1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * cgroups support for the BFQ I/O scheduler. 4 */ 5#include <linux/module.h> 6#include <linux/slab.h> 7#include <linux/blkdev.h> 8#include <linux/cgroup.h> 9#include <linux/elevator.h> 10#include <linux/ktime.h> 11#include <linux/rbtree.h> 12#include <linux/ioprio.h> 13#include <linux/sbitmap.h> 14#include <linux/delay.h> 15 16#include "bfq-iosched.h" 17 18#ifdef CONFIG_BFQ_CGROUP_DEBUG 19static int bfq_stat_init(struct bfq_stat *stat, gfp_t gfp) 20{ 21 int ret; 22 23 ret = percpu_counter_init(&stat->cpu_cnt, 0, gfp); 24 if (ret) 25 return ret; 26 27 atomic64_set(&stat->aux_cnt, 0); 28 return 0; 29} 30 31static void bfq_stat_exit(struct bfq_stat *stat) 32{ 33 percpu_counter_destroy(&stat->cpu_cnt); 34} 35 36/** 37 * bfq_stat_add - add a value to a bfq_stat 38 * @stat: target bfq_stat 39 * @val: value to add 40 * 41 * Add @val to @stat. The caller must ensure that IRQ on the same CPU 42 * don't re-enter this function for the same counter. 43 */ 44static inline void bfq_stat_add(struct bfq_stat *stat, uint64_t val) 45{ 46 percpu_counter_add_batch(&stat->cpu_cnt, val, BLKG_STAT_CPU_BATCH); 47} 48 49/** 50 * bfq_stat_read - read the current value of a bfq_stat 51 * @stat: bfq_stat to read 52 */ 53static inline uint64_t bfq_stat_read(struct bfq_stat *stat) 54{ 55 return percpu_counter_sum_positive(&stat->cpu_cnt); 56} 57 58/** 59 * bfq_stat_reset - reset a bfq_stat 60 * @stat: bfq_stat to reset 61 */ 62static inline void bfq_stat_reset(struct bfq_stat *stat) 63{ 64 percpu_counter_set(&stat->cpu_cnt, 0); 65 atomic64_set(&stat->aux_cnt, 0); 66} 67 68/** 69 * bfq_stat_add_aux - add a bfq_stat into another's aux count 70 * @to: the destination bfq_stat 71 * @from: the source 72 * 73 * Add @from's count including the aux one to @to's aux count. 74 */ 75static inline void bfq_stat_add_aux(struct bfq_stat *to, 76 struct bfq_stat *from) 77{ 78 atomic64_add(bfq_stat_read(from) + atomic64_read(&from->aux_cnt), 79 &to->aux_cnt); 80} 81 82/** 83 * blkg_prfill_stat - prfill callback for bfq_stat 84 * @sf: seq_file to print to 85 * @pd: policy private data of interest 86 * @off: offset to the bfq_stat in @pd 87 * 88 * prfill callback for printing a bfq_stat. 89 */ 90static u64 blkg_prfill_stat(struct seq_file *sf, struct blkg_policy_data *pd, 91 int off) 92{ 93 return __blkg_prfill_u64(sf, pd, bfq_stat_read((void *)pd + off)); 94} 95 96/* bfqg stats flags */ 97enum bfqg_stats_flags { 98 BFQG_stats_waiting = 0, 99 BFQG_stats_idling, 100 BFQG_stats_empty, 101}; 102 103#define BFQG_FLAG_FNS(name) \ 104static void bfqg_stats_mark_##name(struct bfqg_stats *stats) \ 105{ \ 106 stats->flags |= (1 << BFQG_stats_##name); \ 107} \ 108static void bfqg_stats_clear_##name(struct bfqg_stats *stats) \ 109{ \ 110 stats->flags &= ~(1 << BFQG_stats_##name); \ 111} \ 112static int bfqg_stats_##name(struct bfqg_stats *stats) \ 113{ \ 114 return (stats->flags & (1 << BFQG_stats_##name)) != 0; \ 115} \ 116 117BFQG_FLAG_FNS(waiting) 118BFQG_FLAG_FNS(idling) 119BFQG_FLAG_FNS(empty) 120#undef BFQG_FLAG_FNS 121 122/* This should be called with the scheduler lock held. */ 123static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats) 124{ 125 u64 now; 126 127 if (!bfqg_stats_waiting(stats)) 128 return; 129 130 now = ktime_get_ns(); 131 if (now > stats->start_group_wait_time) 132 bfq_stat_add(&stats->group_wait_time, 133 now - stats->start_group_wait_time); 134 bfqg_stats_clear_waiting(stats); 135} 136 137/* This should be called with the scheduler lock held. */ 138static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg, 139 struct bfq_group *curr_bfqg) 140{ 141 struct bfqg_stats *stats = &bfqg->stats; 142 143 if (bfqg_stats_waiting(stats)) 144 return; 145 if (bfqg == curr_bfqg) 146 return; 147 stats->start_group_wait_time = ktime_get_ns(); 148 bfqg_stats_mark_waiting(stats); 149} 150 151/* This should be called with the scheduler lock held. */ 152static void bfqg_stats_end_empty_time(struct bfqg_stats *stats) 153{ 154 u64 now; 155 156 if (!bfqg_stats_empty(stats)) 157 return; 158 159 now = ktime_get_ns(); 160 if (now > stats->start_empty_time) 161 bfq_stat_add(&stats->empty_time, 162 now - stats->start_empty_time); 163 bfqg_stats_clear_empty(stats); 164} 165 166void bfqg_stats_update_dequeue(struct bfq_group *bfqg) 167{ 168 bfq_stat_add(&bfqg->stats.dequeue, 1); 169} 170 171void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) 172{ 173 struct bfqg_stats *stats = &bfqg->stats; 174 175 if (blkg_rwstat_total(&stats->queued)) 176 return; 177 178 /* 179 * group is already marked empty. This can happen if bfqq got new 180 * request in parent group and moved to this group while being added 181 * to service tree. Just ignore the event and move on. 182 */ 183 if (bfqg_stats_empty(stats)) 184 return; 185 186 stats->start_empty_time = ktime_get_ns(); 187 bfqg_stats_mark_empty(stats); 188} 189 190void bfqg_stats_update_idle_time(struct bfq_group *bfqg) 191{ 192 struct bfqg_stats *stats = &bfqg->stats; 193 194 if (bfqg_stats_idling(stats)) { 195 u64 now = ktime_get_ns(); 196 197 if (now > stats->start_idle_time) 198 bfq_stat_add(&stats->idle_time, 199 now - stats->start_idle_time); 200 bfqg_stats_clear_idling(stats); 201 } 202} 203 204void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) 205{ 206 struct bfqg_stats *stats = &bfqg->stats; 207 208 stats->start_idle_time = ktime_get_ns(); 209 bfqg_stats_mark_idling(stats); 210} 211 212void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) 