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 "blk.h" 14 #include "blk-cgroup-rwstat.h" 15 #include "blk-stat.h" 16 #include "blk-throttle.h" 17 18 /* Max dispatch from a group in 1 round */ 19 #define THROTL_GRP_QUANTUM 8 20 21 /* Total max dispatch from all groups in one round */ 22 #define THROTL_QUANTUM 32 23 24 /* Throttling is performed over a slice and after that slice is renewed */ 25 #define DFL_THROTL_SLICE_HD (HZ / 10) 26 #define DFL_THROTL_SLICE_SSD (HZ / 50) 27 #define MAX_THROTL_SLICE (HZ) 28 29 /* A workqueue to queue throttle related work */ 30 static struct workqueue_struct *kthrotld_workqueue; 31 32 #define rb_entry_tg(node) rb_entry((node), struct throtl_grp, rb_node) 33 34 struct throtl_data 35 { 36 /* service tree for active throtl groups */ 37 struct throtl_service_queue service_queue; 38 39 struct request_queue *queue; 40 41 /* Total Number of queued bios on READ and WRITE lists */ 42 unsigned int nr_queued[2]; 43 44 unsigned int throtl_slice; 45 46 /* Work for dispatching throttled bios */ 47 struct work_struct dispatch_work; 48 49 bool track_bio_latency; 50 }; 51 52 static void throtl_pending_timer_fn(struct timer_list *t); 53 54 static inline struct blkcg_gq *tg_to_blkg(struct throtl_grp *tg) 55 { 56 return pd_to_blkg(&tg->pd); 57 } 58 59 /** 60 * sq_to_tg - return the throl_grp the specified service queue belongs to 61 * @sq: the throtl_service_queue of interest 62 * 63 * Return the throtl_grp @sq belongs to. If @sq is the top-level one 64 * embedded in throtl_data, %NULL is returned. 65 */ 66 static struct throtl_grp *sq_to_tg(struct throtl_service_queue *sq) 67 { 68 if (sq && sq->parent_sq) 69 return container_of(sq, struct throtl_grp, service_queue); 70 else 71 return NULL; 72 } 73 74 /** 75 * sq_to_td - return throtl_data the specified service queue belongs to 76 * @sq: the throtl_service_queue of interest 77 * 78 * A service_queue can be embedded in either a throtl_grp or throtl_data. 79 * Determine the associated throtl_data accordingly and return it. 80 */ 81 static struct throtl_data *sq_to_td(struct throtl_service_queue *sq) 82 { 83 struct throtl_grp *tg = sq_to_tg(sq); 84 85 if (tg) 86 return tg->td; 87 else 88 return container_of(sq, struct throtl_data, service_queue); 89 } 90 91 static uint64_t tg_bps_limit(struct throtl_grp *tg, int rw) 92 { 93 struct blkcg_gq *blkg = tg_to_blkg(tg); 94 95 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) 96 return U64_MAX; 97 98 return tg->bps[rw]; 99 } 100 101 static unsigned int tg_iops_limit(struct throtl_grp *tg, int rw) 102 { 103 struct blkcg_gq *blkg = tg_to_blkg(tg); 104 105 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && !blkg->parent) 106 return UINT_MAX; 107 108 return tg->iops[rw]; 109 } 110 111 /** 112 * throtl_log - log debug message via blktrace 113 * @sq: the service_queue being reported 114 * @fmt: printf format string 115 * @args: printf args 116 * 117 * The messages are prefixed with "throtl BLKG_NAME" if @sq belongs to a 118 * throtl_grp; otherwise, just "throtl". 119 */ 120 #define throtl_log(sq, fmt, args...) do { \ 121 struct throtl_grp *__tg = sq_to_tg((sq)); \ 122 struct throtl_data *__td = sq_to_td((sq)); \ 123 \ 124 (void)__td; \ 125 if (likely(!blk_trace_note_message_enabled(__td->queue))) \ 126 break; \ 127 if ((__tg)) { \ 128 blk_add_cgroup_trace_msg(__td->queue, \ 129 &tg_to_blkg(__tg)->blkcg->css, "throtl " fmt, ##args);\ 130 } else { \ 131 blk_add_trace_msg(__td->queue, "throtl " fmt, ##args); \ 132 } \ 133 } while (0) 134 135 static inline unsigned int throtl_bio_data_size(struct bio *bio) 136 { 137 /* assume it's one sector */ 138 if (unlikely(bio_op(bio) == REQ_OP_DISCARD)) 139 return 512; 140 return bio->bi_iter.bi_size; 141 } 142 143 static void throtl_qnode_init(struct throtl_qnode *qn, struct throtl_grp *tg) 144 { 145 INIT_LIST_HEAD(&qn->node); 146 bio_list_init(&qn->bios_bps); 147 bio_list_init(&qn->bios_iops); 148 qn->tg = tg; 149 } 150 151 /** 152 * throtl_qnode_add_bio - add a bio to a throtl_qnode and activate it 153 * @bio: bio being added 154 * @qn: qnode to add bio to 155 * @sq: the service_queue @qn belongs to 156 * 157 * Add @bio to @qn and put @qn on @sq->queued if it's not already on. 158 * @qn->tg's reference count is bumped when @qn is activated. See the 159 * comment on top of throtl_qnode definition for details. 160 */ 161 static void throtl_qnode_add_bio(struct bio *bio, struct throtl_qnode *qn, 162 struct throtl_service_queue *sq) 163 { 164 bool rw = bio_data_dir(bio); 165 166 /* 167 * Split bios have already been throttled by bps, so they are 168 * directly queued into the iops path. 169 */ 170 if (bio_flagged(bio, BIO_TG_BPS_THROTTLED) || 171 bio_flagged(bio, BIO_BPS_THROTTLED)) { 172 bio_list_add(&qn->bios_iops, bio); 173 sq->nr_queued_iops[rw]++; 174 } else { 175 bio_list_add(&qn->bios_bps, bio); 176 sq->nr_queued_bps[rw]++; 177 } 178 179 if (list_empty(&qn->node)) { 180 list_add_tail(&qn->node, &sq->queued[rw]); 181 blkg_get(tg_to_blkg(qn->tg)); 182 } 183 } 184 185 /** 186 * throtl_peek_queued - peek the first bio on a qnode list 187 * @queued: the qnode list to peek 188 * 189 * Always take a bio from the head of the iops queue first. If the queue is 190 * empty, we then take it from the bps queue to maintain the overall idea of 191 * fetching bios from the head. 192 */ 193 static struct bio *throtl_peek_queued(struct list_head *queued) 194 { 195 struct throtl_qnode *qn; 196 struct bio *bio; 197 198 if (list_empty(queued)) 199 return NULL; 200 201 qn = list_first_entry(queued, struct throtl_qnode, node); 202 bio = bio_list_peek(&qn->bios_iops); 203 if (!bio) 204 bio = bio_list_peek(&qn->bios_bps); 205 WARN_ON_ONCE(!bio); 206 return bio; 207 } 208 209 /** 210 * throtl_pop_queued - pop the first bio form a qnode list 211 * @sq: the service_queue to pop a bio from 212 * @tg_to_put: optional out argument for throtl_grp to put 213 * @rw: read/write 214 * 215 * Pop the first bio from the qnode list @sq->queued. Note that we firstly 216 * focus on the iops list because bios are ultimately dispatched from it. 217 * After popping, the first qnode is removed from @sq->queued if empty or moved 218 * to the end of @sq->queued so that the popping order is round-robin. 