Linux Audio

Check our new training course

Open-source upstreaming

Need help get the support for your hardware in upstream Linux?
Loading...
v4.17
 
   1/*
   2 * cgroups support for the BFQ I/O scheduler.
   3 *
   4 *  This program is free software; you can redistribute it and/or
   5 *  modify it under the terms of the GNU General Public License as
   6 *  published by the Free Software Foundation; either version 2 of the
   7 *  License, or (at your option) any later version.
   8 *
   9 *  This program is distributed in the hope that it will be useful,
  10 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12 *  General Public License for more details.
  13 */
  14#include <linux/module.h>
  15#include <linux/slab.h>
  16#include <linux/blkdev.h>
  17#include <linux/cgroup.h>
  18#include <linux/elevator.h>
  19#include <linux/ktime.h>
  20#include <linux/rbtree.h>
  21#include <linux/ioprio.h>
  22#include <linux/sbitmap.h>
  23#include <linux/delay.h>
  24
  25#include "bfq-iosched.h"
  26
  27#if defined(CONFIG_BFQ_GROUP_IOSCHED) &&  defined(CONFIG_DEBUG_BLK_CGROUP)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  28
  29/* bfqg stats flags */
  30enum bfqg_stats_flags {
  31	BFQG_stats_waiting = 0,
  32	BFQG_stats_idling,
  33	BFQG_stats_empty,
  34};
  35
  36#define BFQG_FLAG_FNS(name)						\
  37static void bfqg_stats_mark_##name(struct bfqg_stats *stats)	\
  38{									\
  39	stats->flags |= (1 << BFQG_stats_##name);			\
  40}									\
  41static void bfqg_stats_clear_##name(struct bfqg_stats *stats)	\
  42{									\
  43	stats->flags &= ~(1 << BFQG_stats_##name);			\
  44}									\
  45static int bfqg_stats_##name(struct bfqg_stats *stats)		\
  46{									\
  47	return (stats->flags & (1 << BFQG_stats_##name)) != 0;		\
  48}									\
  49
  50BFQG_FLAG_FNS(waiting)
  51BFQG_FLAG_FNS(idling)
  52BFQG_FLAG_FNS(empty)
  53#undef BFQG_FLAG_FNS
  54
  55/* This should be called with the scheduler lock held. */
  56static void bfqg_stats_update_group_wait_time(struct bfqg_stats *stats)
  57{
  58	unsigned long long now;
  59
  60	if (!bfqg_stats_waiting(stats))
  61		return;
  62
  63	now = sched_clock();
  64	if (time_after64(now, stats->start_group_wait_time))
  65		blkg_stat_add(&stats->group_wait_time,
  66			      now - stats->start_group_wait_time);
  67	bfqg_stats_clear_waiting(stats);
  68}
  69
  70/* This should be called with the scheduler lock held. */
  71static void bfqg_stats_set_start_group_wait_time(struct bfq_group *bfqg,
  72						 struct bfq_group *curr_bfqg)
  73{
  74	struct bfqg_stats *stats = &bfqg->stats;
  75
  76	if (bfqg_stats_waiting(stats))
  77		return;
  78	if (bfqg == curr_bfqg)
  79		return;
  80	stats->start_group_wait_time = sched_clock();
  81	bfqg_stats_mark_waiting(stats);
  82}
  83
  84/* This should be called with the scheduler lock held. */
  85static void bfqg_stats_end_empty_time(struct bfqg_stats *stats)
  86{
  87	unsigned long long now;
  88
  89	if (!bfqg_stats_empty(stats))
  90		return;
  91
  92	now = sched_clock();
  93	if (time_after64(now, stats->start_empty_time))
  94		blkg_stat_add(&stats->empty_time,
  95			      now - stats->start_empty_time);
  96	bfqg_stats_clear_empty(stats);
  97}
  98
  99void bfqg_stats_update_dequeue(struct bfq_group *bfqg)
 100{
 101	blkg_stat_add(&bfqg->stats.dequeue, 1);
 102}
 103
 104void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg)
 105{
 106	struct bfqg_stats *stats = &bfqg->stats;
 107
 108	if (blkg_rwstat_total(&stats->queued))
 109		return;
 110
 111	/*
 112	 * group is already marked empty. This can happen if bfqq got new
 113	 * request in parent group and moved to this group while being added
 114	 * to service tree. Just ignore the event and move on.
 115	 */
 116	if (bfqg_stats_empty(stats))
 117		return;
 118
 119	stats->start_empty_time = sched_clock();
 120	bfqg_stats_mark_empty(stats);
 121}
 122
 123void bfqg_stats_update_idle_time(struct bfq_group *bfqg)
 124{
 125	struct bfqg_stats *stats = &bfqg->stats;
 126
 127	if (bfqg_stats_idling(stats)) {
 128		unsigned long long now = sched_clock();
 129
 130		if (time_after64(now, stats->start_idle_time))
 131			blkg_stat_add(&stats->idle_time,
 132				      now - stats->start_idle_time);
 133		bfqg_stats_clear_idling(stats);
 134	}
 135}
 136
 137void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg)
 138{
 139	struct bfqg_stats *stats = &bfqg->stats;
 140
 141	stats->start_idle_time = sched_clock();
 142	bfqg_stats_mark_idling(stats);
 143}
 144
 145void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg)
 146{
 147	struct bfqg_stats *stats = &bfqg->stats;
 148
 149	blkg_stat_add(&stats->avg_queue_size_sum,
 150		      blkg_rwstat_total(&stats->queued));
 151	blkg_stat_add(&stats->avg_queue_size_samples, 1);
 152	bfqg_stats_update_group_wait_time(stats);
 153}
 154
 155void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
 156			      unsigned int op)
 157{
 158	blkg_rwstat_add(&bfqg->stats.queued, op, 1);
 159	bfqg_stats_end_empty_time(&bfqg->stats);
 160	if (!(bfqq == ((struct bfq_data *)bfqg->bfqd)->in_service_queue))
 161		bfqg_stats_set_start_group_wait_time(bfqg, bfqq_group(bfqq));
 162}
 163
 164void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op)
 165{
 166	blkg_rwstat_add(&bfqg->stats.queued, op, -1);
 167}
 168
 169void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op)
 170{
 171	blkg_rwstat_add(&bfqg->stats.merged, op, 1);
 172}
 173
 174void bfqg_stats_update_completion(struct bfq_group *bfqg, uint64_t start_time,
 175				  uint64_t io_start_time, unsigned int op)
 176{
 177	struct bfqg_stats *stats = &bfqg->stats;
 178	unsigned long long now = sched_clock();
 179
 180	if (time_after64(now, io_start_time))
 181		blkg_rwstat_add(&stats->service_time, op,
 182				now - io_start_time);
 183	if (time_after64(io_start_time, start_time))
 184		blkg_rwstat_add(&stats->wait_time, op,
 185				io_start_time - start_time);
 186}
 187
 188#else /* CONFIG_BFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 189
 190void bfqg_stats_update_io_add(struct bfq_group *bfqg, struct bfq_queue *bfqq,
 191			      unsigned int op) { }
 192void bfqg_stats_update_io_remove(struct bfq_group *bfqg, unsigned int op) { }
 193void bfqg_stats_update_io_merged(struct bfq_group *bfqg, unsigned int op) { }
 194void bfqg_stats_update_completion(struct bfq_group *bfqg, uint64_t start_time,
 195				  uint64_t io_start_time, unsigned int op) { }
 196void bfqg_stats_update_dequeue(struct bfq_group *bfqg) { }
 197void bfqg_stats_set_start_empty_time(struct bfq_group *bfqg) { }
 198void bfqg_stats_update_idle_time(struct bfq_group *bfqg) { }
 199void bfqg_stats_set_start_idle_time(struct bfq_group *bfqg) { }
 200void bfqg_stats_update_avg_queue_size(struct bfq_group *bfqg) { }
 201
 202#endif /* CONFIG_BFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
 203
 204#ifdef CONFIG_BFQ_GROUP_IOSCHED
 205
 206/*
 207 * blk-cgroup policy-related handlers
 208 * The following functions help in converting between blk-cgroup
 209 * internal structures and BFQ-specific structures.
 210 */
 211
 212static struct bfq_group *pd_to_bfqg(struct blkg_policy_data *pd)
 213{
 214	return pd ? container_of(pd, struct bfq_group, pd) : NULL;
 215}
 216
 217struct blkcg_gq *bfqg_to_blkg(struct bfq_group *bfqg)
 218{
 219	return pd_to_blkg(&bfqg->pd);
 220}
 221
 222static struct bfq_group *blkg_to_bfqg(struct blkcg_gq *blkg)
 223{
 224	return pd_to_bfqg(blkg_to_pd(blkg, &blkcg_policy_bfq));
 225}
 226
 227/*
 228 * bfq_group handlers
 229 * The following functions help in navigating the bfq_group hierarchy
 230 * by allowing to find the parent of a bfq_group or the bfq_group
 231 * associated to a bfq_queue.
 232 */
 233
 234static struct bfq_group *bfqg_parent(struct bfq_group *bfqg)
 235{
 236	struct blkcg_gq *pblkg = bfqg_to_blkg(bfqg)->parent;
 237
 238	return pblkg ? blkg_to_bfqg(pblkg) : NULL;
 239}
 240
 241struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
 242{
 243	struct bfq_entity *group_entity = bfqq->entity.parent;
 244
 245	return group_entity ? container_of(group_entity, struct bfq_group,
 246					   entity) :
 247			      bfqq->bfqd->root_group;
 248}
 249
 250/*
 251 * The following two functions handle get and put of a bfq_group by
 252 * wrapping the related blk-cgroup hooks.
 253 */
 254
 255static void bfqg_get(struct bfq_group *bfqg)
 256{
 257	bfqg->ref++;
 258}
 259
 260static void bfqg_put(struct bfq_group *bfqg)
 261{
 262	bfqg->ref--;
 263
 264	if (bfqg->ref == 0)
 265		kfree(bfqg);
 266}
 267
 268static void bfqg_and_blkg_get(struct bfq_group *bfqg)
 269{
 270	/* see comments in bfq_bic_update_cgroup for why refcounting bfqg */
 271	bfqg_get(bfqg);
 272
 273	blkg_get(bfqg_to_blkg(bfqg));
 274}
 275
 276void bfqg_and_blkg_put(struct bfq_group *bfqg)
 277{
 
