX)
1553                 snprintf(bufs[2], sizeof(bufs[2]), "%u",
1554                         tg->iops_conf[READ][off]);
1555         if (tg->iops_conf[WRITE][off] != UINT_MAX)
1556                 snprintf(bufs[3], sizeof(bufs[3]), "%u",
1557                         tg->iops_conf[WRITE][off]);
1558         if (off == LIMIT_LOW) {
1559                 if (tg->idletime_threshold_conf == ULONG_MAX)
1560                         strcpy(idle_time, " idle=max");
1561                 else
1562                         snprintf(idle_time, sizeof(idle_time), " idle=%lu",
1563                                 tg->idletime_threshold_conf);
1564 
1565                 if (tg->latency_target_conf == ULONG_MAX)
1566                         strcpy(latency_time, " latency=max");
1567                 else
1568                         snprintf(latency_time, sizeof(latency_time),
1569                                 " latency=%lu", tg->latency_target_conf);
1570         }
1571 
1572         seq_printf(sf, "%s rbps=%s wbps=%s riops=%s wiops=%s%s%s\n",
1573                    dname, bufs[0], bufs[1], bufs[2], bufs[3], idle_time,
1574                    latency_time);
1575         return 0;
1576 }
1577 
1578 static int tg_print_limit(struct seq_file *sf, void *v)
1579 {
1580         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit,
1581                           &blkcg_policy_throtl, seq_cft(sf)->private, false);
1582         return 0;
1583 }
1584 
1585 static ssize_t tg_set_limit(struct kernfs_open_file *of,
1586                           char *buf, size_t nbytes, loff_t off)
1587 {
1588         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1589         struct blkg_conf_ctx ctx;
1590         struct throtl_grp *tg;
1591         u64 v[4];
1592         unsigned long idle_time;
1593         unsigned long latency_time;
1594         int ret;
1595         int index = of_cft(of)->private;
1596 
1597         ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, buf, &ctx);
1598         if (ret)
1599                 return ret;
1600 
1601         tg = blkg_to_tg(ctx.blkg);
1602 
1603         v[0] = tg->bps_conf[READ][index];
1604         v[1] = tg->bps_conf[WRITE][index];
1605         v[2] = tg->iops_conf[READ][index];
1606         v[3] = tg->iops_conf[WRITE][index];
1607 
1608         idle_time = tg->idletime_threshold_conf;
1609         latency_time = tg->latency_target_conf;
1610         while (true) {
1611                 char tok[27];   /* wiops=18446744073709551616 */
1612                 char *p;
1613                 u64 val = U64_MAX;
1614                 int len;
1615 
1616                 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1)
1617                         break;
1618                 if (tok[0] == '\0')
1619                         break;
1620                 ctx.body += len;
1621 
1622                 ret = -EINVAL;
1623                 p = tok;
1624                 strsep(&p, "=");
1625                 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max")))
1626                         goto out_finish;
1627 
1628                 ret = -ERANGE;
1629                 if (!val)
1630                         goto out_finish;
1631 
1632                 ret = -EINVAL;
1633                 if (!strcmp(tok, "rbps"))
1634                         v[0] = val;
1635                 else if (!strcmp(tok, "wbps"))
1636                         v[1] = val;
1637                 else if (!strcmp(tok, "riops"))
1638                         v[2] = min_t(u64, val, UINT_MAX);
1639                 else if (!strcmp(tok, "wiops"))
1640                         v[3] = min_t(u64, val, UINT_MAX);
1641                 else if (off == LIMIT_LOW && !strcmp(tok, "idle"))
1642                         idle_time = val;
1643                 else if (off == LIMIT_LOW && !strcmp(tok, "latency"))
1644                         latency_time = val;
1645                 else
1646                         goto out_finish;
1647         }
1648 
1649         tg->bps_conf[READ][index] = v[0];
1650         tg->bps_conf[WRITE][index] = v[1];
1651         tg->iops_conf[READ][index] = v[2];
1652         tg->iops_conf[WRITE][index] = v[3];
1653 
1654         if (index == LIMIT_MAX) {
1655                 tg->bps[READ][index] = v[0];
1656                 tg->bps[WRITE][index] = v[1];
1657                 tg->iops[READ][index] = v[2];
1658                 tg->iops[WRITE][index] = v[3];
1659         }
1660         tg->bps[READ][LIMIT_LOW] = min(tg->bps_conf[READ][LIMIT_LOW],
1661                 tg->bps_conf[READ][LIMIT_MAX]);