219 * 220 * When the first qnode is removed, its associated throtl_grp should be put 221 * too. If @tg_to_put is NULL, this function automatically puts it; 222 * otherwise, *@tg_to_put is set to the throtl_grp to put and the caller is 223 * responsible for putting it. 224 */ 225 static struct bio *throtl_pop_queued(struct throtl_service_queue *sq, 226 struct throtl_grp **tg_to_put, bool rw) 227 { 228 struct list_head *queued = &sq->queued[rw]; 229 struct throtl_qnode *qn; 230 struct bio *bio; 231 232 if (list_empty(queued)) 233 return NULL; 234 235 qn = list_first_entry(queued, struct throtl_qnode, node); 236 bio = bio_list_pop(&qn->bios_iops); 237 if (bio) { 238 sq->nr_queued_iops[rw]--; 239 } else { 240 bio = bio_list_pop(&qn->bios_bps); 241 if (bio) 242 sq->nr_queued_bps[rw]--; 243 } 244 WARN_ON_ONCE(!bio); 245 246 if (bio_list_empty(&qn->bios_bps) && bio_list_empty(&qn->bios_iops)) { 247 list_del_init(&qn->node); 248 if (tg_to_put) 249 *tg_to_put = qn->tg; 250 else 251 blkg_put(tg_to_blkg(qn->tg)); 252 } else { 253 list_move_tail(&qn->node, queued); 254 } 255 256 return bio; 257 } 258 259 /* init a service_queue, assumes the caller zeroed it */ 260 static void throtl_service_queue_init(struct throtl_service_queue *sq) 261 { 262 INIT_LIST_HEAD(&sq->queued[READ]); 263 INIT_LIST_HEAD(&sq->queued[WRITE]); 264 sq->pending_tree = RB_ROOT_CACHED; 265 timer_setup(&sq->pending_timer, throtl_pending_timer_fn, 0); 266 } 267 268 static struct blkg_policy_data *throtl_pd_alloc(struct gendisk *disk, 269 struct blkcg *blkcg, gfp_t gfp) 270 { 271 struct throtl_grp *tg; 272 int rw; 273 274 tg = kzalloc_node(sizeof(*tg), gfp, disk->node_id); 275 if (!tg) 276 return NULL; 277 278 if (blkg_rwstat_init(&tg->stat_bytes, gfp)) 279 goto err_free_tg; 280 281 if (blkg_rwstat_init(&tg->stat_ios, gfp)) 282 goto err_exit_stat_bytes; 283 284 throtl_service_queue_init(&tg->service_queue); 285 286 for (rw = READ; rw <= WRITE; rw++) { 287 throtl_qnode_init(&tg->qnode_on_self[rw], tg); 288 throtl_qnode_init(&tg->qnode_on_parent[rw], tg); 289 } 290 291 RB_CLEAR_NODE(&tg->rb_node); 292 tg->bps[READ] = U64_MAX; 293 tg->bps[WRITE] = U64_MAX; 294 tg->iops[READ] = UINT_MAX; 295 tg->iops[WRITE] = UINT_MAX; 296 297 return &tg->pd; 298 299 err_exit_stat_bytes: 300 blkg_rwstat_exit(&tg->stat_bytes); 301 err_free_tg: 302 kfree(tg); 303 return NULL; 304 } 305 306 static void throtl_pd_init(struct blkg_policy_data *pd) 307 { 308 struct throtl_grp *tg = pd_to_tg(pd); 309 struct blkcg_gq *blkg = tg_to_blkg(tg); 310 struct throtl_data *td = blkg->q->td; 311 struct throtl_service_queue *sq = &tg->service_queue; 312 313 /* 314 * If on the default hierarchy, we switch to properly hierarchical 315 * behavior where limits on a given throtl_grp are applied to the 316 * whole subtree rather than just the group itself. e.g. If 16M 317 * read_bps limit is set on a parent group, summary bps of 318 * parent group and its subtree groups can't exceed 16M for the 319 * device. 320 * 321 * If not on the default hierarchy, the broken flat hierarchy 322 * behavior is retained where all throtl_grps are treated as if 323 * they're all separate root groups right below throtl_data. 324 * Limits of a group don't interact with limits of other groups 325 * regardless of the position of the group in the hierarchy. 326 */ 327 sq->parent_sq = &td->service_queue; 328 if (cgroup_subsys_on_dfl(io_cgrp_subsys) && blkg->parent) 329 sq->parent_sq = &blkg_to_tg(blkg->parent)->service_queue; 330 tg->td = td; 331 } 332 333 /* 334 * Set has_rules[] if @tg or any of its parents have limits configured. 335 * This doesn't require walking up to the top of the hierarchy as the 336 * parent's has_rules[] is guaranteed to be correct. 337 */ 338 static void tg_update_has_rules(struct throtl_grp *tg) 339 { 340 struct throtl_grp *parent_tg = sq_to_tg(tg->service_queue.parent_sq); 341 int rw; 342 343 for (rw = READ; rw <= WRITE; rw++) { 344 tg->has_rules_iops[rw] = 345 (parent_tg && parent_tg->has_rules_iops[rw]) || 346 tg_iops_limit(tg, rw) != UINT_MAX; 347 tg->has_rules_bps[rw] = 348 (parent_tg && parent_tg->has_rules_bps[rw]) || 349 tg_bps_limit(tg, rw) != U64_MAX; 350 } 351 } 352 353 static void throtl_pd_online(struct blkg_policy_data *pd) 354 { 355 struct throtl_grp *tg = pd_to_tg(pd); 356 /* 357 * We don't want new groups to escape the limits of its ancestors. 358 * Update has_rules[] after a new group is brought online. 359 */ 360 tg_update_has_rules(tg); 361 } 362 363 static void throtl_pd_free(struct blkg_policy_data *pd) 364 { 365 struct throtl_grp *tg = pd_to_tg(pd); 366 367 timer_delete_sync(&tg->service_queue.pending_timer); 368 blkg_rwstat_exit(&tg->stat_bytes); 369 blkg_rwstat_exit(&tg->stat_ios); 370 kfree(tg); 371 } 372 373 static struct throtl_grp * 374 throtl_rb_first(struct throtl_service_queue *parent_sq) 375 { 376 struct rb_node *n; 377 378 n = rb_first_cached(&parent_sq->pending_tree); 379 WARN_ON_ONCE(!n); 380 if (!n) 381 return NULL; 382 return rb_entry_tg(n); 383 } 384 385 static void throtl_rb_erase(struct rb_node *n, 386 struct throtl_service_queue *parent_sq) 387 { 388 rb_erase_cached(n, &parent_sq->pending_tree); 389 RB_CLEAR_NODE(n); 390 } 391 392 static void update_min_dispatch_time(struct throtl_service_queue *parent_sq) 393 { 394 struct throtl_grp *tg; 395 396 tg = throtl_rb_first(parent_sq); 397 if (!tg) 398 return; 399 400 parent_sq->first_pending_disptime = tg->disptime; 401 } 402 403 static void tg_service_queue_add(struct throtl_grp *tg) 404 { 405 struct throtl_service_queue *parent_sq = tg->service_queue.parent_sq; 406 struct rb_node **node = &parent_sq->pending_tree.rb_root.rb_node; 407 struct rb_node *parent = NULL; 408 struct throtl_grp *__tg; 409 unsigned long key = tg->disptime; 410 bool leftmost = true; 411 412 while (*node != NULL) { 413 parent = *node; 414 __tg = rb_entry_tg(parent); 415 416 if (time_before(key, __tg->disptime)) 417 node = &parent->rb_left; 418 else { 419 node = &parent->rb_right; 420 leftmost = false; 421 } 422 } 423 424 rb_link_node(&tg->rb_node, parent, node); 425 rb_insert_color_cached(&tg->rb_node, &parent_sq->pending_tree, 426 leftmost); 427 } 428 429 static void throtl_enqueue_tg(struct throtl_grp *tg) 430 { 431 if (!(tg->flags & THROTL_TG_PENDING)) { 432 tg_service_queue_add(tg); 433 tg->flags |= THROTL_TG_PENDING; 434 tg->service_queue.parent_sq->nr_pending++; 435 } 436 } 437 438 static void throtl_dequeue_tg(struct throtl_grp *tg) 439 { 440 if (tg->flags & THROTL_TG_PENDING) { 441 struct throtl_service_queue *parent_sq = 442 tg->service_queue.parent_sq; 443 444 throtl_rb_erase(&tg->rb_node, parent_sq); 445 --parent_sq->nr_pending; 446 tg->flags &= ~THROTL_TG_PENDING; 447 } 448 } 449 450 /* Call with queue lock held */ 451 static void throtl_schedule_pending_timer(struct throtl_service_queue *sq, 452 unsigned long expires) 453 { 454 unsigned long max_expire = jiffies + 8 * sq_to_td(sq)->throtl_slice; 455 456 /* 457 * Since we are adjusting the throttle limit dynamically, the sleep 458 * time calculated according to previous limit might be invalid. It's 459 * possible the cgroup sleep time is very long and no other cgroups 460 * have IO running so notify the limit changes. Make sure the cgroup 461 * doesn't sleep too long to avoid the missed notification. 462 */ 463 if (time_after(expires, max_expire)) 464 expires = max_expire; 465 mod_timer(&sq->pending_timer, expires); 466 throtl_log(sq, "schedule timer. delay=%lu jiffies=%lu", 467 expires - jiffies, jiffies); 468 } 469 470 /** 471 * throtl_schedule_next_dispatch - schedule the next dispatch cycle 472 * @sq: the service_queue to schedule dispatch for 473 * @force: force scheduling 474 * 475 * Arm @sq->pending_timer so that the next dispatch cycle starts on the 476 * dispatch time of the first pending child. Returns %true if either timer 477 * is armed or there's no pending child left. %false if the current 478 * dispatch window is still open and the caller should continue 479 * dispatching. 