 
 278	bfqg_put(bfqg);
 
 279
 280	blkg_put(bfqg_to_blkg(bfqg));
 
 
 
 
 
 
 
 
 281}
 282
 283/* @stats = 0 */
 284static void bfqg_stats_reset(struct bfqg_stats *stats)
 285{
 286#ifdef CONFIG_DEBUG_BLK_CGROUP
 287	/* queued stats shouldn't be cleared */
 288	blkg_rwstat_reset(&stats->merged);
 289	blkg_rwstat_reset(&stats->service_time);
 290	blkg_rwstat_reset(&stats->wait_time);
 291	blkg_stat_reset(&stats->time);
 292	blkg_stat_reset(&stats->avg_queue_size_sum);
 293	blkg_stat_reset(&stats->avg_queue_size_samples);
 294	blkg_stat_reset(&stats->dequeue);
 295	blkg_stat_reset(&stats->group_wait_time);
 296	blkg_stat_reset(&stats->idle_time);
 297	blkg_stat_reset(&stats->empty_time);
 298#endif
 299}
 300
 301/* @to += @from */
 302static void bfqg_stats_add_aux(struct bfqg_stats *to, struct bfqg_stats *from)
 303{
 304	if (!to || !from)
 305		return;
 306
 307#ifdef CONFIG_DEBUG_BLK_CGROUP
 308	/* queued stats shouldn't be cleared */
 309	blkg_rwstat_add_aux(&to->merged, &from->merged);
 310	blkg_rwstat_add_aux(&to->service_time, &from->service_time);
 311	blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
 312	blkg_stat_add_aux(&from->time, &from->time);
 313	blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
 314	blkg_stat_add_aux(&to->avg_queue_size_samples,
 315			  &from->avg_queue_size_samples);
 316	blkg_stat_add_aux(&to->dequeue, &from->dequeue);
 317	blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
 318	blkg_stat_add_aux(&to->idle_time, &from->idle_time);
 319	blkg_stat_add_aux(&to->empty_time, &from->empty_time);
 320#endif
 321}
 322
 323/*
 324 * Transfer @bfqg's stats to its parent's aux counts so that the ancestors'
 325 * recursive stats can still account for the amount used by this bfqg after
 326 * it's gone.
 327 */
 328static void bfqg_stats_xfer_dead(struct bfq_group *bfqg)
 329{
 330	struct bfq_group *parent;
 331
 332	if (!bfqg) /* root_group */
 333		return;
 334
 335	parent = bfqg_parent(bfqg);
 336
 337	lockdep_assert_held(bfqg_to_blkg(bfqg)->q->queue_lock);
 338
 339	if (unlikely(!parent))
 340		return;
 341
 342	bfqg_stats_add_aux(&parent->stats, &bfqg->stats);
 343	bfqg_stats_reset(&bfqg->stats);
 344}
 345
 346void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
 347{
 348	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
 349
 350	entity->weight = entity->new_weight;
 351	entity->orig_weight = entity->new_weight;
 352	if (bfqq) {
 353		bfqq->ioprio = bfqq->new_ioprio;
 354		bfqq->ioprio_class = bfqq->new_ioprio_class;
 355		/*
 356		 * Make sure that bfqg and its associated blkg do not
 357		 * disappear before entity.
 358		 */
 359		bfqg_and_blkg_get(bfqg);
 360	}
 361	entity->parent = bfqg->my_entity; /* NULL for root group */
 362	entity->sched_data = &bfqg->sched_data;
 363}
 364
 365static void bfqg_stats_exit(struct bfqg_stats *stats)
 366{
 367#ifdef CONFIG_DEBUG_BLK_CGROUP
 
 
 368	blkg_rwstat_exit(&stats->merged);
 369	blkg_rwstat_exit(&stats->service_time);
 370	blkg_rwstat_exit(&stats->wait_time);
 371	blkg_rwstat_exit(&stats->queued);
 372	blkg_stat_exit(&stats->time);
 373	blkg_stat_exit(&stats->avg_queue_size_sum);
 374	blkg_stat_exit(&stats->avg_queue_size_samples);
 375	blkg_stat_exit(&stats->dequeue);
 376	blkg_stat_exit(&stats->group_wait_time);
 377	blkg_stat_exit(&stats->idle_time);
 378	blkg_stat_exit(&stats->empty_time);
 379#endif
 380}
 381
 382static int bfqg_stats_init(struct bfqg_stats *stats, gfp_t gfp)
 383{
 384#ifdef CONFIG_DEBUG_BLK_CGROUP
 
 
 
 
 385	if (blkg_rwstat_init(&stats->merged, gfp) ||
 386	    blkg_rwstat_init(&stats->service_time, gfp) ||
 387	    blkg_rwstat_init(&stats->wait_time, gfp) ||
 388	    blkg_rwstat_init(&stats->queued, gfp) ||
 389	    blkg_stat_init(&stats->time, gfp) ||
 390	    blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
 391	    blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
 392	    blkg_stat_init(&stats->dequeue, gfp) ||
 393	    blkg_stat_init(&stats->group_wait_time, gfp) ||
 394	    blkg_stat_init(&stats->idle_time, gfp) ||
 395	    blkg_stat_init(&stats->empty_time, gfp)) {
 396		bfqg_stats_exit(stats);
 397		return -ENOMEM;
 398	}
 399#endif
 400
 401	return 0;
 402}
 403
 404static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
 405{
 406	return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
 407}
 408
 409static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
 410{
 411	return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
 412}
 413
 414static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
 415{
 416	struct bfq_group_data *bgd;
 417
 418	bgd = kzalloc(sizeof(*bgd), gfp);
 419	if (!bgd)
 420		return NULL;
 421	return &bgd->pd;
 422}
 423
 424static void bfq_cpd_init(struct blkcg_policy_data *cpd)
 425{
 426	struct bfq_group_data *d = cpd_to_bfqgd(cpd);
 427
 428	d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
 429		CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
 430}
 431
 432static void bfq_cpd_free(struct blkcg_policy_data *cpd)
 433{
 434	kfree(cpd_to_bfqgd(cpd));
 435}
 436
 437static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, int node)
 
 438{
 439	struct bfq_group *bfqg;
 440
 441	bfqg = kzalloc_node(sizeof(*bfqg), gfp, node);
 442	if (!bfqg)
 443		return NULL;
 444
 445	if (bfqg_stats_init(&bfqg->stats, gfp)) {
 446		kfree(bfqg);
 447		return NULL;
 448	}
 449
 450	/* see comments in bfq_bic_update_cgroup for why refcounting */
 451	bfqg_get(bfqg);
 452	return &bfqg->pd;
 453}
 454
 455static void bfq_pd_init(struct blkg_policy_data *pd)
 456{
 457	struct blkcg_gq *blkg = pd_to_blkg(pd);
 458	struct bfq_group *bfqg = blkg_to_bfqg(blkg);
 459	struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
 460	struct bfq_entity *entity = &bfqg->entity;
 461	struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
 462
 463	entity->orig_weight = entity->weight = entity->new_weight = d->weight;
 464	entity->my_sched_data = &bfqg->sched_data;
 465	bfqg->my_entity = entity; /*
 466				   * the root_group's will be set to NULL
 467				   * in bfq_init_queue()
 468				   */
 469	bfqg->bfqd = bfqd;
 470	bfqg->active_entities = 0;
 471	bfqg->rq_pos_tree = RB_ROOT;
 472}
 473
 474static void bfq_pd_free(struct blkg_policy_data *pd)
 475{
 476	struct bfq_group *bfqg = pd_to_bfqg(pd);
 477
 478	bfqg_stats_exit(&bfqg->stats);
 479	bfqg_put(bfqg);
 480}
 481
 482static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
 483{
 484	struct bfq_group *bfqg = pd_to_bfqg(pd);
 485
 486	bfqg_stats_reset(&bfqg->stats);
 487}
 488
 489static void bfq_group_set_parent(struct bfq_group *bfqg,
 490					struct bfq_group *parent)
 491{
 492	struct bfq_entity *entity;
 493
 494	entity = &bfqg->entity;
 495	entity->parent = parent->my_entity;
 496	entity->sched_data = &parent->sched_data;
 497}
 498
 499static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
 500					 struct blkcg *blkcg)
 501{
 502	struct blkcg_gq *blkg;
 503
 504	blkg = blkg_lookup(blkcg, bfqd->queue);
 505	if (likely(blkg))
 506		return blkg_to_bfqg(blkg);
 507	return NULL;
 508}
 509
 510struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
 511				     struct blkcg *blkcg)
 512{
 513	struct bfq_group *bfqg, *parent;
 514	struct bfq_entity *entity;
 515
 516	bfqg = bfq_lookup_bfqg(bfqd, blkcg);
 517
 518	if (unlikely(!bfqg))
 519		return NULL;
 520
 521	/*
 522	 * Update chain of bfq_groups as we might be handling a leaf group
 523	 * which, along with some of its relatives, has not been hooked yet
 524	 * to the private hierarchy of BFQ.
 525	 */
 526	entity = &bfqg->entity;
 527	for_each_entity(entity) {
 528		bfqg = container_of(entity, struct bfq_group, entity);
 529		if (bfqg != bfqd->root_group) {
 530			parent = bfqg_parent(bfqg);
 