213{ 214 struct bfqg_stats *stats = &bfqg->stats; 215 216 bfq_stat_add(&stats->avg_queue_size_sum, 217 blkg_rwstat_total(&stats->queued)); 218 bfq_stat_add(&stats->avg_queue_size_samples, 1); 219 bfqg_stats_update_group_wait_time(stats); 220} 221 222void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq, 223 unsigned int op) 224{ 225 blkg_rwstat_add(&bfqg->stats.queued, op, 1); 226 bfqg_stats_end_empty_time(&bfqg->stats); 227 if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue)) 228 bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq)); 229} 230 231void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) 232{ 233 blkg_rwstat_add(&bfqg->stats.queued, op, -1); 234} 235 236void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) 237{ 238 blkg_rwstat_add(&bfqg->stats.merged, op, 1); 239} 240 241void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns, 242 u64 io_start_time_ns, unsigned int op) 243{ 244 struct bfqg_stats *stats = &bfqg->stats; 245 u64 now = ktime_get_ns(); 246 247 if (now > io_start_time_ns) 248 blkg_rwstat_add(&stats->service_time, op, 249 now - io_start_time_ns); 250 if (io_start_time_ns > start_time_ns) 251 blkg_rwstat_add(&stats->wait_time, op, 252 io_start_time_ns - start_time_ns); 253} 254 255#else /* CONFIG_BFQ_CGROUP_DEBUG */ 256 257void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq, 258 unsigned int op) { } 259void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { } 260void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { } 261void bfqg_stats_update_completion(struct bfq_group *bfqg, u64 start_time_ns, 262 u64 io_start_time_ns, unsigned int op) { } 263void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { } 264void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { } 265void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { } 266void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { } 267void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { } 268 269#endif /* CONFIG_BFQ_CGROUP_DEBUG */ 270 271#ifdef CONFIG_BFQ_GROUP_IOSCHED 272 273/* 274 * blk-cgroup policy-related handlers 275 * The following functions help in converting between blk-cgroup 276 * internal structures and BFQ-specific structures. 277 */ 278 279static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd) 280{ 281 return pd ? container_of(pd, struct bfq_group, pd) : NULL; 282} 283 284struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg) 285{ 286 return pd_to_blkg(&bfqg->pd); 287} 288 289static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg) 290{ 291 return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq)); 292} 293 294/* 295 * bfq_group handlers 296 * The following functions help in navigating the bfq_group hierarchy 297 * by allowing to find the parent of a bfq_group or the bfq_group 298 * associated to a bfq_queue. 299 */ 300 301static struct bfq_group *bfqg_parent(struct bfq_group *bfqg) 302{ 303 struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent; 304 305 return pblkg ? blkg_to_bfqg(pblkg) : NULL; 306} 307 308struct bfq_group *bfqq_group(struct bfq_queue *bfqq) 309{ 310 struct bfq_entity *group_entity = bfqq->entity.parent; 311 312 return group_entity ? container_of(group_entity, struct bfq_group, 313 entity) : 314 bfqq->bfqd->root_group; 315} 316 317/* 318 * The following two functions handle get and put of a bfq_group by 319 * wrapping the related blk-cgroup hooks. 320 */ 321 322static void bfqg_get(struct bfq_group *bfqg) 323{ 324 bfqg->ref++; 325} 326 327static void bfqg_put(struct bfq_group *bfqg) 328{ 329 bfqg->ref--; 330 331 if (bfqg->ref == 0) 332 kfree(bfqg); 333} 334 335static void bfqg_and_blkg_get(struct bfq_group *bfqg) 336{ 337 /* see comments in bfq_bic_update_cgroup for why refcounting bfqg */ 338 bfqg_get(bfqg); 339 340 blkg_get(bfqg_to_blkg(bfqg)); 341} 342 343void bfqg_and_blkg_put(struct bfq_group *bfqg) 344{ 345 blkg_put(bfqg_to_blkg(bfqg)); 346 347 bfqg_put(bfqg); 348} 349 350void bfqg_stats_update_legacy_io(struct request_queue *q, struct request *rq) 351{ 352 struct bfq_group *bfqg = blkg_to_bfqg(rq->bio->bi_blkg); 353 354 if (!bfqg) 355 return; 356 357 blkg_rwstat_add(&bfqg->stats.bytes, rq->cmd_flags, blk_rq_bytes(rq)); 358 blkg_rwstat_add(&bfqg->stats.ios, rq->cmd_flags, 1); 359} 360 361/* @stats = 0 */ 362static void bfqg_stats_reset(struct bfqg_stats *stats) 363{ 364#ifdef CONFIG_BFQ_CGROUP_DEBUG 365 /* queued stats shouldn't be cleared */ 366 blkg_rwstat_reset(&stats->merged); 367 blkg_rwstat_reset(&stats->service_time); 368 blkg_rwstat_reset(&stats->wait_time); 369 bfq_stat_reset(&stats->time); 370 bfq_stat_reset(&stats->avg_queue_size_sum); 371 bfq_stat_reset(&stats->avg_queue_size_samples); 372 bfq_stat_reset(&stats->dequeue); 373 bfq_stat_reset(&stats->group_wait_time); 374 bfq_stat_reset(&stats->idle_time); 375 bfq_stat_reset(&stats->empty_time); 376#endif 377} 378 379/* @to += @from */ 380static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from) 381{ 382 if (!to || !from) 383 return; 384 385#ifdef CONFIG_BFQ_CGROUP_DEBUG 386 /* queued stats shouldn't be cleared */ 387 blkg_rwstat_add_aux(&to->merged, &from->merged); 388 blkg_rwstat_add_aux(&to->service_time, &from->service_time); 389 blkg_rwstat_add_aux(&to->wait_time, &from->wait_time); 390 bfq_stat_add_aux(&from->time, &from->time); 391 bfq_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum); 392 bfq_stat_add_aux(&to->avg_queue_size_samples, 393 &from->avg_queue_size_samples); 394 bfq_stat_add_aux(&to->dequeue, &from->dequeue); 395 