1662         tg->bps[WRITE][LIMIT_LOW] = min(tg->bps_conf[WRITE][LIMIT_LOW],
1663                 tg->bps_conf[WRITE][LIMIT_MAX]);
1664         tg->iops[READ][LIMIT_LOW] = min(tg->iops_conf[READ][LIMIT_LOW],
1665                 tg->iops_conf[READ][LIMIT_MAX]);
1666         tg->iops[WRITE][LIMIT_LOW] = min(tg->iops_conf[WRITE][LIMIT_LOW],
1667                 tg->iops_conf[WRITE][LIMIT_MAX]);
1668         tg->idletime_threshold_conf = idle_time;
1669         tg->latency_target_conf = latency_time;
1670 
1671         /* force user to configure all settings for low limit  */
1672         if (!(tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW] ||
1673               tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW]) ||
1674             tg->idletime_threshold_conf == DFL_IDLE_THRESHOLD ||
1675             tg->latency_target_conf == DFL_LATENCY_TARGET) {
1676                 tg->bps[READ][LIMIT_LOW] = 0;
1677                 tg->bps[WRITE][LIMIT_LOW] = 0;
1678                 tg->iops[READ][LIMIT_LOW] = 0;
1679                 tg->iops[WRITE][LIMIT_LOW] = 0;
1680                 tg->idletime_threshold = DFL_IDLE_THRESHOLD;
1681                 tg->latency_target = DFL_LATENCY_TARGET;
1682         } else if (index == LIMIT_LOW) {
1683                 tg->idletime_threshold = tg->idletime_threshold_conf;
1684                 tg->latency_target = tg->latency_target_conf;
1685         }
1686 
1687         blk_throtl_update_limit_valid(tg->td);
1688         if (tg->td->limit_valid[LIMIT_LOW]) {
1689                 if (index == LIMIT_LOW)
1690                         tg->td->limit_index = LIMIT_LOW;
1691         } else
1692                 tg->td->limit_index = LIMIT_MAX;
1693         tg_conf_updated(tg, index == LIMIT_LOW &&
1694                 tg->td->limit_valid[LIMIT_LOW]);
1695         ret = 0;
1696 out_finish:
1697         blkg_conf_finish(&ctx);
1698         return ret ?: nbytes;
1699 }
1700 
1701 static struct cftype throtl_files[] = {
1702 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
1703         {
1704                 .name = "low",
1705                 .flags = CFTYPE_NOT_ON_ROOT,
1706                 .seq_show = tg_print_limit,
1707                 .write = tg_set_limit,
1708                 .private = LIMIT_LOW,
1709         },
1710 #endif
1711         {
1712                 .name = "max",
1713                 .flags = CFTYPE_NOT_ON_ROOT,
1714                 .seq_show = tg_print_limit,
1715                 .write = tg_set_limit,
1716                 .private = LIMIT_MAX,
1717         },
1718         { }     /* terminate */
1719 };
1720 
1721 static void throtl_shutdown_wq(struct request_queue *q)
1722 {
1723         struct throtl_data *td = q->td;
1724 
1725         cancel_work_sync(&td->dispatch_work);
1726 }
1727 
1728 static struct blkcg_policy blkcg_policy_throtl = {
1729         .dfl_cftypes            = throtl_files,
1730         .legacy_cftypes         = throtl_legacy_files,
1731 
1732         .pd_alloc_fn            = throtl_pd_alloc,
1733         .pd_init_fn             = throtl_pd_init,
1734         .pd_online_fn           = throtl_pd_online,
1735         .pd_offline_fn          = throtl_pd_offline,
1736         .pd_free_fn             = throtl_pd_free,
1737 };
1738 
1739 static unsigned long __tg_last_low_overflow_time(struct throtl_grp *tg)
1740 {
1741         unsigned long rtime = jiffies, wtime = jiffies;
1742 
1743         if (tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW])
1744                 rtime = tg->last_low_overflow_time[READ];
1745         if (tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW])
1746                 wtime = tg->last_low_overflow_time[WRITE];
1747         return min(rtime, wtime);
1748 }
1749 
1750 /* tg should not be an intermediate node */
1751 static unsigned long tg_last_low_overflow_time(struct throtl_grp *tg)
1752 {
1753         struct throtl_service_queue *parent_sq;
1754         struct throtl_grp *parent = tg;
1755         unsigned long ret = __tg_last_low_overflow_time(tg);
1756 
1757         while (true) {
1758                 parent_sq = parent->service_queue.parent_sq;
1759                 parent = sq_to_tg(parent_sq);
1760                 if (!parent)
1761                         break;
1762 
1763                 /*
1764                  * The parent doesn't have low limit, it always reaches low
1765                  * limit. Its overflow time is useless for children
1766                  */
1767                 if (!parent->bps[READ][LIMIT_LOW] &&
1768                     !parent->iops[READ][LIMIT_LOW] &&
