FILE
cgroup.c
File preview with syntax highlighting.
linux-6.19.14/kernel/cgroup/cgroup.c
1// SPDX-License-Identifier: GPL-2.0 2/* 3 * Generic process-grouping system. 4 * 5 * Based originally on the cpuset system, extracted by Paul Menage 6 * Copyright (C) 2006 Google, Inc 7 * 8 * Notifications support 9 * Copyright (C) 2009 Nokia Corporation 10 * Author: Kirill A. Shutemov 11 * 12 * Copyright notices from the original cpuset code: 13 * -------------------------------------------------- 14 * Copyright (C) 2003 BULL SA. 15 * Copyright (C) 2004-2006 Silicon Graphics, Inc. 16 * 17 * Portions derived from Patrick Mochel's sysfs code. 18 * sysfs is Copyright (c) 2001-3 Patrick Mochel 19 * 20 * 2003-10-10 Written by Simon Derr. 21 * 2003-10-22 Updates by Stephen Hemminger. 22 * 2004 May-July Rework by Paul Jackson. 23 * --------------------------------------------------- 24 */ 25 26#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 27 28#include "cgroup-internal.h" 29 30#include <linux/bpf-cgroup.h> 31#include <linux/cred.h> 32#include <linux/errno.h> 33#include <linux/init_task.h> 34#include <linux/kernel.h> 35#include <linux/magic.h> 36#include <linux/mutex.h> 37#include <linux/mount.h> 38#include <linux/pagemap.h> 39#include <linux/proc_fs.h> 40#include <linux/rcupdate.h> 41#include <linux/sched.h> 42#include <linux/sched/task.h> 43#include <linux/slab.h> 44#include <linux/spinlock.h> 45#include <linux/percpu-rwsem.h> 46#include <linux/string.h> 47#include <linux/hashtable.h> 48#include <linux/idr.h> 49#include <linux/kthread.h> 50#include <linux/atomic.h> 51#include <linux/cpuset.h> 52#include <linux/proc_ns.h> 53#include <linux/nsproxy.h> 54#include <linux/file.h> 55#include <linux/fs_parser.h> 56#include <linux/sched/cputime.h> 57#include <linux/sched/deadline.h> 58#include <linux/psi.h> 59#include <linux/nstree.h> 60#include <linux/irq_work.h> 61#include <net/sock.h> 62 63#define CREATE_TRACE_POINTS 64#include <trace/events/cgroup.h> 65 66#define CGROUP_FILE_NAME_MAX (MAX_CGROUP_TYPE_NAMELEN + \ 67 MAX_CFTYPE_NAME + 2) 68/* let's not notify more than 100 times per second */ 69#define CGROUP_FILE_NOTIFY_MIN_INTV DIV_ROUND_UP(HZ, 100) 70 71/* 72 * To avoid confusing the compiler (and generating warnings) with code 73 * that attempts to access what would be a 0-element array (i.e. sized 74 * to a potentially empty array when CGROUP_SUBSYS_COUNT == 0), this 75 * constant expression can be added. 76 */ 77#define CGROUP_HAS_SUBSYS_CONFIG (CGROUP_SUBSYS_COUNT > 0) 78 79/* 80 * cgroup_mutex is the master lock. Any modification to cgroup or its 81 * hierarchy must be performed while holding it. 82 * 83 * css_set_lock protects task->cgroups pointer, the list of css_set 84 * objects, and the chain of tasks off each css_set. 85 * 86 * These locks are exported if CONFIG_PROVE_RCU so that accessors in 87 * cgroup.h can use them for lockdep annotations. 88 */ 89DEFINE_MUTEX(cgroup_mutex); 90DEFINE_SPINLOCK(css_set_lock); 91 92#if (defined CONFIG_PROVE_RCU || defined CONFIG_LOCKDEP) 93EXPORT_SYMBOL_GPL(cgroup_mutex); 94EXPORT_SYMBOL_GPL(css_set_lock); 95#endif 96 97struct blocking_notifier_head cgroup_lifetime_notifier = 98 BLOCKING_NOTIFIER_INIT(cgroup_lifetime_notifier); 99 100DEFINE_SPINLOCK(trace_cgroup_path_lock); 101char trace_cgroup_path[TRACE_CGROUP_PATH_LEN]; 102static bool cgroup_debug __read_mostly; 103 104/* 105 * Protects cgroup_idr and css_idr so that IDs can be released without 106 * grabbing cgroup_mutex. 107 */ 108static DEFINE_SPINLOCK(cgroup_idr_lock); 109 110/* 111 * Protects cgroup_file->kn for !self csses. It synchronizes notifications 112 * against file removal/re-creation across css hiding. 113 */ 114static DEFINE_SPINLOCK(cgroup_file_kn_lock); 115 116DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem); 117 118#define cgroup_assert_mutex_or_rcu_locked() \ 119 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 120 !lockdep_is_held(&cgroup_mutex), \ 121 "cgroup_mutex or RCU read lock required"); 122 123/* 124 * cgroup destruction makes heavy use of work items and there can be a lot 125 * of concurrent destructions. Use a separate workqueue so that cgroup 126 * destruction work items don't end up filling up max_active of system_percpu_wq 127 * which may lead to deadlock. 128 * 129 * A cgroup destruction should enqueue work sequentially to: 130 * cgroup_offline_wq: use for css offline work 131 * cgroup_release_wq: use for css release work 132 * cgroup_free_wq: use for free work 133 * 134 * Rationale for using separate workqueues: 135 * The cgroup root free work may depend on completion of other css offline 136 * operations. If all tasks were enqueued to a single workqueue, this could 137 * create a deadlock scenario where: 138 * - Free work waits for other css offline work to complete. 139 * - But other css offline work is queued after free work in the same queue. 140 * 141 * Example deadlock scenario with single workqueue (cgroup_destroy_wq): 142 * 1. umount net_prio 143 * 2. net_prio root destruction enqueues work to cgroup_destroy_wq (CPUx) 144 * 3. perf_event CSS A offline enqueues work to same cgroup_destroy_wq (CPUx) 145 * 4. net_prio cgroup_destroy_root->cgroup_lock_and_drain_offline. 146 * 5. net_prio root destruction blocks waiting for perf_event CSS A offline, 147 * which can never complete as it's behind in the same queue and 148 * workqueue's max_active is 1. 149 */ 150static struct workqueue_struct *cgroup_offline_wq; 151static struct workqueue_struct *cgroup_release_wq; 152static struct workqueue_struct *cgroup_free_wq; 153 154/* generate an array of cgroup subsystem pointers */ 155#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys, 156struct cgroup_subsys *cgroup_subsys[] = { 157#include <linux/cgroup_subsys.h> 158}; 159#undef SUBSYS 160 161/* array of cgroup subsystem names */ 162#define SUBSYS(_x) [_x ## _cgrp_id] = #_x, 163static const char *cgroup_subsys_name[] = { 164#include <linux/cgroup_subsys.h> 165}; 166#undef SUBSYS 167 168/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */ 169#define SUBSYS(_x) \ 170 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key); \ 171 DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key); \ 172 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key); \ 173 EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key); 174#include <linux/cgroup_subsys.h> 175#undef SUBSYS 176 177#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key, 178static struct static_key_true *cgroup_subsys_enabled_key[] = { 179#include <linux/cgroup_subsys.h> 180}; 181#undef SUBSYS 182 183#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key, 184static struct static_key_true *cgroup_subsys_on_dfl_key[] = { 185#include <linux/cgroup_subsys.h> 186}; 187#undef SUBSYS 188 189static DEFINE_PER_CPU(struct css_rstat_cpu, root_rstat_cpu); 190static DEFINE_PER_CPU(struct cgroup_rstat_base_cpu, root_rstat_base_cpu); 191 192/* the default hierarchy */ 193struct cgroup_root cgrp_dfl_root = { 194 .cgrp.self.rstat_cpu = &root_rstat_cpu, 195 .cgrp.rstat_base_cpu = &root_rstat_base_cpu, 196}; 197EXPORT_SYMBOL_GPL(cgrp_dfl_root); 198 199/* 200 * The default hierarchy always exists but is hidden until mounted for the 201 * first time. This is for backward compatibility. 202 */ 203bool cgrp_dfl_visible; 204 205/* some controllers are not supported in the default hierarchy */ 206static u16 cgrp_dfl_inhibit_ss_mask; 207 208/* some controllers are implicitly enabled on the default hierarchy */ 209static u16 cgrp_dfl_implicit_ss_mask; 210 211/* some controllers can be threaded on the default hierarchy */ 212static u16 cgrp_dfl_threaded_ss_mask; 213 214/* The list of hierarchy roots */ 215LIST_HEAD(cgroup_roots); 216static int cgroup_root_count; 217 218/* hierarchy ID allocation and mapping, protected by cgroup_mutex */ 219static DEFINE_IDR(cgroup_hierarchy_idr); 220 221/* 222 * Assign a monotonically increasing serial number to csses. It guarantees 223 * cgroups with bigger numbers are newer than those with smaller numbers. 224 * Also, as csses are always appended to the parent's ->children list, it 225 * guarantees that sibling csses are always sorted in the ascending serial 226 * number order on the list. Protected by cgroup_mutex. 227 */ 228static u64 css_serial_nr_next = 1; 229 230/* 231 * These bitmasks identify subsystems with specific features to avoid 232 * having to do iterative checks repeatedly. 233 */ 234static u16 have_fork_callback __read_mostly; 235static u16 have_exit_callback __read_mostly; 236static u16 have_release_callback __read_mostly; 237static u16 have_canfork_callback __read_mostly; 238 239static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS); 240 241/* 242 * Write protected by cgroup_mutex and write-lock of cgroup_threadgroup_rwsem, 243 * read protected by either. 244 * 245 * Can only be turned on, but not turned off. 246 */ 247bool cgroup_enable_per_threadgroup_rwsem __read_mostly; 248 249/* cgroup namespace for init task */ 250struct cgroup_namespace init_cgroup_ns = { 251 .ns = NS_COMMON_INIT(init_cgroup_ns), 252 .user_ns = &init_user_ns, 253 .root_cset = &init_css_set, 254}; 255 256static struct file_system_type cgroup2_fs_type; 257static struct cftype cgroup_base_files[]; 258static struct cftype cgroup_psi_files[]; 259 260/* cgroup optional features */ 261enum cgroup_opt_features { 262#ifdef CONFIG_PSI 263 OPT_FEATURE_PRESSURE, 264#endif 265 OPT_FEATURE_COUNT 266}; 267 268static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = { 269#ifdef CONFIG_PSI 270 "pressure", 271#endif 272}; 273 274static u16 cgroup_feature_disable_mask __read_mostly; 275 276static int cgroup_apply_control(struct cgroup *cgrp); 277static void cgroup_finalize_control(struct cgroup *cgrp, int ret); 278static void css_task_iter_skip(struct css_task_iter *it, 279 struct task_struct *task); 280static int cgroup_destroy_locked(struct cgroup *cgrp); 281static struct cgroup_subsys_state *css_create(struct cgroup *cgrp, 282 struct cgroup_subsys *ss); 283static void css_release(struct percpu_ref *ref); 284static void kill_css(struct cgroup_subsys_state *css); 285static int cgroup_addrm_files(struct cgroup_subsys_state *css, 286 struct cgroup *cgrp, struct cftype cfts[], 287 bool is_add); 288static void cgroup_rt_init(void); 289 290#ifdef CONFIG_DEBUG_CGROUP_REF 291#define CGROUP_REF_FN_ATTRS noinline 292#define CGROUP_REF_EXPORT(fn) EXPORT_SYMBOL_GPL(fn); 293#include <linux/cgroup_refcnt.h> 294#endif 295 296/** 297 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID 298 * @ssid: subsys ID of interest 299 * 300 * cgroup_subsys_enabled() can only be used with literal subsys names which 301 * is fine for individual subsystems but unsuitable for cgroup core. This 302 * is slower static_key_enabled() based test indexed by @ssid. 303 */ 304bool cgroup_ssid_enabled(int ssid) 305{ 306 if (!CGROUP_HAS_SUBSYS_CONFIG) 307 return false; 308 309 return static_key_enabled(cgroup_subsys_enabled_key[ssid]); 310} 311 312/** 313 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy 314 * @cgrp: the cgroup of interest 315 * 316 * The default hierarchy is the v2 interface of cgroup and this function 317 * can be used to test whether a cgroup is on the default hierarchy for 318 * cases where a subsystem should behave differently depending on the 319 * interface version. 320 * 321 * List of changed behaviors: 322 * 323 * - Mount options "noprefix", "xattr", "clone_children", "release_agent" 324 * and "name" are disallowed. 325 * 326 * - When mounting an existing superblock, mount options should match. 327 * 328 * - rename(2) is disallowed. 329 * 330 * - "tasks" is removed. Everything should be at process granularity. Use 331 * "cgroup.procs" instead. 332 * 333 * - "cgroup.procs" is not sorted. pids will be unique unless they got 334 * recycled in-between reads. 335 * 336 * - "release_agent" and "notify_on_release" are removed. Replacement 337 * notification mechanism will be implemented. 338 * 339 * - "cgroup.clone_children" is removed. 340 * 341 * - "cgroup.subtree_populated" is available. Its value is 0 if the cgroup 342 * and its descendants contain no task; otherwise, 1. The file also 343 * generates kernfs notification which can be monitored through poll and 344 * [di]notify when the value of the file changes. 345 * 346 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and 347 * take masks of ancestors with non-empty cpus/mems, instead of being 348 * moved to an ancestor. 349 * 350 * - cpuset: a task can be moved into an empty cpuset, and again it takes 351 * masks of ancestors. 352 * 353 * - blkcg: blk-throttle becomes properly hierarchical. 354 */ 355bool cgroup_on_dfl(const struct cgroup *cgrp) 356{ 357 return cgrp->root == &cgrp_dfl_root; 358} 359 360/* IDR wrappers which synchronize using cgroup_idr_lock */ 361static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end, 362 gfp_t gfp_mask) 363{ 364 int ret; 365 366 idr_preload(gfp_mask); 367 spin_lock_bh(&cgroup_idr_lock); 368 ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM); 369 spin_unlock_bh(&cgroup_idr_lock); 370 idr_preload_end(); 371 return ret; 372} 373 374static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id) 375{ 376 void *ret; 377 378 spin_lock_bh(&cgroup_idr_lock); 379 ret = idr_replace(idr, ptr, id); 380 spin_unlock_bh(&cgroup_idr_lock); 381 return ret; 382} 383 384static void cgroup_idr_remove(struct idr *idr, int id) 385{ 386 spin_lock_bh(&cgroup_idr_lock); 387 idr_remove(idr, id); 388 spin_unlock_bh(&cgroup_idr_lock); 389} 390 391static bool cgroup_has_tasks(struct cgroup *cgrp) 392{ 393 return cgrp->nr_populated_csets; 394} 395 396static bool cgroup_is_threaded(struct cgroup *cgrp) 397{ 398 return cgrp->dom_cgrp != cgrp; 399} 400 401/* can @cgrp host both domain and threaded children? */ 402static bool cgroup_is_mixable(struct cgroup *cgrp) 403{ 404 /* 405 * Root isn't under domain level resource control exempting it from 406 * the no-internal-process constraint, so it can serve as a thread 407 * root and a parent of resource domains at the same time. 408 */ 409 return !cgroup_parent(cgrp); 410} 411 412/* can @cgrp become a thread root? Should always be true for a thread root */ 413static bool cgroup_can_be_thread_root(struct cgroup *cgrp) 414{ 415 /* mixables don't care */ 416 if (cgroup_is_mixable(cgrp)) 417 return true; 418 419 /* domain roots can't be nested under threaded */ 420 if (cgroup_is_threaded(cgrp)) 421 return false; 422 423 /* can only have either domain or threaded children */ 424 if (cgrp->nr_populated_domain_children) 425 return false; 426 427 /* and no domain controllers can be enabled */ 428 if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask) 429 return false; 430 431 return true; 432} 433 434/* is @cgrp root of a threaded subtree? */ 435static bool cgroup_is_thread_root(struct cgroup *cgrp) 436{ 437 /* thread root should be a domain */ 438 if (cgroup_is_threaded(cgrp)) 439 return false; 440 441 /* a domain w/ threaded children is a thread root */ 442 if (cgrp->nr_threaded_children) 443 return true; 444 445 /* 446 * A domain which has tasks and explicit threaded controllers 447 * enabled is a thread root. 448 */ 449 if (cgroup_has_tasks(cgrp) && 450 (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask)) 451 return true; 452 453 return false; 454} 455 456/* a domain which isn't connected to the root w/o brekage can't be used */ 457static bool cgroup_is_valid_domain(struct cgroup *cgrp) 458{ 459 /* the cgroup itself can be a thread root */ 460 if (cgroup_is_threaded(cgrp)) 461 return false; 462 463 /* but the ancestors can't be unless mixable */ 464 while ((cgrp = cgroup_parent(cgrp))) { 465 if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp)) 466 return false; 467 if (cgroup_is_threaded(cgrp)) 468 return false; 469 } 470 471 return true; 472} 473 474/* subsystems visibly enabled on a cgroup */ 475static u16 cgroup_control(struct cgroup *cgrp) 476{ 477 struct cgroup *parent = cgroup_parent(cgrp); 478 u16 root_ss_mask = cgrp->root->subsys_mask; 479 480 if (parent) { 481 u16 ss_mask = parent->subtree_control; 482 483 /* threaded cgroups can only have threaded controllers */ 484 if (cgroup_is_threaded(cgrp)) 485 ss_mask &= cgrp_dfl_threaded_ss_mask; 486 return ss_mask; 487 } 488 489 if (cgroup_on_dfl(cgrp)) 490 root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask | 491 cgrp_dfl_implicit_ss_mask); 492 return root_ss_mask; 493} 494 495/* subsystems enabled on a cgroup */ 496static u16 cgroup_ss_mask(struct cgroup *cgrp) 497{ 498 struct cgroup *parent = cgroup_parent(cgrp); 499 500 if (parent) { 501 u16 ss_mask = parent->subtree_ss_mask; 502 503 /* threaded cgroups can only have threaded controllers */ 504 if (cgroup_is_threaded(cgrp)) 505 ss_mask &= cgrp_dfl_threaded_ss_mask; 506 return ss_mask; 507 } 508 509 return cgrp->root->subsys_mask; 510} 511 512/** 513 * cgroup_css - obtain a cgroup's css for the specified subsystem 514 * @cgrp: the cgroup of interest 515 * @ss: the subsystem of interest (%NULL returns @cgrp->self) 516 * 517 * Return @cgrp's css (cgroup_subsys_state) associated with @ss. This 518 * function must be called either under cgroup_mutex or rcu_read_lock() and 519 * the caller is responsible for pinning the returned css if it wants to 520 * keep accessing it outside the said locks. This function may return 521 * %NULL if @cgrp doesn't have @subsys_id enabled. 