 531			if (!parent)
 532				parent = bfqd->root_group;
 533			bfq_group_set_parent(bfqg, parent);
 534		}
 535	}
 536
 537	return bfqg;
 538}
 539
 540/**
 541 * bfq_bfqq_move - migrate @bfqq to @bfqg.
 542 * @bfqd: queue descriptor.
 543 * @bfqq: the queue to move.
 544 * @bfqg: the group to move to.
 545 *
 546 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
 547 * it on the new one.  Avoid putting the entity on the old group idle tree.
 548 *
 549 * Must be called under the scheduler lock, to make sure that the blkg
 550 * owning @bfqg does not disappear (see comments in
 551 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
 552 * objects).
 553 */
 554void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 555		   struct bfq_group *bfqg)
 556{
 557	struct bfq_entity *entity = &bfqq->entity;
 558
 
 
 
 
 
 
 559	/* If bfqq is empty, then bfq_bfqq_expire also invokes
 560	 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
 561	 * from data structures related to current group. Otherwise we
 562	 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
 563	 * we do below.
 564	 */
 565	if (bfqq == bfqd->in_service_queue)
 566		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
 567				false, BFQQE_PREEMPTED);
 568
 569	if (bfq_bfqq_busy(bfqq))
 570		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
 571	else if (entity->on_st)
 572		bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
 573	bfqg_and_blkg_put(bfqq_group(bfqq));
 574
 575	entity->parent = bfqg->my_entity;
 576	entity->sched_data = &bfqg->sched_data;
 577	/* pin down bfqg and its associated blkg  */
 578	bfqg_and_blkg_get(bfqg);
 579
 580	if (bfq_bfqq_busy(bfqq)) {
 581		bfq_pos_tree_add_move(bfqd, bfqq);
 
 582		bfq_activate_bfqq(bfqd, bfqq);
 583	}
 584
 585	if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
 586		bfq_schedule_dispatch(bfqd);
 
 
 587}
 588
 589/**
 590 * __bfq_bic_change_cgroup - move @bic to @cgroup.
 591 * @bfqd: the queue descriptor.
 592 * @bic: the bic to move.
 593 * @blkcg: the blk-cgroup to move to.
 594 *
 595 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
 596 * sure that the reference to cgroup is valid across the call (see
 597 * comments in bfq_bic_update_cgroup on this issue)
 598 *
 599 * NOTE: an alternative approach might have been to store the current
 600 * cgroup in bfqq and getting a reference to it, reducing the lookup
 601 * time here, at the price of slightly more complex code.
 602 */
 603static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
 604						struct bfq_io_cq *bic,
 605						struct blkcg *blkcg)
 606{
 607	struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
 608	struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
 609	struct bfq_group *bfqg;
 610	struct bfq_entity *entity;
 611
 612	bfqg = bfq_find_set_group(bfqd, blkcg);
 613
 614	if (unlikely(!bfqg))
 615		bfqg = bfqd->root_group;
 616
 617	if (async_bfqq) {
 618		entity = &async_bfqq->entity;
 619
 620		if (entity->sched_data != &bfqg->sched_data) {
 621			bic_set_bfqq(bic, NULL, 0);
 622			bfq_log_bfqq(bfqd, async_bfqq,
 623				     "bic_change_group: %p %d",
 624				     async_bfqq, async_bfqq->ref);
 625			bfq_put_queue(async_bfqq);
 626		}
 627	}
 628
 629	if (sync_bfqq) {
 630		entity = &sync_bfqq->entity;
 631		if (entity->sched_data != &bfqg->sched_data)
 632			bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
 633	}
 634
 635	return bfqg;
 636}
 637
 638void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
 639{
 640	struct bfq_data *bfqd = bic_to_bfqd(bic);
 641	struct bfq_group *bfqg = NULL;
 642	uint64_t serial_nr;
 643
 644	rcu_read_lock();
 645	serial_nr = bio_blkcg(bio)->css.serial_nr;
 646
 647	/*
 648	 * Check whether blkcg has changed.  The condition may trigger
 649	 * spuriously on a newly created cic but there's no harm.
 650	 */
 651	if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
 652		goto out;
 653
 654	bfqg = __bfq_bic_change_cgroup(bfqd, bic, bio_blkcg(bio));
 655	/*
 656	 * Update blkg_path for bfq_log_* functions. We cache this
 657	 * path, and update it here, for the following
 658	 * reasons. Operations on blkg objects in blk-cgroup are
 659	 * protected with the request_queue lock, and not with the
 660	 * lock that protects the instances of this scheduler
 661	 * (bfqd->lock). This exposes BFQ to the following sort of
 662	 * race.
 663	 *
 664	 * The blkg_lookup performed in bfq_get_queue, protected
 665	 * through rcu, may happen to return the address of a copy of
 666	 * the original blkg. If this is the case, then the
 667	 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
 668	 * the blkg, is useless: it does not prevent blk-cgroup code
 669	 * from destroying both the original blkg and all objects
 670	 * directly or indirectly referred by the copy of the
 671	 * blkg.
 672	 *
 673	 * On the bright side, destroy operations on a blkg invoke, as
 674	 * a first step, hooks of the scheduler associated with the
 675	 * blkg. And these hooks are executed with bfqd->lock held for
 676	 * BFQ. As a consequence, for any blkg associated with the
 677	 * request queue this instance of the scheduler is attached
 678	 * to, we are guaranteed that such a blkg is not destroyed, and
 679	 * that all the pointers it contains are consistent, while we
 680	 * are holding bfqd->lock. A blkg_lookup performed with
 681	 * bfqd->lock held then returns a fully consistent blkg, which
 682	 * remains consistent until this lock is held.
 683	 *
 684	 * Thanks to the last fact, and to the fact that: (1) bfqg has
 685	 * been obtained through a blkg_lookup in the above
 686	 * assignment, and (2) bfqd->lock is being held, here we can
 687	 * safely use the policy data for the involved blkg (i.e., the
 688	 * field bfqg->pd) to get to the blkg associated with bfqg,
 689	 * and then we can safely use any field of blkg. After we
 690	 * release bfqd->lock, even just getting blkg through this
 691	 * bfqg may cause dangling references to be traversed, as
 692	 * bfqg->pd may not exist any more.
 693	 *
 694	 * In view of the above facts, here we cache, in the bfqg, any
 695	 * blkg data we may need for this bic, and for its associated
 696	 * bfq_queue. As of now, we need to cache only the path of the
 697	 * blkg, which is used in the bfq_log_* functions.
 698	 *
 699	 * Finally, note that bfqg itself needs to be protected from
 700	 * destruction on the blkg_free of the original blkg (which
 701	 * invokes bfq_pd_free). We use an additional private
 702	 * refcounter for bfqg, to let it disappear only after no
 703	 * bfq_queue refers to it any longer.
 704	 */
 705	blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
 706	bic->blkcg_serial_nr = serial_nr;
 707out:
 708	rcu_read_unlock();
 709}
 710
 711/**
 712 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
 713 * @st: the service tree being flushed.
 714 */
 715static void bfq_flush_idle_tree(struct bfq_service_tree *st)
 716{
 717	struct bfq_entity *entity = st->first_idle;
 718
 719	for (; entity ; entity = st->first_idle)
 720		__bfq_deactivate_entity(entity, false);
 721}
 722
 723/**
 724 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
 725 * @bfqd: the device data structure with the root group.
 726 * @entity: the entity to move.
 
 727 */
 728static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
 729				     struct bfq_entity *entity)
 
 730{
 731	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
 
 
 
 
 
 
 
 
 
 732
 
 
 
 
 
 733	bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
 734}
 735
 736/**
 737 * bfq_reparent_active_entities - move to the root group all active
 738 *                                entities.
 739 * @bfqd: the device data structure with the root group.
 740 * @bfqg: the group to move from.
 741 * @st: the service tree with the entities.
 742 */
 743static void bfq_reparent_active_entities(struct bfq_data *bfqd,
 744					 struct bfq_group *bfqg,
 745					 struct bfq_service_tree *st)
 
 746{
 747	struct rb_root *active = &st->active;
 748	struct bfq_entity *entity = NULL;
 749
 750	if (!RB_EMPTY_ROOT(&st->active))
 751		entity = bfq_entity_of(rb_first(active));
 752
 753	for (; entity ; entity = bfq_entity_of(rb_first(active)))
 754		bfq_reparent_leaf_entity(bfqd, entity);
 755
 756	if (bfqg->sched_data.in_service_entity)
 757		bfq_reparent_leaf_entity(bfqd,
 758			bfqg->sched_data.in_service_entity);
 
 759}
 760
 761/**
 762 * bfq_pd_offline - deactivate the entity associated with @pd,
 763 *		    and reparent its children entities.
 764 * @pd: descriptor of the policy going offline.
 765 *
 766 * blkio already grabs the queue_lock for us, so no need to use
 767 * RCU-based magic
 768 */
 769static void bfq_pd_offline(struct blkg_policy_data *pd)
 770{
 771	struct bfq_service_tree *st;
 772	struct bfq_group *bfqg = pd_to_bfqg(pd);
 773	struct bfq_data *bfqd = bfqg->bfqd;
 774	struct bfq_entity *entity = bfqg->my_entity;
 775	unsigned long flags;
 776	int i;
 777
 778	spin_lock_irqsave(&bfqd->lock, flags);
 779
 780	if (!entity) /* root group */
 781		goto put_async_queues;
 782
 783	/*
 784	 * Empty all service_trees belonging to this group before
 785	 * deactivating the group itself.
 786	 */
 787	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
 788		st = bfqg->sched_data.service_tree + i;
 789
 790		/*
 791		 * The idle tree may still contain bfq_queues belonging
 792		 * to exited task because they never migrated to a different
 793		 * cgroup from the one being destroyed now.
 794		 */
 795		bfq_flush_idle_tree(st);
 796
 797		/*
 798		 * It may happen that some queues are still active
 799		 * (busy) upon group destruction (if the corresponding
 800		 * processes have been forced to terminate). We move
 801		 * all the leaf entities corresponding to these queues
 802		 * to the root_group.
 803		 * Also, it may happen that the group has an entity
 804		 * in service, which is disconnected from the active
 805		 * tree: it must be moved, too.
 806		 * There is no need to put the sync queues, as the
 807		 * scheduler has taken no reference.
 808		 */
 809		bfq_reparent_active_entities(bfqd, bfqg, st);
 
 
 