480 * 481 * If @force is %true, the dispatch timer is always scheduled and this 482 * function is guaranteed to return %true. This is to be used when the 483 * caller can't dispatch itself and needs to invoke pending_timer 484 * unconditionally. Note that forced scheduling is likely to induce short 485 * delay before dispatch starts even if @sq->first_pending_disptime is not 486 * in the future and thus shouldn't be used in hot paths. 487 */ 488 static bool throtl_schedule_next_dispatch(struct throtl_service_queue *sq, 489 bool force) 490 { 491 /* any pending children left? */ 492 if (!sq->nr_pending) 493 return true; 494 495 update_min_dispatch_time(sq); 496 497 /* is the next dispatch time in the future? */ 498 if (force || time_after(sq->first_pending_disptime, jiffies)) { 499 throtl_schedule_pending_timer(sq, sq->first_pending_disptime); 500 return true; 501 } 502 503 /* tell the caller to continue dispatching */ 504 return false; 505 } 506 507 static inline void throtl_start_new_slice_with_credit(struct throtl_grp *tg, 508 bool rw, unsigned long start) 509 { 510 tg->bytes_disp[rw] = 0; 511 tg->io_disp[rw] = 0; 512 513 /* 514 * Previous slice has expired. We must have trimmed it after last 515 * bio dispatch. That means since start of last slice, we never used 516 * that bandwidth. Do try to make use of that bandwidth while giving 517 * credit. 518 */ 519 if (time_after(start, tg->slice_start[rw])) 520 tg->slice_start[rw] = start; 521 522 tg->slice_end[rw] = jiffies + tg->td->throtl_slice; 523 throtl_log(&tg->service_queue, 524 "[%c] new slice with credit start=%lu end=%lu jiffies=%lu", 525 rw == READ ? 'R' : 'W', tg->slice_start[rw], 526 tg->slice_end[rw], jiffies); 527 } 528 529 static inline void throtl_start_new_slice(struct throtl_grp *tg, bool rw, 530 bool clear) 531 { 532 if (clear) { 533 tg->bytes_disp[rw] = 0; 534 tg->io_disp[rw] = 0; 535 } 536 tg->slice_start[rw] = jiffies; 537 tg->slice_end[rw] = jiffies + tg->td->throtl_slice; 538 539 throtl_log(&tg->service_queue, 540 "[%c] new slice start=%lu end=%lu jiffies=%lu", 541 rw == READ ? 'R' : 'W', tg->slice_start[rw], 542 tg->slice_end[rw], jiffies); 543 } 544 545 static inline void throtl_set_slice_end(struct throtl_grp *tg, bool rw, 546 unsigned long jiffy_end) 547 { 548 tg->slice_end[rw] = roundup(jiffy_end, tg->td->throtl_slice); 549 } 550 551 static inline void throtl_extend_slice(struct throtl_grp *tg, bool rw, 552 unsigned long jiffy_end) 553 { 554 if (!time_before(tg->slice_end[rw], jiffy_end)) 555 return; 556 557 throtl_set_slice_end(tg, rw, jiffy_end); 558 throtl_log(&tg->service_queue, 559 "[%c] extend slice start=%lu end=%lu jiffies=%lu", 560 rw == READ ? 'R' : 'W', tg->slice_start[rw], 561 tg->slice_end[rw], jiffies); 562 } 563 564 /* Determine if previously allocated or extended slice is complete or not */ 565 static bool throtl_slice_used(struct throtl_grp *tg, bool rw) 566 { 567 if (time_in_range(jiffies, tg->slice_start[rw], tg->slice_end[rw])) 568 return false; 569 570 return true; 571 } 572 573 static unsigned int sq_queued(struct throtl_service_queue *sq, int type) 574 { 575 return sq->nr_queued_bps[type] + sq->nr_queued_iops[type]; 576 } 577 578 static unsigned int calculate_io_allowed(u32 iops_limit, 579 unsigned long jiffy_elapsed) 580 { 581 unsigned int io_allowed; 582 u64 tmp; 583 584 /* 585 * jiffy_elapsed should not be a big value as minimum iops can be 586 * 1 then at max jiffy elapsed should be equivalent of 1 second as we 587 * will allow dispatch after 1 second and after that slice should 588 * have been trimmed. 589 */ 590 591 tmp = (u64)iops_limit * jiffy_elapsed; 592 do_div(tmp, HZ); 593 594 if (tmp > UINT_MAX) 595 io_allowed = UINT_MAX; 596 else 597 io_allowed = tmp; 598 599 return io_allowed; 600 } 601 602 static u64 calculate_bytes_allowed(u64 bps_limit, unsigned long jiffy_elapsed) 603 { 604 /* 605 * Can result be wider than 64 bits? 606 * We check against 62, not 64, due to ilog2 truncation. 607 */ 608 if (ilog2(bps_limit) + ilog2(jiffy_elapsed) - ilog2(HZ) > 62) 609 return U64_MAX; 610 return mul_u64_u64_div_u64(bps_limit, (u64)jiffy_elapsed, (u64)HZ); 611 } 612 613 static long long throtl_trim_bps(struct throtl_grp *tg, bool rw, 614 unsigned long time_elapsed) 615 { 616 u64 bps_limit = tg_bps_limit(tg, rw); 617 long long bytes_trim; 618 619 if (bps_limit == U64_MAX) 620 return 0; 621 622 /* Need to consider the case of bytes_allowed overflow. */ 623 bytes_trim = calculate_bytes_allowed(bps_limit, time_elapsed); 624 if (bytes_trim <= 0 || tg->bytes_disp[rw] < bytes_trim) { 625 bytes_trim = tg->bytes_disp[rw]; 626 tg->bytes_disp[rw] = 0; 627 } else { 628 tg->bytes_disp[rw] -= bytes_trim; 629 } 630 631 return bytes_trim; 632 } 633 634 static int throtl_trim_iops(struct throtl_grp *tg, bool rw, 635 unsigned long time_elapsed) 636 { 637 u32 iops_limit = tg_iops_limit(tg, rw); 638 int io_trim; 639 640 if (iops_limit == UINT_MAX) 641 return 0; 642 643 /* Need to consider the case of io_allowed overflow. */ 644 io_trim = calculate_io_allowed(iops_limit, time_elapsed); 645 if (io_trim <= 0 || tg->io_disp[rw] < io_trim) { 646 io_trim = tg->io_disp[rw]; 647 tg->io_disp[rw] = 0; 648 } else { 649 tg->io_disp[rw] -= io_trim; 650 } 651 652 return io_trim; 653 } 654 655 /* Trim the used slices and adjust slice start accordingly */ 656 static inline void throtl_trim_slice(struct throtl_grp *tg, bool rw) 657 { 658 unsigned long time_elapsed; 659 long long bytes_trim; 660 int io_trim; 661 662 BUG_ON(time_before(tg->slice_end[rw], tg->slice_start[rw])); 663 664 /* 665 * If bps are unlimited (-1), then time slice don't get 666 * renewed. Don't try to trim the slice if slice is used. A new 667 * slice will start when appropriate. 668 */ 669 if (throtl_slice_used(tg, rw)) 670 return; 671 672 /* 673 * A bio has been dispatched. Also adjust slice_end. It might happen 674 * that initially cgroup limit was very low resulting in high 675 * slice_end, but later limit was bumped up and bio was dispatched 676 * sooner, then we need to reduce slice_end. A high bogus slice_end 677 * is bad because it does not allow new slice to start. 678 */ 679 throtl_set_slice_end(tg, rw, jiffies + tg->td->throtl_slice); 680 681 time_elapsed = rounddown(jiffies - tg->slice_start[rw], 682 tg->td->throtl_slice); 683 /* Don't trim slice until at least 2 slices are used */ 684 if (time_elapsed < tg->td->throtl_slice * 2) 685 return; 686 687 /* 688 * The bio submission time may be a few jiffies more than the expected 689 * waiting time, due to 'extra_bytes' can't be divided in 690 * tg_within_bps_limit(), and also due to timer wakeup delay. In this 691 * case, adjust slice_start will discard the extra wait time, causing 692 * lower rate than expected. Therefore, other than the above rounddown, 693 * one extra slice is preserved for deviation. 