522 */ 523static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp, 524 struct cgroup_subsys *ss) 525{ 526 if (CGROUP_HAS_SUBSYS_CONFIG && ss) 527 return rcu_dereference_check(cgrp->subsys[ss->id], 528 lockdep_is_held(&cgroup_mutex)); 529 else 530 return &cgrp->self; 531} 532 533/** 534 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss 535 * @cgrp: the cgroup of interest 536 * @ss: the subsystem of interest (%NULL returns @cgrp->self) 537 * 538 * Similar to cgroup_css() but returns the effective css, which is defined 539 * as the matching css of the nearest ancestor including self which has @ss 540 * enabled. If @ss is associated with the hierarchy @cgrp is on, this 541 * function is guaranteed to return non-NULL css. 542 */ 543static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp, 544 struct cgroup_subsys *ss) 545{ 546 lockdep_assert_held(&cgroup_mutex); 547 548 if (!ss) 549 return &cgrp->self; 550 551 /* 552 * This function is used while updating css associations and thus 553 * can't test the csses directly. Test ss_mask. 554 */ 555 while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) { 556 cgrp = cgroup_parent(cgrp); 557 if (!cgrp) 558 return NULL; 559 } 560 561 return cgroup_css(cgrp, ss); 562} 563 564/** 565 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem 566 * @cgrp: the cgroup of interest 567 * @ss: the subsystem of interest 568 * 569 * Find and get the effective css of @cgrp for @ss. The effective css is 570 * defined as the matching css of the nearest ancestor including self which 571 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on, 572 * the root css is returned, so this function always returns a valid css. 573 * 574 * The returned css is not guaranteed to be online, and therefore it is the 575 * callers responsibility to try get a reference for it. 576 */ 577struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp, 578 struct cgroup_subsys *ss) 579{ 580 struct cgroup_subsys_state *css; 581 582 if (!CGROUP_HAS_SUBSYS_CONFIG) 583 return NULL; 584 585 do { 586 css = cgroup_css(cgrp, ss); 587 588 if (css) 589 return css; 590 cgrp = cgroup_parent(cgrp); 591 } while (cgrp); 592 593 return init_css_set.subsys[ss->id]; 594} 595 596/** 597 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem 598 * @cgrp: the cgroup of interest 599 * @ss: the subsystem of interest 600 * 601 * Find and get the effective css of @cgrp for @ss. The effective css is 602 * defined as the matching css of the nearest ancestor including self which 603 * has @ss enabled. If @ss is not mounted on the hierarchy @cgrp is on, 604 * the root css is returned, so this function always returns a valid css. 605 * The returned css must be put using css_put(). 606 */ 607struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp, 608 struct cgroup_subsys *ss) 609{ 610 struct cgroup_subsys_state *css; 611 612 if (!CGROUP_HAS_SUBSYS_CONFIG) 613 return NULL; 614 615 rcu_read_lock(); 616 617 do { 618 css = cgroup_css(cgrp, ss); 619 620 if (css && css_tryget_online(css)) 621 goto out_unlock; 622 cgrp = cgroup_parent(cgrp); 623 } while (cgrp); 624 625 css = init_css_set.subsys[ss->id]; 626 css_get(css); 627out_unlock: 628 rcu_read_unlock(); 629 return css; 630} 631EXPORT_SYMBOL_GPL(cgroup_get_e_css); 632 633static void cgroup_get_live(struct cgroup *cgrp) 634{ 635 WARN_ON_ONCE(cgroup_is_dead(cgrp)); 636 cgroup_get(cgrp); 637} 638 639/** 640 * __cgroup_task_count - count the number of tasks in a cgroup. The caller 641 * is responsible for taking the css_set_lock. 642 * @cgrp: the cgroup in question 643 */ 644int __cgroup_task_count(const struct cgroup *cgrp) 645{ 646 int count = 0; 647 struct cgrp_cset_link *link; 648 649 lockdep_assert_held(&css_set_lock); 650 651 list_for_each_entry(link, &cgrp->cset_links, cset_link) 652 count += link->cset->nr_tasks; 653 654 return count; 655} 656 657/** 658 * cgroup_task_count - count the number of tasks in a cgroup. 659 * @cgrp: the cgroup in question 660 */ 661int cgroup_task_count(const struct cgroup *cgrp) 662{ 663 int count; 664 665 spin_lock_irq(&css_set_lock); 666 count = __cgroup_task_count(cgrp); 667 spin_unlock_irq(&css_set_lock); 668 669 return count; 670} 671 672static struct cgroup *kn_priv(struct kernfs_node *kn) 673{ 674 struct kernfs_node *parent; 675 /* 676 * The parent can not be replaced due to KERNFS_ROOT_INVARIANT_PARENT. 677 * Therefore it is always safe to dereference this pointer outside of a 678 * RCU section. 679 */ 680 parent = rcu_dereference_check(kn->__parent, 681 kernfs_root_flags(kn) & KERNFS_ROOT_INVARIANT_PARENT); 682 return parent->priv; 683} 684 685struct cgroup_subsys_state *of_css(struct kernfs_open_file *of) 686{ 687 struct cgroup *cgrp = kn_priv(of->kn); 688 struct cftype *cft = of_cft(of); 689 690 /* 691 * This is open and unprotected implementation of cgroup_css(). 692 * seq_css() is only called from a kernfs file operation which has 693 * an active reference on the file. Because all the subsystem 694 * files are drained before a css is disassociated with a cgroup, 695 * the matching css from the cgroup's subsys table is guaranteed to 696 * be and stay valid until the enclosing operation is complete. 697 */ 698 if (CGROUP_HAS_SUBSYS_CONFIG && cft->ss) 699 return rcu_dereference_raw(cgrp->subsys[cft->ss->id]); 700 else 701 return &cgrp->self; 702} 703EXPORT_SYMBOL_GPL(of_css); 704 705/** 706 * for_each_css - iterate all css's of a cgroup 707 * @css: the iteration cursor 708 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end 709 * @cgrp: the target cgroup to iterate css's of 710 * 711 * Should be called under cgroup_mutex. 712 */ 713#define for_each_css(css, ssid, cgrp) \ 714 for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++) \ 715 if (!((css) = rcu_dereference_check( \ 716 (cgrp)->subsys[(ssid)], \ 717 lockdep_is_held(&cgroup_mutex)))) { } \ 718 else 719 720/** 721 * do_each_subsys_mask - filter for_each_subsys with a bitmask 722 * @ss: the iteration cursor 723 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end 724 * @ss_mask: the bitmask 725 * 726 * The block will only run for cases where the ssid-th bit (1 << ssid) of 727 * @ss_mask is set. 728 */ 729#define do_each_subsys_mask(ss, ssid, ss_mask) do { \ 730 unsigned long __ss_mask = (ss_mask); \ 731 if (!CGROUP_HAS_SUBSYS_CONFIG) { \ 732 (ssid) = 0; \ 733 break; \ 734 } \ 735 for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) { \ 736 (ss) = cgroup_subsys[ssid]; \ 737 { 738 739#define while_each_subsys_mask() \ 740 } \ 741 } \ 742} while (false) 743 744/* iterate over child cgrps, lock should be held throughout iteration */ 745#define cgroup_for_each_live_child(child, cgrp) \ 746 list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \ 747 if (({ lockdep_assert_held(&cgroup_mutex); \ 748 cgroup_is_dead(child); })) \ 749 ; \ 750 else 751 752/* walk live descendants in pre order */ 753#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) \ 754 css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL)) \ 755 if (({ lockdep_assert_held(&cgroup_mutex); \ 756 (dsct) = (d_css)->cgroup; \ 757 cgroup_is_dead(dsct); })) \ 758 ; \ 759 else 760 761/* walk live descendants in postorder */ 762#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) \ 763 css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \ 764 if (({ lockdep_assert_held(&cgroup_mutex); \ 765 (dsct) = (d_css)->cgroup; \ 766 cgroup_is_dead(dsct); })) \ 767 ; \ 768 else 769 770/* 771 * The default css_set - used by init and its children prior to any 772 * hierarchies being mounted. It contains a pointer to the root state 773 * for each subsystem. Also used to anchor the list of css_sets. Not 774 * reference-counted, to improve performance when child cgroups 775 * haven't been created. 776 */ 777struct css_set init_css_set = { 778 .refcount = REFCOUNT_INIT(1), 779 .dom_cset = &init_css_set, 780 .tasks = LIST_HEAD_INIT(init_css_set.tasks), 781 .mg_tasks = LIST_HEAD_INIT(init_css_set.mg_tasks), 782 .dying_tasks = LIST_HEAD_INIT(init_css_set.dying_tasks), 783 .task_iters = LIST_HEAD_INIT(init_css_set.task_iters), 784 .threaded_csets = LIST_HEAD_INIT(init_css_set.threaded_csets), 785 .cgrp_links = LIST_HEAD_INIT(init_css_set.cgrp_links), 786 .mg_src_preload_node = LIST_HEAD_INIT(init_css_set.mg_src_preload_node), 787 .mg_dst_preload_node = LIST_HEAD_INIT(init_css_set.mg_dst_preload_node), 788 .mg_node = LIST_HEAD_INIT(init_css_set.mg_node), 789 790 /* 791 * The following field is re-initialized when this cset gets linked 792 * in cgroup_init(). However, let's initialize the field 793 * statically too so that the default cgroup can be accessed safely 794 * early during boot. 795 */ 796 .dfl_cgrp = &cgrp_dfl_root.cgrp, 797}; 798 799static int css_set_count = 1; /* 1 for init_css_set */ 800 801static bool css_set_threaded(struct css_set *cset) 802{ 803 return cset->dom_cset != cset; 804} 805 806/** 807 * css_set_populated - does a css_set contain any tasks? 808 * @cset: target css_set 809 * 810 * css_set_populated() should be the same as !!cset->nr_tasks at steady 811 * state. However, css_set_populated() can be called while a task is being 812 * added to or removed from the linked list before the nr_tasks is 813 * properly updated. Hence, we can't just look at ->nr_tasks here. 814 */ 815static bool css_set_populated(struct css_set *cset) 816{ 817 lockdep_assert_held(&css_set_lock); 818 819 return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks); 820} 821 822/** 823 * cgroup_update_populated - update the populated count of a cgroup 824 * @cgrp: the target cgroup 825 * @populated: inc or dec populated count 826 * 827 * One of the css_sets associated with @cgrp is either getting its first 828 * task or losing the last. Update @cgrp->nr_populated_* accordingly. The 829 * count is propagated towards root so that a given cgroup's 830 * nr_populated_children is zero iff none of its descendants contain any 831 * tasks. 832 * 833 * @cgrp's interface file "cgroup.populated" is zero if both 834 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and 835 * 1 otherwise. When the sum changes from or to zero, userland is notified 836 * that the content of the interface file has changed. This can be used to 837 * detect when @cgrp and its descendants become populated or empty. 838 */ 839static void cgroup_update_populated(struct cgroup *cgrp, bool populated) 840{ 841 struct cgroup *child = NULL; 842 int adj = populated ? 1 : -1; 843 844 lockdep_assert_held(&css_set_lock); 845 846 do { 847 bool was_populated = cgroup_is_populated(cgrp); 848 849 if (!child) { 850 cgrp->nr_populated_csets += adj; 851 } else { 852 if (cgroup_is_threaded(child)) 853 cgrp->nr_populated_threaded_children += adj; 854 else 855 cgrp->nr_populated_domain_children += adj; 856 } 857 858 if (was_populated == cgroup_is_populated(cgrp)) 859 break; 860 861 cgroup1_check_for_release(cgrp); 862 TRACE_CGROUP_PATH(notify_populated, cgrp, 863 cgroup_is_populated(cgrp)); 864 cgroup_file_notify(&cgrp->events_file); 865 866 child = cgrp; 867 cgrp = cgroup_parent(cgrp); 868 } while (cgrp); 869} 870 871/** 872 * css_set_update_populated - update populated state of a css_set 873 * @cset: target css_set 874 * @populated: whether @cset is populated or depopulated 875 * 876 * @cset is either getting the first task or losing the last. Update the 877 * populated counters of all associated cgroups accordingly. 878 */ 879static void css_set_update_populated(struct css_set *cset, bool populated) 880{ 881 struct cgrp_cset_link *link; 882 883 lockdep_assert_held(&css_set_lock); 884 885 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) 886 cgroup_update_populated(link->cgrp, populated); 887} 888 889/* 890 * @task is leaving, advance task iterators which are pointing to it so 891 * that they can resume at the next position. Advancing an iterator might 892 * remove it from the list, use safe walk. See css_task_iter_skip() for 893 * details. 894 */ 895static void css_set_skip_task_iters(struct css_set *cset, 896 struct task_struct *task) 897{ 898 struct css_task_iter *it, *pos; 899 900 list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node) 901 css_task_iter_skip(it, task); 902} 903 904/** 905 * css_set_move_task - move a task from one css_set to another 906 * @task: task being moved 907 * @from_cset: css_set @task currently belongs to (may be NULL) 908 * @to_cset: new css_set @task is being moved to (may be NULL) 909 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks 910 * 911 * Move @task from @from_cset to @to_cset. If @task didn't belong to any 912 * css_set, @from_cset can be NULL. If @task is being disassociated 913 * instead of moved, @to_cset can be NULL. 914 * 915 * This function automatically handles populated counter updates and 916 * css_task_iter adjustments but the caller is responsible for managing 917 * @from_cset and @to_cset's reference counts. 918 */ 919static void css_set_move_task(struct task_struct *task, 920 struct css_set *from_cset, struct css_set *to_cset, 921 bool use_mg_tasks) 922{ 923 lockdep_assert_held(&css_set_lock); 924 925 if (to_cset && !css_set_populated(to_cset)) 926 css_set_update_populated(to_cset, true); 927 928 if (from_cset) { 929 WARN_ON_ONCE(list_empty(&task->cg_list)); 930 931 css_set_skip_task_iters(from_cset, task); 932 list_del_init(&task->cg_list); 933 if (!css_set_populated(from_cset)) 934 css_set_update_populated(from_cset, false); 935 } else { 936 WARN_ON_ONCE(!list_empty(&task->cg_list)); 937 } 938 939 if (to_cset) { 940 /* 941 * We are synchronized through cgroup_threadgroup_rwsem 942 * against PF_EXITING setting such that we can't race 943 * against cgroup_task_dead()/cgroup_task_free() dropping 944 * the css_set. 945 */ 946 WARN_ON_ONCE(task->flags & PF_EXITING); 947 948 cgroup_move_task(task, to_cset); 949 list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks : 950 &to_cset->tasks); 951 } 952} 953 954/* 955 * hash table for cgroup groups. This improves the performance to find 956 * an existing css_set. This hash doesn't (currently) take into 957 * account cgroups in empty hierarchies. 