1769                     !parent->bps[WRITE][LIMIT_LOW] &&
1770                     !parent->iops[WRITE][LIMIT_LOW])
1771                         continue;
1772                 if (time_after(__tg_last_low_overflow_time(parent), ret))
1773                         ret = __tg_last_low_overflow_time(parent);
1774         }
1775         return ret;
1776 }
1777 
1778 static bool throtl_tg_is_idle(struct throtl_grp *tg)
1779 {
1780         /*
1781          * cgroup is idle if:
1782          * - single idle is too long, longer than a fixed value (in case user
1783          *   configure a too big threshold) or 4 times of idletime threshold
1784          * - average think time is more than threshold
1785          * - IO latency is largely below threshold
1786          */
1787         unsigned long time;
1788         bool ret;
1789 
1790         time = min_t(unsigned long, MAX_IDLE_TIME, 4 * tg->idletime_threshold);
1791         ret = tg->latency_target == DFL_LATENCY_TARGET ||
1792               tg->idletime_threshold == DFL_IDLE_THRESHOLD ||
1793               (ktime_get_ns() >> 10) - tg->last_finish_time > time ||
1794               tg->avg_idletime > tg->idletime_threshold ||
1795               (tg->latency_target && tg->bio_cnt &&
1796                 tg->bad_bio_cnt * 5 < tg->bio_cnt);
1797         throtl_log(&tg->service_queue,
1798                 "avg_idle=%ld, idle_threshold=%ld, bad_bio=%d, total_bio=%d, is_idle=%d, scale=%d",
1799                 tg->avg_idletime, tg->idletime_threshold, tg->bad_bio_cnt,
1800                 tg->bio_cnt, ret, tg->td->scale);
1801         return ret;
1802 }
1803 
1804 static bool throtl_tg_can_upgrade(struct throtl_grp *tg)
1805 {
1806         struct throtl_service_queue *sq = &tg->service_queue;
1807         bool read_limit, write_limit;
1808 
1809         /*
1810          * if cgroup reaches low limit (if low limit is 0, the cgroup always
1811          * reaches), it's ok to upgrade to next limit
1812          */
1813         read_limit = tg->bps[READ][LIMIT_LOW] || tg->iops[READ][LIMIT_LOW];
1814         write_limit = tg->bps[WRITE][LIMIT_LOW] || tg->iops[WRITE][LIMIT_LOW];
1815         if (!read_limit && !write_limit)
1816                 return true;
1817         if (read_limit && sq->nr_queued[READ] &&
1818             (!write_limit || sq->nr_queued[WRITE]))
1819                 return true;
1820         if (write_limit && sq->nr_queued[WRITE] &&
1821             (!read_limit || sq->nr_queued[READ]))
1822                 return true;
1823 
1824         if (time_after_eq(jiffies,
1825                 tg_last_low_overflow_time(tg) + tg->td->throtl_slice) &&
1826             throtl_tg_is_idle(tg))
1827                 return true;
1828         return false;
1829 }
1830 
1831 static bool throtl_hierarchy_can_upgrade(struct throtl_grp *tg)
1832 {
1833         while (true) {
1834                 if (throtl_tg_can_upgrade(tg))
1835                         return true;
1836                 tg = sq_to_tg(tg->service_queue.parent_sq);
1837                 if (!tg || !tg_to_blkg(tg)->parent)
1838                         return false;
1839         }
1840         return false;
1841 }
1842 
1843 static bool throtl_can_upgrade(struct throtl_data *td,
1844         struct throtl_grp *this_tg)
1845 {
1846         struct cgroup_subsys_state *pos_css;
1847         struct blkcg_gq *blkg;
1848 
1849         if (td->limit_index != LIMIT_LOW)
1850                 return false;
1851 
1852         if (time_before(jiffies, td->low_downgrade_time + td->throtl_slice))
1853                 return false;
1854 
1855         rcu_read_lock();
1856         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1857                 struct throtl_grp *tg = blkg_to_tg(blkg);
1858 
1859                 if (tg == this_tg)
1860                         continue;
1861                 if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1862                         continue;
1863                 if (!throtl_hierarchy_can_upgrade(tg)) {
1864                         rcu_read_unlock();
1865                         return false;
1866                 }
1867         }
1868         rcu_read_unlock();
1869         return true;
1870 }
1871 
1872 static void throtl_upgrade_check(struct throtl_grp *tg)
1873 {
1874         unsigned long now = jiffies;
1875 
1876         if (tg->td->limit_index != LIMIT_LOW)
1877                 return;
1878 
1879         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1880                 return;
1881 
1882         tg->last_check_time = now;
1883 