694 */ 695 time_elapsed -= tg->td->throtl_slice; 696 bytes_trim = throtl_trim_bps(tg, rw, time_elapsed); 697 io_trim = throtl_trim_iops(tg, rw, time_elapsed); 698 if (!bytes_trim && !io_trim) 699 return; 700 701 tg->slice_start[rw] += time_elapsed; 702 703 throtl_log(&tg->service_queue, 704 "[%c] trim slice nr=%lu bytes=%lld io=%d start=%lu end=%lu jiffies=%lu", 705 rw == READ ? 'R' : 'W', time_elapsed / tg->td->throtl_slice, 706 bytes_trim, io_trim, tg->slice_start[rw], tg->slice_end[rw], 707 jiffies); 708 } 709 710 static void __tg_update_carryover(struct throtl_grp *tg, bool rw, 711 long long *bytes, int *ios) 712 { 713 unsigned long jiffy_elapsed = jiffies - tg->slice_start[rw]; 714 u64 bps_limit = tg_bps_limit(tg, rw); 715 u32 iops_limit = tg_iops_limit(tg, rw); 716 long long bytes_allowed; 717 int io_allowed; 718 719 /* 720 * If the queue is empty, carryover handling is not needed. In such cases, 721 * tg->[bytes/io]_disp should be reset to 0 to avoid impacting the dispatch 722 * of subsequent bios. The same handling applies when the previous BPS/IOPS 723 * limit was set to max. 724 */ 725 if (sq_queued(&tg->service_queue, rw) == 0) { 726 tg->bytes_disp[rw] = 0; 727 tg->io_disp[rw] = 0; 728 return; 729 } 730 731 /* 732 * If config is updated while bios are still throttled, calculate and 733 * accumulate how many bytes/ios are waited across changes. And use the 734 * calculated carryover (@bytes/@ios) to update [bytes/io]_disp, which 735 * will be used to calculate new wait time under new configuration. 736 * And we need to consider the case of bytes/io_allowed overflow. 737 */ 738 if (bps_limit != U64_MAX) { 739 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed); 740 if (bytes_allowed > 0) 741 *bytes = bytes_allowed - tg->bytes_disp[rw]; 742 } 743 if (iops_limit != UINT_MAX) { 744 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed); 745 if (io_allowed > 0) 746 *ios = io_allowed - tg->io_disp[rw]; 747 } 748 749 tg->bytes_disp[rw] = -*bytes; 750 tg->io_disp[rw] = -*ios; 751 } 752 753 static void tg_update_carryover(struct throtl_grp *tg) 754 { 755 long long bytes[2] = {0}; 756 int ios[2] = {0}; 757 758 __tg_update_carryover(tg, READ, &bytes[READ], &ios[READ]); 759 __tg_update_carryover(tg, WRITE, &bytes[WRITE], &ios[WRITE]); 760 761 /* see comments in struct throtl_grp for meaning of carryover. */ 762 throtl_log(&tg->service_queue, "%s: %lld %lld %d %d\n", __func__, 763 bytes[READ], bytes[WRITE], ios[READ], ios[WRITE]); 764 } 765 766 static unsigned long tg_within_iops_limit(struct throtl_grp *tg, struct bio *bio, 767 u32 iops_limit) 768 { 769 bool rw = bio_data_dir(bio); 770 int io_allowed; 771 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; 772 773 jiffy_elapsed = jiffies - tg->slice_start[rw]; 774 775 /* Round up to the next throttle slice, wait time must be nonzero */ 776 jiffy_elapsed_rnd = roundup(jiffy_elapsed + 1, tg->td->throtl_slice); 777 io_allowed = calculate_io_allowed(iops_limit, jiffy_elapsed_rnd); 778 if (io_allowed > 0 && tg->io_disp[rw] + 1 <= io_allowed) 779 return 0; 780 781 /* Calc approx time to dispatch */ 782 jiffy_wait = jiffy_elapsed_rnd - jiffy_elapsed; 783 784 /* make sure at least one io can be dispatched after waiting */ 785 jiffy_wait = max(jiffy_wait, HZ / iops_limit + 1); 786 return jiffy_wait; 787 } 788 789 static unsigned long tg_within_bps_limit(struct throtl_grp *tg, struct bio *bio, 790 u64 bps_limit) 791 { 792 bool rw = bio_data_dir(bio); 793 long long bytes_allowed; 794 u64 extra_bytes; 795 unsigned long jiffy_elapsed, jiffy_wait, jiffy_elapsed_rnd; 796 unsigned int bio_size = throtl_bio_data_size(bio); 797 798 jiffy_elapsed = jiffy_elapsed_rnd = jiffies - tg->slice_start[rw]; 799 800 /* Slice has just started. Consider one slice interval */ 801 if (!jiffy_elapsed) 802 jiffy_elapsed_rnd = tg->td->throtl_slice; 803 804 jiffy_elapsed_rnd = roundup(jiffy_elapsed_rnd, tg->td->throtl_slice); 805 bytes_allowed = calculate_bytes_allowed(bps_limit, jiffy_elapsed_rnd); 806 /* Need to consider the case of bytes_allowed overflow. */ 807 if ((bytes_allowed > 0 && tg->bytes_disp[rw] + bio_size <= bytes_allowed) 808 || bytes_allowed < 0) 809 return 0; 810 811 /* Calc approx time to dispatch */ 812 extra_bytes = tg->bytes_disp[rw] + bio_size - bytes_allowed; 813 jiffy_wait = div64_u64(extra_bytes * HZ, bps_limit); 814 815 if (!jiffy_wait) 816 jiffy_wait = 1; 817 818 /* 819 * This wait time is without taking into consideration the rounding 820 * up we did. Add that time also. 821 */ 822 jiffy_wait = jiffy_wait + (jiffy_elapsed_rnd - jiffy_elapsed); 823 return jiffy_wait; 824 } 825 826 static void throtl_charge_bps_bio(struct throtl_grp *tg, struct bio *bio) 827 { 828 unsigned int bio_size = throtl_bio_data_size(bio); 829 830 /* Charge the bio to the group */ 831 if (!bio_flagged(bio, BIO_BPS_THROTTLED) && 832 !bio_flagged(bio, BIO_TG_BPS_THROTTLED)) { 833 bio_set_flag(bio, BIO_TG_BPS_THROTTLED); 834 tg->bytes_disp[bio_data_dir(bio)] += bio_size; 835 } 836 } 837 838 static void throtl_charge_iops_bio(struct throtl_grp *tg, struct bio *bio) 839 { 840 bio_clear_flag(bio, BIO_TG_BPS_THROTTLED); 841 tg->io_disp[bio_data_dir(bio)]++; 842 } 843 844 /* 845 * If previous slice expired, start a new one otherwise renew/extend existing 846 * slice to make sure it is at least throtl_slice interval long since now. New 847 * slice is started only for empty throttle group. If there is queued bio, that 848 * means there should be an active slice and it should be extended instead. 849 */ 850 static void tg_update_slice(struct throtl_grp *tg, bool rw) 851 { 852 if (throtl_slice_used(tg, rw) && 853 sq_queued(&tg->service_queue, rw) == 0) 854 throtl_start_new_slice(tg, rw, true); 855 else 856 throtl_extend_slice(tg, rw, jiffies + tg->td->throtl_slice); 857 } 858 859 static unsigned long tg_dispatch_bps_time(struct throtl_grp *tg, struct bio *bio) 860 { 861 bool rw = bio_data_dir(bio); 862 u64 bps_limit = tg_bps_limit(tg, rw); 863 unsigned long bps_wait; 864 865 /* no need to throttle if this bio's bytes have been accounted */ 866 if (bps_limit == U64_MAX || tg->flags & THROTL_TG_CANCELING || 867 bio_flagged(bio, BIO_BPS_THROTTLED) || 868 bio_flagged(bio, BIO_TG_BPS_THROTTLED)) 869 return 0; 870 871 tg_update_slice(tg, rw); 872 bps_wait = tg_within_bps_limit(tg, bio, bps_limit); 873 throtl_extend_slice(tg, rw, jiffies + bps_wait); 874 875 return bps_wait; 876 } 877 878 static unsigned long tg_dispatch_iops_time(struct throtl_grp *tg, struct bio *bio) 879 { 880 bool rw = bio_data_dir(bio); 881 u32 iops_limit = tg_iops_limit(tg, rw); 882 unsigned long iops_wait; 883 884 if (iops_limit == UINT_MAX || tg->flags & THROTL_TG_CANCELING) 885 return 0; 886 887 tg_update_slice(tg, rw); 888 iops_wait = tg_within_iops_limit(tg, bio, iops_limit); 889 throtl_extend_slice(tg, rw, jiffies + iops_wait); 890 891 return iops_wait; 892 } 893 894 /* 895 * Returns approx number of jiffies to wait before this bio is with-in IO rate 896 * and can be moved to other queue or dispatched. 