 
 
 
 
 
 
 
 
 
 
 810	}
 811
 812	__bfq_deactivate_entity(entity, false);
 813
 814put_async_queues:
 815	bfq_put_async_queues(bfqd, bfqg);
 816
 817	spin_unlock_irqrestore(&bfqd->lock, flags);
 818	/*
 819	 * @blkg is going offline and will be ignored by
 820	 * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
 821	 * that they don't get lost.  If IOs complete after this point, the
 822	 * stats for them will be lost.  Oh well...
 823	 */
 824	bfqg_stats_xfer_dead(bfqg);
 825}
 826
 827void bfq_end_wr_async(struct bfq_data *bfqd)
 828{
 829	struct blkcg_gq *blkg;
 830
 831	list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
 832		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
 833
 834		bfq_end_wr_async_queues(bfqd, bfqg);
 835	}
 836	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
 837}
 838
 839static int bfq_io_show_weight(struct seq_file *sf, void *v)
 840{
 841	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
 842	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
 843	unsigned int val = 0;
 844
 845	if (bfqgd)
 846		val = bfqgd->weight;
 847
 848	seq_printf(sf, "%u\n", val);
 849
 850	return 0;
 851}
 852
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 853static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
 854				    struct cftype *cftype,
 855				    u64 val)
 856{
 857	struct blkcg *blkcg = css_to_blkcg(css);
 858	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
 859	struct blkcg_gq *blkg;
 860	int ret = -ERANGE;
 861
 862	if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
 863		return ret;
 864
 865	ret = 0;
 866	spin_lock_irq(&blkcg->lock);
 867	bfqgd->weight = (unsigned short)val;
 868	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
 869		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
 870
 871		if (!bfqg)
 872			continue;
 873		/*
 874		 * Setting the prio_changed flag of the entity
 875		 * to 1 with new_weight == weight would re-set
 876		 * the value of the weight to its ioprio mapping.
 877		 * Set the flag only if necessary.
 878		 */
 879		if ((unsigned short)val != bfqg->entity.new_weight) {
 880			bfqg->entity.new_weight = (unsigned short)val;
 881			/*
 882			 * Make sure that the above new value has been
 883			 * stored in bfqg->entity.new_weight before
 884			 * setting the prio_changed flag. In fact,
 885			 * this flag may be read asynchronously (in
 886			 * critical sections protected by a different
 887			 * lock than that held here), and finding this
 888			 * flag set may cause the execution of the code
 889			 * for updating parameters whose value may
 890			 * depend also on bfqg->entity.new_weight (in
 891			 * __bfq_entity_update_weight_prio).
 892			 * This barrier makes sure that the new value
 893			 * of bfqg->entity.new_weight is correctly
 894			 * seen in that code.
 895			 */
 896			smp_wmb();
 897			bfqg->entity.prio_changed = 1;
 898		}
 899	}
 900	spin_unlock_irq(&blkcg->lock);
 901
 902	return ret;
 903}
 904
 905static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
 906				 char *buf, size_t nbytes,
 907				 loff_t off)
 908{
 909	u64 weight;
 910	/* First unsigned long found in the file is used */
 911	int ret = kstrtoull(strim(buf), 0, &weight);
 
 
 912
 
 913	if (ret)
 914		return ret;
 915
 916	return bfq_io_set_weight_legacy(of_css(of), NULL, weight);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 917}
 918
 919#ifdef CONFIG_DEBUG_BLK_CGROUP
 920static int bfqg_print_stat(struct seq_file *sf, void *v)
 
 921{
 922	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
 923			  &blkcg_policy_bfq, seq_cft(sf)->private, false);
 924	return 0;
 
 
 
 
 
 
 
 
 
 
 
 925}
 926
 927static int bfqg_print_rwstat(struct seq_file *sf, void *v)
 928{
 929	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
 930			  &blkcg_policy_bfq, seq_cft(sf)->private, true);
 931	return 0;
 932}
 933
 934static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
 935				      struct blkg_policy_data *pd, int off)
 936{
 937	u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
 938					  &blkcg_policy_bfq, off);
 939	return __blkg_prfill_u64(sf, pd, sum);
 940}
 941
 942static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
 943					struct blkg_policy_data *pd, int off)
 944{
 945	struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
 946							   &blkcg_policy_bfq,
 947							   off);
 948	return __blkg_prfill_rwstat(sf, pd, &sum);
 949}
 950
 951static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
 952{
 953	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 954			  bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
 955			  seq_cft(sf)->private, false);
 956	return 0;
 957}
 958
 959static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 960{
 961	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 962			  bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
 963			  seq_cft(sf)->private, true);
 964	return 0;
 965}
 966
 967static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
 968			       int off)
 969{
 970	u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
 
 971
 972	return __blkg_prfill_u64(sf, pd, sum >> 9);
 973}
 974
 975static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
 976{
 977	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 978			  bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
 979	return 0;
 980}
 981
 982static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
 983					 struct blkg_policy_data *pd, int off)
 984{
 985	struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
 986					offsetof(struct blkcg_gq, stat_bytes));
 987	u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
 988		atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
 989
 990	return __blkg_prfill_u64(sf, pd, sum >> 9);
 
 
 
 
 991}
 992
 993static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
 994{
 995	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
 996			  bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
 997			  false);
 998	return 0;
 999}
1000
1001static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1002				      struct blkg_policy_data *pd, int off)
1003{
1004	struct bfq_group *bfqg = pd_to_bfqg(pd);
1005	u64 samples = blkg_stat_read(&bfqg->stats.avg_queue_size_samples);
1006	u64 v = 0;
1007
1008	if (samples) {
1009		v = blkg_stat_read(&bfqg->stats.avg_queue_size_sum);
1010		v = div64_u64(v, samples);
1011	}
1012	__blkg_prfill_u64(sf, pd, v);
1013	return 0;
1014}
1015
1016/* print avg_queue_size */
1017static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1018{
1019	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1020			  bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1021			  0, false);
1022	return 0;
1023}
1024#endif /* CONFIG_DEBUG_BLK_CGROUP */
1025
1026struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1027{
1028	int ret;
1029
1030	ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1031	if (ret)
1032		return NULL;
1033
1034	return blkg_to_bfqg(bfqd->queue->root_blkg);
1035}
1036
1037struct blkcg_policy blkcg_policy_bfq = {
1038	.dfl_cftypes		= bfq_blkg_files,
1039	.legacy_cftypes		= bfq_blkcg_legacy_files,
1040
1041	.cpd_alloc_fn		= bfq_cpd_alloc,
1042	.cpd_init_fn		= bfq_cpd_init,
1043	.cpd_bind_fn	        = bfq_cpd_init,
1044	.cpd_free_fn		= bfq_cpd_free,
1045
1046	.pd_alloc_fn		= bfq_pd_alloc,
1047	.pd_init_fn		= bfq_pd_init,
1048	.pd_offline_fn		= bfq_pd_offline,
1049	.pd_free_fn		= bfq_pd_free,
1050	.pd_reset_stats_fn	= bfq_pd_reset_stats,
1051};
1052
1053struct cftype bfq_blkcg_legacy_files[] = {
1054	{
1055		.name = "bfq.weight",
1056		.flags = CFTYPE_NOT_ON_ROOT,
1057		.seq_show = bfq_io_show_weight,
1058		.write_u64 = bfq_io_set_weight_legacy,
1059	},
 
 
 
 
 