1884         if (!time_after_eq(now,
1885              __tg_last_low_overflow_time(tg) + tg->td->throtl_slice))
1886                 return;
1887 
1888         if (throtl_can_upgrade(tg->td, NULL))
1889                 throtl_upgrade_state(tg->td);
1890 }
1891 
1892 static void throtl_upgrade_state(struct throtl_data *td)
1893 {
1894         struct cgroup_subsys_state *pos_css;
1895         struct blkcg_gq *blkg;
1896 
1897         throtl_log(&td->service_queue, "upgrade to max");
1898         td->limit_index = LIMIT_MAX;
1899         td->low_upgrade_time = jiffies;
1900         td->scale = 0;
1901         rcu_read_lock();
1902         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg) {
1903                 struct throtl_grp *tg = blkg_to_tg(blkg);
1904                 struct throtl_service_queue *sq = &tg->service_queue;
1905 
1906                 tg->disptime = jiffies - 1;
1907                 throtl_select_dispatch(sq);
1908                 throtl_schedule_next_dispatch(sq, true);
1909         }
1910         rcu_read_unlock();
1911         throtl_select_dispatch(&td->service_queue);
1912         throtl_schedule_next_dispatch(&td->service_queue, true);
1913         queue_work(kthrotld_workqueue, &td->dispatch_work);
1914 }
1915 
1916 static void throtl_downgrade_state(struct throtl_data *td, int new)
1917 {
1918         td->scale /= 2;
1919 
1920         throtl_log(&td->service_queue, "downgrade, scale %d", td->scale);
1921         if (td->scale) {
1922                 td->low_upgrade_time = jiffies - td->scale * td->throtl_slice;
1923                 return;
1924         }
1925 
1926         td->limit_index = new;
1927         td->low_downgrade_time = jiffies;
1928 }
1929 
1930 static bool throtl_tg_can_downgrade(struct throtl_grp *tg)
1931 {
1932         struct throtl_data *td = tg->td;
1933         unsigned long now = jiffies;
1934 
1935         /*
1936          * If cgroup is below low limit, consider downgrade and throttle other
1937          * cgroups
1938          */
1939         if (time_after_eq(now, td->low_upgrade_time + td->throtl_slice) &&
1940             time_after_eq(now, tg_last_low_overflow_time(tg) +
1941                                         td->throtl_slice) &&
1942             (!throtl_tg_is_idle(tg) ||
1943              !list_empty(&tg_to_blkg(tg)->blkcg->css.children)))
1944                 return true;
1945         return false;
1946 }
1947 
1948 static bool throtl_hierarchy_can_downgrade(struct throtl_grp *tg)
1949 {
1950         while (true) {
1951                 if (!throtl_tg_can_downgrade(tg))
1952                         return false;
1953                 tg = sq_to_tg(tg->service_queue.parent_sq);
1954                 if (!tg || !tg_to_blkg(tg)->parent)
1955                         break;
1956         }
1957         return true;
1958 }
1959 
1960 static void throtl_downgrade_check(struct throtl_grp *tg)
1961 {
1962         uint64_t bps;
1963         unsigned int iops;
1964         unsigned long elapsed_time;
1965         unsigned long now = jiffies;
1966 
1967         if (tg->td->limit_index != LIMIT_MAX ||
1968             !tg->td->limit_valid[LIMIT_LOW])
1969                 return;
1970         if (!list_empty(&tg_to_blkg(tg)->blkcg->css.children))
1971                 return;
1972         if (time_after(tg->last_check_time + tg->td->throtl_slice, now))
1973                 return;
1974 
1975         elapsed_time = now - tg->last_check_time;
1976         tg->last_check_time = now;
1977 
1978         if (time_before(now, tg_last_low_overflow_time(tg) +
1979                         tg->td->throtl_slice))
1980                 return;
1981 
1982         if (tg->bps[READ][LIMIT_LOW]) {
1983                 bps = tg->last_bytes_disp[READ] * HZ;
1984                 do_div(bps, elapsed_time);
1985                 if (bps >= tg->bps[READ][LIMIT_LOW])
1986                         tg->last_low_overflow_time[READ] = now;
1987         }
1988 
1989         if (tg->bps[WRITE][LIMIT_LOW]) {
1990                 bps = tg->last_bytes_disp[WRITE] * HZ;
1991                 do_div(bps, elapsed_time);
1992                 if (bps >= tg->bps[WRITE][LIMIT_LOW])
1993                         tg->last_low_overflow_time[WRITE] = now;
1994         }
1995 
1996         if (tg->iops[READ][LIMIT_LOW]) {
1997                 iops = tg->last_io_disp[READ] * HZ / elapsed_time;
1998                 if (iops >= tg->iops[READ][LIMIT_LOW])
1999                         tg->last_low_overflow_time[READ] = now;
2000         }
2001 
2002         if (tg->iops[WRITE][LIMIT_LOW]) {