897 */ 898 static unsigned long tg_dispatch_time(struct throtl_grp *tg, struct bio *bio) 899 { 900 bool rw = bio_data_dir(bio); 901 unsigned long wait; 902 903 /* 904 * Currently whole state machine of group depends on first bio 905 * queued in the group bio list. So one should not be calling 906 * this function with a different bio if there are other bios 907 * queued. 908 */ 909 BUG_ON(sq_queued(&tg->service_queue, rw) && 910 bio != throtl_peek_queued(&tg->service_queue.queued[rw])); 911 912 wait = tg_dispatch_bps_time(tg, bio); 913 if (wait != 0) 914 return wait; 915 916 /* 917 * Charge bps here because @bio will be directly placed into the 918 * iops queue afterward. 919 */ 920 throtl_charge_bps_bio(tg, bio); 921 922 return tg_dispatch_iops_time(tg, bio); 923 } 924 925 /** 926 * throtl_add_bio_tg - add a bio to the specified throtl_grp 927 * @bio: bio to add 928 * @qn: qnode to use 929 * @tg: the target throtl_grp 930 * 931 * Add @bio to @tg's service_queue using @qn. If @qn is not specified, 932 * tg->qnode_on_self[] is used. 933 */ 934 static void throtl_add_bio_tg(struct bio *bio, struct throtl_qnode *qn, 935 struct throtl_grp *tg) 936 { 937 struct throtl_service_queue *sq = &tg->service_queue; 938 bool rw = bio_data_dir(bio); 939 940 if (!qn) 941 qn = &tg->qnode_on_self[rw]; 942 943 /* 944 * If @tg doesn't currently have any bios queued in the same 945 * direction, queueing @bio can change when @tg should be 946 * dispatched. Mark that @tg was empty. This is automatically 947 * cleared on the next tg_update_disptime(). 948 */ 949 if (sq_queued(sq, rw) == 0) 950 tg->flags |= THROTL_TG_WAS_EMPTY; 951 952 throtl_qnode_add_bio(bio, qn, sq); 953 954 /* 955 * Since we have split the queues, when the iops queue is 956 * previously empty and a new @bio is added into the first @qn, 957 * we also need to update the @tg->disptime. 958 */ 959 if (bio_flagged(bio, BIO_BPS_THROTTLED) && 960 bio == throtl_peek_queued(&sq->queued[rw])) 961 tg->flags |= THROTL_TG_IOPS_WAS_EMPTY; 962 963 throtl_enqueue_tg(tg); 964 } 965 966 static void tg_update_disptime(struct throtl_grp *tg) 967 { 968 struct throtl_service_queue *sq = &tg->service_queue; 969 unsigned long read_wait = -1, write_wait = -1, min_wait, disptime; 970 struct bio *bio; 971 972 bio = throtl_peek_queued(&sq->queued[READ]); 973 if (bio) 974 read_wait = tg_dispatch_time(tg, bio); 975 976 bio = throtl_peek_queued(&sq->queued[WRITE]); 977 if (bio) 978 write_wait = tg_dispatch_time(tg, bio); 979 980 min_wait = min(read_wait, write_wait); 981 disptime = jiffies + min_wait; 982 983 /* Update dispatch time */ 984 throtl_rb_erase(&tg->rb_node, tg->service_queue.parent_sq); 985 tg->disptime = disptime; 986 tg_service_queue_add(tg); 987 988 /* see throtl_add_bio_tg() */ 989 tg->flags &= ~THROTL_TG_WAS_EMPTY; 990 tg->flags &= ~THROTL_TG_IOPS_WAS_EMPTY; 991 } 992 993 static void start_parent_slice_with_credit(struct throtl_grp *child_tg, 994 struct throtl_grp *parent_tg, bool rw) 995 { 996 if (throtl_slice_used(parent_tg, rw)) { 997 throtl_start_new_slice_with_credit(parent_tg, rw, 998 child_tg->slice_start[rw]); 999 } 1000 1001 } 1002 1003 static void tg_dispatch_one_bio(struct throtl_grp *tg, bool rw) 1004 { 1005 struct throtl_service_queue *sq = &tg->service_queue; 1006 struct throtl_service_queue *parent_sq = sq->parent_sq; 1007 struct throtl_grp *parent_tg = sq_to_tg(parent_sq); 1008 struct throtl_grp *tg_to_put = NULL; 1009 struct bio *bio; 1010 1011 /* 1012 * @bio is being transferred from @tg to @parent_sq. Popping a bio 1013 * from @tg may put its reference and @parent_sq might end up 1014 * getting released prematurely. Remember the tg to put and put it 1015 * after @bio is transferred to @parent_sq. 1016 */ 1017 bio = throtl_pop_queued(sq, &tg_to_put, rw); 1018 1019 throtl_charge_iops_bio(tg, bio); 1020 1021 /* 1022 * If our parent is another tg, we just need to transfer @bio to 1023 * the parent using throtl_add_bio_tg(). If our parent is 1024 * @td->service_queue, @bio is ready to be issued. Put it on its 1025 * bio_lists[] and decrease total number queued. The caller is 1026 * responsible for issuing these bios. 1027 */ 1028 if (parent_tg) { 1029 throtl_add_bio_tg(bio, &tg->qnode_on_parent[rw], parent_tg); 1030 start_parent_slice_with_credit(tg, parent_tg, rw); 1031 } else { 1032 bio_set_flag(bio, BIO_BPS_THROTTLED); 1033 throtl_qnode_add_bio(bio, &tg->qnode_on_parent[rw], 1034 parent_sq); 1035 BUG_ON(tg->td->nr_queued[rw] <= 0); 1036 tg->td->nr_queued[rw]--; 1037 } 1038 1039 throtl_trim_slice(tg, rw); 1040 1041 if (tg_to_put) 1042 blkg_put(tg_to_blkg(tg_to_put)); 1043 } 1044 1045 static int throtl_dispatch_tg(struct throtl_grp *tg) 1046 { 1047 struct throtl_service_queue *sq = &tg->service_queue; 1048 unsigned int nr_reads = 0, nr_writes = 0; 1049 unsigned int max_nr_reads = THROTL_GRP_QUANTUM * 3 / 4; 1050 unsigned int max_nr_writes = THROTL_GRP_QUANTUM - max_nr_reads; 1051 struct bio *bio; 1052 1053 /* Try to dispatch 75% READS and 25% WRITES */ 1054 1055 while ((bio = throtl_peek_queued(&sq->queued[READ])) && 1056 tg_dispatch_time(tg, bio) == 0) { 1057 1058 tg_dispatch_one_bio(tg, READ); 1059 nr_reads++; 1060 1061 if (nr_reads >= max_nr_reads) 1062 break; 1063 } 1064 1065 while ((bio = throtl_peek_queued(&sq->queued[WRITE])) && 1066 tg_dispatch_time(tg, bio) == 0) { 1067 1068 tg_dispatch_one_bio(tg, WRITE); 1069 nr_writes++; 1070 1071 if (nr_writes >= max_nr_writes) 1072 break; 1073 } 1074 1075 return nr_reads + nr_writes; 1076 } 1077 1078 static int throtl_select_dispatch(struct throtl_service_queue *parent_sq) 1079 { 1080 unsigned int nr_disp = 0; 1081 1082 while (1) { 1083 struct throtl_grp *tg; 1084 struct throtl_service_queue *sq; 1085 1086 if (!parent_sq->nr_pending) 1087 break; 1088 1089 tg = throtl_rb_first(parent_sq); 1090 if (!tg) 1091 break; 1092 1093 if (time_before(jiffies, tg->disptime)) 1094 break; 1095 1096 nr_disp += throtl_dispatch_tg(tg); 1097 1098 sq = &tg->service_queue; 1099 if (sq_queued(sq, READ) || sq_queued(sq, WRITE)) 1100 tg_update_disptime(tg); 1101 else 1102 throtl_dequeue_tg(tg); 1103 1104 if (nr_disp >= THROTL_QUANTUM) 1105 break; 1106 } 1107 1108 return nr_disp; 1109 } 1110 1111 /** 1112 * throtl_pending_timer_fn - timer function for service_queue->pending_timer 1113 * @t: the pending_timer member of the throtl_service_queue being serviced 1114 * 1115 * This timer is armed when a child throtl_grp with active bio's become 1116 * pending and queued on the service_queue's pending_tree and expires when 1117 * the first child throtl_grp should be dispatched. This function 1118 * dispatches bio's from the children throtl_grps to the parent 1119 * service_queue. 1120 * 1121 * If the parent's parent is another throtl_grp, dispatching is propagated 1122 * by either arming its pending_timer or repeating dispatch directly. If 1123 * the top-level service_tree is reached, throtl_data->dispatch_work is 1124 * kicked so that the ready bio's are issued. 