 
1060
1061	/* statistics, covers only the tasks in the bfqg */
1062	{
1063		.name = "bfq.io_service_bytes",
1064		.private = (unsigned long)&blkcg_policy_bfq,
1065		.seq_show = blkg_print_stat_bytes,
1066	},
1067	{
1068		.name = "bfq.io_serviced",
1069		.private = (unsigned long)&blkcg_policy_bfq,
1070		.seq_show = blkg_print_stat_ios,
1071	},
1072#ifdef CONFIG_DEBUG_BLK_CGROUP
1073	{
1074		.name = "bfq.time",
1075		.private = offsetof(struct bfq_group, stats.time),
1076		.seq_show = bfqg_print_stat,
1077	},
1078	{
1079		.name = "bfq.sectors",
1080		.seq_show = bfqg_print_stat_sectors,
1081	},
1082	{
1083		.name = "bfq.io_service_time",
1084		.private = offsetof(struct bfq_group, stats.service_time),
1085		.seq_show = bfqg_print_rwstat,
1086	},
1087	{
1088		.name = "bfq.io_wait_time",
1089		.private = offsetof(struct bfq_group, stats.wait_time),
1090		.seq_show = bfqg_print_rwstat,
1091	},
1092	{
1093		.name = "bfq.io_merged",
1094		.private = offsetof(struct bfq_group, stats.merged),
1095		.seq_show = bfqg_print_rwstat,
1096	},
1097	{
1098		.name = "bfq.io_queued",
1099		.private = offsetof(struct bfq_group, stats.queued),
1100		.seq_show = bfqg_print_rwstat,
1101	},
1102#endif /* CONFIG_DEBUG_BLK_CGROUP */
1103
1104	/* the same statictics which cover the bfqg and its descendants */
1105	{
1106		.name = "bfq.io_service_bytes_recursive",
1107		.private = (unsigned long)&blkcg_policy_bfq,
1108		.seq_show = blkg_print_stat_bytes_recursive,
1109	},
1110	{
1111		.name = "bfq.io_serviced_recursive",
1112		.private = (unsigned long)&blkcg_policy_bfq,
1113		.seq_show = blkg_print_stat_ios_recursive,
1114	},
1115#ifdef CONFIG_DEBUG_BLK_CGROUP
1116	{
1117		.name = "bfq.time_recursive",
1118		.private = offsetof(struct bfq_group, stats.time),
1119		.seq_show = bfqg_print_stat_recursive,
1120	},
1121	{
1122		.name = "bfq.sectors_recursive",
1123		.seq_show = bfqg_print_stat_sectors_recursive,
1124	},
1125	{
1126		.name = "bfq.io_service_time_recursive",
1127		.private = offsetof(struct bfq_group, stats.service_time),
1128		.seq_show = bfqg_print_rwstat_recursive,
1129	},
1130	{
1131		.name = "bfq.io_wait_time_recursive",
1132		.private = offsetof(struct bfq_group, stats.wait_time),
1133		.seq_show = bfqg_print_rwstat_recursive,
1134	},
1135	{
1136		.name = "bfq.io_merged_recursive",
1137		.private = offsetof(struct bfq_group, stats.merged),
1138		.seq_show = bfqg_print_rwstat_recursive,
1139	},
1140	{
1141		.name = "bfq.io_queued_recursive",
1142		.private = offsetof(struct bfq_group, stats.queued),
1143		.seq_show = bfqg_print_rwstat_recursive,
1144	},
1145	{
1146		.name = "bfq.avg_queue_size",
1147		.seq_show = bfqg_print_avg_queue_size,
1148	},
1149	{
1150		.name = "bfq.group_wait_time",
1151		.private = offsetof(struct bfq_group, stats.group_wait_time),
1152		.seq_show = bfqg_print_stat,
1153	},
1154	{
1155		.name = "bfq.idle_time",
1156		.private = offsetof(struct bfq_group, stats.idle_time),
1157		.seq_show = bfqg_print_stat,
1158	},
1159	{
1160		.name = "bfq.empty_time",
1161		.private = offsetof(struct bfq_group, stats.empty_time),
1162		.seq_show = bfqg_print_stat,
1163	},
1164	{
1165		.name = "bfq.dequeue",
1166		.private = offsetof(struct bfq_group, stats.dequeue),
1167		.seq_show = bfqg_print_stat,
1168	},
1169#endif	/* CONFIG_DEBUG_BLK_CGROUP */
1170	{ }	/* terminate */
1171};
1172
1173struct cftype bfq_blkg_files[] = {
1174	{
1175		.name = "bfq.weight",
1176		.flags = CFTYPE_NOT_ON_ROOT,
1177		.seq_show = bfq_io_show_weight,
1178		.write = bfq_io_set_weight,
1179	},
1180	{} /* terminate */
1181};
1182
1183#else	/* CONFIG_BFQ_GROUP_IOSCHED */
1184
1185void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1186		   struct bfq_group *bfqg) {}
1187
1188void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1189{
1190	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1191
1192	entity->weight = entity->new_weight;
1193	entity->orig_weight = entity->new_weight;
1194	if (bfqq) {
1195		bfqq->ioprio = bfqq->new_ioprio;
1196		bfqq->ioprio_class = bfqq->new_ioprio_class;
1197	}
1198	entity->sched_data = &bfqg->sched_data;
1199}
1200
1201void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1202
1203void bfq_end_wr_async(struct bfq_data *bfqd)
1204{
1205	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1206}
1207
1208struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1209{
1210	return bfqd->root_group;
1211}
1212
1213struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1214{
1215	return bfqq->bfqd->root_group;
1216}
 
 
 
 
1217
1218struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1219{
1220	struct bfq_group *bfqg;
1221	int i;
1222
1223	bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1224	if (!bfqg)
1225		return NULL;
1226
1227	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1228		bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1229
1230	return bfqg;
1231}
1232#endif	/* CONFIG_BFQ_GROUP_IOSCHED */
v5.9
   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		return -ENOMEM;
 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		bfqg_stats_exit(stats);
 481		return -ENOMEM;
 482	}
 483#endif
 484
 485	return 0;
 486}
 487
 488static struct bfq_group_data *cpd_to_bfqgd(struct blkcg_policy_data *cpd)
 489{
 490	return cpd ? container_of(cpd, struct bfq_group_data, pd) : NULL;
 491}
 492
 493static struct bfq_group_data *blkcg_to_bfqgd(struct blkcg *blkcg)
 494{
 495	return cpd_to_bfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_bfq));
 496}
 497
 498static struct blkcg_policy_data *bfq_cpd_alloc(gfp_t gfp)
 499{
 500	struct bfq_group_data *bgd;
 501
 502	bgd = kzalloc(sizeof(*bgd), gfp);
 503	if (!bgd)
 504		return NULL;
 505	return &bgd->pd;
 506}
 507
 508static void bfq_cpd_init(struct blkcg_policy_data *cpd)
 509{
 510	struct bfq_group_data *d = cpd_to_bfqgd(cpd);
 511
 512	d->weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
 513		CGROUP_WEIGHT_DFL : BFQ_WEIGHT_LEGACY_DFL;
 514}
 515
 516static void bfq_cpd_free(struct blkcg_policy_data *cpd)
 517{
 518	kfree(cpd_to_bfqgd(cpd));
 519}
 520
 521static struct blkg_policy_data *bfq_pd_alloc(gfp_t gfp, struct request_queue *q,
 522					     struct blkcg *blkcg)
 523{
 524	struct bfq_group *bfqg;
 525
 526	bfqg = kzalloc_node(sizeof(*bfqg), gfp, q->node);
 527	if (!bfqg)
 528		return NULL;
 529
 530	if (bfqg_stats_init(&bfqg->stats, gfp)) {
 531		kfree(bfqg);
 532		return NULL;
 533	}
 534
 535	/* see comments in bfq_bic_update_cgroup for why refcounting */
 536	bfqg_get(bfqg);
 537	return &bfqg->pd;
 538}
 539
 540static void bfq_pd_init(struct blkg_policy_data *pd)
 541{
 542	struct blkcg_gq *blkg = pd_to_blkg(pd);
 543	struct bfq_group *bfqg = blkg_to_bfqg(blkg);
 544	struct bfq_data *bfqd = blkg->q->elevator->elevator_data;
 545	struct bfq_entity *entity = &bfqg->entity;
 546	struct bfq_group_data *d = blkcg_to_bfqgd(blkg->blkcg);
 547
 548	entity->orig_weight = entity->weight = entity->new_weight = d->weight;
 549	entity->my_sched_data = &bfqg->sched_data;
 550	bfqg->my_entity = entity; /*
 551				   * the root_group's will be set to NULL
 552				   * in bfq_init_queue()
 553				   */
 554	bfqg->bfqd = bfqd;
 555	bfqg->active_entities = 0;
 556	bfqg->rq_pos_tree = RB_ROOT;
 557}
 558
 559static void bfq_pd_free(struct blkg_policy_data *pd)
 560{
 561	struct bfq_group *bfqg = pd_to_bfqg(pd);
 562
 563	bfqg_stats_exit(&bfqg->stats);
 564	bfqg_put(bfqg);
 565}
 566
 567static void bfq_pd_reset_stats(struct blkg_policy_data *pd)
 568{
 569	struct bfq_group *bfqg = pd_to_bfqg(pd);
 570
 571	bfqg_stats_reset(&bfqg->stats);
 572}
 573
 574static void bfq_group_set_parent(struct bfq_group *bfqg,
 575					struct bfq_group *parent)
 576{
 577	struct bfq_entity *entity;
 578
 579	entity = &bfqg->entity;
 580	entity->parent = parent->my_entity;
 581	entity->sched_data = &parent->sched_data;
 582}
 583
 584static struct bfq_group *bfq_lookup_bfqg(struct bfq_data *bfqd,
 585					 struct blkcg *blkcg)
 586{
 587	struct blkcg_gq *blkg;
 588
 589	blkg = blkg_lookup(blkcg, bfqd->queue);
 590	if (likely(blkg))
 591		return blkg_to_bfqg(blkg);
 592	return NULL;
 593}
 594
 595struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd,
 596				     struct blkcg *blkcg)
 597{
 598	struct bfq_group *bfqg, *parent;
 599	struct bfq_entity *entity;
 600
 601	bfqg = bfq_lookup_bfqg(bfqd, blkcg);
 602
 603	if (unlikely(!bfqg))
 604		return NULL;
 605
 606	/*
 607	 * Update chain of bfq_groups as we might be handling a leaf group
 608	 * which, along with some of its relatives, has not been hooked yet
 609	 * to the private hierarchy of BFQ.
 610	 */
 611	entity = &bfqg->entity;
 612	for_each_entity(entity) {
 613		struct bfq_group *curr_bfqg = container_of(entity,
 614						struct bfq_group, entity);
 615		if (curr_bfqg != bfqd->root_group) {
 616			parent = bfqg_parent(curr_bfqg);
 617			if (!parent)
 618				parent = bfqd->root_group;
 619			bfq_group_set_parent(curr_bfqg, parent);
 620		}
 621	}
 622
 623	return bfqg;
 624}
 625
 626/**
 627 * bfq_bfqq_move - migrate @bfqq to @bfqg.
 628 * @bfqd: queue descriptor.
 629 * @bfqq: the queue to move.
 630 * @bfqg: the group to move to.
 631 *
 632 * Move @bfqq to @bfqg, deactivating it from its old group and reactivating
 633 * it on the new one.  Avoid putting the entity on the old group idle tree.
 634 *
 635 * Must be called under the scheduler lock, to make sure that the blkg
 636 * owning @bfqg does not disappear (see comments in
 637 * bfq_bic_update_cgroup on guaranteeing the consistency of blkg
 638 * objects).
 639 */
 640void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
 641		   struct bfq_group *bfqg)
 642{
 643	struct bfq_entity *entity = &bfqq->entity;
 644
 645	/*
 646	 * Get extra reference to prevent bfqq from being freed in
 647	 * next possible expire or deactivate.
 648	 */
 649	bfqq->ref++;
 650
 651	/* If bfqq is empty, then bfq_bfqq_expire also invokes
 652	 * bfq_del_bfqq_busy, thereby removing bfqq and its entity
 653	 * from data structures related to current group. Otherwise we
 654	 * need to remove bfqq explicitly with bfq_deactivate_bfqq, as
 655	 * we do below.
 656	 */
 657	if (bfqq == bfqd->in_service_queue)
 658		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
 659				false, BFQQE_PREEMPTED);
 660
 661	if (bfq_bfqq_busy(bfqq))
 662		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
 663	else if (entity->on_st_or_in_serv)
 664		bfq_put_idle_entity(bfq_entity_service_tree(entity), entity);
 665	bfqg_and_blkg_put(bfqq_group(bfqq));
 666
 667	entity->parent = bfqg->my_entity;
 668	entity->sched_data = &bfqg->sched_data;
 669	/* pin down bfqg and its associated blkg  */
 670	bfqg_and_blkg_get(bfqg);
 671
 672	if (bfq_bfqq_busy(bfqq)) {
 673		if (unlikely(!bfqd->nonrot_with_queueing))
 674			bfq_pos_tree_add_move(bfqd, bfqq);
 675		bfq_activate_bfqq(bfqd, bfqq);
 676	}
 677
 678	if (!bfqd->in_service_queue && !bfqd->rq_in_driver)
 679		bfq_schedule_dispatch(bfqd);
 680	/* release extra ref taken above, bfqq may happen to be freed now */
 681	bfq_put_queue(bfqq);
 682}
 683
 684/**
 685 * __bfq_bic_change_cgroup - move @bic to @cgroup.
 686 * @bfqd: the queue descriptor.
 687 * @bic: the bic to move.
 688 * @blkcg: the blk-cgroup to move to.
 689 *
 690 * Move bic to blkcg, assuming that bfqd->lock is held; which makes
 691 * sure that the reference to cgroup is valid across the call (see
 692 * comments in bfq_bic_update_cgroup on this issue)
 693 *
 694 * NOTE: an alternative approach might have been to store the current
 695 * cgroup in bfqq and getting a reference to it, reducing the lookup
 696 * time here, at the price of slightly more complex code.
 697 */
 698static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd,
 699						struct bfq_io_cq *bic,
 700						struct blkcg *blkcg)
 701{
 702	struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0);
 703	struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1);
 704	struct bfq_group *bfqg;
 705	struct bfq_entity *entity;
 706
 707	bfqg = bfq_find_set_group(bfqd, blkcg);
 708
 709	if (unlikely(!bfqg))
 710		bfqg = bfqd->root_group;
 711
 712	if (async_bfqq) {
 713		entity = &async_bfqq->entity;
 714
 715		if (entity->sched_data != &bfqg->sched_data) {
 716			bic_set_bfqq(bic, NULL, 0);
 717			bfq_release_process_ref(bfqd, async_bfqq);
 