958 */ 959#define CSS_SET_HASH_BITS 7 960static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS); 961 962static unsigned long css_set_hash(struct cgroup_subsys_state **css) 963{ 964 unsigned long key = 0UL; 965 struct cgroup_subsys *ss; 966 int i; 967 968 for_each_subsys(ss, i) 969 key += (unsigned long)css[i]; 970 key = (key >> 16) ^ key; 971 972 return key; 973} 974 975void put_css_set_locked(struct css_set *cset) 976{ 977 struct cgrp_cset_link *link, *tmp_link; 978 struct cgroup_subsys *ss; 979 int ssid; 980 981 lockdep_assert_held(&css_set_lock); 982 983 if (!refcount_dec_and_test(&cset->refcount)) 984 return; 985 986 WARN_ON_ONCE(!list_empty(&cset->threaded_csets)); 987 988 /* This css_set is dead. Unlink it and release cgroup and css refs */ 989 for_each_subsys(ss, ssid) { 990 list_del(&cset->e_cset_node[ssid]); 991 css_put(cset->subsys[ssid]); 992 } 993 hash_del(&cset->hlist); 994 css_set_count--; 995 996 list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) { 997 list_del(&link->cset_link); 998 list_del(&link->cgrp_link); 999 if (cgroup_parent(link->cgrp)) 1000 cgroup_put(link->cgrp); 1001 kfree(link); 1002 } 1003 1004 if (css_set_threaded(cset)) { 1005 list_del(&cset->threaded_csets_node); 1006 put_css_set_locked(cset->dom_cset); 1007 } 1008 1009 kfree_rcu(cset, rcu_head); 1010} 1011 1012/** 1013 * compare_css_sets - helper function for find_existing_css_set(). 1014 * @cset: candidate css_set being tested 1015 * @old_cset: existing css_set for a task 1016 * @new_cgrp: cgroup that's being entered by the task 1017 * @template: desired set of css pointers in css_set (pre-calculated) 1018 * 1019 * Returns true if "cset" matches "old_cset" except for the hierarchy 1020 * which "new_cgrp" belongs to, for which it should match "new_cgrp". 1021 */ 1022static bool compare_css_sets(struct css_set *cset, 1023 struct css_set *old_cset, 1024 struct cgroup *new_cgrp, 1025 struct cgroup_subsys_state *template[]) 1026{ 1027 struct cgroup *new_dfl_cgrp; 1028 struct list_head *l1, *l2; 1029 1030 /* 1031 * On the default hierarchy, there can be csets which are 1032 * associated with the same set of cgroups but different csses. 1033 * Let's first ensure that csses match. 1034 */ 1035 if (memcmp(template, cset->subsys, sizeof(cset->subsys))) 1036 return false; 1037 1038 1039 /* @cset's domain should match the default cgroup's */ 1040 if (cgroup_on_dfl(new_cgrp)) 1041 new_dfl_cgrp = new_cgrp; 1042 else 1043 new_dfl_cgrp = old_cset->dfl_cgrp; 1044 1045 if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp) 1046 return false; 1047 1048 /* 1049 * Compare cgroup pointers in order to distinguish between 1050 * different cgroups in hierarchies. As different cgroups may 1051 * share the same effective css, this comparison is always 1052 * necessary. 1053 */ 1054 l1 = &cset->cgrp_links; 1055 l2 = &old_cset->cgrp_links; 1056 while (1) { 1057 struct cgrp_cset_link *link1, *link2; 1058 struct cgroup *cgrp1, *cgrp2; 1059 1060 l1 = l1->next; 1061 l2 = l2->next; 1062 /* See if we reached the end - both lists are equal length. */ 1063 if (l1 == &cset->cgrp_links) { 1064 BUG_ON(l2 != &old_cset->cgrp_links); 1065 break; 1066 } else { 1067 BUG_ON(l2 == &old_cset->cgrp_links); 1068 } 1069 /* Locate the cgroups associated with these links. */ 1070 link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link); 1071 link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link); 1072 cgrp1 = link1->cgrp; 1073 cgrp2 = link2->cgrp; 1074 /* Hierarchies should be linked in the same order. */ 1075 BUG_ON(cgrp1->root != cgrp2->root); 1076 1077 /* 1078 * If this hierarchy is the hierarchy of the cgroup 1079 * that's changing, then we need to check that this 1080 * css_set points to the new cgroup; if it's any other 1081 * hierarchy, then this css_set should point to the 1082 * same cgroup as the old css_set. 1083 */ 1084 if (cgrp1->root == new_cgrp->root) { 1085 if (cgrp1 != new_cgrp) 1086 return false; 1087 } else { 1088 if (cgrp1 != cgrp2) 1089 return false; 1090 } 1091 } 1092 return true; 1093} 1094 1095/** 1096 * find_existing_css_set - init css array and find the matching css_set 1097 * @old_cset: the css_set that we're using before the cgroup transition 1098 * @cgrp: the cgroup that we're moving into 1099 * @template: out param for the new set of csses, should be clear on entry 1100 */ 1101static struct css_set *find_existing_css_set(struct css_set *old_cset, 1102 struct cgroup *cgrp, 1103 struct cgroup_subsys_state **template) 1104{ 1105 struct cgroup_root *root = cgrp->root; 1106 struct cgroup_subsys *ss; 1107 struct css_set *cset; 1108 unsigned long key; 1109 int i; 1110 1111 /* 1112 * Build the set of subsystem state objects that we want to see in the 1113 * new css_set. While subsystems can change globally, the entries here 1114 * won't change, so no need for locking. 1115 */ 1116 for_each_subsys(ss, i) { 1117 if (root->subsys_mask & (1UL << i)) { 1118 /* 1119 * @ss is in this hierarchy, so we want the 1120 * effective css from @cgrp. 1121 */ 1122 template[i] = cgroup_e_css_by_mask(cgrp, ss); 1123 } else { 1124 /* 1125 * @ss is not in this hierarchy, so we don't want 1126 * to change the css. 1127 */ 1128 template[i] = old_cset->subsys[i]; 1129 } 1130 } 1131 1132 key = css_set_hash(template); 1133 hash_for_each_possible(css_set_table, cset, hlist, key) { 1134 if (!compare_css_sets(cset, old_cset, cgrp, template)) 1135 continue; 1136 1137 /* This css_set matches what we need */ 1138 return cset; 1139 } 1140 1141 /* No existing cgroup group matched */ 1142 return NULL; 1143} 1144 1145static void free_cgrp_cset_links(struct list_head *links_to_free) 1146{ 1147 struct cgrp_cset_link *link, *tmp_link; 1148 1149 list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) { 1150 list_del(&link->cset_link); 1151 kfree(link); 1152 } 1153} 1154 1155/** 1156 * allocate_cgrp_cset_links - allocate cgrp_cset_links 1157 * @count: the number of links to allocate 1158 * @tmp_links: list_head the allocated links are put on 1159 * 1160 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links 1161 * through ->cset_link. Returns 0 on success or -errno. 1162 */ 1163static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links) 1164{ 1165 struct cgrp_cset_link *link; 1166 int i; 1167 1168 INIT_LIST_HEAD(tmp_links); 1169 1170 for (i = 0; i < count; i++) { 1171 link = kzalloc(sizeof(*link), GFP_KERNEL); 1172 if (!link) { 1173 free_cgrp_cset_links(tmp_links); 1174 return -ENOMEM; 1175 } 1176 list_add(&link->cset_link, tmp_links); 1177 } 1178 return 0; 1179} 1180 1181/** 1182 * link_css_set - a helper function to link a css_set to a cgroup 1183 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links() 1184 * @cset: the css_set to be linked 1185 * @cgrp: the destination cgroup 1186 */ 1187static void link_css_set(struct list_head *tmp_links, struct css_set *cset, 1188 struct cgroup *cgrp) 1189{ 1190 struct cgrp_cset_link *link; 1191 1192 BUG_ON(list_empty(tmp_links)); 1193 1194 if (cgroup_on_dfl(cgrp)) 1195 cset->dfl_cgrp = cgrp; 1196 1197 link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link); 1198 link->cset = cset; 1199 link->cgrp = cgrp; 1200 1201 /* 1202 * Always add links to the tail of the lists so that the lists are 1203 * in chronological order. 1204 */ 1205 list_move_tail(&link->cset_link, &cgrp->cset_links); 1206 list_add_tail(&link->cgrp_link, &cset->cgrp_links); 1207 1208 if (cgroup_parent(cgrp)) 1209 cgroup_get_live(cgrp); 1210} 1211 1212/** 1213 * find_css_set - return a new css_set with one cgroup updated 1214 * @old_cset: the baseline css_set 1215 * @cgrp: the cgroup to be updated 1216 * 1217 * Return a new css_set that's equivalent to @old_cset, but with @cgrp 1218 * substituted into the appropriate hierarchy. 1219 */ 1220static struct css_set *find_css_set(struct css_set *old_cset, 1221 struct cgroup *cgrp) 1222{ 1223 struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { }; 1224 struct css_set *cset; 1225 struct list_head tmp_links; 1226 struct cgrp_cset_link *link; 1227 struct cgroup_subsys *ss; 1228 unsigned long key; 1229 int ssid; 1230 1231 lockdep_assert_held(&cgroup_mutex); 1232 1233 /* First see if we already have a cgroup group that matches 1234 * the desired set */ 1235 spin_lock_irq(&css_set_lock); 1236 cset = find_existing_css_set(old_cset, cgrp, template); 1237 if (cset) 1238 get_css_set(cset); 1239 spin_unlock_irq(&css_set_lock); 1240 1241 if (cset) 1242 return cset; 1243 1244 cset = kzalloc(sizeof(*cset), GFP_KERNEL); 1245 if (!cset) 1246 return NULL; 1247 1248 /* Allocate all the cgrp_cset_link objects that we'll need */ 1249 if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) { 1250 kfree(cset); 1251 return NULL; 1252 } 1253 1254 refcount_set(&cset->refcount, 1); 1255 cset->dom_cset = cset; 1256 INIT_LIST_HEAD(&cset->tasks); 1257 INIT_LIST_HEAD(&cset->mg_tasks); 1258 INIT_LIST_HEAD(&cset->dying_tasks); 1259 INIT_LIST_HEAD(&cset->task_iters); 1260 INIT_LIST_HEAD(&cset->threaded_csets); 1261 INIT_HLIST_NODE(&cset->hlist); 1262 INIT_LIST_HEAD(&cset->cgrp_links); 1263 INIT_LIST_HEAD(&cset->mg_src_preload_node); 1264 INIT_LIST_HEAD(&cset->mg_dst_preload_node); 1265 INIT_LIST_HEAD(&cset->mg_node); 1266 1267 /* Copy the set of subsystem state objects generated in 1268 * find_existing_css_set() */ 1269 memcpy(cset->subsys, template, sizeof(cset->subsys)); 1270 1271 spin_lock_irq(&css_set_lock); 1272 /* Add reference counts and links from the new css_set. */ 1273 list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) { 1274 struct cgroup *c = link->cgrp; 1275 1276 if (c->root == cgrp->root) 1277 c = cgrp; 1278 link_css_set(&tmp_links, cset, c); 1279 } 1280 1281 BUG_ON(!list_empty(&tmp_links)); 1282 1283 css_set_count++; 1284 1285 /* Add @cset to the hash table */ 1286 key = css_set_hash(cset->subsys); 1287 hash_add(css_set_table, &cset->hlist, key); 1288 1289 for_each_subsys(ss, ssid) { 1290 struct cgroup_subsys_state *css = cset->subsys[ssid]; 1291 1292 list_add_tail(&cset->e_cset_node[ssid], 1293 &css->cgroup->e_csets[ssid]); 1294 css_get(css); 1295 } 1296 1297 spin_unlock_irq(&css_set_lock); 1298 1299 /* 1300 * If @cset should be threaded, look up the matching dom_cset and 1301 * link them up. We first fully initialize @cset then look for the 1302 * dom_cset. It's simpler this way and safe as @cset is guaranteed 1303 * to stay empty until we return. 1304 */ 1305 if (cgroup_is_threaded(cset->dfl_cgrp)) { 1306 struct css_set *dcset; 1307 1308 dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp); 1309 if (!dcset) { 1310 put_css_set(cset); 1311 return NULL; 1312 } 1313 1314 spin_lock_irq(&css_set_lock); 1315 cset->dom_cset = dcset; 1316 list_add_tail(&cset->threaded_csets_node, 1317 &dcset->threaded_csets); 1318 spin_unlock_irq(&css_set_lock); 1319 } 1320 1321 return cset; 1322} 1323 1324struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root) 1325{ 1326 struct cgroup *root_cgrp = kernfs_root_to_node(kf_root)->priv; 1327 1328 return root_cgrp->root; 1329} 1330 1331void cgroup_favor_dynmods(struct cgroup_root *root, bool favor) 1332{ 1333 bool favoring = root->flags & CGRP_ROOT_FAVOR_DYNMODS; 1334 1335 /* 1336 * see the comment above CGRP_ROOT_FAVOR_DYNMODS definition. 1337 * favordynmods can flip while task is between 1338 * cgroup_threadgroup_change_begin() and end(), so down_write global 1339 * cgroup_threadgroup_rwsem to synchronize them. 1340 * 1341 * Once cgroup_enable_per_threadgroup_rwsem is enabled, holding 1342 * cgroup_threadgroup_rwsem doesn't exlude tasks between 1343 * cgroup_thread_group_change_begin() and end() and thus it's unsafe to 1344 * turn off. As the scenario is unlikely, simply disallow disabling once 1345 * enabled and print out a warning. 1346 */ 1347 percpu_down_write(&cgroup_threadgroup_rwsem); 1348 if (favor && !favoring) { 1349 cgroup_enable_per_threadgroup_rwsem = true; 1350 rcu_sync_enter(&cgroup_threadgroup_rwsem.rss); 1351 root->flags |= CGRP_ROOT_FAVOR_DYNMODS; 1352 } else if (!favor && favoring) { 1353 if (cgroup_enable_per_threadgroup_rwsem) 1354 pr_warn_once("cgroup favordynmods: per threadgroup rwsem mechanism can't be disabled\n"); 1355 rcu_sync_exit(&cgroup_threadgroup_rwsem.rss); 1356 root->flags &= ~CGRP_ROOT_FAVOR_DYNMODS; 1357 } 1358 percpu_up_write(&cgroup_threadgroup_rwsem); 1359} 1360 1361static int cgroup_init_root_id(struct cgroup_root *root) 1362{ 1363 int id; 1364 1365 lockdep_assert_held(&cgroup_mutex); 1366 1367 id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL); 1368 if (id < 0) 1369 return id; 1370 1371 root->hierarchy_id = id; 1372 return 0; 1373} 1374 1375static void cgroup_exit_root_id(struct cgroup_root *root) 1376{ 1377 lockdep_assert_held(&cgroup_mutex); 1378 1379 idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id); 1380} 1381 1382void cgroup_free_root(struct cgroup_root *root) 1383{ 1384 kfree_rcu(root, rcu); 1385} 1386 1387static void cgroup_destroy_root(struct cgroup_root *root) 1388{ 1389 struct cgroup *cgrp = &root->cgrp; 1390 struct cgrp_cset_link *link, *tmp_link; 1391 int ret; 1392 1393 trace_cgroup_destroy_root(root); 1394 1395 cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp); 1396 1397 BUG_ON(atomic_read(&root->nr_cgrps)); 1398 BUG_ON(!list_empty(&cgrp->self.children)); 1399 1400 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier, 1401 CGROUP_LIFETIME_OFFLINE, cgrp); 1402 WARN_ON_ONCE(notifier_to_errno(ret)); 1403 1404 /* Rebind all subsystems back to the default hierarchy */ 1405 WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask)); 1406 1407 /* 1408 * Release all the links from cset_links to this hierarchy's 1409 * root cgroup 1410 */ 1411 spin_lock_irq(&css_set_lock); 1412 1413 list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) { 1414 list_del(&link->cset_link); 1415 list_del(&link->cgrp_link); 1416 kfree(link); 1417 } 1418 1419 spin_unlock_irq(&css_set_lock); 1420 1421 WARN_ON_ONCE(list_empty(&root->root_list)); 1422 list_del_rcu(&root->root_list); 1423 cgroup_root_count--; 1424 1425 if (!have_favordynmods) 1426 cgroup_favor_dynmods(root, false); 1427 1428 cgroup_exit_root_id(root); 1429 1430 cgroup_unlock(); 1431 1432 kernfs_destroy_root(root->kf_root); 1433 cgroup_free_root(root); 1434} 1435 1436/* 1437 * Returned cgroup is without refcount but it's valid as long as cset pins it. 1438 */ 1439static inline struct cgroup *__cset_cgroup_from_root(struct css_set *cset, 1440 struct cgroup_root *root) 1441{ 1442 struct cgroup *res_cgroup = NULL; 1443 1444 if (cset == &init_css_set) { 1445 res_cgroup = &root->cgrp; 1446 } else if (root == &cgrp_dfl_root) { 1447 res_cgroup = cset->dfl_cgrp; 1448 } else { 1449 struct cgrp_cset_link *link; 1450 lockdep_assert_held(&css_set_lock); 1451 1452 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) { 1453 struct cgroup *c = link->cgrp; 1454 1455 if (c->root == root) { 1456 res_cgroup = c; 1457 break; 1458 } 1459 } 1460 } 1461 1462 /* 1463 * If cgroup_mutex is not held, the cgrp_cset_link will be freed 1464 * before we remove the cgroup root from the root_list. Consequently, 1465 * when accessing a cgroup root, the cset_link may have already been 1466 * freed, resulting in a NULL res_cgroup. However, by holding the 1467 * cgroup_mutex, we ensure that res_cgroup can't be NULL. 1468 * If we don't hold cgroup_mutex in the caller, we must do the NULL 1469 * check. 