1125 */ 1126 static void throtl_pending_timer_fn(struct timer_list *t) 1127 { 1128 struct throtl_service_queue *sq = timer_container_of(sq, t, 1129 pending_timer); 1130 struct throtl_grp *tg = sq_to_tg(sq); 1131 struct throtl_data *td = sq_to_td(sq); 1132 struct throtl_service_queue *parent_sq; 1133 struct request_queue *q; 1134 bool dispatched; 1135 int ret; 1136 1137 /* throtl_data may be gone, so figure out request queue by blkg */ 1138 if (tg) 1139 q = tg->pd.blkg->q; 1140 else 1141 q = td->queue; 1142 1143 spin_lock_irq(&q->queue_lock); 1144 1145 if (!q->root_blkg) 1146 goto out_unlock; 1147 1148 again: 1149 parent_sq = sq->parent_sq; 1150 dispatched = false; 1151 1152 while (true) { 1153 unsigned int __maybe_unused bio_cnt_r = sq_queued(sq, READ); 1154 unsigned int __maybe_unused bio_cnt_w = sq_queued(sq, WRITE); 1155 1156 throtl_log(sq, "dispatch nr_queued=%u read=%u write=%u", 1157 bio_cnt_r + bio_cnt_w, bio_cnt_r, bio_cnt_w); 1158 1159 ret = throtl_select_dispatch(sq); 1160 if (ret) { 1161 throtl_log(sq, "bios disp=%u", ret); 1162 dispatched = true; 1163 } 1164 1165 if (throtl_schedule_next_dispatch(sq, false)) 1166 break; 1167 1168 /* this dispatch windows is still open, relax and repeat */ 1169 spin_unlock_irq(&q->queue_lock); 1170 cpu_relax(); 1171 spin_lock_irq(&q->queue_lock); 1172 } 1173 1174 if (!dispatched) 1175 goto out_unlock; 1176 1177 if (parent_sq) { 1178 /* @parent_sq is another throl_grp, propagate dispatch */ 1179 if (tg->flags & THROTL_TG_WAS_EMPTY || 1180 tg->flags & THROTL_TG_IOPS_WAS_EMPTY) { 1181 tg_update_disptime(tg); 1182 if (!throtl_schedule_next_dispatch(parent_sq, false)) { 1183 /* window is already open, repeat dispatching */ 1184 sq = parent_sq; 1185 tg = sq_to_tg(sq); 1186 goto again; 1187 } 1188 } 1189 } else { 1190 /* reached the top-level, queue issuing */ 1191 queue_work(kthrotld_workqueue, &td->dispatch_work); 1192 } 1193 out_unlock: 1194 spin_unlock_irq(&q->queue_lock); 1195 } 1196 1197 /** 1198 * blk_throtl_dispatch_work_fn - work function for throtl_data->dispatch_work 1199 * @work: work item being executed 1200 * 1201 * This function is queued for execution when bios reach the bio_lists[] 1202 * of throtl_data->service_queue. Those bios are ready and issued by this 1203 * function. 1204 */ 1205 static void blk_throtl_dispatch_work_fn(struct work_struct *work) 1206 { 1207 struct throtl_data *td = container_of(work, struct throtl_data, 1208 dispatch_work); 1209 struct throtl_service_queue *td_sq = &td->service_queue; 1210 struct request_queue *q = td->queue; 1211 struct bio_list bio_list_on_stack; 1212 struct bio *bio; 1213 struct blk_plug plug; 1214 int rw; 1215 1216 bio_list_init(&bio_list_on_stack); 1217 1218 spin_lock_irq(&q->queue_lock); 1219 for (rw = READ; rw <= WRITE; rw++) 1220 while ((bio = throtl_pop_queued(td_sq, NULL, rw))) 1221 bio_list_add(&bio_list_on_stack, bio); 1222 spin_unlock_irq(&q->queue_lock); 1223 1224 if (!bio_list_empty(&bio_list_on_stack)) { 1225 blk_start_plug(&plug); 1226 while ((bio = bio_list_pop(&bio_list_on_stack))) 1227 submit_bio_noacct_nocheck(bio); 1228 blk_finish_plug(&plug); 1229 } 1230 } 1231 1232 static u64 tg_prfill_conf_u64(struct seq_file *sf, struct blkg_policy_data *pd, 1233 int off) 1234 { 1235 struct throtl_grp *tg = pd_to_tg(pd); 1236 u64 v = *(u64 *)((void *)tg + off); 1237 1238 if (v == U64_MAX) 1239 return 0; 1240 return __blkg_prfill_u64(sf, pd, v); 1241 } 1242 1243 static u64 tg_prfill_conf_uint(struct seq_file *sf, struct blkg_policy_data *pd, 1244 int off) 1245 { 1246 struct throtl_grp *tg = pd_to_tg(pd); 1247 unsigned int v = *(unsigned int *)((void *)tg + off); 1248 1249 if (v == UINT_MAX) 1250 return 0; 1251 return __blkg_prfill_u64(sf, pd, v); 1252 } 1253 1254 static int tg_print_conf_u64(struct seq_file *sf, void *v) 1255 { 1256 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_u64, 1257 &blkcg_policy_throtl, seq_cft(sf)->private, false); 1258 return 0; 1259 } 1260 1261 static int tg_print_conf_uint(struct seq_file *sf, void *v) 1262 { 1263 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_conf_uint, 1264 &blkcg_policy_throtl, seq_cft(sf)->private, false); 1265 return 0; 1266 } 1267 1268 static void tg_conf_updated(struct throtl_grp *tg, bool global) 1269 { 1270 struct throtl_service_queue *sq = &tg->service_queue; 1271 struct cgroup_subsys_state *pos_css; 1272 struct blkcg_gq *blkg; 1273 1274 throtl_log(&tg->service_queue, 1275 "limit change rbps=%llu wbps=%llu riops=%u wiops=%u", 1276 tg_bps_limit(tg, READ), tg_bps_limit(tg, WRITE), 1277 tg_iops_limit(tg, READ), tg_iops_limit(tg, WRITE)); 1278 1279 rcu_read_lock(); 1280 /* 1281 * Update has_rules[] flags for the updated tg's subtree. A tg is 1282 * considered to have rules if either the tg itself or any of its 1283 * ancestors has rules. This identifies groups without any 1284 * restrictions in the whole hierarchy and allows them to bypass 1285 * blk-throttle. 1286 */ 1287 blkg_for_each_descendant_pre(blkg, pos_css, 1288 global ? tg->td->queue->root_blkg : tg_to_blkg(tg)) { 1289 struct throtl_grp *this_tg = blkg_to_tg(blkg); 1290 1291 tg_update_has_rules(this_tg); 1292 /* ignore root/second level */ 1293 if (!cgroup_subsys_on_dfl(io_cgrp_subsys) || !blkg->parent || 1294 !blkg->parent->parent) 1295 continue; 1296 } 1297 rcu_read_unlock(); 1298 1299 /* 1300 * We're already holding queue_lock and know @tg is valid. Let's 1301 * apply the new config directly. 1302 * 1303 * Restart the slices for both READ and WRITES. It might happen 1304 * that a group's limit are dropped suddenly and we don't want to 1305 * account recently dispatched IO with new low rate. 1306 */ 1307 throtl_start_new_slice(tg, READ, false); 1308 throtl_start_new_slice(tg, WRITE, false); 1309 1310 if (tg->flags & THROTL_TG_PENDING) { 1311 tg_update_disptime(tg); 1312 throtl_schedule_next_dispatch(sq->parent_sq, true); 1313 } 1314 } 1315 1316 static int blk_throtl_init(struct gendisk *disk) 1317 { 1318 struct request_queue *q = disk->queue; 1319 struct throtl_data *td; 1320 unsigned int memflags; 1321 int ret; 1322 1323 td = kzalloc_node(sizeof(*td), GFP_KERNEL, q->node); 1324 if (!td) 1325 return -ENOMEM; 1326 1327 INIT_WORK(&td->dispatch_work, blk_throtl_dispatch_work_fn); 1328 throtl_service_queue_init(&td->service_queue); 1329 1330 /* 1331 * Freeze queue before activating policy, to synchronize with IO path, 1332 * which is protected by 'q_usage_counter'. 