2003                 iops = tg->last_io_disp[WRITE] * HZ / elapsed_time;
2004                 if (iops >= tg->iops[WRITE][LIMIT_LOW])
2005                         tg->last_low_overflow_time[WRITE] = now;
2006         }
2007 
2008         /*
2009          * If cgroup is below low limit, consider downgrade and throttle other
2010          * cgroups
2011          */
2012         if (throtl_hierarchy_can_downgrade(tg))
2013                 throtl_downgrade_state(tg->td, LIMIT_LOW);
2014 
2015         tg->last_bytes_disp[READ] = 0;
2016         tg->last_bytes_disp[WRITE] = 0;
2017         tg->last_io_disp[READ] = 0;
2018         tg->last_io_disp[WRITE] = 0;
2019 }
2020 
2021 static void blk_throtl_update_idletime(struct throtl_grp *tg)
2022 {
2023         unsigned long now = ktime_get_ns() >> 10;
2024         unsigned long last_finish_time = tg->last_finish_time;
2025 
2026         if (now <= last_finish_time || last_finish_time == 0 ||
2027             last_finish_time == tg->checked_last_finish_time)
2028                 return;
2029 
2030         tg->avg_idletime = (tg->avg_idletime * 7 + now - last_finish_time) >> 3;
2031         tg->checked_last_finish_time = last_finish_time;
2032 }
2033 
2034 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2035 static void throtl_update_latency_buckets(struct throtl_data *td)
2036 {
2037         struct avg_latency_bucket avg_latency[2][LATENCY_BUCKET_SIZE];
2038         int i, cpu, rw;
2039         unsigned long last_latency[2] = { 0 };
2040         unsigned long latency[2];
2041 
2042         if (!blk_queue_nonrot(td->queue))
2043                 return;
2044         if (time_before(jiffies, td->last_calculate_time + HZ))
2045                 return;
2046         td->last_calculate_time = jiffies;
2047 
2048         memset(avg_latency, 0, sizeof(avg_latency));
2049         for (rw = READ; rw <= WRITE; rw++) {
2050                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2051                         struct latency_bucket *tmp = &td->tmp_buckets[rw][i];
2052 
2053                         for_each_possible_cpu(cpu) {
2054                                 struct latency_bucket *bucket;
2055 
2056                                 /* this isn't race free, but ok in practice */
2057                                 bucket = per_cpu_ptr(td->latency_buckets[rw],
2058                                         cpu);
2059                                 tmp->total_latency += bucket[i].total_latency;
2060                                 tmp->samples += bucket[i].samples;
2061                                 bucket[i].total_latency = 0;
2062                                 bucket[i].samples = 0;
2063                         }
2064 
2065                         if (tmp->samples >= 32) {
2066                                 int samples = tmp->samples;
2067 
2068                                 latency[rw] = tmp->total_latency;
2069 
2070                                 tmp->total_latency = 0;
2071                                 tmp->samples = 0;
2072                                 latency[rw] /= samples;
2073                                 if (latency[rw] == 0)
2074                                         continue;
2075                                 avg_latency[rw][i].latency = latency[rw];
2076                         }
2077                 }
2078         }
2079 
2080         for (rw = READ; rw <= WRITE; rw++) {
2081                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2082                         if (!avg_latency[rw][i].latency) {
2083                                 if (td->avg_buckets[rw][i].latency < last_latency[rw])
2084                                         td->avg_buckets[rw][i].latency =
2085                                                 last_latency[rw];
2086                                 continue;
2087                         }
2088 
2089                         if (!td->avg_buckets[rw][i].valid)
2090                                 latency[rw] = avg_latency[rw][i].latency;
2091                         else
2092                                 latency[rw] = (td->avg_buckets[rw][i].latency * 7 +
2093                                         avg_latency[rw][i].latency) >> 3;
2094 
2095                         td->avg_buckets[rw][i].latency = max(latency[rw],
2096                                 last_latency[rw]);
2097                         td->avg_buckets[rw][i].valid = true;
2098                         last_latency[rw] = td->avg_buckets[rw][i].latency;
2099                 }
2100         }
2101 
2102         for (i = 0; i < LATENCY_BUCKET_SIZE; i++)