 
 
 718		}
 719	}
 720
 721	if (sync_bfqq) {
 722		entity = &sync_bfqq->entity;
 723		if (entity->sched_data != &bfqg->sched_data)
 724			bfq_bfqq_move(bfqd, sync_bfqq, bfqg);
 725	}
 726
 727	return bfqg;
 728}
 729
 730void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio)
 731{
 732	struct bfq_data *bfqd = bic_to_bfqd(bic);
 733	struct bfq_group *bfqg = NULL;
 734	uint64_t serial_nr;
 735
 736	rcu_read_lock();
 737	serial_nr = __bio_blkcg(bio)->css.serial_nr;
 738
 739	/*
 740	 * Check whether blkcg has changed.  The condition may trigger
 741	 * spuriously on a newly created cic but there's no harm.
 742	 */
 743	if (unlikely(!bfqd) || likely(bic->blkcg_serial_nr == serial_nr))
 744		goto out;
 745
 746	bfqg = __bfq_bic_change_cgroup(bfqd, bic, __bio_blkcg(bio));
 747	/*
 748	 * Update blkg_path for bfq_log_* functions. We cache this
 749	 * path, and update it here, for the following
 750	 * reasons. Operations on blkg objects in blk-cgroup are
 751	 * protected with the request_queue lock, and not with the
 752	 * lock that protects the instances of this scheduler
 753	 * (bfqd->lock). This exposes BFQ to the following sort of
 754	 * race.
 755	 *
 756	 * The blkg_lookup performed in bfq_get_queue, protected
 757	 * through rcu, may happen to return the address of a copy of
 758	 * the original blkg. If this is the case, then the
 759	 * bfqg_and_blkg_get performed in bfq_get_queue, to pin down
 760	 * the blkg, is useless: it does not prevent blk-cgroup code
 761	 * from destroying both the original blkg and all objects
 762	 * directly or indirectly referred by the copy of the
 763	 * blkg.
 764	 *
 765	 * On the bright side, destroy operations on a blkg invoke, as
 766	 * a first step, hooks of the scheduler associated with the
 767	 * blkg. And these hooks are executed with bfqd->lock held for
 768	 * BFQ. As a consequence, for any blkg associated with the
 769	 * request queue this instance of the scheduler is attached
 770	 * to, we are guaranteed that such a blkg is not destroyed, and
 771	 * that all the pointers it contains are consistent, while we
 772	 * are holding bfqd->lock. A blkg_lookup performed with
 773	 * bfqd->lock held then returns a fully consistent blkg, which
 774	 * remains consistent until this lock is held.
 775	 *
 776	 * Thanks to the last fact, and to the fact that: (1) bfqg has
 777	 * been obtained through a blkg_lookup in the above
 778	 * assignment, and (2) bfqd->lock is being held, here we can
 779	 * safely use the policy data for the involved blkg (i.e., the
 780	 * field bfqg->pd) to get to the blkg associated with bfqg,
 781	 * and then we can safely use any field of blkg. After we
 782	 * release bfqd->lock, even just getting blkg through this
 783	 * bfqg may cause dangling references to be traversed, as
 784	 * bfqg->pd may not exist any more.
 785	 *
 786	 * In view of the above facts, here we cache, in the bfqg, any
 787	 * blkg data we may need for this bic, and for its associated
 788	 * bfq_queue. As of now, we need to cache only the path of the
 789	 * blkg, which is used in the bfq_log_* functions.
 790	 *
 791	 * Finally, note that bfqg itself needs to be protected from
 792	 * destruction on the blkg_free of the original blkg (which
 793	 * invokes bfq_pd_free). We use an additional private
 794	 * refcounter for bfqg, to let it disappear only after no
 795	 * bfq_queue refers to it any longer.
 796	 */
 797	blkg_path(bfqg_to_blkg(bfqg), bfqg->blkg_path, sizeof(bfqg->blkg_path));
 798	bic->blkcg_serial_nr = serial_nr;
 799out:
 800	rcu_read_unlock();
 801}
 802
 803/**
 804 * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st.
 805 * @st: the service tree being flushed.
 806 */
 807static void bfq_flush_idle_tree(struct bfq_service_tree *st)
 808{
 809	struct bfq_entity *entity = st->first_idle;
 810
 811	for (; entity ; entity = st->first_idle)
 812		__bfq_deactivate_entity(entity, false);
 813}
 814
 815/**
 816 * bfq_reparent_leaf_entity - move leaf entity to the root_group.
 817 * @bfqd: the device data structure with the root group.
 818 * @entity: the entity to move, if entity is a leaf; or the parent entity
 819 *	    of an active leaf entity to move, if entity is not a leaf.
 820 */
 821static void bfq_reparent_leaf_entity(struct bfq_data *bfqd,
 822				     struct bfq_entity *entity,
 823				     int ioprio_class)
 824{
 825	struct bfq_queue *bfqq;
 826	struct bfq_entity *child_entity = entity;
 827
 828	while (child_entity->my_sched_data) { /* leaf not reached yet */
 829		struct bfq_sched_data *child_sd = child_entity->my_sched_data;
 830		struct bfq_service_tree *child_st = child_sd->service_tree +
 831			ioprio_class;
 832		struct rb_root *child_active = &child_st->active;
 833
 834		child_entity = bfq_entity_of(rb_first(child_active));
 835
 836		if (!child_entity)
 837			child_entity = child_sd->in_service_entity;
 838	}
 839
 840	bfqq = bfq_entity_to_bfqq(child_entity);
 841	bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
 842}
 843
 844/**
 845 * bfq_reparent_active_queues - move to the root group all active queues.
 
 846 * @bfqd: the device data structure with the root group.
 847 * @bfqg: the group to move from.
 848 * @st: the service tree to start the search from.
 849 */
 850static void bfq_reparent_active_queues(struct bfq_data *bfqd,
 851				       struct bfq_group *bfqg,
 852				       struct bfq_service_tree *st,
 853				       int ioprio_class)
 854{
 855	struct rb_root *active = &st->active;
 856	struct bfq_entity *entity;
 
 
 
 857
 858	while ((entity = bfq_entity_of(rb_first(active))))
 859		bfq_reparent_leaf_entity(bfqd, entity, ioprio_class);
 860
 861	if (bfqg->sched_data.in_service_entity)
 862		bfq_reparent_leaf_entity(bfqd,
 863					 bfqg->sched_data.in_service_entity,
 864					 ioprio_class);
 865}
 866
 867/**
 868 * bfq_pd_offline - deactivate the entity associated with @pd,
 869 *		    and reparent its children entities.
 870 * @pd: descriptor of the policy going offline.
 871 *
 872 * blkio already grabs the queue_lock for us, so no need to use
 873 * RCU-based magic
 874 */
 875static void bfq_pd_offline(struct blkg_policy_data *pd)
 876{
 877	struct bfq_service_tree *st;
 878	struct bfq_group *bfqg = pd_to_bfqg(pd);
 879	struct bfq_data *bfqd = bfqg->bfqd;
 880	struct bfq_entity *entity = bfqg->my_entity;
 881	unsigned long flags;
 882	int i;
 883
 884	spin_lock_irqsave(&bfqd->lock, flags);
 885
 886	if (!entity) /* root group */
 887		goto put_async_queues;
 888
 889	/*
 890	 * Empty all service_trees belonging to this group before
 891	 * deactivating the group itself.
 892	 */
 893	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) {
 894		st = bfqg->sched_data.service_tree + i;
 895
 896		/*
 
 
 
 
 
 
 