1333 */ 1334 memflags = blk_mq_freeze_queue(disk->queue); 1335 blk_mq_quiesce_queue(disk->queue); 1336 1337 q->td = td; 1338 td->queue = q; 1339 1340 /* activate policy */ 1341 ret = blkcg_activate_policy(disk, &blkcg_policy_throtl); 1342 if (ret) { 1343 q->td = NULL; 1344 kfree(td); 1345 goto out; 1346 } 1347 1348 if (blk_queue_nonrot(q)) 1349 td->throtl_slice = DFL_THROTL_SLICE_SSD; 1350 else 1351 td->throtl_slice = DFL_THROTL_SLICE_HD; 1352 td->track_bio_latency = !queue_is_mq(q); 1353 if (!td->track_bio_latency) 1354 blk_stat_enable_accounting(q); 1355 1356 out: 1357 blk_mq_unquiesce_queue(disk->queue); 1358 blk_mq_unfreeze_queue(disk->queue, memflags); 1359 1360 return ret; 1361 } 1362 1363 1364 static ssize_t tg_set_conf(struct kernfs_open_file *of, 1365 char *buf, size_t nbytes, loff_t off, bool is_u64) 1366 { 1367 struct blkcg *blkcg = css_to_blkcg(of_css(of)); 1368 struct blkg_conf_ctx ctx; 1369 struct throtl_grp *tg; 1370 int ret; 1371 u64 v; 1372 1373 blkg_conf_init(&ctx, buf); 1374 1375 ret = blkg_conf_open_bdev(&ctx); 1376 if (ret) 1377 goto out_finish; 1378 1379 if (!blk_throtl_activated(ctx.bdev->bd_queue)) { 1380 ret = blk_throtl_init(ctx.bdev->bd_disk); 1381 if (ret) 1382 goto out_finish; 1383 } 1384 1385 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx); 1386 if (ret) 1387 goto out_finish; 1388 1389 ret = -EINVAL; 1390 if (sscanf(ctx.body, "%llu", &v) != 1) 1391 goto out_finish; 1392 if (!v) 1393 v = U64_MAX; 1394 1395 tg = blkg_to_tg(ctx.blkg); 1396 tg_update_carryover(tg); 1397 1398 if (is_u64) 1399 *(u64 *)((void *)tg + of_cft(of)->private) = v; 1400 else 1401 *(unsigned int *)((void *)tg + of_cft(of)->private) = v; 1402 1403 tg_conf_updated(tg, false); 1404 ret = 0; 1405 out_finish: 1406 blkg_conf_exit(&ctx); 1407 return ret ?: nbytes; 1408 } 1409 1410 static ssize_t tg_set_conf_u64(struct kernfs_open_file *of, 1411 char *buf, size_t nbytes, loff_t off) 1412 { 1413 return tg_set_conf(of, buf, nbytes, off, true); 1414 } 1415 1416 static ssize_t tg_set_conf_uint(struct kernfs_open_file *of, 1417 char *buf, size_t nbytes, loff_t off) 1418 { 1419 return tg_set_conf(of, buf, nbytes, off, false); 1420 } 1421 1422 static int tg_print_rwstat(struct seq_file *sf, void *v) 1423 { 1424 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1425 blkg_prfill_rwstat, &blkcg_policy_throtl, 1426 seq_cft(sf)->private, true); 1427 return 0; 1428 } 1429 1430 static u64 tg_prfill_rwstat_recursive(struct seq_file *sf, 1431 struct blkg_policy_data *pd, int off) 1432 { 1433 struct blkg_rwstat_sample sum; 1434 1435 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_throtl, off, 1436 &sum); 1437 return __blkg_prfill_rwstat(sf, pd, &sum); 1438 } 1439 1440 static int tg_print_rwstat_recursive(struct seq_file *sf, void *v) 1441 { 1442 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1443 tg_prfill_rwstat_recursive, &blkcg_policy_throtl, 1444 seq_cft(sf)->private, true); 1445 return 0; 1446 } 1447 1448 static struct cftype throtl_legacy_files[] = { 1449 { 1450 .name = "throttle.read_bps_device", 1451 .private = offsetof(struct throtl_grp, bps[READ]), 1452 .seq_show = tg_print_conf_u64, 1453 .write = tg_set_conf_u64, 1454 }, 1455 { 1456 .name = "throttle.write_bps_device", 1457 .private = offsetof(struct throtl_grp, bps[WRITE]), 1458 .seq_show = tg_print_conf_u64, 1459 .write = tg_set_conf_u64, 1460 }, 1461 { 1462 .name = "throttle.read_iops_device", 1463 .private = offsetof(struct throtl_grp, iops[READ]), 1464 .seq_show = tg_print_conf_uint, 1465 .write = tg_set_conf_uint, 1466 }, 1467 { 1468 .name = "throttle.write_iops_device", 1469 .private = offsetof(struct throtl_grp, iops[WRITE]), 1470 .seq_show = tg_print_conf_uint, 1471 .write = tg_set_conf_uint, 1472 }, 1473 { 1474 .name = "throttle.io_service_bytes", 1475 .private = offsetof(struct throtl_grp, stat_bytes), 1476 .seq_show = tg_print_rwstat, 1477 }, 1478 { 1479 .name = "throttle.io_service_bytes_recursive", 1480 .private = offsetof(struct throtl_grp, stat_bytes), 1481 .seq_show = tg_print_rwstat_recursive, 1482 }, 1483 { 1484 .name = "throttle.io_serviced", 1485 .private = offsetof(struct throtl_grp, stat_ios), 1486 .seq_show = tg_print_rwstat, 1487 }, 1488 { 1489 .name = "throttle.io_serviced_recursive", 1490 .private = offsetof(struct throtl_grp, stat_ios), 1491 .seq_show = tg_print_rwstat_recursive, 1492 }, 1493 { } /* terminate */ 1494 }; 1495 1496 static u64 tg_prfill_limit(struct seq_file *sf, struct blkg_policy_data *pd, 1497 int off) 1498 { 1499 struct throtl_grp *tg = pd_to_tg(pd); 1500 const char *dname = blkg_dev_name(pd->blkg); 1501 u64 bps_dft; 1502 unsigned int iops_dft; 1503 1504 if (!dname) 1505 return 0; 1506 1507 bps_dft = U64_MAX; 1508 iops_dft = UINT_MAX; 1509 1510 if (tg->bps[READ] == bps_dft && 1511 tg->bps[WRITE] == bps_dft && 1512 tg->iops[READ] == iops_dft && 1513 tg->iops[WRITE] == iops_dft) 1514 return 0; 1515 1516 seq_printf(sf, "%s", dname); 1517 if (tg->bps[READ] == U64_MAX) 1518 seq_printf(sf, " rbps=max"); 1519 else 1520 seq_printf(sf, " rbps=%llu", tg->bps[READ]); 1521 1522 if (tg->bps[WRITE] == U64_MAX) 1523 seq_printf(sf, " wbps=max"); 1524 else 1525 seq_printf(sf, " wbps=%llu", tg->bps[WRITE]); 1526 1527 if (tg->iops[READ] == UINT_MAX) 1528 seq_printf(sf, " riops=max"); 1529 else 1530 seq_printf(sf, " riops=%u", tg->iops[READ]); 1531 1532 if (tg->iops[WRITE] == UINT_MAX) 1533 seq_printf(sf, " wiops=max"); 1534 else 1535 seq_printf(sf, " wiops=%u", tg->iops[WRITE]); 1536 1537 seq_printf(sf, "\n"); 1538 return 0; 1539 } 1540 1541 static int tg_print_limit(struct seq_file *sf, void *v) 1542 { 1543 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), tg_prfill_limit, 1544 &blkcg_policy_throtl, seq_cft(sf)->private, false); 1545 return 0; 1546 } 1547 1548 static ssize_t tg_set_limit(struct kernfs_open_file *of, 1549 char *buf, size_t nbytes, loff_t off) 1550 { 1551 struct blkcg *blkcg = css_to_blkcg(of_css(of)); 1552 struct blkg_conf_ctx ctx; 1553 struct throtl_grp *tg; 1554 u64 v[4]; 1555 int ret; 1556 1557 blkg_conf_init(&ctx, buf); 1558 1559 ret = blkg_conf_open_bdev(&ctx); 1560 if (ret) 1561 goto out_finish; 1562 1563 if (!blk_throtl_activated(ctx.bdev->bd_queue)) { 1564 ret = blk_throtl_init(ctx.bdev->bd_disk); 1565 if (ret) 1566 goto out_finish; 1567 } 1568 1569 ret = blkg_conf_prep(blkcg, &blkcg_policy_throtl, &ctx); 1570 if (ret) 1571 goto out_finish; 1572 1573 tg = blkg_to_tg(ctx.blkg); 1574 tg_update_carryover(tg); 1575 1576 v[0] = tg->bps[READ]; 1577 v[1] = tg->bps[WRITE]; 1578 v[2] = tg->iops[READ]; 1579 v[3] = tg->iops[WRITE]; 1580 1581 while (true) { 1582 char tok[27]; /* wiops=18446744073709551616 */ 1583 char *p; 1584 u64 val = U64_MAX; 1585 int len; 1586 1587 if (sscanf(ctx.body, "%26s%n", tok, &len) != 1) 1588 break; 1589 if (tok[0] == '\0') 1590 break; 1591 ctx.body += len; 1592 1593 ret = -EINVAL; 1594 p = tok; 1595 strsep(&p, "="); 1596 if (!p || (sscanf(p, "%llu", &val) != 1 && strcmp(p, "max"))) 1597 goto out_finish; 1598 1599 ret = -ERANGE; 1600 if (!val) 1601 goto out_finish; 1602 1603 ret = -EINVAL; 1604 if (!strcmp(tok, "rbps")) 1605 v[0] = val; 1606 else if (!strcmp(tok, "wbps")) 1607 v[1] = val; 1608 else if (!strcmp(tok, "riops")) 1609 v[2] = min_t(u64, val, UINT_MAX); 1610 else if (!strcmp(tok, "wiops")) 1611 v[3] = min_t(u64, val, UINT_MAX); 1612 else 1613 goto out_finish; 1614 } 1615 1616 tg->bps[READ] = v[0]; 1617 tg->bps[WRITE] = v[1]; 1618 tg->iops[READ] = v[2]; 1619 tg->iops[WRITE] = v[3]; 1620 1621 tg_conf_updated(tg, false); 1622 ret = 0; 1623 out_finish: 1624 blkg_conf_exit(&ctx); 1625 return ret ?: nbytes; 1626 } 1627 1628 static struct cftype throtl_files[] = { 1629 { 1630 .name = "max", 1631 .flags = CFTYPE_NOT_ON_ROOT, 1632 .seq_show = tg_print_limit, 1633 .write = tg_set_limit, 1634 }, 1635 { } /* terminate */ 1636 }; 1637 1638 static void throtl_shutdown_wq(struct request_queue *q) 1639 { 1640 struct throtl_data *td = q->td; 1641 1642 cancel_work_sync(&td->dispatch_work); 1643 } 1644 1645 static void tg_flush_bios(struct throtl_grp *tg) 1646 { 1647 struct throtl_service_queue *sq = &tg->service_queue; 1648 1649 if (tg->flags & THROTL_TG_CANCELING) 1650 return; 1651 /* 1652 * Set the flag to make sure throtl_pending_timer_fn() won't 1653 * stop until all throttled bios are dispatched. 