bfq_stat_add_aux(&to->group_wait_time, &from->group_wait_time); 396 bfq_stat_add_aux(&to->idle_time, &from->idle_time); 397 bfq_stat_add_aux(&to->empty_time, &from->empty_time); 398#endif 399} 400 401/* 402 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors' 403 * recursive stats can still account for the amount used by this bfqg after 404 * it's gone. 405 */ 406static void bfqg_stats_xfer_dead(struct bfq_group *bfqg) 407{ 408 struct bfq_group *parent; 409 410 if (!bfqg) /* root_group */ 411 return; 412 413 parent = bfqg_parent(bfqg); 414 415 lockdep_assert_held(&bfqg_to_blkg(bfqg)->q->queue_lock); 416 417 if (unlikely(!parent)) 418 return; 419 420 bfqg_stats_add_aux(&parent->stats, &bfqg->stats); 421 bfqg_stats_reset(&bfqg->stats); 422} 423 424void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg) 425{ 426 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); 427 428 entity->weight = entity->new_weight; 429 entity->orig_weight = entity->new_weight; 430 if (bfqq) { 431 bfqq->ioprio = bfqq->new_ioprio; 432 bfqq->ioprio_class = bfqq->new_ioprio_class; 433 /* 434 * Make sure that bfqg and its associated blkg do not 435 * disappear before entity. 436 */ 437 bfqg_and_blkg_get(bfqg); 438 } 439 entity->parent = bfqg->my_entity; /* NULL for root group */ 440 entity->sched_data = &bfqg->sched_data; 441} 442 443static void bfqg_stats_exit(struct bfqg_stats *stats) 444{ 445 blkg_rwstat_exit(&stats->bytes); 446 blkg_rwstat_exit(&stats->ios); 447#ifdef CONFIG_BFQ_CGROUP_DEBUG 448 blkg_rwstat_exit(&stats->merged); 449 blkg_rwstat_exit(&stats->service_time); 450 blkg_rwstat_exit(&stats->wait_time); 451 blkg_rwstat_exit(&stats->queued); 452 bfq_stat_exit(&stats->time); 453 bfq_stat_exit(&stats->avg_queue_size_sum); 454 bfq_stat_exit(&stats->avg_queue_size_samples); 455 bfq_stat_exit(&stats->dequeue); 456 bfq_stat_exit(&stats->group_wait_time); 457 bfq_stat_exit(&stats->idle_time); 458 bfq_stat_exit(&stats->empty_time); 459#endif 460} 461 462static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp) 463{ 464 if (blkg_rwstat_init(&stats->bytes, gfp) || 465 blkg_rwstat_init(&stats->ios, gfp)) 466 goto error; 467 468#ifdef CONFIG_BFQ_CGROUP_DEBUG 469 if (blkg_rwstat_init(&stats->merged, gfp) || 470 blkg_rwstat_init(&stats->service_time, gfp) || 471 blkg_rwstat_init(&stats->wait_time, gfp) || 472 blkg_rwstat_init(&stats->queued, gfp) || 473 bfq_stat_init(&stats->time, gfp) || 474 bfq_stat_init(&stats->avg_queue_size_sum, gfp) || 475 bfq_stat_init(&stats->avg_queue_size_samples, gfp) || 476 bfq_stat_init(&stats->dequeue, gfp) || 477 bfq_stat_init(&stats->group_wait_time, gfp) || 478 bfq_stat_init(&stats->idle_time, gfp) || 479 bfq_stat_init(&stats->empty_time, gfp)) 480 goto error; 481#endif 482 483 return 0; 484 485error: 486 bfqg_stats_exit(stats); 487 return -ENOMEM; 488} 489 490static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd) 491{ 492 return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL; 493} 494 495static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg) 496{ 497 return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq)); 498} 499 500static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp) 501{ 502 struct bfq_group_data *bgd; 503 504 bgd = kzalloc(sizeof(*bgd), gfp); 505 if (!bgd) 506 return NULL; 507 return &bgd->pd; 508} 509 510static void bfq_cpd_init(struct blkcg_policy_data *cpd) 511{ 512 struct bfq_group_data *d = cpd_to_bfqgd(cpd); 513 514 d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ? 515 CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL; 516} 517 518static void bfq_cpd_free(struct blkcg_policy_data *cpd) 519{ 520 kfree(cpd_to_bfqgd(cpd)); 521} 522 523static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q, 524 struct blkcg *blkcg) 525{ 526 struct bfq_group *bfqg; 527 528 bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node); 529 if (!bfqg) 530 return NULL; 531 532 if (bfqg_stats_init(&bfqg->stats, gfp)) { 533 kfree(bfqg); 534 return NULL; 535 } 536 537 /* see comments in bfq_bic_update_cgroup for why refcounting */ 538 bfqg_get(bfqg); 539 return &bfqg->pd; 540} 541 542static void bfq_pd_init(struct blkg_policy_data *pd) 543{ 544 struct blkcg_gq *blkg = pd_to_blkg(pd); 545 struct bfq_group *bfqg = blkg_to_bfqg(blkg); 546 struct bfq_data *bfqd = blkg->q->elevator->elevator_data; 547 struct bfq_entity *entity = &bfqg->entity; 548 struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg); 549 550 entity->orig_weight = entity->weight = entity->new_weight = d->weight; 551 entity->my_sched_data = &bfqg->sched_data; 552 bfqg->my_entity = entity; /* 553 * the root_group's will be set to NULL 554 * in bfq_init_queue() 555 */ 556 bfqg->bfqd = bfqd; 557 bfqg->active_entities = 0; 558 bfqg->online = true; 559 bfqg->rq_pos_tree = RB_ROOT; 560} 561 562static void bfq_pd_free(struct blkg_policy_data *pd) 563{ 564 struct bfq_group *bfqg = pd_to_bfqg(pd); 565 566 bfqg_stats_exit(&bfqg->stats); 567 bfqg_put(bfqg); 568} 569 570static void bfq_pd_reset_stats(struct blkg_policy_data *pd) 571{ 572 struct bfq_group *bfqg = pd_to_bfqg(pd); 573 574 bfqg_stats_reset(&bfqg->stats); 575} 576 577static void bfq_group_set_parent(struct bfq_group *bfqg, 578 struct bfq_group *parent) 579{ 580 struct bfq_entity *entity; 581 582 entity = &bfqg->entity; 583 entity->parent = parent->my_entity; 584 entity->sched_data = &parent->sched_data; 585} 586 587static void bfq_link_bfqg(struct bfq_data *bfqd, struct bfq_group *bfqg) 588{ 589 struct bfq_group *parent; 590 struct