1470 */ 1471 return res_cgroup; 1472} 1473 1474/* 1475 * look up cgroup associated with current task's cgroup namespace on the 1476 * specified hierarchy 1477 */ 1478static struct cgroup * 1479current_cgns_cgroup_from_root(struct cgroup_root *root) 1480{ 1481 struct cgroup *res = NULL; 1482 struct css_set *cset; 1483 1484 lockdep_assert_held(&css_set_lock); 1485 1486 rcu_read_lock(); 1487 1488 cset = current->nsproxy->cgroup_ns->root_cset; 1489 res = __cset_cgroup_from_root(cset, root); 1490 1491 rcu_read_unlock(); 1492 1493 /* 1494 * The namespace_sem is held by current, so the root cgroup can't 1495 * be umounted. Therefore, we can ensure that the res is non-NULL. 1496 */ 1497 WARN_ON_ONCE(!res); 1498 return res; 1499} 1500 1501/* 1502 * Look up cgroup associated with current task's cgroup namespace on the default 1503 * hierarchy. 1504 * 1505 * Unlike current_cgns_cgroup_from_root(), this doesn't need locks: 1506 * - Internal rcu_read_lock is unnecessary because we don't dereference any rcu 1507 * pointers. 1508 * - css_set_lock is not needed because we just read cset->dfl_cgrp. 1509 * - As a bonus returned cgrp is pinned with the current because it cannot 1510 * switch cgroup_ns asynchronously. 1511 */ 1512static struct cgroup *current_cgns_cgroup_dfl(void) 1513{ 1514 struct css_set *cset; 1515 1516 if (current->nsproxy) { 1517 cset = current->nsproxy->cgroup_ns->root_cset; 1518 return __cset_cgroup_from_root(cset, &cgrp_dfl_root); 1519 } else { 1520 /* 1521 * NOTE: This function may be called from bpf_cgroup_from_id() 1522 * on a task which has already passed exit_nsproxy_namespaces() 1523 * and nsproxy == NULL. Fall back to cgrp_dfl_root which will 1524 * make all cgroups visible for lookups. 1525 */ 1526 return &cgrp_dfl_root.cgrp; 1527 } 1528} 1529 1530/* look up cgroup associated with given css_set on the specified hierarchy */ 1531static struct cgroup *cset_cgroup_from_root(struct css_set *cset, 1532 struct cgroup_root *root) 1533{ 1534 lockdep_assert_held(&css_set_lock); 1535 1536 return __cset_cgroup_from_root(cset, root); 1537} 1538 1539/* 1540 * Return the cgroup for "task" from the given hierarchy. Must be 1541 * called with css_set_lock held to prevent task's groups from being modified. 1542 * Must be called with either cgroup_mutex or rcu read lock to prevent the 1543 * cgroup root from being destroyed. 1544 */ 1545struct cgroup *task_cgroup_from_root(struct task_struct *task, 1546 struct cgroup_root *root) 1547{ 1548 /* 1549 * No need to lock the task - since we hold css_set_lock the 1550 * task can't change groups. 1551 */ 1552 return cset_cgroup_from_root(task_css_set(task), root); 1553} 1554 1555/* 1556 * A task must hold cgroup_mutex to modify cgroups. 1557 * 1558 * Any task can increment and decrement the count field without lock. 1559 * So in general, code holding cgroup_mutex can't rely on the count 1560 * field not changing. However, if the count goes to zero, then only 1561 * cgroup_attach_task() can increment it again. Because a count of zero 1562 * means that no tasks are currently attached, therefore there is no 1563 * way a task attached to that cgroup can fork (the other way to 1564 * increment the count). So code holding cgroup_mutex can safely 1565 * assume that if the count is zero, it will stay zero. Similarly, if 1566 * a task holds cgroup_mutex on a cgroup with zero count, it 1567 * knows that the cgroup won't be removed, as cgroup_rmdir() 1568 * needs that mutex. 1569 * 1570 * A cgroup can only be deleted if both its 'count' of using tasks 1571 * is zero, and its list of 'children' cgroups is empty. Since all 1572 * tasks in the system use _some_ cgroup, and since there is always at 1573 * least one task in the system (init, pid == 1), therefore, root cgroup 1574 * always has either children cgroups and/or using tasks. So we don't 1575 * need a special hack to ensure that root cgroup cannot be deleted. 1576 * 1577 * P.S. One more locking exception. RCU is used to guard the 1578 * update of a tasks cgroup pointer by cgroup_attach_task() 1579 */ 1580 1581static struct kernfs_syscall_ops cgroup_kf_syscall_ops; 1582 1583static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft, 1584 char *buf) 1585{ 1586 struct cgroup_subsys *ss = cft->ss; 1587 1588 if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) && 1589 !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) { 1590 const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : ""; 1591 1592 snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s", 1593 dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name, 1594 cft->name); 1595 } else { 1596 strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX); 1597 } 1598 return buf; 1599} 1600 1601/** 1602 * cgroup_file_mode - deduce file mode of a control file 1603 * @cft: the control file in question 1604 * 1605 * S_IRUGO for read, S_IWUSR for write. 1606 */ 1607static umode_t cgroup_file_mode(const struct cftype *cft) 1608{ 1609 umode_t mode = 0; 1610 1611 if (cft->read_u64 || cft->read_s64 || cft->seq_show) 1612 mode |= S_IRUGO; 1613 1614 if (cft->write_u64 || cft->write_s64 || cft->write) { 1615 if (cft->flags & CFTYPE_WORLD_WRITABLE) 1616 mode |= S_IWUGO; 1617 else 1618 mode |= S_IWUSR; 1619 } 1620 1621 return mode; 1622} 1623 1624/** 1625 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask 1626 * @subtree_control: the new subtree_control mask to consider 1627 * @this_ss_mask: available subsystems 1628 * 1629 * On the default hierarchy, a subsystem may request other subsystems to be 1630 * enabled together through its ->depends_on mask. In such cases, more 1631 * subsystems than specified in "cgroup.subtree_control" may be enabled. 1632 * 1633 * This function calculates which subsystems need to be enabled if 1634 * @subtree_control is to be applied while restricted to @this_ss_mask. 1635 */ 1636static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask) 1637{ 1638 u16 cur_ss_mask = subtree_control; 1639 struct cgroup_subsys *ss; 1640 int ssid; 1641 1642 lockdep_assert_held(&cgroup_mutex); 1643 1644 cur_ss_mask |= cgrp_dfl_implicit_ss_mask; 1645 1646 while (true) { 1647 u16 new_ss_mask = cur_ss_mask; 1648 1649 do_each_subsys_mask(ss, ssid, cur_ss_mask) { 1650 new_ss_mask |= ss->depends_on; 1651 } while_each_subsys_mask(); 1652 1653 /* 1654 * Mask out subsystems which aren't available. This can 1655 * happen only if some depended-upon subsystems were bound 1656 * to non-default hierarchies. 1657 */ 1658 new_ss_mask &= this_ss_mask; 1659 1660 if (new_ss_mask == cur_ss_mask) 1661 break; 1662 cur_ss_mask = new_ss_mask; 1663 } 1664 1665 return cur_ss_mask; 1666} 1667 1668/** 1669 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods 1670 * @kn: the kernfs_node being serviced 1671 * 1672 * This helper undoes cgroup_kn_lock_live() and should be invoked before 1673 * the method finishes if locking succeeded. Note that once this function 1674 * returns the cgroup returned by cgroup_kn_lock_live() may become 1675 * inaccessible any time. If the caller intends to continue to access the 1676 * cgroup, it should pin it before invoking this function. 1677 */ 1678void cgroup_kn_unlock(struct kernfs_node *kn) 1679{ 1680 struct cgroup *cgrp; 1681 1682 if (kernfs_type(kn) == KERNFS_DIR) 1683 cgrp = kn->priv; 1684 else 1685 cgrp = kn_priv(kn); 1686 1687 cgroup_unlock(); 1688 1689 kernfs_unbreak_active_protection(kn); 1690 cgroup_put(cgrp); 1691} 1692 1693/** 1694 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods 1695 * @kn: the kernfs_node being serviced 1696 * @drain_offline: perform offline draining on the cgroup 1697 * 1698 * This helper is to be used by a cgroup kernfs method currently servicing 1699 * @kn. It breaks the active protection, performs cgroup locking and 1700 * verifies that the associated cgroup is alive. Returns the cgroup if 1701 * alive; otherwise, %NULL. A successful return should be undone by a 1702 * matching cgroup_kn_unlock() invocation. If @drain_offline is %true, the 1703 * cgroup is drained of offlining csses before return. 1704 * 1705 * Any cgroup kernfs method implementation which requires locking the 1706 * associated cgroup should use this helper. It avoids nesting cgroup 1707 * locking under kernfs active protection and allows all kernfs operations 1708 * including self-removal. 1709 */ 1710struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline) 1711{ 1712 struct cgroup *cgrp; 1713 1714 if (kernfs_type(kn) == KERNFS_DIR) 1715 cgrp = kn->priv; 1716 else 1717 cgrp = kn_priv(kn); 1718 1719 /* 1720 * We're gonna grab cgroup_mutex which nests outside kernfs 1721 * active_ref. cgroup liveliness check alone provides enough 1722 * protection against removal. Ensure @cgrp stays accessible and 1723 * break the active_ref protection. 1724 */ 1725 if (!cgroup_tryget(cgrp)) 1726 return NULL; 1727 kernfs_break_active_protection(kn); 1728 1729 if (drain_offline) 1730 cgroup_lock_and_drain_offline(cgrp); 1731 else 1732 cgroup_lock(); 1733 1734 if (!cgroup_is_dead(cgrp)) 1735 return cgrp; 1736 1737 cgroup_kn_unlock(kn); 1738 return NULL; 1739} 1740 1741static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft) 1742{ 1743 char name[CGROUP_FILE_NAME_MAX]; 1744 1745 lockdep_assert_held(&cgroup_mutex); 1746 1747 if (cft->file_offset) { 1748 struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss); 1749 struct cgroup_file *cfile = (void *)css + cft->file_offset; 1750 1751 spin_lock_irq(&cgroup_file_kn_lock); 1752 cfile->kn = NULL; 1753 spin_unlock_irq(&cgroup_file_kn_lock); 1754 1755 timer_delete_sync(&cfile->notify_timer); 1756 } 1757 1758 kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name)); 1759} 1760 1761/** 1762 * css_clear_dir - remove subsys files in a cgroup directory 1763 * @css: target css 1764 */ 1765static void css_clear_dir(struct cgroup_subsys_state *css) 1766{ 1767 struct cgroup *cgrp = css->cgroup; 1768 struct cftype *cfts; 1769 1770 if (!(css->flags & CSS_VISIBLE)) 1771 return; 1772 1773 css->flags &= ~CSS_VISIBLE; 1774 1775 if (css_is_self(css)) { 1776 if (cgroup_on_dfl(cgrp)) { 1777 cgroup_addrm_files(css, cgrp, 1778 cgroup_base_files, false); 1779 if (cgroup_psi_enabled()) 1780 cgroup_addrm_files(css, cgrp, 1781 cgroup_psi_files, false); 1782 } else { 1783 cgroup_addrm_files(css, cgrp, 1784 cgroup1_base_files, false); 1785 } 1786 } else { 1787 list_for_each_entry(cfts, &css->ss->cfts, node) 1788 cgroup_addrm_files(css, cgrp, cfts, false); 1789 } 1790} 1791 1792/** 1793 * css_populate_dir - create subsys files in a cgroup directory 1794 * @css: target css 1795 * 1796 * On failure, no file is added. 1797 */ 1798static int css_populate_dir(struct cgroup_subsys_state *css) 1799{ 1800 struct cgroup *cgrp = css->cgroup; 1801 struct cftype *cfts, *failed_cfts; 1802 int ret; 1803 1804 if (css->flags & CSS_VISIBLE) 1805 return 0; 1806 1807 if (css_is_self(css)) { 1808 if (cgroup_on_dfl(cgrp)) { 1809 ret = cgroup_addrm_files(css, cgrp, 1810 cgroup_base_files, true); 1811 if (ret < 0) 1812 return ret; 1813 1814 if (cgroup_psi_enabled()) { 1815 ret = cgroup_addrm_files(css, cgrp, 1816 cgroup_psi_files, true); 1817 if (ret < 0) { 1818 cgroup_addrm_files(css, cgrp, 1819 cgroup_base_files, false); 1820 return ret; 1821 } 1822 } 1823 } else { 1824 ret = cgroup_addrm_files(css, cgrp, 1825 cgroup1_base_files, true); 1826 if (ret < 0) 1827 return ret; 1828 } 1829 } else { 1830 list_for_each_entry(cfts, &css->ss->cfts, node) { 1831 ret = cgroup_addrm_files(css, cgrp, cfts, true); 1832 if (ret < 0) { 1833 failed_cfts = cfts; 1834 goto err; 1835 } 1836 } 1837 } 1838 1839 css->flags |= CSS_VISIBLE; 1840 1841 return 0; 1842err: 1843 list_for_each_entry(cfts, &css->ss->cfts, node) { 1844 if (cfts == failed_cfts) 1845 break; 1846 cgroup_addrm_files(css, cgrp, cfts, false); 1847 } 1848 return ret; 1849} 1850 1851int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask) 1852{ 1853 struct cgroup *dcgrp = &dst_root->cgrp; 1854 struct cgroup_subsys *ss; 1855 int ssid, ret; 1856 u16 dfl_disable_ss_mask = 0; 1857 1858 lockdep_assert_held(&cgroup_mutex); 1859 1860 do_each_subsys_mask(ss, ssid, ss_mask) { 1861 /* 1862 * If @ss has non-root csses attached to it, can't move. 1863 * If @ss is an implicit controller, it is exempt from this 1864 * rule and can be stolen. 1865 */ 1866 if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) && 1867 !ss->implicit_on_dfl) 1868 return -EBUSY; 1869 1870 /* can't move between two non-dummy roots either */ 1871 if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root) 1872 return -EBUSY; 1873 1874 /* 1875 * Collect ssid's that need to be disabled from default 1876 * hierarchy. 1877 */ 1878 if (ss->root == &cgrp_dfl_root) 1879 dfl_disable_ss_mask |= 1 << ssid; 1880 1881 } while_each_subsys_mask(); 1882 1883 if (dfl_disable_ss_mask) { 1884 struct cgroup *scgrp = &cgrp_dfl_root.cgrp; 1885 1886 /* 1887 * Controllers from default hierarchy that need to be rebound 1888 * are all disabled together in one go. 1889 */ 1890 cgrp_dfl_root.subsys_mask &= ~dfl_disable_ss_mask; 1891 WARN_ON(cgroup_apply_control(scgrp)); 1892 cgroup_finalize_control(scgrp, 0); 1893 } 1894 1895 do_each_subsys_mask(ss, ssid, ss_mask) { 1896 struct cgroup_root *src_root = ss->root; 1897 struct cgroup *scgrp = &src_root->cgrp; 1898 struct cgroup_subsys_state *css = cgroup_css(scgrp, ss); 1899 struct css_set *cset, *cset_pos; 1900 struct css_task_iter *it; 1901 1902 WARN_ON(!css || cgroup_css(dcgrp, ss)); 1903 1904 if (src_root != &cgrp_dfl_root) { 1905 /* disable from the source */ 1906 src_root->subsys_mask &= ~(1 << ssid); 1907 WARN_ON(cgroup_apply_control(scgrp)); 1908 cgroup_finalize_control(scgrp, 0); 1909 } 1910 1911 /* rebind */ 1912 RCU_INIT_POINTER(scgrp->subsys[ssid], NULL); 1913 rcu_assign_pointer(dcgrp->subsys[ssid], css); 1914 ss->root = dst_root; 1915 1916 spin_lock_irq(&css_set_lock); 1917 css->cgroup = dcgrp; 1918 WARN_ON(!list_empty(&dcgrp->e_csets[ss->id])); 1919 list_for_each_entry_safe(cset, cset_pos, &scgrp->e_csets[ss->id], 1920 e_cset_node[ss->id]) { 1921 list_move_tail(&cset->e_cset_node[ss->id], 1922 &dcgrp->e_csets[ss->id]); 1923 /* 1924 * all css_sets of scgrp together in same order to dcgrp, 1925 * patch in-flight iterators to preserve correct iteration. 