2103                 throtl_log(&td->service_queue,
2104                         "Latency bucket %d: read latency=%ld, read valid=%d, "
2105                         "write latency=%ld, write valid=%d", i,
2106                         td->avg_buckets[READ][i].latency,
2107                         td->avg_buckets[READ][i].valid,
2108                         td->avg_buckets[WRITE][i].latency,
2109                         td->avg_buckets[WRITE][i].valid);
2110 }
2111 #else
2112 static inline void throtl_update_latency_buckets(struct throtl_data *td)
2113 {
2114 }
2115 #endif
2116 
2117 bool blk_throtl_bio(struct request_queue *q, struct blkcg_gq *blkg,
2118                     struct bio *bio)
2119 {
2120         struct throtl_qnode *qn = NULL;
2121         struct throtl_grp *tg = blkg_to_tg(blkg ?: q->root_blkg);
2122         struct throtl_service_queue *sq;
2123         bool rw = bio_data_dir(bio);
2124         bool throttled = false;
2125         struct throtl_data *td = tg->td;
2126 
2127         WARN_ON_ONCE(!rcu_read_lock_held());
2128 
2129         /* see throtl_charge_bio() */
2130         if (bio_flagged(bio, BIO_THROTTLED) || !tg->has_rules[rw])
2131                 goto out;
2132 
2133         spin_lock_irq(&q->queue_lock);
2134 
2135         throtl_update_latency_buckets(td);
2136 
2137         blk_throtl_update_idletime(tg);
2138 
2139         sq = &tg->service_queue;
2140 
2141 again:
2142         while (true) {
2143                 if (tg->last_low_overflow_time[rw] == 0)
2144                         tg->last_low_overflow_time[rw] = jiffies;
2145                 throtl_downgrade_check(tg);
2146                 throtl_upgrade_check(tg);
2147                 /* throtl is FIFO - if bios are already queued, should queue */
2148                 if (sq->nr_queued[rw])
2149                         break;
2150 
2151                 /* if above limits, break to queue */
2152                 if (!tg_may_dispatch(tg, bio, NULL)) {
2153                         tg->last_low_overflow_time[rw] = jiffies;
2154                         if (throtl_can_upgrade(td, tg)) {
2155                                 throtl_upgrade_state(td);
2156                                 goto again;
2157                         }
2158                         break;
2159                 }
2160 
2161                 /* within limits, let's charge and dispatch directly */
2162                 throtl_charge_bio(tg, bio);
2163 
2164                 /*
2165                  * We need to trim slice even when bios are not being queued
2166                  * otherwise it might happen that a bio is not queued for
2167                  * a long time and slice keeps on extending and trim is not
2168                  * called for a long time. Now if limits are reduced suddenly
2169                  * we take into account all the IO dispatched so far at new
2170                  * low rate and * newly queued IO gets a really long dispatch
2171                  * time.
2172                  *
2173                  * So keep on trimming slice even if bio is not queued.
2174                  */
2175                 throtl_trim_slice(tg, rw);
2176 
2177                 /*
2178                  * @bio passed through this layer without being throttled.
2179                  * Climb up the ladder.  If we''re already at the top, it
2180                  * can be executed directly.
2181                  */
2182                 qn = &tg->qnode_on_parent[rw];
2183                 sq = sq->parent_sq;
2184                 tg = sq_to_tg(sq);
2185                 if (!tg)
2186                         goto out_unlock;
2187         }
2188 
2189         /* out-of-limit, queue to @tg */
2190         throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d",
2191                    rw == READ ? 'R' : 'W',
2192                    tg->bytes_disp[rw], bio->bi_iter.bi_size,
2193                    tg_bps_limit(tg, rw),
2194                    tg->io_disp[rw], tg_iops_limit(tg, rw),
2195                    sq->nr_queued[READ], sq->nr_queued[WRITE]);
2196 
2197         tg->last_low_overflow_time[rw] = jiffies;
2198 
2199         td->nr_queued[rw]++;
2200         throtl_add_bio_tg(bio, qn, tg);
2201         throttled = true;
2202 
2203         /*
2204          * Update @tg's dispatch time and force schedule dispatch if @tg
2205          * was empty before @bio.  The forced scheduling isn't likely to
2206          * cause undue delay as @bio is likely to be dispatched directly if
2207          * its @tg's disptime is not in the future.