bfq_entity *entity; 591 592 /* 593 * Update chain of bfq_groups as we might be handling a leaf group 594 * which, along with some of its relatives, has not been hooked yet 595 * to the private hierarchy of BFQ. 596 */ 597 entity = &bfqg->entity; 598 for_each_entity(entity) { 599 struct bfq_group *curr_bfqg = container_of(entity, 600 struct bfq_group, entity); 601 if (curr_bfqg != bfqd->root_group) { 602 parent = bfqg_parent(curr_bfqg); 603 if (!parent) 604 parent = bfqd->root_group; 605 bfq_group_set_parent(curr_bfqg, parent); 606 } 607 } 608} 609 610struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio) 611{ 612 struct blkcg_gq *blkg = bio->bi_blkg; 613 struct bfq_group *bfqg; 614 615 while (blkg) { 616 if (!blkg->online) { 617 blkg = blkg->parent; 618 continue; 619 } 620 bfqg = blkg_to_bfqg(blkg); 621 if (bfqg->online) { 622 bio_associate_blkg_from_css(bio, &blkg->blkcg->css); 623 return bfqg; 624 } 625 blkg = blkg->parent; 626 } 627 bio_associate_blkg_from_css(bio, 628 &bfqg_to_blkg(bfqd->root_group)->blkcg->css); 629 return bfqd->root_group; 630} 631 632/** 633 * bfq_bfqq_move - migrate @bfqq to @bfqg. 634 * @bfqd: queue descriptor. 635 * @bfqq: the queue to move. 636 * @bfqg: the group to move to. 637 * 638 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating 639 * it on the new one. Avoid putting the entity on the old group idle tree. 640 * 641 * Must be called under the scheduler lock, to make sure that the blkg 642 * owning @bfqg does not disappear (see comments in 643 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg 644 * objects). 645 */ 646void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, 647 struct bfq_group *bfqg) 648{ 649 struct bfq_entity *entity = &bfqq->entity; 650 651 /* 652 * oom_bfqq is not allowed to move, oom_bfqq will hold ref to root_group 653 * until elevator exit. 654 */ 655 if (bfqq == &bfqd->oom_bfqq) 656 return; 657 /* 658 * Get extra reference to prevent bfqq from being freed in 659 * next possible expire or deactivate. 660 */ 661 bfqq->ref++; 662 663 /* If bfqq is empty, then bfq_bfqq_expire also invokes 664 * bfq_del_bfqq_busy, thereby removing bfqq and its entity 665 * from data structures related to current group. Otherwise we 666 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as 667 * we do below. 668 */ 669 if (bfqq == bfqd->in_service_queue) 670 bfq_bfqq_expire(bfqd, bfqd->in_service_queue, 671 false, BFQQE_PREEMPTED); 672 673 if (bfq_bfqq_busy(bfqq)) 674 bfq_deactivate_bfqq(bfqd, bfqq, false, false); 675 else if (entity->on_st_or_in_serv) 676 bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); 677 bfqg_and_blkg_put(bfqq_group(bfqq)); 678 679 entity->parent = bfqg->my_entity; 680 entity->sched_data = &bfqg->sched_data; 681 /* pin down bfqg and its associated blkg */ 682 bfqg_and_blkg_get(bfqg); 683 684 if (bfq_bfqq_busy(bfqq)) { 685 if (unlikely(!bfqd->nonrot_with_queueing)) 686 bfq_pos_tree_add_move(bfqd, bfqq); 687 bfq_activate_bfqq(bfqd, bfqq); 688 } 689 690 if (!bfqd->in_service_queue && !bfqd->rq_in_driver) 691 bfq_schedule_dispatch(bfqd); 692 /* release extra ref taken above, bfqq may happen to be freed now */ 693 bfq_put_queue(bfqq); 694} 695 696/** 697 * __bfq_bic_change_cgroup - move @bic to @cgroup. 698 * @bfqd: the queue descriptor. 699 * @bic: the bic to move. 700 * @blkcg: the blk-cgroup to move to. 701 * 702 * Move bic to blkcg, assuming that bfqd->lock is held; which makes 703 * sure that the reference to cgroup is valid across the call (see 704 * comments in bfq_bic_update_cgroup on this issue) 705 */ 706static void *__bfq_bic_change_cgroup(struct bfq_data *bfqd, 707 struct bfq_io_cq *bic, 708 struct bfq_group *bfqg) 709{ 710 struct bfq_queue *async_bfqq = bic_to_bfqq(bic, false); 711 struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, true); 712 struct bfq_entity *entity; 713 714 if (async_bfqq) { 715 entity = &async_bfqq->entity; 716 717 if (entity->sched_data != &bfqg->sched_data) { 718 bic_set_bfqq(bic, NULL, false); 719 bfq_release_process_ref(bfqd, async_bfqq); 720 } 721 } 722 723 if (sync_bfqq) { 724 if (!sync_bfqq->new_bfqq && !bfq_bfqq_coop(sync_bfqq)) { 725 /* We are the only user of this bfqq, just move it */ 726 if (sync_bfqq->entity.sched_data != &bfqg->sched_data) 727 bfq_bfqq_move(bfqd, sync_bfqq, bfqg); 728 } else { 729 struct bfq_queue *bfqq; 730 731 /* 732 * The queue was merged to a different queue. Check 733 * that the merge chain still belongs to the same 734 * cgroup. 735 */ 736 for (bfqq = sync_bfqq; bfqq; bfqq = bfqq->new_bfqq) 737 if (bfqq->entity.sched_data != 738 &bfqg->sched_data) 739 break; 740 if (bfqq) { 741 /* 742 * Some queue changed cgroup so the merge is 743 * not valid anymore. We cannot easily just 744 * cancel the merge (by clearing new_bfqq) as 745 * there may be other processes using this 746 * queue and holding refs to all queues below 747 * sync_bfqq->new_bfqq. Similarly if the merge 748 * already happened, we need to detach from 749 * bfqq now so that we cannot merge bio to a 750 * request from the old cgroup. 751 */ 752 bfq_put_cooperator(sync_bfqq); 753 bic_set_bfqq(bic, NULL, true); 754 bfq_release_process_ref(bfqd, sync_bfqq); 755 } 756 } 757 } 758 759 return bfqg; 760} 761 762void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) 763{ 764 struct bfq_data *bfqd = bic_to_bfqd(bic); 765 struct bfq_group *bfqg = bfq_bio_bfqg(bfqd, bio); 766 uint64_t serial_nr; 767 768 serial_nr = bfqg_to_blkg(bfqg)->blkcg->css.serial_nr; 769 770 /* 771 * Check whether blkcg has changed. The condition may trigger 772 * spuriously on a newly created cic but there's no harm. 773 */ 774 if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr)) 775 return; 776 777 /* 778 * New cgroup for this process. Make sure it is linked to bfq internal 779 * cgroup hierarchy. 