1926 * since the iterator is always advanced right away and 1927 * finished when it->cset_pos meets it->cset_head, so only 1928 * update it->cset_head is enough here. 1929 */ 1930 list_for_each_entry(it, &cset->task_iters, iters_node) 1931 if (it->cset_head == &scgrp->e_csets[ss->id]) 1932 it->cset_head = &dcgrp->e_csets[ss->id]; 1933 } 1934 spin_unlock_irq(&css_set_lock); 1935 1936 /* default hierarchy doesn't enable controllers by default */ 1937 dst_root->subsys_mask |= 1 << ssid; 1938 if (dst_root == &cgrp_dfl_root) { 1939 static_branch_enable(cgroup_subsys_on_dfl_key[ssid]); 1940 } else { 1941 dcgrp->subtree_control |= 1 << ssid; 1942 static_branch_disable(cgroup_subsys_on_dfl_key[ssid]); 1943 } 1944 1945 ret = cgroup_apply_control(dcgrp); 1946 if (ret) 1947 pr_warn("partial failure to rebind %s controller (err=%d)\n", 1948 ss->name, ret); 1949 1950 if (ss->bind) 1951 ss->bind(css); 1952 } while_each_subsys_mask(); 1953 1954 kernfs_activate(dcgrp->kn); 1955 return 0; 1956} 1957 1958int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node, 1959 struct kernfs_root *kf_root) 1960{ 1961 int len = 0; 1962 char *buf = NULL; 1963 struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root); 1964 struct cgroup *ns_cgroup; 1965 1966 buf = kmalloc(PATH_MAX, GFP_KERNEL); 1967 if (!buf) 1968 return -ENOMEM; 1969 1970 spin_lock_irq(&css_set_lock); 1971 ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot); 1972 len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX); 1973 spin_unlock_irq(&css_set_lock); 1974 1975 if (len == -E2BIG) 1976 len = -ERANGE; 1977 else if (len > 0) { 1978 seq_escape(sf, buf, " \t\n\\"); 1979 len = 0; 1980 } 1981 kfree(buf); 1982 return len; 1983} 1984 1985enum cgroup2_param { 1986 Opt_nsdelegate, 1987 Opt_favordynmods, 1988 Opt_memory_localevents, 1989 Opt_memory_recursiveprot, 1990 Opt_memory_hugetlb_accounting, 1991 Opt_pids_localevents, 1992 nr__cgroup2_params 1993}; 1994 1995static const struct fs_parameter_spec cgroup2_fs_parameters[] = { 1996 fsparam_flag("nsdelegate", Opt_nsdelegate), 1997 fsparam_flag("favordynmods", Opt_favordynmods), 1998 fsparam_flag("memory_localevents", Opt_memory_localevents), 1999 fsparam_flag("memory_recursiveprot", Opt_memory_recursiveprot), 2000 fsparam_flag("memory_hugetlb_accounting", Opt_memory_hugetlb_accounting), 2001 fsparam_flag("pids_localevents", Opt_pids_localevents), 2002 {} 2003}; 2004 2005static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param) 2006{ 2007 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2008 struct fs_parse_result result; 2009 int opt; 2010 2011 opt = fs_parse(fc, cgroup2_fs_parameters, param, &result); 2012 if (opt < 0) 2013 return opt; 2014 2015 switch (opt) { 2016 case Opt_nsdelegate: 2017 ctx->flags |= CGRP_ROOT_NS_DELEGATE; 2018 return 0; 2019 case Opt_favordynmods: 2020 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS; 2021 return 0; 2022 case Opt_memory_localevents: 2023 ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS; 2024 return 0; 2025 case Opt_memory_recursiveprot: 2026 ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT; 2027 return 0; 2028 case Opt_memory_hugetlb_accounting: 2029 ctx->flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING; 2030 return 0; 2031 case Opt_pids_localevents: 2032 ctx->flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS; 2033 return 0; 2034 } 2035 return -EINVAL; 2036} 2037 2038struct cgroup_of_peak *of_peak(struct kernfs_open_file *of) 2039{ 2040 struct cgroup_file_ctx *ctx = of->priv; 2041 2042 return &ctx->peak; 2043} 2044 2045static void apply_cgroup_root_flags(unsigned int root_flags) 2046{ 2047 if (current->nsproxy->cgroup_ns == &init_cgroup_ns) { 2048 if (root_flags & CGRP_ROOT_NS_DELEGATE) 2049 cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE; 2050 else 2051 cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE; 2052 2053 cgroup_favor_dynmods(&cgrp_dfl_root, 2054 root_flags & CGRP_ROOT_FAVOR_DYNMODS); 2055 2056 if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS) 2057 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS; 2058 else 2059 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS; 2060 2061 if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT) 2062 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT; 2063 else 2064 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT; 2065 2066 if (root_flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING) 2067 cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING; 2068 else 2069 cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING; 2070 2071 if (root_flags & CGRP_ROOT_PIDS_LOCAL_EVENTS) 2072 cgrp_dfl_root.flags |= CGRP_ROOT_PIDS_LOCAL_EVENTS; 2073 else 2074 cgrp_dfl_root.flags &= ~CGRP_ROOT_PIDS_LOCAL_EVENTS; 2075 } 2076} 2077 2078static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root) 2079{ 2080 if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) 2081 seq_puts(seq, ",nsdelegate"); 2082 if (cgrp_dfl_root.flags & CGRP_ROOT_FAVOR_DYNMODS) 2083 seq_puts(seq, ",favordynmods"); 2084 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS) 2085 seq_puts(seq, ",memory_localevents"); 2086 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT) 2087 seq_puts(seq, ",memory_recursiveprot"); 2088 if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_HUGETLB_ACCOUNTING) 2089 seq_puts(seq, ",memory_hugetlb_accounting"); 2090 if (cgrp_dfl_root.flags & CGRP_ROOT_PIDS_LOCAL_EVENTS) 2091 seq_puts(seq, ",pids_localevents"); 2092 return 0; 2093} 2094 2095static int cgroup_reconfigure(struct fs_context *fc) 2096{ 2097 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2098 2099 apply_cgroup_root_flags(ctx->flags); 2100 return 0; 2101} 2102 2103static void init_cgroup_housekeeping(struct cgroup *cgrp) 2104{ 2105 struct cgroup_subsys *ss; 2106 int ssid; 2107 2108 INIT_LIST_HEAD(&cgrp->self.sibling); 2109 INIT_LIST_HEAD(&cgrp->self.children); 2110 INIT_LIST_HEAD(&cgrp->cset_links); 2111 INIT_LIST_HEAD(&cgrp->pidlists); 2112 mutex_init(&cgrp->pidlist_mutex); 2113 cgrp->self.cgroup = cgrp; 2114 cgrp->self.flags |= CSS_ONLINE; 2115 cgrp->dom_cgrp = cgrp; 2116 cgrp->max_descendants = INT_MAX; 2117 cgrp->max_depth = INT_MAX; 2118 prev_cputime_init(&cgrp->prev_cputime); 2119 2120 for_each_subsys(ss, ssid) 2121 INIT_LIST_HEAD(&cgrp->e_csets[ssid]); 2122 2123#ifdef CONFIG_CGROUP_BPF 2124 for (int i = 0; i < ARRAY_SIZE(cgrp->bpf.revisions); i++) 2125 cgrp->bpf.revisions[i] = 1; 2126#endif 2127 2128 init_waitqueue_head(&cgrp->offline_waitq); 2129 init_waitqueue_head(&cgrp->dying_populated_waitq); 2130 INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent); 2131} 2132 2133void init_cgroup_root(struct cgroup_fs_context *ctx) 2134{ 2135 struct cgroup_root *root = ctx->root; 2136 struct cgroup *cgrp = &root->cgrp; 2137 2138 INIT_LIST_HEAD_RCU(&root->root_list); 2139 atomic_set(&root->nr_cgrps, 1); 2140 cgrp->root = root; 2141 init_cgroup_housekeeping(cgrp); 2142 2143 /* DYNMODS must be modified through cgroup_favor_dynmods() */ 2144 root->flags = ctx->flags & ~CGRP_ROOT_FAVOR_DYNMODS; 2145 if (ctx->release_agent) 2146 strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX); 2147 if (ctx->name) 2148 strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN); 2149 if (ctx->cpuset_clone_children) 2150 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags); 2151} 2152 2153int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask) 2154{ 2155 LIST_HEAD(tmp_links); 2156 struct cgroup *root_cgrp = &root->cgrp; 2157 struct kernfs_syscall_ops *kf_sops; 2158 struct css_set *cset; 2159 int i, ret; 2160 2161 lockdep_assert_held(&cgroup_mutex); 2162 2163 ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release, 2164 0, GFP_KERNEL); 2165 if (ret) 2166 goto out; 2167 2168 /* 2169 * We're accessing css_set_count without locking css_set_lock here, 2170 * but that's OK - it can only be increased by someone holding 2171 * cgroup_lock, and that's us. Later rebinding may disable 2172 * controllers on the default hierarchy and thus create new csets, 2173 * which can't be more than the existing ones. Allocate 2x. 2174 */ 2175 ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links); 2176 if (ret) 2177 goto cancel_ref; 2178 2179 ret = cgroup_init_root_id(root); 2180 if (ret) 2181 goto cancel_ref; 2182 2183 kf_sops = root == &cgrp_dfl_root ? 2184 &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops; 2185 2186 root->kf_root = kernfs_create_root(kf_sops, 2187 KERNFS_ROOT_CREATE_DEACTIVATED | 2188 KERNFS_ROOT_SUPPORT_EXPORTOP | 2189 KERNFS_ROOT_SUPPORT_USER_XATTR | 2190 KERNFS_ROOT_INVARIANT_PARENT, 2191 root_cgrp); 2192 if (IS_ERR(root->kf_root)) { 2193 ret = PTR_ERR(root->kf_root); 2194 goto exit_root_id; 2195 } 2196 root_cgrp->kn = kernfs_root_to_node(root->kf_root); 2197 WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1); 2198 root_cgrp->ancestors[0] = root_cgrp; 2199 2200 ret = css_populate_dir(&root_cgrp->self); 2201 if (ret) 2202 goto destroy_root; 2203 2204 ret = css_rstat_init(&root_cgrp->self); 2205 if (ret) 2206 goto destroy_root; 2207 2208 ret = rebind_subsystems(root, ss_mask); 2209 if (ret) 2210 goto exit_stats; 2211 2212 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier, 2213 CGROUP_LIFETIME_ONLINE, root_cgrp); 2214 WARN_ON_ONCE(notifier_to_errno(ret)); 2215 2216 trace_cgroup_setup_root(root); 2217 2218 /* 2219 * There must be no failure case after here, since rebinding takes 2220 * care of subsystems' refcounts, which are explicitly dropped in 2221 * the failure exit path. 2222 */ 2223 list_add_rcu(&root->root_list, &cgroup_roots); 2224 cgroup_root_count++; 2225 2226 /* 2227 * Link the root cgroup in this hierarchy into all the css_set 2228 * objects. 2229 */ 2230 spin_lock_irq(&css_set_lock); 2231 hash_for_each(css_set_table, i, cset, hlist) { 2232 link_css_set(&tmp_links, cset, root_cgrp); 2233 if (css_set_populated(cset)) 2234 cgroup_update_populated(root_cgrp, true); 2235 } 2236 spin_unlock_irq(&css_set_lock); 2237 2238 BUG_ON(!list_empty(&root_cgrp->self.children)); 2239 BUG_ON(atomic_read(&root->nr_cgrps) != 1); 2240 2241 ret = 0; 2242 goto out; 2243 2244exit_stats: 2245 css_rstat_exit(&root_cgrp->self); 2246destroy_root: 2247 kernfs_destroy_root(root->kf_root); 2248 root->kf_root = NULL; 2249exit_root_id: 2250 cgroup_exit_root_id(root); 2251cancel_ref: 2252 percpu_ref_exit(&root_cgrp->self.refcnt); 2253out: 2254 free_cgrp_cset_links(&tmp_links); 2255 return ret; 2256} 2257 2258int cgroup_do_get_tree(struct fs_context *fc) 2259{ 2260 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2261 int ret; 2262 2263 ctx->kfc.root = ctx->root->kf_root; 2264 if (fc->fs_type == &cgroup2_fs_type) 2265 ctx->kfc.magic = CGROUP2_SUPER_MAGIC; 2266 else 2267 ctx->kfc.magic = CGROUP_SUPER_MAGIC; 2268 ret = kernfs_get_tree(fc); 2269 2270 /* 2271 * In non-init cgroup namespace, instead of root cgroup's dentry, 2272 * we return the dentry corresponding to the cgroupns->root_cgrp. 2273 */ 2274 if (!ret && ctx->ns != &init_cgroup_ns) { 2275 struct dentry *nsdentry; 2276 struct super_block *sb = fc->root->d_sb; 2277 struct cgroup *cgrp; 2278 2279 cgroup_lock(); 2280 spin_lock_irq(&css_set_lock); 2281 2282 cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root); 2283 2284 spin_unlock_irq(&css_set_lock); 2285 cgroup_unlock(); 2286 2287 nsdentry = kernfs_node_dentry(cgrp->kn, sb); 2288 dput(fc->root); 2289 if (IS_ERR(nsdentry)) { 2290 deactivate_locked_super(sb); 2291 ret = PTR_ERR(nsdentry); 2292 nsdentry = NULL; 2293 } 2294 fc->root = nsdentry; 2295 } 2296 2297 if (!ctx->kfc.new_sb_created) 2298 cgroup_put(&ctx->root->cgrp); 2299 2300 return ret; 2301} 2302 2303/* 2304 * Destroy a cgroup filesystem context. 2305 */ 2306static void cgroup_fs_context_free(struct fs_context *fc) 2307{ 2308 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2309 2310 kfree(ctx->name); 2311 kfree(ctx->release_agent); 2312 put_cgroup_ns(ctx->ns); 2313 kernfs_free_fs_context(fc); 2314 kfree(ctx); 2315} 2316 2317static int cgroup_get_tree(struct fs_context *fc) 2318{ 2319 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2320 int ret; 2321 2322 WRITE_ONCE(cgrp_dfl_visible, true); 2323 cgroup_get_live(&cgrp_dfl_root.cgrp); 2324 ctx->root = &cgrp_dfl_root; 2325 2326 ret = cgroup_do_get_tree(fc); 2327 if (!ret) 2328 apply_cgroup_root_flags(ctx->flags); 2329 return ret; 2330} 2331 2332static const struct fs_context_operations cgroup_fs_context_ops = { 2333 .free = cgroup_fs_context_free, 2334 .parse_param = cgroup2_parse_param, 2335 .get_tree = cgroup_get_tree, 2336 .reconfigure = cgroup_reconfigure, 2337}; 2338 2339static const struct fs_context_operations cgroup1_fs_context_ops = { 2340 .free = cgroup_fs_context_free, 2341 .parse_param = cgroup1_parse_param, 2342 .get_tree = cgroup1_get_tree, 2343 .reconfigure = cgroup1_reconfigure, 2344}; 2345 2346/* 2347 * Initialise the cgroup filesystem creation/reconfiguration context. Notably, 2348 * we select the namespace we're going to use. 2349 */ 2350static int cgroup_init_fs_context(struct fs_context *fc) 2351{ 2352 struct cgroup_fs_context *ctx; 2353 2354 ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL); 2355 if (!ctx) 2356 return -ENOMEM; 2357 2358 ctx->ns = current->nsproxy->cgroup_ns; 2359 get_cgroup_ns(ctx->ns); 2360 fc->fs_private = &ctx->kfc; 2361 if (fc->fs_type == &cgroup2_fs_type) 2362 fc->ops = &cgroup_fs_context_ops; 2363 else 2364 fc->ops = &cgroup1_fs_context_ops; 2365 put_user_ns(fc->user_ns); 2366 fc->user_ns = get_user_ns(ctx->ns->user_ns); 2367 fc->global = true; 2368 2369 if (have_favordynmods) 2370 ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS; 2371 2372 return 0; 2373} 2374 2375static void cgroup_kill_sb(struct super_block *sb) 2376{ 2377 struct kernfs_root *kf_root = kernfs_root_from_sb(sb); 2378 struct cgroup_root *root = cgroup_root_from_kf(kf_root); 2379 2380 /* 2381 * If @root doesn't have any children, start killing it. 2382 * This prevents new mounts by disabling percpu_ref_tryget_live(). 2383 * 2384 * And don't kill the default root. 2385 */ 2386 if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root && 2387 !percpu_ref_is_dying(&root->cgrp.self.refcnt)) 2388 percpu_ref_kill(&root->cgrp.self.refcnt); 2389 cgroup_put(&root->cgrp); 2390 kernfs_kill_sb(sb); 2391} 2392 2393struct file_system_type cgroup_fs_type = { 2394 .name = "cgroup", 2395 .init_fs_context = cgroup_init_fs_context, 2396 .parameters = cgroup1_fs_parameters, 2397 .kill_sb = cgroup_kill_sb, 2398 .fs_flags = FS_USERNS_MOUNT, 2399}; 2400 2401static struct file_system_type cgroup2_fs_type = { 2402 .name = "cgroup2", 2403 .init_fs_context = cgroup_init_fs_context, 2404 .parameters = cgroup2_fs_parameters, 2405 .kill_sb = cgroup_kill_sb, 2406 .fs_flags = FS_USERNS_MOUNT, 2407}; 2408 2409#ifdef CONFIG_CPUSETS_V1 2410enum cpuset_param { 2411 Opt_cpuset_v2_mode, 2412}; 2413 2414static const struct fs_parameter_spec cpuset_fs_parameters[] = { 2415 fsparam_flag ("cpuset_v2_mode", Opt_cpuset_v2_mode), 2416 {} 2417}; 2418 2419static int cpuset_parse_param(struct fs_context *fc, struct fs_parameter *param) 2420{ 2421 struct cgroup_fs_context *ctx = cgroup_fc2context(fc); 2422 struct fs_parse_result result; 2423 int opt; 2424 2425 opt = fs_parse(fc, cpuset_fs_parameters, param, &result); 2426 if (opt < 0) 2427 return opt; 2428 2429 switch (opt) { 2430 case Opt_cpuset_v2_mode: 2431 ctx->flags |= CGRP_ROOT_CPUSET_V2_MODE; 2432 return 0; 2433 } 2434 return -EINVAL; 2435} 2436 2437static const struct fs_context_operations cpuset_fs_context_ops = { 2438 .get_tree = cgroup1_get_tree, 2439 .free = cgroup_fs_context_free, 2440 .parse_param = cpuset_parse_param, 2441}; 2442 2443/* 2444 * This is ugly, but preserves the userspace API for existing cpuset 2445 * users. If someone tries to mount the "cpuset" filesystem, we 2446 * silently switch it to mount "cgroup" instead 2447 */ 2448static int cpuset_init_fs_context(struct fs_context *fc) 2449{ 2450 char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER); 2451 struct cgroup_fs_context *ctx; 2452 int err; 2453 2454 err = cgroup_init_fs_context(fc); 2455 if (err) { 2456 kfree(agent); 2457 return err; 2458 } 2459 2460 fc->ops = &cpuset_fs_context_ops; 2461 2462 ctx = cgroup_fc2context(fc); 2463 ctx->subsys_mask = 1 << cpuset_cgrp_id; 2464 ctx->flags |= CGRP_ROOT_NOPREFIX; 2465 ctx->release_agent = agent; 2466 2467 get_filesystem(&cgroup_fs_type); 2468 put_filesystem(fc->fs_type); 2469 fc->fs_type = &cgroup_fs_type; 2470 2471 return 0; 2472} 2473 2474static struct file_system_type cpuset_fs_type = { 2475 .name = "cpuset", 2476 .init_fs_context = cpuset_init_fs_context, 2477 .parameters = cpuset_fs_parameters, 2478 .fs_flags = FS_USERNS_MOUNT, 2479}; 2480#endif 2481 2482int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen, 2483 struct cgroup_namespace *ns) 2484{ 2485 struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root); 2486 2487 return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen); 2488} 2489 2490int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen, 2491 struct cgroup_namespace *ns) 2492{ 2493 int ret; 2494 2495 cgroup_lock(); 2496 spin_lock_irq(&css_set_lock); 2497 2498 ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns); 2499 2500 spin_unlock_irq(&css_set_lock); 2501 cgroup_unlock(); 2502 2503 return ret; 2504} 2505EXPORT_SYMBOL_GPL(cgroup_path_ns); 2506 2507/** 2508 * cgroup_attach_lock - Lock for ->attach() 2509 * @lock_mode: whether acquire and acquire which rwsem 2510 * @tsk: thread group to lock 2511 * 2512 * cgroup migration sometimes needs to stabilize threadgroups against forks and 2513 * exits by write-locking cgroup_threadgroup_rwsem. However, some ->attach() 2514 * implementations (e.g. cpuset), also need to disable CPU hotplug. 