2208          */
2209         if (tg->flags & THROTL_TG_WAS_EMPTY) {
2210                 tg_update_disptime(tg);
2211                 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true);
2212         }
2213 
2214 out_unlock:
2215         spin_unlock_irq(&q->queue_lock);
2216 out:
2217         bio_set_flag(bio, BIO_THROTTLED);
2218 
2219 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2220         if (throttled || !td->track_bio_latency)
2221                 bio->bi_issue.value |= BIO_ISSUE_THROTL_SKIP_LATENCY;
2222 #endif
2223         return throttled;
2224 }
2225 
2226 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2227 static void throtl_track_latency(struct throtl_data *td, sector_t size,
2228         int op, unsigned long time)
2229 {
2230         struct latency_bucket *latency;
2231         int index;
2232 
2233         if (!td || td->limit_index != LIMIT_LOW ||
2234             !(op == REQ_OP_READ || op == REQ_OP_WRITE) ||
2235             !blk_queue_nonrot(td->queue))
2236                 return;
2237 
2238         index = request_bucket_index(size);
2239 
2240         latency = get_cpu_ptr(td->latency_buckets[op]);
2241         latency[index].total_latency += time;
2242         latency[index].samples++;
2243         put_cpu_ptr(td->latency_buckets[op]);
2244 }
2245 
2246 void blk_throtl_stat_add(struct request *rq, u64 time_ns)
2247 {
2248         struct request_queue *q = rq->q;
2249         struct throtl_data *td = q->td;
2250 
2251         throtl_track_latency(td, blk_rq_stats_sectors(rq), req_op(rq),
2252                              time_ns >> 10);
2253 }
2254 
2255 void blk_throtl_bio_endio(struct bio *bio)
2256 {
2257         struct blkcg_gq *blkg;
2258         struct throtl_grp *tg;
2259         u64 finish_time_ns;
2260         unsigned long finish_time;
2261         unsigned long start_time;
2262         unsigned long lat;
2263         int rw = bio_data_dir(bio);
2264 
2265         blkg = bio->bi_blkg;
2266         if (!blkg)
2267                 return;
2268         tg = blkg_to_tg(blkg);
2269 
2270         finish_time_ns = ktime_get_ns();
2271         tg->last_finish_time = finish_time_ns >> 10;
2272 
2273         start_time = bio_issue_time(&bio->bi_issue) >> 10;
2274         finish_time = __bio_issue_time(finish_time_ns) >> 10;
2275         if (!start_time || finish_time <= start_time)
2276                 return;
2277 
2278         lat = finish_time - start_time;
2279         /* this is only for bio based driver */
2280         if (!(bio->bi_issue.value & BIO_ISSUE_THROTL_SKIP_LATENCY))
2281                 throtl_track_latency(tg->td, bio_issue_size(&bio->bi_issue),
2282                                      bio_op(bio), lat);
2283 
2284         if (tg->latency_target && lat >= tg->td->filtered_latency) {
2285                 int bucket;
2286                 unsigned int threshold;
2287 
2288                 bucket = request_bucket_index(bio_issue_size(&bio->bi_issue));
2289                 threshold = tg->td->avg_buckets[rw][bucket].latency +
2290                         tg->latency_target;
2291                 if (lat > threshold)
2292                         tg->bad_bio_cnt++;
2293                 /*
2294                  * Not race free, could get wrong count, which means cgroups
2295                  * will be throttled
2296                  */
2297                 tg->bio_cnt++;
2298         }
2299 
2300         if (time_after(jiffies, tg->bio_cnt_reset_time) || tg->bio_cnt > 1024) {
2301                 tg->bio_cnt_reset_time = tg->td->throtl_slice + jiffies;
2302                 tg->bio_cnt /= 2;
2303                 tg->bad_bio_cnt /= 2;
2304         }
2305 }
2306 #endif
2307 
2308 /*
2309  * Dispatch all bios from all children tg's queued on @parent_sq.  On
2310  * return, @parent_sq is guaranteed to not have any active children tg's
2311  * and all bios from previously active tg's are on @parent_sq->bio_lists[].
2312  */
2313 static void tg_drain_bios(struct throtl_service_queue *parent_sq)
2314 {
2315         struct throtl_grp *tg;
2316 
2317         while ((tg = throtl_rb_first(parent_sq))) {
2318                 struct throtl_service_queue *sq = &tg->service_queue;
2319                 struct bio *bio;
2320 
2321                 throtl_dequeue_tg(tg);
2322 
2323                 while ((bio = throtl_peek_queued(&sq->queued[READ])))
2324                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2325                 while ((bio = throtl_peek_queued(&sq->queued[WRITE])))
2326                         tg_dispatch_one_bio(tg, bio_data_dir(bio));
2327         }
2328 }
2329 
2330 /**
2331  * blk_throtl_drain - drain throttled bios
2332  * @q: request_queue to drain throttled bios for
2333  *
2334  * Dispatch all currently throttled bios on @q through ->make_request_fn().
2335  */
2336 void blk_throtl_drain(struct request_queue *q)
2337         __releases(&q->queue_lock) __acquires(&q->queue_lock)
2338 {
2339         struct throtl_data *td = q->td;
2340         struct blkcg_gq *blkg;
2341         struct cgroup_subsys_state *pos_css;
2342         struct bio *bio;
2343         int rw;
2344 
2345         rcu_read_lock();
2346 
2347         /*
2348          * Drain each tg while doing post-order walk on the blkg tree, so
2349          * that all bios are propagated to td->service_queue.  It'd be
2350          * better to walk service_queue tree directly but blkg walk is
2351          * easier.