780 */ 781 bfq_link_bfqg(bfqd, bfqg); 782 __bfq_bic_change_cgroup(bfqd, bic, bfqg); 783 /* 784 * Update blkg_path for bfq_log_* functions. We cache this 785 * path, and update it here, for the following 786 * reasons. Operations on blkg objects in blk-cgroup are 787 * protected with the request_queue lock, and not with the 788 * lock that protects the instances of this scheduler 789 * (bfqd->lock). This exposes BFQ to the following sort of 790 * race. 791 * 792 * The blkg_lookup performed in bfq_get_queue, protected 793 * through rcu, may happen to return the address of a copy of 794 * the original blkg. If this is the case, then the 795 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down 796 * the blkg, is useless: it does not prevent blk-cgroup code 797 * from destroying both the original blkg and all objects 798 * directly or indirectly referred by the copy of the 799 * blkg. 800 * 801 * On the bright side, destroy operations on a blkg invoke, as 802 * a first step, hooks of the scheduler associated with the 803 * blkg. And these hooks are executed with bfqd->lock held for 804 * BFQ. As a consequence, for any blkg associated with the 805 * request queue this instance of the scheduler is attached 806 * to, we are guaranteed that such a blkg is not destroyed, and 807 * that all the pointers it contains are consistent, while we 808 * are holding bfqd->lock. A blkg_lookup performed with 809 * bfqd->lock held then returns a fully consistent blkg, which 810 * remains consistent until this lock is held. 811 * 812 * Thanks to the last fact, and to the fact that: (1) bfqg has 813 * been obtained through a blkg_lookup in the above 814 * assignment, and (2) bfqd->lock is being held, here we can 815 * safely use the policy data for the involved blkg (i.e., the 816 * field bfqg->pd) to get to the blkg associated with bfqg, 817 * and then we can safely use any field of blkg. After we 818 * release bfqd->lock, even just getting blkg through this 819 * bfqg may cause dangling references to be traversed, as 820 * bfqg->pd may not exist any more. 821 * 822 * In view of the above facts, here we cache, in the bfqg, any 823 * blkg data we may need for this bic, and for its associated 824 * bfq_queue. As of now, we need to cache only the path of the 825 * blkg, which is used in the bfq_log_* functions. 826 * 827 * Finally, note that bfqg itself needs to be protected from 828 * destruction on the blkg_free of the original blkg (which 829 * invokes bfq_pd_free). We use an additional private 830 * refcounter for bfqg, to let it disappear only after no 831 * bfq_queue refers to it any longer. 832 */ 833 blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path)); 834 bic->blkcg_serial_nr = serial_nr; 835} 836 837/** 838 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. 839 * @st: the service tree being flushed. 840 */ 841static void bfq_flush_idle_tree(struct bfq_service_tree *st) 842{ 843 struct bfq_entity *entity = st->first_idle; 844 845 for (; entity ; entity = st->first_idle) 846 __bfq_deactivate_entity(entity, false); 847} 848 849/** 850 * bfq_reparent_leaf_entity - move leaf entity to the root_group. 851 * @bfqd: the device data structure with the root group. 852 * @entity: the entity to move, if entity is a leaf; or the parent entity 853 * of an active leaf entity to move, if entity is not a leaf. 854 */ 855static void bfq_reparent_leaf_entity(struct bfq_data *bfqd, 856 struct bfq_entity *entity, 857 int ioprio_class) 858{ 859 struct bfq_queue *bfqq; 860 struct bfq_entity *child_entity = entity; 861 862 while (child_entity->my_sched_data) { /* leaf not reached yet */ 863 struct bfq_sched_data *child_sd = child_entity->my_sched_data; 864 struct bfq_service_tree *child_st = child_sd->service_tree + 865 ioprio_class; 866 struct rb_root *child_active = &child_st->active; 867 868 child_entity = bfq_entity_of(rb_first(child_active)); 869 870 if (!child_entity) 871 child_entity = child_sd->in_service_entity; 872 } 873 874 bfqq = bfq_entity_to_bfqq(child_entity); 875 bfq_bfqq_move(bfqd, bfqq, bfqd->root_group); 876} 877 878/** 879 * bfq_reparent_active_queues - move to the root group all active queues. 880 * @bfqd: the device data structure with the root group. 881 * @bfqg: the group to move from. 882 * @st: the service tree to start the search from. 883 */ 884static void bfq_reparent_active_queues(struct bfq_data *bfqd, 885 struct bfq_group *bfqg, 886 struct bfq_service_tree *st, 887 int ioprio_class) 888{ 889 struct rb_root *active = &st->active; 890 struct bfq_entity *entity; 891 892 while ((entity = bfq_entity_of(rb_first(active)))) 893 bfq_reparent_leaf_entity(bfqd, entity, ioprio_class); 894 895 if (bfqg->sched_data.in_service_entity) 896 bfq_reparent_leaf_entity(bfqd, 897 bfqg->sched_data.in_service_entity, 898 ioprio_class); 899} 900 901/** 902 * bfq_pd_offline - deactivate the entity associated with @pd, 903 * and reparent its children entities. 904 * @pd: descriptor of the policy going offline. 