 897		 * It may happen that some queues are still active
 898		 * (busy) upon group destruction (if the corresponding
 899		 * processes have been forced to terminate). We move
 900		 * all the leaf entities corresponding to these queues
 901		 * to the root_group.
 902		 * Also, it may happen that the group has an entity
 903		 * in service, which is disconnected from the active
 904		 * tree: it must be moved, too.
 905		 * There is no need to put the sync queues, as the
 906		 * scheduler has taken no reference.
 907		 */
 908		bfq_reparent_active_queues(bfqd, bfqg, st, i);
 909
 910		/*
 911		 * The idle tree may still contain bfq_queues
 912		 * belonging to exited task because they never
 913		 * migrated to a different cgroup from the one being
 914		 * destroyed now. In addition, even
 915		 * bfq_reparent_active_queues() may happen to add some
 916		 * entities to the idle tree. It happens if, in some
 917		 * of the calls to bfq_bfqq_move() performed by
 918		 * bfq_reparent_active_queues(), the queue to move is
 919		 * empty and gets expired.
 920		 */
 921		bfq_flush_idle_tree(st);
 922	}
 923
 924	__bfq_deactivate_entity(entity, false);
 925
 926put_async_queues:
 927	bfq_put_async_queues(bfqd, bfqg);
 928
 929	spin_unlock_irqrestore(&bfqd->lock, flags);
 930	/*
 931	 * @blkg is going offline and will be ignored by
 932	 * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
 933	 * that they don't get lost.  If IOs complete after this point, the
 934	 * stats for them will be lost.  Oh well...
 935	 */
 936	bfqg_stats_xfer_dead(bfqg);
 937}
 938
 939void bfq_end_wr_async(struct bfq_data *bfqd)
 940{
 941	struct blkcg_gq *blkg;
 942
 943	list_for_each_entry(blkg, &bfqd->queue->blkg_list, q_node) {
 944		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
 945
 946		bfq_end_wr_async_queues(bfqd, bfqg);
 947	}
 948	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
 949}
 950
 951static int bfq_io_show_weight_legacy(struct seq_file *sf, void *v)
 952{
 953	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
 954	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
 955	unsigned int val = 0;
 956
 957	if (bfqgd)
 958		val = bfqgd->weight;
 959
 960	seq_printf(sf, "%u\n", val);
 961
 962	return 0;
 963}
 964
 965static u64 bfqg_prfill_weight_device(struct seq_file *sf,
 966				     struct blkg_policy_data *pd, int off)
 967{
 968	struct bfq_group *bfqg = pd_to_bfqg(pd);
 969
 970	if (!bfqg->entity.dev_weight)
 971		return 0;
 972	return __blkg_prfill_u64(sf, pd, bfqg->entity.dev_weight);
 973}
 974
 975static int bfq_io_show_weight(struct seq_file *sf, void *v)
 976{
 977	struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
 978	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
 979
 980	seq_printf(sf, "default %u\n", bfqgd->weight);
 981	blkcg_print_blkgs(sf, blkcg, bfqg_prfill_weight_device,
 982			  &blkcg_policy_bfq, 0, false);
 983	return 0;
 984}
 985
 986static void bfq_group_set_weight(struct bfq_group *bfqg, u64 weight, u64 dev_weight)
 987{
 988	weight = dev_weight ?: weight;
 989
 990	bfqg->entity.dev_weight = dev_weight;
 991	/*
 992	 * Setting the prio_changed flag of the entity
 993	 * to 1 with new_weight == weight would re-set
 994	 * the value of the weight to its ioprio mapping.
 995	 * Set the flag only if necessary.
 996	 */
 997	if ((unsigned short)weight != bfqg->entity.new_weight) {
 998		bfqg->entity.new_weight = (unsigned short)weight;
 999		/*
1000		 * Make sure that the above new value has been
1001		 * stored in bfqg->entity.new_weight before
1002		 * setting the prio_changed flag. In fact,
1003		 * this flag may be read asynchronously (in
1004		 * critical sections protected by a different
1005		 * lock than that held here), and finding this
1006		 * flag set may cause the execution of the code
1007		 * for updating parameters whose value may
1008		 * depend also on bfqg->entity.new_weight (in
1009		 * __bfq_entity_update_weight_prio).
1010		 * This barrier makes sure that the new value
1011		 * of bfqg->entity.new_weight is correctly
1012		 * seen in that code.
1013		 */
1014		smp_wmb();
1015		bfqg->entity.prio_changed = 1;
1016	}
1017}
1018
1019static int bfq_io_set_weight_legacy(struct cgroup_subsys_state *css,
1020				    struct cftype *cftype,
1021				    u64 val)
1022{
1023	struct blkcg *blkcg = css_to_blkcg(css);
1024	struct bfq_group_data *bfqgd = blkcg_to_bfqgd(blkcg);
1025	struct blkcg_gq *blkg;
1026	int ret = -ERANGE;
1027
1028	if (val < BFQ_MIN_WEIGHT || val > BFQ_MAX_WEIGHT)
1029		return ret;
1030
1031	ret = 0;
1032	spin_lock_irq(&blkcg->lock);
1033	bfqgd->weight = (unsigned short)val;
1034	hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1035		struct bfq_group *bfqg = blkg_to_bfqg(blkg);
1036
1037		if (bfqg)
1038			bfq_group_set_weight(bfqg, val, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1039	}
1040	spin_unlock_irq(&blkcg->lock);
1041
1042	return ret;
1043}
1044
1045static ssize_t bfq_io_set_device_weight(struct kernfs_open_file *of,
1046					char *buf, size_t nbytes,
1047					loff_t off)
1048{
1049	int ret;
1050	struct blkg_conf_ctx ctx;
1051	struct blkcg *blkcg = css_to_blkcg(of_css(of));
1052	struct bfq_group *bfqg;
1053	u64 v;
1054
1055	ret = blkg_conf_prep(blkcg, &blkcg_policy_bfq, buf, &ctx);
1056	if (ret)
1057		return ret;
1058
1059	if (sscanf(ctx.body, "%llu", &v) == 1) {
1060		/* require "default" on dfl */
1061		ret = -ERANGE;
1062		if (!v)
1063			goto out;
1064	} else if (!strcmp(strim(ctx.body), "default")) {
1065		v = 0;
1066	} else {
1067		ret = -EINVAL;
1068		goto out;
1069	}
1070
1071	bfqg = blkg_to_bfqg(ctx.blkg);
1072
1073	ret = -ERANGE;
1074	if (!v || (v >= BFQ_MIN_WEIGHT && v <= BFQ_MAX_WEIGHT)) {
1075		bfq_group_set_weight(bfqg, bfqg->entity.weight, v);
1076		ret = 0;
1077	}
1078out:
1079	blkg_conf_finish(&ctx);
1080	return ret ?: nbytes;
1081}
1082
1083static ssize_t bfq_io_set_weight(struct kernfs_open_file *of,
1084				 char *buf, size_t nbytes,
1085				 loff_t off)
1086{
1087	char *endp;
1088	int ret;
1089	u64 v;
1090
1091	buf = strim(buf);
1092
1093	/* "WEIGHT" or "default WEIGHT" sets the default weight */
1094	v = simple_strtoull(buf, &endp, 0);
1095	if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
1096		ret = bfq_io_set_weight_legacy(of_css(of), NULL, v);
1097		return ret ?: nbytes;
1098	}
1099
1100	return bfq_io_set_device_weight(of, buf, nbytes, off);
1101}
1102
1103static int bfqg_print_rwstat(struct seq_file *sf, void *v)
1104{
1105	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1106			  &blkcg_policy_bfq, seq_cft(sf)->private, true);
1107	return 0;
1108}
1109
 
 
 
 
 
 
 
 
1110static u64 bfqg_prfill_rwstat_recursive(struct seq_file *sf,
1111					struct blkg_policy_data *pd, int off)
1112{
1113	struct blkg_rwstat_sample sum;
1114
1115	blkg_rwstat_recursive_sum(pd_to_blkg(pd), &blkcg_policy_bfq, off, &sum);
1116	return __blkg_prfill_rwstat(sf, pd, &sum);
1117}
1118
1119static int bfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1120{
1121	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1122			  bfqg_prfill_rwstat_recursive, &blkcg_policy_bfq,
1123			  seq_cft(sf)->private, true);
1124	return 0;
1125}
1126
1127#ifdef CONFIG_BFQ_CGROUP_DEBUG
1128static int bfqg_print_stat(struct seq_file *sf, void *v)
1129{
1130	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1131			  &blkcg_policy_bfq, seq_cft(sf)->private, false);
1132	return 0;
1133}
1134
1135static u64 bfqg_prfill_stat_recursive(struct seq_file *sf,
1136				      struct blkg_policy_data *pd, int off)
1137{
1138	struct blkcg_gq *blkg = pd_to_blkg(pd);
1139	struct blkcg_gq *pos_blkg;
1140	struct cgroup_subsys_state *pos_css;
1141	u64 sum = 0;
1142
1143	lockdep_assert_held(&blkg->q->queue_lock);
1144
1145	rcu_read_lock();
1146	blkg_for_each_descendant_pre(pos_blkg, pos_css, blkg) {
1147		struct bfq_stat *stat;
1148
1149		if (!pos_blkg->online)
1150			continue;
1151
1152		stat = (void *)blkg_to_pd(pos_blkg, &blkcg_policy_bfq) + off;
1153		sum += bfq_stat_read(stat) + atomic64_read(&stat->aux_cnt);
1154	}
1155	rcu_read_unlock();
1156
1157	return __blkg_prfill_u64(sf, pd, sum);
1158}
1159
1160static int bfqg_print_stat_recursive(struct seq_file *sf, void *v)
1161{
1162	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1163			  bfqg_prfill_stat_recursive, &blkcg_policy_bfq,
1164			  seq_cft(sf)->private, false);
1165	return 0;
1166}
1167
1168static u64 bfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1169			       int off)
1170{
1171	struct bfq_group *bfqg = blkg_to_bfqg(pd->blkg);
1172	u64 sum = blkg_rwstat_total(&bfqg->stats.bytes);
1173
1174	return __blkg_prfill_u64(sf, pd, sum >> 9);
1175}
1176
1177static int bfqg_print_stat_sectors(struct seq_file *sf, void *v)
1178{
1179	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1180			  bfqg_prfill_sectors, &blkcg_policy_bfq, 0, false);
1181	return 0;
1182}
1183
1184static u64 bfqg_prfill_sectors_recursive(struct seq_file *sf,
1185					 struct blkg_policy_data *pd, int off)
1186{
1187	struct blkg_rwstat_sample tmp;
 