1654 */ 1655 tg->flags |= THROTL_TG_CANCELING; 1656 1657 /* 1658 * Do not dispatch cgroup without THROTL_TG_PENDING or cgroup 1659 * will be inserted to service queue without THROTL_TG_PENDING 1660 * set in tg_update_disptime below. Then IO dispatched from 1661 * child in tg_dispatch_one_bio will trigger double insertion 1662 * and corrupt the tree. 1663 */ 1664 if (!(tg->flags & THROTL_TG_PENDING)) 1665 return; 1666 1667 /* 1668 * Update disptime after setting the above flag to make sure 1669 * throtl_select_dispatch() won't exit without dispatching. 1670 */ 1671 tg_update_disptime(tg); 1672 1673 throtl_schedule_pending_timer(sq, jiffies + 1); 1674 } 1675 1676 static void throtl_pd_offline(struct blkg_policy_data *pd) 1677 { 1678 tg_flush_bios(pd_to_tg(pd)); 1679 } 1680 1681 struct blkcg_policy blkcg_policy_throtl = { 1682 .dfl_cftypes = throtl_files, 1683 .legacy_cftypes = throtl_legacy_files, 1684 1685 .pd_alloc_fn = throtl_pd_alloc, 1686 .pd_init_fn = throtl_pd_init, 1687 .pd_online_fn = throtl_pd_online, 1688 .pd_offline_fn = throtl_pd_offline, 1689 .pd_free_fn = throtl_pd_free, 1690 }; 1691 1692 void blk_throtl_cancel_bios(struct gendisk *disk) 1693 { 1694 struct request_queue *q = disk->queue; 1695 struct cgroup_subsys_state *pos_css; 1696 struct blkcg_gq *blkg; 1697 1698 if (!blk_throtl_activated(q)) 1699 return; 1700 1701 spin_lock_irq(&q->queue_lock); 1702 /* 1703 * queue_lock is held, rcu lock is not needed here technically. 1704 * However, rcu lock is still held to emphasize that following 1705 * path need RCU protection and to prevent warning from lockdep. 1706 */ 1707 rcu_read_lock(); 1708 blkg_for_each_descendant_post(blkg, pos_css, q->root_blkg) { 1709 /* 1710 * disk_release will call pd_offline_fn to cancel bios. 1711 * However, disk_release can't be called if someone get 1712 * the refcount of device and issued bios which are 1713 * inflight after del_gendisk. 1714 * Cancel bios here to ensure no bios are inflight after 1715 * del_gendisk. 1716 */ 1717 tg_flush_bios(blkg_to_tg(blkg)); 1718 } 1719 rcu_read_unlock(); 1720 spin_unlock_irq(&q->queue_lock); 1721 } 1722 1723 static bool tg_within_limit(struct throtl_grp *tg, struct bio *bio, bool rw) 1724 { 1725 struct throtl_service_queue *sq = &tg->service_queue; 1726 1727 /* 1728 * For a split bio, we need to specifically distinguish whether the 1729 * iops queue is empty. 1730 */ 1731 if (bio_flagged(bio, BIO_BPS_THROTTLED)) 1732 return sq->nr_queued_iops[rw] == 0 && 1733 tg_dispatch_iops_time(tg, bio) == 0; 1734 1735 /* 1736 * Throtl is FIFO - if bios are already queued, should queue. 1737 * If the bps queue is empty and @bio is within the bps limit, charge 1738 * bps here for direct placement into the iops queue. 1739 */ 1740 if (sq_queued(&tg->service_queue, rw)) { 1741 if (sq->nr_queued_bps[rw] == 0 && 1742 tg_dispatch_bps_time(tg, bio) == 0) 1743 throtl_charge_bps_bio(tg, bio); 1744 1745 return false; 1746 } 1747 1748 return tg_dispatch_time(tg, bio) == 0; 1749 } 1750 1751 bool __blk_throtl_bio(struct bio *bio) 1752 { 1753 struct request_queue *q = bdev_get_queue(bio->bi_bdev); 1754 struct blkcg_gq *blkg = bio->bi_blkg; 1755 struct throtl_qnode *qn = NULL; 1756 struct throtl_grp *tg = blkg_to_tg(blkg); 1757 struct throtl_service_queue *sq; 1758 bool rw = bio_data_dir(bio); 1759 bool throttled = false; 1760 struct throtl_data *td = tg->td; 1761 1762 rcu_read_lock(); 1763 spin_lock_irq(&q->queue_lock); 1764 sq = &tg->service_queue; 1765 1766 while (true) { 1767 if (tg_within_limit(tg, bio, rw)) { 1768 /* within limits, let's charge and dispatch directly */ 1769 throtl_charge_iops_bio(tg, bio); 1770 1771 /* 1772 * We need to trim slice even when bios are not being 1773 * queued otherwise it might happen that a bio is not 1774 * queued for a long time and slice keeps on extending 1775 * and trim is not called for a long time. Now if limits 1776 * are reduced suddenly we take into account all the IO 1777 * dispatched so far at new low rate and * newly queued 1778 * IO gets a really long dispatch time. 1779 * 1780 * So keep on trimming slice even if bio is not queued. 1781 */ 1782 throtl_trim_slice(tg, rw); 1783 } else if (bio_issue_as_root_blkg(bio)) { 1784 /* 1785 * IOs which may cause priority inversions are 1786 * dispatched directly, even if they're over limit. 1787 * 1788 * Charge and dispatch directly, and our throttle 1789 * control algorithm is adaptive, and extra IO bytes 1790 * will be throttled for paying the debt 1791 */ 1792 throtl_charge_bps_bio(tg, bio); 1793 throtl_charge_iops_bio(tg, bio); 1794 } else { 1795 /* if above limits, break to queue */ 1796 break; 1797 } 1798 1799 /* 1800 * @bio passed through this layer without being throttled. 1801 * Climb up the ladder. If we're already at the top, it 1802 * can be executed directly. 1803 */ 1804 qn = &tg->qnode_on_parent[rw]; 1805 sq = sq->parent_sq; 1806 tg = sq_to_tg(sq); 1807 if (!tg) { 1808 bio_set_flag(bio, BIO_BPS_THROTTLED); 1809 goto out_unlock; 1810 } 1811 } 1812 1813 /* out-of-limit, queue to @tg */ 1814 throtl_log(sq, "[%c] bio. bdisp=%llu sz=%u bps=%llu iodisp=%u iops=%u queued=%d/%d", 1815 rw == READ ? 'R' : 'W', 1816 tg->bytes_disp[rw], bio->bi_iter.bi_size, 1817 tg_bps_limit(tg, rw), 1818 tg->io_disp[rw], tg_iops_limit(tg, rw), 1819 sq_queued(sq, READ), sq_queued(sq, WRITE)); 1820 1821 td->nr_queued[rw]++; 1822 throtl_add_bio_tg(bio, qn, tg); 1823 throttled = true; 1824 1825 /* 1826 * Update @tg's dispatch time and force schedule dispatch if @tg 1827 * was empty before @bio, or the iops queue is empty and @bio will 1828 * add to. The forced scheduling isn't likely to cause undue 1829 * delay as @bio is likely to be dispatched directly if its @tg's 1830 * disptime is not in the future. 1831 */ 1832 if (tg->flags & THROTL_TG_WAS_EMPTY || 1833 tg->flags & THROTL_TG_IOPS_WAS_EMPTY) { 1834 tg_update_disptime(tg); 1835 throtl_schedule_next_dispatch(tg->service_queue.parent_sq, true); 1836 } 1837 1838 out_unlock: 1839 spin_unlock_irq(&q->queue_lock); 1840 1841 rcu_read_unlock(); 1842 return throttled; 1843 } 1844 1845 void blk_throtl_exit(struct gendisk *disk) 1846 { 1847 struct request_queue *q = disk->queue; 1848 1849 if (!blk_throtl_activated(q)) 1850 return; 1851 1852 timer_delete_sync(&q->td->service_queue.pending_timer); 1853 throtl_shutdown_wq(q); 1854 blkcg_deactivate_policy(disk, &blkcg_policy_throtl); 1855 kfree(q->td); 1856 } 1857 1858 static int __init throtl_init(void) 1859 { 1860 kthrotld_workqueue = alloc_workqueue("kthrotld", WQ_MEM_RECLAIM, 0); 1861 if (!kthrotld_workqueue) 1862 panic("Failed to create kthrotld\n"); 1863 1864 return blkcg_policy_register(&blkcg_policy_throtl); 1865 } 1866 1867 module_init(throtl_init); 1868