2515 * Unfortunately, letting ->attach() operations acquire cpus_read_lock() can 2516 * lead to deadlocks. 2517 * 2518 * Bringing up a CPU may involve creating and destroying tasks which requires 2519 * read-locking threadgroup_rwsem, so threadgroup_rwsem nests inside 2520 * cpus_read_lock(). If we call an ->attach() which acquires the cpus lock while 2521 * write-locking threadgroup_rwsem, the locking order is reversed and we end up 2522 * waiting for an on-going CPU hotplug operation which in turn is waiting for 2523 * the threadgroup_rwsem to be released to create new tasks. For more details: 2524 * 2525 * http://lkml.kernel.org/r/20220711174629.uehfmqegcwn2lqzu@wubuntu 2526 * 2527 * Resolve the situation by always acquiring cpus_read_lock() before optionally 2528 * write-locking cgroup_threadgroup_rwsem. This allows ->attach() to assume that 2529 * CPU hotplug is disabled on entry. 2530 * 2531 * When favordynmods is enabled, take per threadgroup rwsem to reduce overhead 2532 * on dynamic cgroup modifications. see the comment above 2533 * CGRP_ROOT_FAVOR_DYNMODS definition. 2534 * 2535 * tsk is not NULL only when writing to cgroup.procs. 2536 */ 2537void cgroup_attach_lock(enum cgroup_attach_lock_mode lock_mode, 2538 struct task_struct *tsk) 2539{ 2540 cpus_read_lock(); 2541 2542 switch (lock_mode) { 2543 case CGRP_ATTACH_LOCK_NONE: 2544 break; 2545 case CGRP_ATTACH_LOCK_GLOBAL: 2546 percpu_down_write(&cgroup_threadgroup_rwsem); 2547 break; 2548 case CGRP_ATTACH_LOCK_PER_THREADGROUP: 2549 down_write(&tsk->signal->cgroup_threadgroup_rwsem); 2550 break; 2551 default: 2552 pr_warn("cgroup: Unexpected attach lock mode."); 2553 break; 2554 } 2555} 2556 2557/** 2558 * cgroup_attach_unlock - Undo cgroup_attach_lock() 2559 * @lock_mode: whether release and release which rwsem 2560 * @tsk: thread group to lock 2561 */ 2562void cgroup_attach_unlock(enum cgroup_attach_lock_mode lock_mode, 2563 struct task_struct *tsk) 2564{ 2565 switch (lock_mode) { 2566 case CGRP_ATTACH_LOCK_NONE: 2567 break; 2568 case CGRP_ATTACH_LOCK_GLOBAL: 2569 percpu_up_write(&cgroup_threadgroup_rwsem); 2570 break; 2571 case CGRP_ATTACH_LOCK_PER_THREADGROUP: 2572 up_write(&tsk->signal->cgroup_threadgroup_rwsem); 2573 break; 2574 default: 2575 pr_warn("cgroup: Unexpected attach lock mode."); 2576 break; 2577 } 2578 2579 cpus_read_unlock(); 2580} 2581 2582/** 2583 * cgroup_migrate_add_task - add a migration target task to a migration context 2584 * @task: target task 2585 * @mgctx: target migration context 2586 * 2587 * Add @task, which is a migration target, to @mgctx->tset. This function 2588 * becomes noop if @task doesn't need to be migrated. @task's css_set 2589 * should have been added as a migration source and @task->cg_list will be 2590 * moved from the css_set's tasks list to mg_tasks one. 2591 */ 2592static void cgroup_migrate_add_task(struct task_struct *task, 2593 struct cgroup_mgctx *mgctx) 2594{ 2595 struct css_set *cset; 2596 2597 lockdep_assert_held(&css_set_lock); 2598 2599 /* @task either already exited or can't exit until the end */ 2600 if (task->flags & PF_EXITING) 2601 return; 2602 2603 /* cgroup_threadgroup_rwsem protects racing against forks */ 2604 WARN_ON_ONCE(list_empty(&task->cg_list)); 2605 2606 cset = task_css_set(task); 2607 if (!cset->mg_src_cgrp) 2608 return; 2609 2610 mgctx->tset.nr_tasks++; 2611 2612 css_set_skip_task_iters(cset, task); 2613 list_move_tail(&task->cg_list, &cset->mg_tasks); 2614 if (list_empty(&cset->mg_node)) 2615 list_add_tail(&cset->mg_node, 2616 &mgctx->tset.src_csets); 2617 if (list_empty(&cset->mg_dst_cset->mg_node)) 2618 list_add_tail(&cset->mg_dst_cset->mg_node, 2619 &mgctx->tset.dst_csets); 2620} 2621 2622/** 2623 * cgroup_taskset_first - reset taskset and return the first task 2624 * @tset: taskset of interest 2625 * @dst_cssp: output variable for the destination css 2626 * 2627 * @tset iteration is initialized and the first task is returned. 2628 */ 2629struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset, 2630 struct cgroup_subsys_state **dst_cssp) 2631{ 2632 tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node); 2633 tset->cur_task = NULL; 2634 2635 return cgroup_taskset_next(tset, dst_cssp); 2636} 2637 2638/** 2639 * cgroup_taskset_next - iterate to the next task in taskset 2640 * @tset: taskset of interest 2641 * @dst_cssp: output variable for the destination css 2642 * 2643 * Return the next task in @tset. Iteration must have been initialized 2644 * with cgroup_taskset_first(). 2645 */ 2646struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset, 2647 struct cgroup_subsys_state **dst_cssp) 2648{ 2649 struct css_set *cset = tset->cur_cset; 2650 struct task_struct *task = tset->cur_task; 2651 2652 while (CGROUP_HAS_SUBSYS_CONFIG && &cset->mg_node != tset->csets) { 2653 if (!task) 2654 task = list_first_entry(&cset->mg_tasks, 2655 struct task_struct, cg_list); 2656 else 2657 task = list_next_entry(task, cg_list); 2658 2659 if (&task->cg_list != &cset->mg_tasks) { 2660 tset->cur_cset = cset; 2661 tset->cur_task = task; 2662 2663 /* 2664 * This function may be called both before and 2665 * after cgroup_migrate_execute(). The two cases 2666 * can be distinguished by looking at whether @cset 2667 * has its ->mg_dst_cset set. 2668 */ 2669 if (cset->mg_dst_cset) 2670 *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid]; 2671 else 2672 *dst_cssp = cset->subsys[tset->ssid]; 2673 2674 return task; 2675 } 2676 2677 cset = list_next_entry(cset, mg_node); 2678 task = NULL; 2679 } 2680 2681 return NULL; 2682} 2683 2684/** 2685 * cgroup_migrate_execute - migrate a taskset 2686 * @mgctx: migration context 2687 * 2688 * Migrate tasks in @mgctx as setup by migration preparation functions. 2689 * This function fails iff one of the ->can_attach callbacks fails and 2690 * guarantees that either all or none of the tasks in @mgctx are migrated. 2691 * @mgctx is consumed regardless of success. 2692 */ 2693static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx) 2694{ 2695 struct cgroup_taskset *tset = &mgctx->tset; 2696 struct cgroup_subsys *ss; 2697 struct task_struct *task, *tmp_task; 2698 struct css_set *cset, *tmp_cset; 2699 int ssid, failed_ssid, ret; 2700 2701 /* check that we can legitimately attach to the cgroup */ 2702 if (tset->nr_tasks) { 2703 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2704 if (ss->can_attach) { 2705 tset->ssid = ssid; 2706 ret = ss->can_attach(tset); 2707 if (ret) { 2708 failed_ssid = ssid; 2709 goto out_cancel_attach; 2710 } 2711 } 2712 } while_each_subsys_mask(); 2713 } 2714 2715 /* 2716 * Now that we're guaranteed success, proceed to move all tasks to 2717 * the new cgroup. There are no failure cases after here, so this 2718 * is the commit point. 2719 */ 2720 spin_lock_irq(&css_set_lock); 2721 list_for_each_entry(cset, &tset->src_csets, mg_node) { 2722 list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) { 2723 struct css_set *from_cset = task_css_set(task); 2724 struct css_set *to_cset = cset->mg_dst_cset; 2725 2726 get_css_set(to_cset); 2727 to_cset->nr_tasks++; 2728 css_set_move_task(task, from_cset, to_cset, true); 2729 from_cset->nr_tasks--; 2730 /* 2731 * If the source or destination cgroup is frozen, 2732 * the task might require to change its state. 2733 */ 2734 cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp, 2735 to_cset->dfl_cgrp); 2736 put_css_set_locked(from_cset); 2737 2738 } 2739 } 2740 spin_unlock_irq(&css_set_lock); 2741 2742 /* 2743 * Migration is committed, all target tasks are now on dst_csets. 2744 * Nothing is sensitive to fork() after this point. Notify 2745 * controllers that migration is complete. 2746 */ 2747 tset->csets = &tset->dst_csets; 2748 2749 if (tset->nr_tasks) { 2750 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2751 if (ss->attach) { 2752 tset->ssid = ssid; 2753 ss->attach(tset); 2754 } 2755 } while_each_subsys_mask(); 2756 } 2757 2758 ret = 0; 2759 goto out_release_tset; 2760 2761out_cancel_attach: 2762 if (tset->nr_tasks) { 2763 do_each_subsys_mask(ss, ssid, mgctx->ss_mask) { 2764 if (ssid == failed_ssid) 2765 break; 2766 if (ss->cancel_attach) { 2767 tset->ssid = ssid; 2768 ss->cancel_attach(tset); 2769 } 2770 } while_each_subsys_mask(); 2771 } 2772out_release_tset: 2773 spin_lock_irq(&css_set_lock); 2774 list_splice_init(&tset->dst_csets, &tset->src_csets); 2775 list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) { 2776 list_splice_tail_init(&cset->mg_tasks, &cset->tasks); 2777 list_del_init(&cset->mg_node); 2778 } 2779 spin_unlock_irq(&css_set_lock); 2780 2781 /* 2782 * Re-initialize the cgroup_taskset structure in case it is reused 2783 * again in another cgroup_migrate_add_task()/cgroup_migrate_execute() 2784 * iteration. 2785 */ 2786 tset->nr_tasks = 0; 2787 tset->csets = &tset->src_csets; 2788 return ret; 2789} 2790 2791/** 2792 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination 2793 * @dst_cgrp: destination cgroup to test 2794 * 2795 * On the default hierarchy, except for the mixable, (possible) thread root 2796 * and threaded cgroups, subtree_control must be zero for migration 2797 * destination cgroups with tasks so that child cgroups don't compete 2798 * against tasks. 2799 */ 2800int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp) 2801{ 2802 /* v1 doesn't have any restriction */ 2803 if (!cgroup_on_dfl(dst_cgrp)) 2804 return 0; 2805 2806 /* verify @dst_cgrp can host resources */ 2807 if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp)) 2808 return -EOPNOTSUPP; 2809 2810 /* 2811 * If @dst_cgrp is already or can become a thread root or is 2812 * threaded, it doesn't matter. 2813 */ 2814 if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp)) 2815 return 0; 2816 2817 /* apply no-internal-process constraint */ 2818 if (dst_cgrp->subtree_control) 2819 return -EBUSY; 2820 2821 return 0; 2822} 2823 2824/** 2825 * cgroup_migrate_finish - cleanup after attach 2826 * @mgctx: migration context 2827 * 2828 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst(). See 2829 * those functions for details. 2830 */ 2831void cgroup_migrate_finish(struct cgroup_mgctx *mgctx) 2832{ 2833 struct css_set *cset, *tmp_cset; 2834 2835 lockdep_assert_held(&cgroup_mutex); 2836 2837 spin_lock_irq(&css_set_lock); 2838 2839 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_src_csets, 2840 mg_src_preload_node) { 2841 cset->mg_src_cgrp = NULL; 2842 cset->mg_dst_cgrp = NULL; 2843 cset->mg_dst_cset = NULL; 2844 list_del_init(&cset->mg_src_preload_node); 2845 put_css_set_locked(cset); 2846 } 2847 2848 list_for_each_entry_safe(cset, tmp_cset, &mgctx->preloaded_dst_csets, 2849 mg_dst_preload_node) { 2850 cset->mg_src_cgrp = NULL; 2851 cset->mg_dst_cgrp = NULL; 2852 cset->mg_dst_cset = NULL; 2853 list_del_init(&cset->mg_dst_preload_node); 2854 put_css_set_locked(cset); 2855 } 2856 2857 spin_unlock_irq(&css_set_lock); 2858} 2859 2860/** 2861 * cgroup_migrate_add_src - add a migration source css_set 2862 * @src_cset: the source css_set to add 2863 * @dst_cgrp: the destination cgroup 2864 * @mgctx: migration context 2865 * 2866 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp. Pin 2867 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned 2868 * up by cgroup_migrate_finish(). 2869 * 2870 * This function may be called without holding cgroup_threadgroup_rwsem 2871 * even if the target is a process. Threads may be created and destroyed 2872 * but as long as cgroup_mutex is not dropped, no new css_set can be put 2873 * into play and the preloaded css_sets are guaranteed to cover all 2874 * migrations. 2875 */ 2876void cgroup_migrate_add_src(struct css_set *src_cset, 2877 struct cgroup *dst_cgrp, 2878 struct cgroup_mgctx *mgctx) 2879{ 2880 struct cgroup *src_cgrp; 2881 2882 lockdep_assert_held(&cgroup_mutex); 2883 lockdep_assert_held(&css_set_lock); 2884 2885 /* 2886 * If ->dead, @src_set is associated with one or more dead cgroups 2887 * and doesn't contain any migratable tasks. Ignore it early so 2888 * that the rest of migration path doesn't get confused by it. 2889 */ 2890 if (src_cset->dead) 2891 return; 2892 2893 if (!list_empty(&src_cset->mg_src_preload_node)) 2894 return; 2895 2896 src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root); 2897 2898 WARN_ON(src_cset->mg_src_cgrp); 2899 WARN_ON(src_cset->mg_dst_cgrp); 2900 WARN_ON(!list_empty(&src_cset->mg_tasks)); 2901 WARN_ON(!list_empty(&src_cset->mg_node)); 2902 2903 src_cset->mg_src_cgrp = src_cgrp; 2904 src_cset->mg_dst_cgrp = dst_cgrp; 2905 get_css_set(src_cset); 2906 list_add_tail(&src_cset->mg_src_preload_node, &mgctx->preloaded_src_csets); 2907} 2908 2909/** 2910 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration 2911 * @mgctx: migration context 2912 * 2913 * Tasks are about to be moved and all the source css_sets have been 2914 * preloaded to @mgctx->preloaded_src_csets. This function looks up and 2915 * pins all destination css_sets, links each to its source, and append them 2916 * to @mgctx->preloaded_dst_csets. 2917 * 2918 * This function must be called after cgroup_migrate_add_src() has been 2919 * called on each migration source css_set. After migration is performed 2920 * using cgroup_migrate(), cgroup_migrate_finish() must be called on 2921 * @mgctx. 2922 */ 2923int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx) 2924{ 2925 struct css_set *src_cset, *tmp_cset; 2926 2927 lockdep_assert_held(&cgroup_mutex); 2928 2929 /* look up the dst cset for each src cset and link it to src */ 2930 list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets, 2931 mg_src_preload_node) { 2932 struct css_set *dst_cset; 2933 struct cgroup_subsys *ss; 2934 int ssid; 2935 2936 dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp); 2937 if (!dst_cset) 2938 return -ENOMEM; 2939 2940 WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset); 2941 2942 /* 2943 * If src cset equals dst, it's noop. Drop the src. 2944 * cgroup_migrate() will skip the cset too. Note that we 2945 * can't handle src == dst as some nodes are used by both. 2946 */ 2947 if (src_cset == dst_cset) { 2948 src_cset->mg_src_cgrp = NULL; 2949 src_cset->mg_dst_cgrp = NULL; 2950 list_del_init(&src_cset->mg_src_preload_node); 2951 put_css_set(src_cset); 2952 put_css_set(dst_cset); 2953 continue; 2954 } 2955 2956 src_cset->mg_dst_cset = dst_cset; 2957 2958 if (list_empty(&dst_cset->mg_dst_preload_node)) 2959 list_add_tail(&dst_cset->mg_dst_preload_node, 2960 &mgctx->preloaded_dst_csets); 2961 else 2962 put_css_set(dst_cset); 2963 2964 for_each_subsys(ss, ssid) 2965 if (src_cset->subsys[ssid] != dst_cset->subsys[ssid]) 2966 mgctx->ss_mask |= 1 << ssid; 2967 } 2968 2969 return 0; 2970} 2971 2972/** 2973 * cgroup_migrate - migrate a process or task to a cgroup 2974 * @leader: the leader of the process or the task to migrate 2975 * @threadgroup: whether @leader points to the whole process or a single task 2976 * @mgctx: migration context 2977 * 2978 * Migrate a process or task denoted by @leader. If migrating a process, 2979 * the caller must be holding cgroup_threadgroup_rwsem. The caller is also 2980 * responsible for invoking cgroup_migrate_add_src() and 2981 * cgroup_migrate_prepare_dst() on the targets before invoking this 2982 * function and following up with cgroup_migrate_finish(). 