 
 
1188
1189	blkg_rwstat_recursive_sum(pd->blkg, &blkcg_policy_bfq,
1190			offsetof(struct bfq_group, stats.bytes), &tmp);
1191
1192	return __blkg_prfill_u64(sf, pd,
1193		(tmp.cnt[BLKG_RWSTAT_READ] + tmp.cnt[BLKG_RWSTAT_WRITE]) >> 9);
1194}
1195
1196static int bfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1197{
1198	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1199			  bfqg_prfill_sectors_recursive, &blkcg_policy_bfq, 0,
1200			  false);
1201	return 0;
1202}
1203
1204static u64 bfqg_prfill_avg_queue_size(struct seq_file *sf,
1205				      struct blkg_policy_data *pd, int off)
1206{
1207	struct bfq_group *bfqg = pd_to_bfqg(pd);
1208	u64 samples = bfq_stat_read(&bfqg->stats.avg_queue_size_samples);
1209	u64 v = 0;
1210
1211	if (samples) {
1212		v = bfq_stat_read(&bfqg->stats.avg_queue_size_sum);
1213		v = div64_u64(v, samples);
1214	}
1215	__blkg_prfill_u64(sf, pd, v);
1216	return 0;
1217}
1218
1219/* print avg_queue_size */
1220static int bfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1221{
1222	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1223			  bfqg_prfill_avg_queue_size, &blkcg_policy_bfq,
1224			  0, false);
1225	return 0;
1226}
1227#endif /* CONFIG_BFQ_CGROUP_DEBUG */
1228
1229struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1230{
1231	int ret;
1232
1233	ret = blkcg_activate_policy(bfqd->queue, &blkcg_policy_bfq);
1234	if (ret)
1235		return NULL;
1236
1237	return blkg_to_bfqg(bfqd->queue->root_blkg);
1238}
1239
1240struct blkcg_policy blkcg_policy_bfq = {
1241	.dfl_cftypes		= bfq_blkg_files,
1242	.legacy_cftypes		= bfq_blkcg_legacy_files,
1243
1244	.cpd_alloc_fn		= bfq_cpd_alloc,
1245	.cpd_init_fn		= bfq_cpd_init,
1246	.cpd_bind_fn	        = bfq_cpd_init,
1247	.cpd_free_fn		= bfq_cpd_free,
1248
1249	.pd_alloc_fn		= bfq_pd_alloc,
1250	.pd_init_fn		= bfq_pd_init,
1251	.pd_offline_fn		= bfq_pd_offline,
1252	.pd_free_fn		= bfq_pd_free,
1253	.pd_reset_stats_fn	= bfq_pd_reset_stats,
1254};
1255
1256struct cftype bfq_blkcg_legacy_files[] = {
1257	{
1258		.name = "bfq.weight",
1259		.flags = CFTYPE_NOT_ON_ROOT,
1260		.seq_show = bfq_io_show_weight_legacy,
1261		.write_u64 = bfq_io_set_weight_legacy,
1262	},
1263	{
1264		.name = "bfq.weight_device",
1265		.flags = CFTYPE_NOT_ON_ROOT,
1266		.seq_show = bfq_io_show_weight,
1267		.write = bfq_io_set_weight,
1268	},
1269
1270	/* statistics, covers only the tasks in the bfqg */
1271	{
1272		.name = "bfq.io_service_bytes",
1273		.private = offsetof(struct bfq_group, stats.bytes),
1274		.seq_show = bfqg_print_rwstat,
1275	},
1276	{
1277		.name = "bfq.io_serviced",
1278		.private = offsetof(struct bfq_group, stats.ios),
1279		.seq_show = bfqg_print_rwstat,
1280	},
1281#ifdef CONFIG_BFQ_CGROUP_DEBUG
1282	{
1283		.name = "bfq.time",
1284		.private = offsetof(struct bfq_group, stats.time),
1285		.seq_show = bfqg_print_stat,
1286	},
1287	{
1288		.name = "bfq.sectors",
1289		.seq_show = bfqg_print_stat_sectors,
1290	},
1291	{
1292		.name = "bfq.io_service_time",
1293		.private = offsetof(struct bfq_group, stats.service_time),
1294		.seq_show = bfqg_print_rwstat,
1295	},
1296	{
1297		.name = "bfq.io_wait_time",
1298		.private = offsetof(struct bfq_group, stats.wait_time),
1299		.seq_show = bfqg_print_rwstat,
1300	},
1301	{
1302		.name = "bfq.io_merged",
1303		.private = offsetof(struct bfq_group, stats.merged),
1304		.seq_show = bfqg_print_rwstat,
1305	},
1306	{
1307		.name = "bfq.io_queued",
1308		.private = offsetof(struct bfq_group, stats.queued),
1309		.seq_show = bfqg_print_rwstat,
1310	},
1311#endif /* CONFIG_BFQ_CGROUP_DEBUG */
1312
1313	/* the same statistics which cover the bfqg and its descendants */
1314	{
1315		.name = "bfq.io_service_bytes_recursive",
1316		.private = offsetof(struct bfq_group, stats.bytes),
1317		.seq_show = bfqg_print_rwstat_recursive,
1318	},
1319	{
1320		.name = "bfq.io_serviced_recursive",
1321		.private = offsetof(struct bfq_group, stats.ios),
1322		.seq_show = bfqg_print_rwstat_recursive,
1323	},
1324#ifdef CONFIG_BFQ_CGROUP_DEBUG
1325	{
1326		.name = "bfq.time_recursive",
1327		.private = offsetof(struct bfq_group, stats.time),
1328		.seq_show = bfqg_print_stat_recursive,
1329	},
1330	{
1331		.name = "bfq.sectors_recursive",
1332		.seq_show = bfqg_print_stat_sectors_recursive,
1333	},
1334	{
1335		.name = "bfq.io_service_time_recursive",
1336		.private = offsetof(struct bfq_group, stats.service_time),
1337		.seq_show = bfqg_print_rwstat_recursive,
1338	},
1339	{
1340		.name = "bfq.io_wait_time_recursive",
1341		.private = offsetof(struct bfq_group, stats.wait_time),
1342		.seq_show = bfqg_print_rwstat_recursive,
1343	},
1344	{
1345		.name = "bfq.io_merged_recursive",
1346		.private = offsetof(struct bfq_group, stats.merged),
1347		.seq_show = bfqg_print_rwstat_recursive,
1348	},
1349	{
1350		.name = "bfq.io_queued_recursive",
1351		.private = offsetof(struct bfq_group, stats.queued),
1352		.seq_show = bfqg_print_rwstat_recursive,
1353	},
1354	{
1355		.name = "bfq.avg_queue_size",
1356		.seq_show = bfqg_print_avg_queue_size,
1357	},
1358	{
1359		.name = "bfq.group_wait_time",
1360		.private = offsetof(struct bfq_group, stats.group_wait_time),
1361		.seq_show = bfqg_print_stat,
1362	},
1363	{
1364		.name = "bfq.idle_time",
1365		.private = offsetof(struct bfq_group, stats.idle_time),
1366		.seq_show = bfqg_print_stat,
1367	},
1368	{
1369		.name = "bfq.empty_time",
1370		.private = offsetof(struct bfq_group, stats.empty_time),
1371		.seq_show = bfqg_print_stat,
1372	},
1373	{
1374		.name = "bfq.dequeue",
1375		.private = offsetof(struct bfq_group, stats.dequeue),
1376		.seq_show = bfqg_print_stat,
1377	},
1378#endif	/* CONFIG_BFQ_CGROUP_DEBUG */
1379	{ }	/* terminate */
1380};
1381
1382struct cftype bfq_blkg_files[] = {
1383	{
1384		.name = "bfq.weight",
1385		.flags = CFTYPE_NOT_ON_ROOT,
1386		.seq_show = bfq_io_show_weight,
1387		.write = bfq_io_set_weight,
1388	},
1389	{} /* terminate */
1390};
1391
1392#else	/* CONFIG_BFQ_GROUP_IOSCHED */
1393
1394void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq,
1395		   struct bfq_group *bfqg) {}
1396
1397void bfq_init_entity(struct bfq_entity *entity, struct bfq_group *bfqg)
1398{
1399	struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity);
1400
1401	entity->weight = entity->new_weight;
1402	entity->orig_weight = entity->new_weight;
1403	if (bfqq) {
1404		bfqq->ioprio = bfqq->new_ioprio;
1405		bfqq->ioprio_class = bfqq->new_ioprio_class;
1406	}
1407	entity->sched_data = &bfqg->sched_data;
1408}
1409
1410void bfq_bic_update_cgroup(struct bfq_io_cq *bic, struct bio *bio) {}
1411
1412void bfq_end_wr_async(struct bfq_data *bfqd)
1413{
1414	bfq_end_wr_async_queues(bfqd, bfqd->root_group);
1415}
1416
1417struct bfq_group *bfq_find_set_group(struct bfq_data *bfqd, struct blkcg *blkcg)
1418{
1419	return bfqd->root_group;
1420}
1421
1422struct bfq_group *bfqq_group(struct bfq_queue *bfqq)
1423{
1424	return bfqq->bfqd->root_group;
1425}
1426
1427void bfqg_and_blkg_get(struct bfq_group *bfqg) {}
1428
1429void bfqg_and_blkg_put(struct bfq_group *bfqg) {}
1430
1431struct bfq_group *bfq_create_group_hierarchy(struct bfq_data *bfqd, int node)
1432{
1433	struct bfq_group *bfqg;
1434	int i;
1435
1436	bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node);
1437	if (!bfqg)
1438		return NULL;
1439
1440	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
1441		bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
1442
1443	return bfqg;
1444}
1445#endif	/* CONFIG_BFQ_GROUP_IOSCHED */