905 * 906 * blkio already grabs the queue_lock for us, so no need to use 907 * RCU-based magic 908 */ 909static void bfq_pd_offline(struct blkg_policy_data *pd) 910{ 911 struct bfq_service_tree *st; 912 struct bfq_group *bfqg = pd_to_bfqg(pd); 913 struct bfq_data *bfqd = bfqg->bfqd; 914 struct bfq_entity *entity = bfqg->my_entity; 915 unsigned long flags; 916 int i; 917 918 spin_lock_irqsave(&bfqd->lock, flags); 919 920 if (!entity) /* root group */ 921 goto put_async_queues; 922 923 /* 924 * Empty all service_trees belonging to this group before 925 * deactivating the group itself. 926 */ 927 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { 928 st = bfqg->sched_data.service_tree + i; 929 930 /* 931 * It may happen that some queues are still active 932 * (busy) upon group destruction (if the corresponding 933 * processes have been forced to terminate). We move 934 * all the leaf entities corresponding to these queues 935 * to the root_group. 936 * Also, it may happen that the group has an entity 937 * in service, which is disconnected from the active 938 * tree: it must be moved, too. 939 * There is no need to put the sync queues, as the 940 * scheduler has taken no reference. 941 */ 942 bfq_reparent_active_queues(bfqd, bfqg, st, i); 943 944 /* 945 * The idle tree may still contain bfq_queues 946 * belonging to exited task because they never 947 * migrated to a different cgroup from the one being 948 * destroyed now. In addition, even 949 * bfq_reparent_active_queues() may happen to add some 950 * entities to the idle tree. It happens if, in some 951 * of the calls to bfq_bfqq_move() performed by 952 * bfq_reparent_active_queues(), the queue to move is 953 * empty and gets expired. 954 */ 955 bfq_flush_idle_tree(st); 956 } 957 958 __bfq_deactivate_entity(entity, false); 959 960put_async_queues: 961 bfq_put_async_queues(bfqd, bfqg); 962 bfqg->online = false; 963 964 spin_unlock_irqrestore(&bfqd->lock, flags); 965 /* 966 * @blkg is going offline and will be ignored by 967 * blkg_[rw]stat_recursive_sum(). Transfer stats to the parent so 968 * that they don't get lost. If IOs complete after this point, the 969 * stats for them will be lost. Oh well... 970 */ 971 bfqg_stats_xfer_dead(bfqg); 972} 973 974void bfq_end_wr_async(struct bfq_data *bfqd) 975{ 976 struct blkcg_gq *blkg; 977 978 list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) { 979 struct bfq_group *bfqg = blkg_to_bfqg(blkg); 980 981 bfq_end_wr_async_queues(bfqd, bfqg); 982 } 983 bfq_end_wr_async_queues(bfqd, bfqd->root_group); 984} 985 986static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v) 987{ 988 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 989 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg); 990 unsigned int val = 0; 991 992 if (bfqgd) 993 val = bfqgd->weight; 994 995 seq_printf(sf, "%u\n", val); 996 997 return 0; 998} 999 1000static u64 bfqg_prfill_weight_device(struct seq_file *sf, 1001 struct blkg_policy_data *pd, int off) 1002{ 1003 struct bfq_group *bfqg = pd_to_bfqg(pd); 1004 1005 if (!bfqg->entity.dev_weight) 1006 return 0; 1007 return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight); 1008} 1009 1010static int bfq_io_show_weight(struct seq_file *sf, void *v) 1011{ 1012 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 1013 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg); 1014 1015 seq_printf(sf, "default %u\n", bfqgd->weight); 1016 blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device, 1017 &blkcg_policy_bfq, 0, false); 1018 return 0; 1019} 1020 1021static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight) 1022{ 1023 weight = dev_weight ?: weight; 1024 1025 bfqg->entity.dev_weight = dev_weight; 1026 /* 1027 * Setting the prio_changed flag of the entity 1028 * to 1 with new_weight == weight would re-set 1029 * the value of the weight to its ioprio mapping. 1030 * Set the flag only if necessary. 1031 */ 1032 if ((unsigned short)weight != bfqg->entity.new_weight) { 1033 bfqg->entity.new_weight = (unsigned short)weight; 1034 /* 1035 * Make sure that the above new value has been 1036 * stored in bfqg->entity.new_weight before 1037 * setting the prio_changed flag. In fact, 1038 * this flag may be read asynchronously (in 1039 * critical sections protected by a different 1040 * lock than that held here), and finding this 1041 * flag set may cause the execution of the code 1042 * for updating parameters whose value may 1043 * depend also on bfqg->entity.new_weight (in 1044 * __bfq_entity_update_weight_prio). 1045 * This barrier makes sure that the new value 1046 * of bfqg->entity.new_weight is correctly 1047 * seen in that code. 1048 */ 1049 smp_wmb(); 1050 bfqg->entity.prio_changed = 1; 1051 } 1052} 1053 1054static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css, 1055 struct cftype *cftype, 1056 u64 val) 1057{ 1058 struct blkcg *blkcg = css_to_blkcg(css); 1059 struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg); 1060 struct blkcg_gq *blkg; 1061 int ret = -ERANGE; 1062 1063 if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT) 1064 return ret; 1065 1066 ret = 0; 1067 spin_lock_irq(&blkcg->lock); 1068 bfqgd->weight = (unsigned short)val; 1069 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { 1070 struct bfq_group *bfqg = blkg_to_bfqg(blkg); 1071 1072 if (bfqg) 1073 bfq_group_set_weight(bfqg, val, 0); 1074 } 1075 spin_unlock_irq(&blkcg->lock); 1076 1077 return ret; 1078} 1079 1080static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of, 1081 char *buf, size_t nbytes, 1082 loff_t off) 1083{ 1084 int ret; 1085 struct blkg_conf_ctx ctx; 1086 struct blkcg *blkcg = css_to_blkcg(of_css(of)); 1087 struct bfq_group *bfqg; 1088 u64 v; 1089 1090 ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx); 