2983 * 2984 * As long as a controller's ->can_attach() doesn't fail, this function is 2985 * guaranteed to succeed. This means that, excluding ->can_attach() 2986 * failure, when migrating multiple targets, the success or failure can be 2987 * decided for all targets by invoking group_migrate_prepare_dst() before 2988 * actually starting migrating. 2989 */ 2990int cgroup_migrate(struct task_struct *leader, bool threadgroup, 2991 struct cgroup_mgctx *mgctx) 2992{ 2993 struct task_struct *task; 2994 2995 /* 2996 * The following thread iteration should be inside an RCU critical 2997 * section to prevent tasks from being freed while taking the snapshot. 2998 * spin_lock_irq() implies RCU critical section here. 2999 */ 3000 spin_lock_irq(&css_set_lock); 3001 task = leader; 3002 do { 3003 cgroup_migrate_add_task(task, mgctx); 3004 if (!threadgroup) 3005 break; 3006 } while_each_thread(leader, task); 3007 spin_unlock_irq(&css_set_lock); 3008 3009 return cgroup_migrate_execute(mgctx); 3010} 3011 3012/** 3013 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup 3014 * @dst_cgrp: the cgroup to attach to 3015 * @leader: the task or the leader of the threadgroup to be attached 3016 * @threadgroup: attach the whole threadgroup? 3017 * 3018 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem. 3019 */ 3020int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader, 3021 bool threadgroup) 3022{ 3023 DEFINE_CGROUP_MGCTX(mgctx); 3024 struct task_struct *task; 3025 int ret = 0; 3026 3027 /* look up all src csets */ 3028 spin_lock_irq(&css_set_lock); 3029 task = leader; 3030 do { 3031 cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx); 3032 if (!threadgroup) 3033 break; 3034 } while_each_thread(leader, task); 3035 spin_unlock_irq(&css_set_lock); 3036 3037 /* prepare dst csets and commit */ 3038 ret = cgroup_migrate_prepare_dst(&mgctx); 3039 if (!ret) 3040 ret = cgroup_migrate(leader, threadgroup, &mgctx); 3041 3042 cgroup_migrate_finish(&mgctx); 3043 3044 if (!ret) 3045 TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup); 3046 3047 return ret; 3048} 3049 3050struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup, 3051 enum cgroup_attach_lock_mode *lock_mode) 3052{ 3053 struct task_struct *tsk; 3054 pid_t pid; 3055 3056 if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0) 3057 return ERR_PTR(-EINVAL); 3058 3059retry_find_task: 3060 rcu_read_lock(); 3061 if (pid) { 3062 tsk = find_task_by_vpid(pid); 3063 if (!tsk) { 3064 tsk = ERR_PTR(-ESRCH); 3065 goto out_unlock_rcu; 3066 } 3067 } else { 3068 tsk = current; 3069 } 3070 3071 if (threadgroup) 3072 tsk = tsk->group_leader; 3073 3074 /* 3075 * kthreads may acquire PF_NO_SETAFFINITY during initialization. 3076 * If userland migrates such a kthread to a non-root cgroup, it can 3077 * become trapped in a cpuset, or RT kthread may be born in a 3078 * cgroup with no rt_runtime allocated. Just say no. 3079 */ 3080 if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) { 3081 tsk = ERR_PTR(-EINVAL); 3082 goto out_unlock_rcu; 3083 } 3084 get_task_struct(tsk); 3085 rcu_read_unlock(); 3086 3087 /* 3088 * If we migrate a single thread, we don't care about threadgroup 3089 * stability. If the thread is `current`, it won't exit(2) under our 3090 * hands or change PID through exec(2). We exclude 3091 * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write callers 3092 * by cgroup_mutex. Therefore, we can skip the global lock. 3093 */ 3094 lockdep_assert_held(&cgroup_mutex); 3095 3096 if (pid || threadgroup) { 3097 if (cgroup_enable_per_threadgroup_rwsem) 3098 *lock_mode = CGRP_ATTACH_LOCK_PER_THREADGROUP; 3099 else 3100 *lock_mode = CGRP_ATTACH_LOCK_GLOBAL; 3101 } else { 3102 *lock_mode = CGRP_ATTACH_LOCK_NONE; 3103 } 3104 3105 cgroup_attach_lock(*lock_mode, tsk); 3106 3107 if (threadgroup) { 3108 if (!thread_group_leader(tsk)) { 3109 /* 3110 * A race with de_thread from another thread's exec() 3111 * may strip us of our leadership. If this happens, 3112 * throw this task away and try again. 3113 */ 3114 cgroup_attach_unlock(*lock_mode, tsk); 3115 put_task_struct(tsk); 3116 goto retry_find_task; 3117 } 3118 } 3119 3120 return tsk; 3121 3122out_unlock_rcu: 3123 rcu_read_unlock(); 3124 return tsk; 3125} 3126 3127void cgroup_procs_write_finish(struct task_struct *task, 3128 enum cgroup_attach_lock_mode lock_mode) 3129{ 3130 cgroup_attach_unlock(lock_mode, task); 3131 3132 /* release reference from cgroup_procs_write_start() */ 3133 put_task_struct(task); 3134} 3135 3136static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask) 3137{ 3138 struct cgroup_subsys *ss; 3139 bool printed = false; 3140 int ssid; 3141 3142 do_each_subsys_mask(ss, ssid, ss_mask) { 3143 if (printed) 3144 seq_putc(seq, ' '); 3145 seq_puts(seq, ss->name); 3146 printed = true; 3147 } while_each_subsys_mask(); 3148 if (printed) 3149 seq_putc(seq, '\n'); 3150} 3151 3152/* show controllers which are enabled from the parent */ 3153static int cgroup_controllers_show(struct seq_file *seq, void *v) 3154{ 3155 struct cgroup *cgrp = seq_css(seq)->cgroup; 3156 3157 cgroup_print_ss_mask(seq, cgroup_control(cgrp)); 3158 return 0; 3159} 3160 3161/* show controllers which are enabled for a given cgroup's children */ 3162static int cgroup_subtree_control_show(struct seq_file *seq, void *v) 3163{ 3164 struct cgroup *cgrp = seq_css(seq)->cgroup; 3165 3166 cgroup_print_ss_mask(seq, cgrp->subtree_control); 3167 return 0; 3168} 3169 3170/** 3171 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy 3172 * @cgrp: root of the subtree to update csses for 3173 * 3174 * @cgrp's control masks have changed and its subtree's css associations 3175 * need to be updated accordingly. This function looks up all css_sets 3176 * which are attached to the subtree, creates the matching updated css_sets 3177 * and migrates the tasks to the new ones. 3178 */ 3179static int cgroup_update_dfl_csses(struct cgroup *cgrp) 3180{ 3181 DEFINE_CGROUP_MGCTX(mgctx); 3182 struct cgroup_subsys_state *d_css; 3183 struct cgroup *dsct; 3184 struct css_set *src_cset; 3185 enum cgroup_attach_lock_mode lock_mode; 3186 bool has_tasks; 3187 int ret; 3188 3189 lockdep_assert_held(&cgroup_mutex); 3190 3191 /* look up all csses currently attached to @cgrp's subtree */ 3192 spin_lock_irq(&css_set_lock); 3193 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 3194 struct cgrp_cset_link *link; 3195 3196 /* 3197 * As cgroup_update_dfl_csses() is only called by 3198 * cgroup_apply_control(). The csses associated with the 3199 * given cgrp will not be affected by changes made to 3200 * its subtree_control file. We can skip them. 3201 */ 3202 if (dsct == cgrp) 3203 continue; 3204 3205 list_for_each_entry(link, &dsct->cset_links, cset_link) 3206 cgroup_migrate_add_src(link->cset, dsct, &mgctx); 3207 } 3208 spin_unlock_irq(&css_set_lock); 3209 3210 /* 3211 * We need to write-lock threadgroup_rwsem while migrating tasks. 3212 * However, if there are no source csets for @cgrp, changing its 3213 * controllers isn't gonna produce any task migrations and the 3214 * write-locking can be skipped safely. 3215 */ 3216 has_tasks = !list_empty(&mgctx.preloaded_src_csets); 3217 3218 if (has_tasks) 3219 lock_mode = CGRP_ATTACH_LOCK_GLOBAL; 3220 else 3221 lock_mode = CGRP_ATTACH_LOCK_NONE; 3222 3223 cgroup_attach_lock(lock_mode, NULL); 3224 3225 /* NULL dst indicates self on default hierarchy */ 3226 ret = cgroup_migrate_prepare_dst(&mgctx); 3227 if (ret) 3228 goto out_finish; 3229 3230 spin_lock_irq(&css_set_lock); 3231 list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, 3232 mg_src_preload_node) { 3233 struct task_struct *task, *ntask; 3234 3235 /* all tasks in src_csets need to be migrated */ 3236 list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list) 3237 cgroup_migrate_add_task(task, &mgctx); 3238 } 3239 spin_unlock_irq(&css_set_lock); 3240 3241 ret = cgroup_migrate_execute(&mgctx); 3242out_finish: 3243 cgroup_migrate_finish(&mgctx); 3244 cgroup_attach_unlock(lock_mode, NULL); 3245 return ret; 3246} 3247 3248/** 3249 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses 3250 * @cgrp: root of the target subtree 3251 * 3252 * Because css offlining is asynchronous, userland may try to re-enable a 3253 * controller while the previous css is still around. This function grabs 3254 * cgroup_mutex and drains the previous css instances of @cgrp's subtree. 3255 */ 3256void cgroup_lock_and_drain_offline(struct cgroup *cgrp) 3257 __acquires(&cgroup_mutex) 3258{ 3259 struct cgroup *dsct; 3260 struct cgroup_subsys_state *d_css; 3261 struct cgroup_subsys *ss; 3262 int ssid; 3263 3264restart: 3265 cgroup_lock(); 3266 3267 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 3268 for_each_subsys(ss, ssid) { 3269 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 3270 DEFINE_WAIT(wait); 3271 3272 if (!css || !percpu_ref_is_dying(&css->refcnt)) 3273 continue; 3274 3275 cgroup_get_live(dsct); 3276 prepare_to_wait(&dsct->offline_waitq, &wait, 3277 TASK_UNINTERRUPTIBLE); 3278 3279 cgroup_unlock(); 3280 schedule(); 3281 finish_wait(&dsct->offline_waitq, &wait); 3282 3283 cgroup_put(dsct); 3284 goto restart; 3285 } 3286 } 3287} 3288 3289/** 3290 * cgroup_save_control - save control masks and dom_cgrp of a subtree 3291 * @cgrp: root of the target subtree 3292 * 3293 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the 3294 * respective old_ prefixed fields for @cgrp's subtree including @cgrp 3295 * itself. 3296 */ 3297static void cgroup_save_control(struct cgroup *cgrp) 3298{ 3299 struct cgroup *dsct; 3300 struct cgroup_subsys_state *d_css; 3301 3302 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 3303 dsct->old_subtree_control = dsct->subtree_control; 3304 dsct->old_subtree_ss_mask = dsct->subtree_ss_mask; 3305 dsct->old_dom_cgrp = dsct->dom_cgrp; 3306 } 3307} 3308 3309/** 3310 * cgroup_propagate_control - refresh control masks of a subtree 3311 * @cgrp: root of the target subtree 3312 * 3313 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches 3314 * ->subtree_control and propagate controller availability through the 3315 * subtree so that descendants don't have unavailable controllers enabled. 3316 */ 3317static void cgroup_propagate_control(struct cgroup *cgrp) 3318{ 3319 struct cgroup *dsct; 3320 struct cgroup_subsys_state *d_css; 3321 3322 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 3323 dsct->subtree_control &= cgroup_control(dsct); 3324 dsct->subtree_ss_mask = 3325 cgroup_calc_subtree_ss_mask(dsct->subtree_control, 3326 cgroup_ss_mask(dsct)); 3327 } 3328} 3329 3330/** 3331 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree 3332 * @cgrp: root of the target subtree 3333 * 3334 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the 3335 * respective old_ prefixed fields for @cgrp's subtree including @cgrp 3336 * itself. 3337 */ 3338static void cgroup_restore_control(struct cgroup *cgrp) 3339{ 3340 struct cgroup *dsct; 3341 struct cgroup_subsys_state *d_css; 3342 3343 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 3344 dsct->subtree_control = dsct->old_subtree_control; 3345 dsct->subtree_ss_mask = dsct->old_subtree_ss_mask; 3346 dsct->dom_cgrp = dsct->old_dom_cgrp; 3347 } 3348} 3349 3350static bool css_visible(struct cgroup_subsys_state *css) 3351{ 3352 struct cgroup_subsys *ss = css->ss; 3353 struct cgroup *cgrp = css->cgroup; 3354 3355 if (cgroup_control(cgrp) & (1 << ss->id)) 3356 return true; 3357 if (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) 3358 return false; 3359 return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl; 3360} 3361 3362/** 3363 * cgroup_apply_control_enable - enable or show csses according to control 3364 * @cgrp: root of the target subtree 3365 * 3366 * Walk @cgrp's subtree and create new csses or make the existing ones 3367 * visible. A css is created invisible if it's being implicitly enabled 3368 * through dependency. An invisible css is made visible when the userland 3369 * explicitly enables it. 3370 * 3371 * Returns 0 on success, -errno on failure. On failure, csses which have 3372 * been processed already aren't cleaned up. The caller is responsible for 3373 * cleaning up with cgroup_apply_control_disable(). 3374 */ 3375static int cgroup_apply_control_enable(struct cgroup *cgrp) 3376{ 3377 struct cgroup *dsct; 3378 struct cgroup_subsys_state *d_css; 3379 struct cgroup_subsys *ss; 3380 int ssid, ret; 3381 3382 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) { 3383 for_each_subsys(ss, ssid) { 3384 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 3385 3386 if (!(cgroup_ss_mask(dsct) & (1 << ss->id))) 3387 continue; 3388 3389 if (!css) { 3390 css = css_create(dsct, ss); 3391 if (IS_ERR(css)) 3392 return PTR_ERR(css); 3393 } 3394 3395 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt)); 3396 3397 if (css_visible(css)) { 3398 ret = css_populate_dir(css); 3399 if (ret) 3400 return ret; 3401 } 3402 } 3403 } 3404 3405 return 0; 3406} 3407 3408/** 3409 * cgroup_apply_control_disable - kill or hide csses according to control 3410 * @cgrp: root of the target subtree 3411 * 3412 * Walk @cgrp's subtree and kill and hide csses so that they match 3413 * cgroup_ss_mask() and cgroup_visible_mask(). 3414 * 3415 * A css is hidden when the userland requests it to be disabled while other 3416 * subsystems are still depending on it. The css must not actively control 3417 * resources and be in the vanilla state if it's made visible again later. 3418 * Controllers which may be depended upon should provide ->css_reset() for 3419 * this purpose. 3420 */ 3421static void cgroup_apply_control_disable(struct cgroup *cgrp) 3422{ 3423 struct cgroup *dsct; 3424 struct cgroup_subsys_state *d_css; 3425 struct cgroup_subsys *ss; 3426 int ssid; 3427 3428 cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) { 3429 for_each_subsys(ss, ssid) { 3430 struct cgroup_subsys_state *css = cgroup_css(dsct, ss); 3431 3432 if (!css) 3433 continue; 3434 3435 WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt)); 3436 3437 if (css->parent && 3438 !(cgroup_ss_mask(dsct) & (1 << ss->id))) { 3439 kill_css(css); 3440 } else if (!css_visible(css)) { 3441 css_clear_dir(css); 3442 if (ss->css_reset) 3443 ss->css_reset(css); 3444 } 3445 } 3446 } 3447} 3448 3449/** 3450 * cgroup_apply_control - apply control mask updates to the subtree 3451 * @cgrp: root of the target subtree 3452 * 3453 * subsystems can be enabled and disabled in a subtree using the following 3454 * steps. 3455 * 3456 * 1. Call cgroup_save_control() to stash the current state. 3457 * 2. Update ->subtree_control masks in the subtree as desired. 3458 * 3. Call cgroup_apply_control() to apply the changes. 3459 * 4. Optionally perform other related operations. 3460 * 5. Call cgroup_finalize_control() to finish up. 3461 * 3462 * This function implements step 3 and propagates the mask changes 3463 * throughout @cgrp's subtree, updates csses accordingly and perform 3464 * process migrations. 3465 */ 3466static int cgroup_apply_control(struct cgroup *cgrp) 3467{ 3468 int ret; 3469 3470 cgroup_propagate_control(cgrp); 3471 3472 ret = cgroup_apply_control_enable(cgrp); 3473 if (ret) 3474 return ret; 3475 3476 /* 3477 * At this point, cgroup_e_css_by_mask() results reflect the new csses 3478 * making the following cgroup_update_dfl_csses() properly update 3479 * css associations of all tasks in the subtree. 