2352          */
2353         blkg_for_each_descendant_post(blkg, pos_css, td->queue->root_blkg)
2354                 tg_drain_bios(&blkg_to_tg(blkg)->service_queue);
2355 
2356         /* finally, transfer bios from top-level tg's into the td */
2357         tg_drain_bios(&td->service_queue);
2358 
2359         rcu_read_unlock();
2360         spin_unlock_irq(&q->queue_lock);
2361 
2362         /* all bios now should be in td->service_queue, issue them */
2363         for (rw = READ; rw <= WRITE; rw++)
2364                 while ((bio = throtl_pop_queued(&td->service_queue.queued[rw],
2365                                                 NULL)))
2366                         generic_make_request(bio);
2367 
2368         spin_lock_irq(&q->queue_lock);
2369 }
2370 
2371 int blk_throtl_init(struct request_queue *q)
2372 {
2373         struct throtl_data *td;
2374         int ret;
2375 
2376         td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node);
2377         if (!td)
2378                 return -ENOMEM;
2379         td->latency_buckets[READ] = __alloc_percpu(sizeof(struct latency_bucket) *
2380                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2381         if (!td->latency_buckets[READ]) {
2382                 kfree(td);
2383                 return -ENOMEM;
2384         }
2385         td->latency_buckets[WRITE] = __alloc_percpu(sizeof(struct latency_bucket) *
2386                 LATENCY_BUCKET_SIZE, __alignof__(u64));
2387         if (!td->latency_buckets[WRITE]) {
2388                 free_percpu(td->latency_buckets[READ]);
2389                 kfree(td);
2390                 return -ENOMEM;
2391         }
2392 
2393         INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn);
2394         throtl_service_queue_init(&td->service_queue);
2395 
2396         q->td = td;
2397         td->queue = q;
2398 
2399         td->limit_valid[LIMIT_MAX] = true;
2400         td->limit_index = LIMIT_MAX;
2401         td->low_upgrade_time = jiffies;
2402         td->low_downgrade_time = jiffies;
2403 
2404         /* activate policy */
2405         ret = blkcg_activate_policy(q, &blkcg_policy_throtl);
2406         if (ret) {
2407                 free_percpu(td->latency_buckets[READ]);
2408                 free_percpu(td->latency_buckets[WRITE]);
2409                 kfree(td);
2410         }
2411         return ret;
2412 }
2413 
2414 void blk_throtl_exit(struct request_queue *q)
2415 {
2416         BUG_ON(!q->td);
2417         throtl_shutdown_wq(q);
2418         blkcg_deactivate_policy(q, &blkcg_policy_throtl);
2419         free_percpu(q->td->latency_buckets[READ]);
2420         free_percpu(q->td->latency_buckets[WRITE]);
2421         kfree(q->td);
2422 }
2423 
2424 void blk_throtl_register_queue(struct request_queue *q)
2425 {
2426         struct throtl_data *td;
2427         int i;
2428 
2429         td = q->td;
2430         BUG_ON(!td);
2431 
2432         if (blk_queue_nonrot(q)) {
2433                 td->throtl_slice = DFL_THROTL_SLICE_SSD;
2434                 td->filtered_latency = LATENCY_FILTERED_SSD;
2435         } else {
2436                 td->throtl_slice = DFL_THROTL_SLICE_HD;
2437                 td->filtered_latency = LATENCY_FILTERED_HD;
2438                 for (i = 0; i < LATENCY_BUCKET_SIZE; i++) {
2439                         td->avg_buckets[READ][i].latency = DFL_HD_BASELINE_LATENCY;
2440                         td->avg_buckets[WRITE][i].latency = DFL_HD_BASELINE_LATENCY;
2441                 }
2442         }
2443 #ifndef CONFIG_BLK_DEV_THROTTLING_LOW
2444         /* if no low limit, use previous default */
2445         td->throtl_slice = DFL_THROTL_SLICE_HD;
2446 #endif
2447 
2448         td->track_bio_latency = !queue_is_mq(q);
2449         if (!td->track_bio_latency)
2450                 blk_stat_enable_accounting(q);
2451 }
2452 
2453 #ifdef CONFIG_BLK_DEV_THROTTLING_LOW
2454 ssize_t blk_throtl_sample_time_show(struct request_queue *q, char *page)
2455 {
2456         if (!q->td)
2457                 return -EINVAL;
2458         return sprintf(page, "%u\n", jiffies_to_msecs(q->td->throtl_slice));
2459 }
2460 
2461 ssize_t blk_throtl_sample_time_store(struct request_queue *q,
2462         const char *page, size_t count)
2463 {
2464         unsigned long v;
2465         unsigned long t;
2466 
2467         if (!q->td)
2468                 return -EINVAL;
2469         if (kstrtoul(page, 10, &v))
2470                 return -EINVAL;
2471         t = msecs_to_jiffies(v);
2472         if (t == 0 || t > MAX_THROTL_SLICE)
2473                 return -EINVAL;
2474         q->td->throtl_slice = t;
2475         return count;
2476 }
2477 #endif
2478 
2479 static int __init throtl_init(void)
2480 {
2481         kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0);
2482         if (!kthrotld_workqueue)
2483                 panic("Failed to create kthrotld\n");
2484 
2485         return blkcg_policy_register(&blkcg_policy_throtl);
2486 }
2487 
2488 module_init(throtl_init);