1091 if (ret) 1092 return ret; 1093 1094 if (sscanf(ctx.body, "%llu", &v) == 1) { 1095 /* require "default" on dfl */ 1096 ret = -ERANGE; 1097 if (!v) 1098 goto out; 1099 } else if (!strcmp(strim(ctx.body), "default")) { 1100 v = 0; 1101 } else { 1102 ret = -EINVAL; 1103 goto out; 1104 } 1105 1106 bfqg = blkg_to_bfqg(ctx.blkg); 1107 1108 ret = -ERANGE; 1109 if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) { 1110 bfq_group_set_weight(bfqg, bfqg->entity.weight, v); 1111 ret = 0; 1112 } 1113out: 1114 blkg_conf_finish(&ctx); 1115 return ret ?: nbytes; 1116} 1117 1118static ssize_t bfq_io_set_weight(struct kernfs_open_file *of, 1119 char *buf, size_t nbytes, 1120 loff_t off) 1121{ 1122 char *endp; 1123 int ret; 1124 u64 v; 1125 1126 buf = strim(buf); 1127 1128 /* "WEIGHT" or "default WEIGHT" sets the default weight */ 1129 v = simple_strtoull(buf, &endp, 0); 1130 if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) { 1131 ret = bfq_io_set_weight_legacy(of_css(of), NULL, v); 1132 return ret ?: nbytes; 1133 } 1134 1135 return bfq_io_set_device_weight(of, buf, nbytes, off); 1136} 1137 1138static int bfqg_print_rwstat(struct seq_file *sf, void *v) 1139{ 1140 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat, 1141 &blkcg_policy_bfq, seq_cft(sf)->private, true); 1142 return 0; 1143} 1144 1145static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf, 1146 struct blkg_policy_data *pd, int off) 1147{ 1148 struct blkg_rwstat_sample sum; 1149 1150 blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum); 1151 return __blkg_prfill_rwstat(sf, pd, &sum); 1152} 1153 1154static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v) 1155{ 1156 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1157 bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq, 1158 seq_cft(sf)->private, true); 1159 return 0; 1160} 1161 1162#ifdef CONFIG_BFQ_CGROUP_DEBUG 1163static int bfqg_print_stat(struct seq_file *sf, void *v) 1164{ 1165 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat, 1166 &blkcg_policy_bfq, seq_cft(sf)->private, false); 1167 return 0; 1168} 1169 1170static u64 bfqg_prfill_stat_recursive(struct seq_file *sf, 1171 struct blkg_policy_data *pd, int off) 1172{ 1173 struct blkcg_gq *blkg = pd_to_blkg(pd); 1174 struct blkcg_gq *pos_blkg; 1175 struct cgroup_subsys_state *pos_css; 1176 u64 sum = 0; 1177 1178 lockdep_assert_held(&blkg->q->queue_lock); 1179 1180 rcu_read_lock(); 1181 blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) { 1182 struct bfq_stat *stat; 1183 1184 if (!pos_blkg->online) 1185 continue; 1186 1187 stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off; 1188 sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt); 1189 } 1190 rcu_read_unlock(); 1191 1192 return __blkg_prfill_u64(sf, pd, sum); 1193} 1194 1195static int bfqg_print_stat_recursive(struct seq_file *sf, void *v) 1196{ 1197 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1198 bfqg_prfill_stat_recursive, &blkcg_policy_bfq, 1199 seq_cft(sf)->private, false); 1200 return 0; 1201} 1202 1203static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd, 1204 int off) 1205{ 1206 struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg); 1207 u64 sum = blkg_rwstat_total(&bfqg->stats.bytes); 1208 1209 return __blkg_prfill_u64(sf, pd, sum >> 9); 1210} 1211 1212static int bfqg_print_stat_sectors(struct seq_file *sf, void *v) 1213{ 1214 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1215 bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false); 1216 return 0; 1217} 1218 1219static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf, 1220 struct blkg_policy_data *pd, int off) 1221{ 1222 struct blkg_rwstat_sample tmp; 1223 1224 blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq, 1225 offsetof(struct bfq_group, stats.bytes), &tmp); 1226 1227 return __blkg_prfill_u64(sf, pd, 1228 (tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9); 1229} 1230 1231static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v) 1232{ 1233 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1234 bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0, 1235 false); 1236 return 0; 1237} 1238 1239static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf, 1240 struct blkg_policy_data *pd, int off) 1241{ 1242 struct bfq_group *bfqg = pd_to_bfqg(pd); 1243 u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples); 1244 u64 v = 0; 1245 1246 if (samples) { 1247 v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum); 1248 v = div64_u64(v, samples); 1249 } 1250 __blkg_prfill_u64(sf, pd, v); 1251 return 0; 1252} 1253 1254/* print avg_queue_size */ 1255static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v) 1256{ 1257 blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), 1258 bfqg_prfill_avg_queue_size, &blkcg_policy_bfq, 1259 0, false); 1260 return 0; 1261} 1262#endif /* CONFIG_BFQ_CGROUP_DEBUG */ 1263 1264struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node) 1265{ 1266 int ret; 1267 1268 ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq); 1269 if (ret) 1270 return NULL; 1271 1272 return blkg_to_bfqg(bfqd->queue->root_blkg); 1273} 1274 1275struct blkcg_policy blkcg_policy_bfq = { 1276 .dfl_cftypes = bfq_blkg_files, 1277 .legacy_cftypes = bfq_blkcg_legacy_files, 1278 1279 .cpd_alloc_fn = bfq_cpd_alloc, 1280 .cpd_init_fn = bfq_cpd_init, 1281 .cpd_bind_fn = bfq_cpd_init, 1282 .cpd_free_fn = bfq_cpd_free, 1283 1284 .pd_alloc_fn = bfq_pd_alloc, 1285 .pd_init_fn = bfq_pd_init, 1286 .pd_offline_fn = bfq_pd_offline, 1287 .pd_free_fn = bfq_pd_free, 1288 .pd_reset_stats_fn = bfq_pd_reset_stats, 1289}; 1290 1291struct cftype bfq_blkcg_legacy_files[] = { 1292 { 1293 .name = "bfq.weight", 1294 .flags = CFTYPE_NOT_ON_ROOT, 1295 .seq_show = bfq_io_show_weight_legacy, 1296 .write_u64 = bfq_io_set_weight_legacy, 1297 }, 1298 { 1299 .name = "bfq.weight_device", 1300 .flags = CFTYPE_NOT_ON_ROOT, 1301 .seq_show = bfq_io_show_weight, 1302 .write = bfq_io_set_weight, 1303 }, 1304 1305 /* statistics, covers only the tasks in the bfqg */ 1306 { 1307 .name = "bfq.io_service_bytes", 1308 .private = offsetof(struct bfq_group, stats.bytes), 1309 .seq_show = bfqg_print_rwstat, 1310 }, 1311 { 1312 .name = "bfq.io_serviced", 1313 .private = offsetof(struct bfq_group, stats.ios), 1314 .seq_show = bfqg_print_rwstat, 1315 }, 1316#ifdef CONFIG_BFQ_CGROUP_DEBUG 1317 { 1318 .name = "bfq.time", 1319 .private = offsetof(struct bfq_group, stats.time), 1320 .seq_show = bfqg_print_stat, 1321 }, 1322 { 1323 .name = "bfq.sectors", 1324 .seq_show = bfqg_print_stat_sectors, 1325 }, 1326 { 1327 .name = "bfq.io_service_time", 1328 .private = offsetof(struct bfq_group, stats.service_time), 1329 .seq_show = bfqg_print_rwstat, 1330 }, 1331 { 1332 .name = "bfq.io_wait_time", 1333 .private = offsetof(struct bfq_group, stats.wait_time), 1334 .seq_show = bfqg_print_rwstat, 1335 }, 1336 { 1337 .name = "bfq.io_merged", 1338 .private = offsetof(struct bfq_group, stats.merged), 1339 .seq_show = bfqg_print_rwstat, 1340 }, 1341 { 1342 .name = "bfq.io_queued", 1343 .private = offsetof(struct bfq_group, stats.queued), 1344 .seq_show = bfqg_print_rwstat, 1345 }, 1346#endif /* CONFIG_BFQ_CGROUP_DEBUG */ 1347 1348 /* the same statistics which cover the bfqg and its descendants */ 1349 { 1350 .name = "bfq.io_service_bytes_recursive", 1351 .private = offsetof(struct bfq_group, stats.bytes), 1352 .seq_show = bfqg_print_rwstat_recursive, 1353 }, 1354 { 1355 .name = "bfq.io_serviced_recursive", 1356 .private = offsetof(struct bfq_group, stats.ios), 1357 .seq_show = bfqg_print_rwstat_recursive, 1358 }, 1359#ifdef CONFIG_BFQ_CGROUP_DEBUG 1360 { 1361 .name = "bfq.time_recursive", 1362 .private = offsetof(struct bfq_group, stats.time), 1363 .seq_show = bfqg_print_stat_recursive, 1364 }, 1365 { 1366 .name = "bfq.sectors_recursive", 1367 .seq_show = bfqg_print_stat_sectors_recursive, 1368 }, 1369 { 1370 .name = "bfq.io_service_time_recursive", 1371 .private = offsetof(struct bfq_group, stats.service_time), 1372 .seq_show = bfqg_print_rwstat_recursive, 1373 }, 1374 { 1375 .name = "bfq.io_wait_time_recursive", 1376 .private = offsetof(struct bfq_group, stats.wait_time), 1377 .seq_show = bfqg_print_rwstat_recursive, 1378 }, 1379 { 1380 .name = "bfq.io_merged_recursive", 1381 .private = offsetof(struct bfq_group, stats.merged), 1382 .seq_show = bfqg_print_rwstat_recursive, 1383 }, 1384 { 1385 .name = "bfq.io_queued_recursive", 1386 .private = offsetof(struct bfq_group, stats.queued), 1387 .seq_show = bfqg_print_rwstat_recursive, 1388 }, 1389 { 1390 .name = "bfq.avg_queue_size", 1391 .seq_show = bfqg_print_avg_queue_size, 1392 }, 1393 { 1394 .name = "bfq.group_wait_time", 1395 .private = offsetof(struct bfq_group, stats.group_wait_time), 1396 .seq_show = bfqg_print_stat, 1397 }, 1398 { 1399 .name = "bfq.idle_time", 1400 .private = offsetof(struct bfq_group, stats.idle_time), 1401 .seq_show = bfqg_print_stat, 1402 }, 1403 { 1404 .name = "bfq.empty_time", 1405 .private = offsetof(struct bfq_group, stats.empty_time), 1406 .seq_show = bfqg_print_stat, 1407 }, 1408 { 1409 .name = "bfq.dequeue", 1410 .private = offsetof(struct bfq_group, stats.dequeue), 1411 .seq_show = bfqg_print_stat, 1412 }, 1413#endif /* CONFIG_BFQ_CGROUP_DEBUG */ 1414 { } /* terminate */ 1415}; 1416 1417struct cftype bfq_blkg_files[] = { 1418 { 1419 .name = "bfq.weight", 1420 .flags = CFTYPE_NOT_ON_ROOT, 1421 .seq_show = bfq_io_show_weight, 1422 .write = bfq_io_set_weight, 1423 }, 1424 {} /* terminate */ 1425}; 1426 1427#else /* CONFIG_BFQ_GROUP_IOSCHED */ 1428 1429void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, 1430 struct bfq_group *bfqg) {} 1431 1432void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg) 1433{ 1434 struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); 1435 1436 entity->weight = entity->new_weight; 1437 entity->orig_weight = entity->new_weight; 1438 if (bfqq) { 1439 bfqq->ioprio = bfqq->new_ioprio; 1440 bfqq->ioprio_class = bfqq->new_ioprio_class; 1441 } 1442 entity->sched_data = &bfqg->sched_data; 1443} 1444 1445void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {} 1446 1447void bfq_end_wr_async(struct bfq_data *bfqd) 1448{ 1449 bfq_end_wr_async_queues(bfqd, bfqd->root_group); 1450} 1451 1452struct bfq_group *bfq_bio_bfqg(struct bfq_data *bfqd, struct bio *bio) 1453{ 1454 return bfqd->root_group; 1455} 1456 1457struct bfq_group *bfqq_group(struct bfq_queue *bfqq) 1458{ 1459 return bfqq->bfqd->root_group; 1460} 1461 1462void bfqg_and_blkg_get(struct bfq_group *bfqg) {} 1463 1464void bfqg_and_blkg_put(struct bfq_group *bfqg) {} 1465 1466struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node) 1467{ 1468 struct bfq_group *bfqg; 1469 int i; 1470 1471 bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); 1472 if (!bfqg) 1473 return NULL; 1474 1475 for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) 1476 bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; 1477 1478 return bfqg; 1479} 1480#endif /* CONFIG_BFQ_GROUP_IOSCHED */ 1481