3480 */ 3481 return cgroup_update_dfl_csses(cgrp); 3482} 3483 3484/** 3485 * cgroup_finalize_control - finalize control mask update 3486 * @cgrp: root of the target subtree 3487 * @ret: the result of the update 3488 * 3489 * Finalize control mask update. See cgroup_apply_control() for more info. 3490 */ 3491static void cgroup_finalize_control(struct cgroup *cgrp, int ret) 3492{ 3493 if (ret) { 3494 cgroup_restore_control(cgrp); 3495 cgroup_propagate_control(cgrp); 3496 } 3497 3498 cgroup_apply_control_disable(cgrp); 3499} 3500 3501static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable) 3502{ 3503 u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask; 3504 3505 /* if nothing is getting enabled, nothing to worry about */ 3506 if (!enable) 3507 return 0; 3508 3509 /* can @cgrp host any resources? */ 3510 if (!cgroup_is_valid_domain(cgrp->dom_cgrp)) 3511 return -EOPNOTSUPP; 3512 3513 /* mixables don't care */ 3514 if (cgroup_is_mixable(cgrp)) 3515 return 0; 3516 3517 if (domain_enable) { 3518 /* can't enable domain controllers inside a thread subtree */ 3519 if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp)) 3520 return -EOPNOTSUPP; 3521 } else { 3522 /* 3523 * Threaded controllers can handle internal competitions 3524 * and are always allowed inside a (prospective) thread 3525 * subtree. 3526 */ 3527 if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp)) 3528 return 0; 3529 } 3530 3531 /* 3532 * Controllers can't be enabled for a cgroup with tasks to avoid 3533 * child cgroups competing against tasks. 3534 */ 3535 if (cgroup_has_tasks(cgrp)) 3536 return -EBUSY; 3537 3538 return 0; 3539} 3540 3541/* change the enabled child controllers for a cgroup in the default hierarchy */ 3542static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of, 3543 char *buf, size_t nbytes, 3544 loff_t off) 3545{ 3546 u16 enable = 0, disable = 0; 3547 struct cgroup *cgrp, *child; 3548 struct cgroup_subsys *ss; 3549 char *tok; 3550 int ssid, ret; 3551 3552 /* 3553 * Parse input - space separated list of subsystem names prefixed 3554 * with either + or -. 3555 */ 3556 buf = strstrip(buf); 3557 while ((tok = strsep(&buf, " "))) { 3558 if (tok[0] == '\0') 3559 continue; 3560 do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) { 3561 if (!cgroup_ssid_enabled(ssid) || 3562 strcmp(tok + 1, ss->name)) 3563 continue; 3564 3565 if (*tok == '+') { 3566 enable |= 1 << ssid; 3567 disable &= ~(1 << ssid); 3568 } else if (*tok == '-') { 3569 disable |= 1 << ssid; 3570 enable &= ~(1 << ssid); 3571 } else { 3572 return -EINVAL; 3573 } 3574 break; 3575 } while_each_subsys_mask(); 3576 if (ssid == CGROUP_SUBSYS_COUNT) 3577 return -EINVAL; 3578 } 3579 3580 cgrp = cgroup_kn_lock_live(of->kn, true); 3581 if (!cgrp) 3582 return -ENODEV; 3583 3584 for_each_subsys(ss, ssid) { 3585 if (enable & (1 << ssid)) { 3586 if (cgrp->subtree_control & (1 << ssid)) { 3587 enable &= ~(1 << ssid); 3588 continue; 3589 } 3590 3591 if (!(cgroup_control(cgrp) & (1 << ssid))) { 3592 ret = -ENOENT; 3593 goto out_unlock; 3594 } 3595 } else if (disable & (1 << ssid)) { 3596 if (!(cgrp->subtree_control & (1 << ssid))) { 3597 disable &= ~(1 << ssid); 3598 continue; 3599 } 3600 3601 /* a child has it enabled? */ 3602 cgroup_for_each_live_child(child, cgrp) { 3603 if (child->subtree_control & (1 << ssid)) { 3604 ret = -EBUSY; 3605 goto out_unlock; 3606 } 3607 } 3608 } 3609 } 3610 3611 if (!enable && !disable) { 3612 ret = 0; 3613 goto out_unlock; 3614 } 3615 3616 ret = cgroup_vet_subtree_control_enable(cgrp, enable); 3617 if (ret) 3618 goto out_unlock; 3619 3620 /* save and update control masks and prepare csses */ 3621 cgroup_save_control(cgrp); 3622 3623 cgrp->subtree_control |= enable; 3624 cgrp->subtree_control &= ~disable; 3625 3626 ret = cgroup_apply_control(cgrp); 3627 cgroup_finalize_control(cgrp, ret); 3628 if (ret) 3629 goto out_unlock; 3630 3631 kernfs_activate(cgrp->kn); 3632out_unlock: 3633 cgroup_kn_unlock(of->kn); 3634 return ret ?: nbytes; 3635} 3636 3637/** 3638 * cgroup_enable_threaded - make @cgrp threaded 3639 * @cgrp: the target cgroup 3640 * 3641 * Called when "threaded" is written to the cgroup.type interface file and 3642 * tries to make @cgrp threaded and join the parent's resource domain. 3643 * This function is never called on the root cgroup as cgroup.type doesn't 3644 * exist on it. 3645 */ 3646static int cgroup_enable_threaded(struct cgroup *cgrp) 3647{ 3648 struct cgroup *parent = cgroup_parent(cgrp); 3649 struct cgroup *dom_cgrp = parent->dom_cgrp; 3650 struct cgroup *dsct; 3651 struct cgroup_subsys_state *d_css; 3652 int ret; 3653 3654 lockdep_assert_held(&cgroup_mutex); 3655 3656 /* noop if already threaded */ 3657 if (cgroup_is_threaded(cgrp)) 3658 return 0; 3659 3660 /* 3661 * If @cgroup is populated or has domain controllers enabled, it 3662 * can't be switched. While the below cgroup_can_be_thread_root() 3663 * test can catch the same conditions, that's only when @parent is 3664 * not mixable, so let's check it explicitly. 3665 */ 3666 if (cgroup_is_populated(cgrp) || 3667 cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask) 3668 return -EOPNOTSUPP; 3669 3670 /* we're joining the parent's domain, ensure its validity */ 3671 if (!cgroup_is_valid_domain(dom_cgrp) || 3672 !cgroup_can_be_thread_root(dom_cgrp)) 3673 return -EOPNOTSUPP; 3674 3675 /* 3676 * The following shouldn't cause actual migrations and should 3677 * always succeed. 3678 */ 3679 cgroup_save_control(cgrp); 3680 3681 cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) 3682 if (dsct == cgrp || cgroup_is_threaded(dsct)) 3683 dsct->dom_cgrp = dom_cgrp; 3684 3685 ret = cgroup_apply_control(cgrp); 3686 if (!ret) 3687 parent->nr_threaded_children++; 3688 3689 cgroup_finalize_control(cgrp, ret); 3690 return ret; 3691} 3692 3693static int cgroup_type_show(struct seq_file *seq, void *v) 3694{ 3695 struct cgroup *cgrp = seq_css(seq)->cgroup; 3696 3697 if (cgroup_is_threaded(cgrp)) 3698 seq_puts(seq, "threaded\n"); 3699 else if (!cgroup_is_valid_domain(cgrp)) 3700 seq_puts(seq, "domain invalid\n"); 3701 else if (cgroup_is_thread_root(cgrp)) 3702 seq_puts(seq, "domain threaded\n"); 3703 else 3704 seq_puts(seq, "domain\n"); 3705 3706 return 0; 3707} 3708 3709static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf, 3710 size_t nbytes, loff_t off) 3711{ 3712 struct cgroup *cgrp; 3713 int ret; 3714 3715 /* only switching to threaded mode is supported */ 3716 if (strcmp(strstrip(buf), "threaded")) 3717 return -EINVAL; 3718 3719 /* drain dying csses before we re-apply (threaded) subtree control */ 3720 cgrp = cgroup_kn_lock_live(of->kn, true); 3721 if (!cgrp) 3722 return -ENOENT; 3723 3724 /* threaded can only be enabled */ 3725 ret = cgroup_enable_threaded(cgrp); 3726 3727 cgroup_kn_unlock(of->kn); 3728 return ret ?: nbytes; 3729} 3730 3731static int cgroup_max_descendants_show(struct seq_file *seq, void *v) 3732{ 3733 struct cgroup *cgrp = seq_css(seq)->cgroup; 3734 int descendants = READ_ONCE(cgrp->max_descendants); 3735 3736 if (descendants == INT_MAX) 3737 seq_puts(seq, "max\n"); 3738 else 3739 seq_printf(seq, "%d\n", descendants); 3740 3741 return 0; 3742} 3743 3744static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of, 3745 char *buf, size_t nbytes, loff_t off) 3746{ 3747 struct cgroup *cgrp; 3748 int descendants; 3749 ssize_t ret; 3750 3751 buf = strstrip(buf); 3752 if (!strcmp(buf, "max")) { 3753 descendants = INT_MAX; 3754 } else { 3755 ret = kstrtoint(buf, 0, &descendants); 3756 if (ret) 3757 return ret; 3758 } 3759 3760 if (descendants < 0) 3761 return -ERANGE; 3762 3763 cgrp = cgroup_kn_lock_live(of->kn, false); 3764 if (!cgrp) 3765 return -ENOENT; 3766 3767 cgrp->max_descendants = descendants; 3768 3769 cgroup_kn_unlock(of->kn); 3770 3771 return nbytes; 3772} 3773 3774static int cgroup_max_depth_show(struct seq_file *seq, void *v) 3775{ 3776 struct cgroup *cgrp = seq_css(seq)->cgroup; 3777 int depth = READ_ONCE(cgrp->max_depth); 3778 3779 if (depth == INT_MAX) 3780 seq_puts(seq, "max\n"); 3781 else 3782 seq_printf(seq, "%d\n", depth); 3783 3784 return 0; 3785} 3786 3787static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of, 3788 char *buf, size_t nbytes, loff_t off) 3789{ 3790 struct cgroup *cgrp; 3791 ssize_t ret; 3792 int depth; 3793 3794 buf = strstrip(buf); 3795 if (!strcmp(buf, "max")) { 3796 depth = INT_MAX; 3797 } else { 3798 ret = kstrtoint(buf, 0, &depth); 3799 if (ret) 3800 return ret; 3801 } 3802 3803 if (depth < 0) 3804 return -ERANGE; 3805 3806 cgrp = cgroup_kn_lock_live(of->kn, false); 3807 if (!cgrp) 3808 return -ENOENT; 3809 3810 cgrp->max_depth = depth; 3811 3812 cgroup_kn_unlock(of->kn); 3813 3814 return nbytes; 3815} 3816 3817static int cgroup_events_show(struct seq_file *seq, void *v) 3818{ 3819 struct cgroup *cgrp = seq_css(seq)->cgroup; 3820 3821 seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp)); 3822 seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags)); 3823 3824 return 0; 3825} 3826 3827static int cgroup_stat_show(struct seq_file *seq, void *v) 3828{ 3829 struct cgroup *cgroup = seq_css(seq)->cgroup; 3830 struct cgroup_subsys_state *css; 3831 int dying_cnt[CGROUP_SUBSYS_COUNT]; 3832 int ssid; 3833 3834 seq_printf(seq, "nr_descendants %d\n", 3835 cgroup->nr_descendants); 3836 3837 /* 3838 * Show the number of live and dying csses associated with each of 3839 * non-inhibited cgroup subsystems that is bound to cgroup v2. 3840 * 3841 * Without proper lock protection, racing is possible. So the 3842 * numbers may not be consistent when that happens. 3843 */ 3844 rcu_read_lock(); 3845 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) { 3846 dying_cnt[ssid] = -1; 3847 if ((BIT(ssid) & cgrp_dfl_inhibit_ss_mask) || 3848 (cgroup_subsys[ssid]->root != &cgrp_dfl_root)) 3849 continue; 3850 css = rcu_dereference_raw(cgroup->subsys[ssid]); 3851 dying_cnt[ssid] = cgroup->nr_dying_subsys[ssid]; 3852 seq_printf(seq, "nr_subsys_%s %d\n", cgroup_subsys[ssid]->name, 3853 css ? (css->nr_descendants + 1) : 0); 3854 } 3855 3856 seq_printf(seq, "nr_dying_descendants %d\n", 3857 cgroup->nr_dying_descendants); 3858 for (ssid = 0; ssid < CGROUP_SUBSYS_COUNT; ssid++) { 3859 if (dying_cnt[ssid] >= 0) 3860 seq_printf(seq, "nr_dying_subsys_%s %d\n", 3861 cgroup_subsys[ssid]->name, dying_cnt[ssid]); 3862 } 3863 rcu_read_unlock(); 3864 return 0; 3865} 3866 3867static int cgroup_core_local_stat_show(struct seq_file *seq, void *v) 3868{ 3869 struct cgroup *cgrp = seq_css(seq)->cgroup; 3870 unsigned int sequence; 3871 u64 freeze_time; 3872 3873 do { 3874 sequence = read_seqcount_begin(&cgrp->freezer.freeze_seq); 3875 freeze_time = cgrp->freezer.frozen_nsec; 3876 /* Add in current freezer interval if the cgroup is freezing. */ 3877 if (test_bit(CGRP_FREEZE, &cgrp->flags)) 3878 freeze_time += (ktime_get_ns() - 3879 cgrp->freezer.freeze_start_nsec); 3880 } while (read_seqcount_retry(&cgrp->freezer.freeze_seq, sequence)); 3881 3882 do_div(freeze_time, NSEC_PER_USEC); 3883 seq_printf(seq, "frozen_usec %llu\n", freeze_time); 3884 3885 return 0; 3886} 3887 3888#ifdef CONFIG_CGROUP_SCHED 3889/** 3890 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem 3891 * @cgrp: the cgroup of interest 3892 * @ss: the subsystem of interest 3893 * 3894 * Find and get @cgrp's css associated with @ss. If the css doesn't exist 3895 * or is offline, %NULL is returned. 3896 */ 3897static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp, 3898 struct cgroup_subsys *ss) 3899{ 3900 struct cgroup_subsys_state *css; 3901 3902 rcu_read_lock(); 3903 css = cgroup_css(cgrp, ss); 3904 if (css && !css_tryget_online(css)) 3905 css = NULL; 3906 rcu_read_unlock(); 3907 3908 return css; 3909} 3910 3911static int cgroup_extra_stat_show(struct seq_file *seq, int ssid) 3912{ 3913 struct cgroup *cgrp = seq_css(seq)->cgroup; 3914 struct cgroup_subsys *ss = cgroup_subsys[ssid]; 3915 struct cgroup_subsys_state *css; 3916 int ret; 3917 3918 if (!ss->css_extra_stat_show) 3919 return 0; 3920 3921 css = cgroup_tryget_css(cgrp, ss); 3922 if (!css) 3923 return 0; 3924 3925 ret = ss->css_extra_stat_show(seq, css); 3926 css_put(css); 3927 return ret; 3928} 3929 3930static int cgroup_local_stat_show(struct seq_file *seq, 3931 struct cgroup *cgrp, int ssid) 3932{ 3933 struct cgroup_subsys *ss = cgroup_subsys[ssid]; 3934 struct cgroup_subsys_state *css; 3935 int ret; 3936 3937 if (!ss->css_local_stat_show) 3938 return 0; 3939 3940 css = cgroup_tryget_css(cgrp, ss); 3941 if (!css) 3942 return 0; 3943 3944 ret = ss->css_local_stat_show(seq, css); 3945 css_put(css); 3946 return ret; 3947} 3948#endif 3949 3950static int cpu_stat_show(struct seq_file *seq, void *v) 3951{ 3952 int ret = 0; 3953 3954 cgroup_base_stat_cputime_show(seq); 3955#ifdef CONFIG_CGROUP_SCHED 3956 ret = cgroup_extra_stat_show(seq, cpu_cgrp_id); 3957#endif 3958 return ret; 3959} 3960 3961static int cpu_local_stat_show(struct seq_file *seq, void *v) 3962{ 3963 struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup; 3964 int ret = 0; 3965 3966#ifdef CONFIG_CGROUP_SCHED 3967 ret = cgroup_local_stat_show(seq, cgrp, cpu_cgrp_id); 3968#endif 3969 return ret; 3970} 3971 3972#ifdef CONFIG_PSI 3973static int cgroup_io_pressure_show(struct seq_file *seq, void *v) 3974{ 3975 struct cgroup *cgrp = seq_css(seq)->cgroup; 3976 struct psi_group *psi = cgroup_psi(cgrp); 3977 3978 return psi_show(seq, psi, PSI_IO); 3979} 3980static int cgroup_memory_pressure_show(struct seq_file *seq, void *v) 3981{ 3982 struct cgroup *cgrp = seq_css(seq)->cgroup; 3983 struct psi_group *psi = cgroup_psi(cgrp); 3984 3985 return psi_show(seq, psi, PSI_MEM); 3986} 3987static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v) 3988{ 3989 struct cgroup *cgrp = seq_css(seq)->cgroup; 3990 struct psi_group *psi = cgroup_psi(cgrp); 3991 3992 return psi_show(seq, psi, PSI_CPU); 3993} 3994 3995static ssize_t pressure_write(struct kernfs_open_file *of, char *buf, 3996 size_t nbytes, enum psi_res res) 3997{ 3998 struct cgroup_file_ctx *ctx = of->priv; 3999 struct psi_trigger *new; 4000 struct cgroup *cgrp; 4001 struct psi_group *psi; 4002 4003 cgrp = cgroup_kn_lock_live(of->kn, false); 4004 if (!cgrp) 4005 return -ENODEV; 4006 4007 cgroup_get(cgrp); 4008 cgroup_kn_unlock(of->kn); 4009 4010 /* Allow only one trigger per file descriptor */ 4011 if (ctx->psi.trigger) { 4012 cgroup_put(cgrp); 4013 return -EBUSY; 4014 } 4015 4016 psi = cgroup_psi(cgrp); 4017 new = psi_trigger_create(psi, buf, res, of->file, of); 4018 if (IS_ERR(new)) { 4019 cgroup_put(cgrp); 4020 return PTR_ERR(new); 4021 } 4022 4023 smp_store_release(&ctx->psi.trigger, new); 4024 cgroup_put(cgrp); 4025 4026 return nbytes; 4027} 4028 4029static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of, 4030 char *buf, size_t nbytes, 4031 loff_t off) 4032{ 4033 return pressure_write(of, buf, nbytes, PSI_IO); 4034} 4035 4036static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of, 4037 char *buf, size_t nbytes, 4038 loff_t off) 4039{ 4040 return pressure_write(of, buf, nbytes, PSI_MEM); 4041} 4042 4043static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of, 4044 char *buf, size_t nbytes, 4045 loff_t off) 4046{ 4047 return pressure_write(of, buf, nbytes, PSI_CPU); 4048} 4049 4050#ifdef CONFIG_IRQ_TIME_ACCOUNTING 4051static int cgroup_irq_pressure_show(struct seq_file *seq, void *v) 4052{ 4053 struct cgroup *cgrp = seq_css(seq)->cgroup; 4054 struct psi_group *psi = cgroup_psi(cgrp); 4055 4056 return psi_show(seq, psi, PSI_IRQ); 4057} 4058 4059static ssize_t cgroup_irq_pressure_write(struct kernfs_open_file *of, 4060 char *buf, size_t nbytes, 4061 loff_t off) 4062{ 4063 return pressure_write(of, buf, nbytes, PSI_IRQ); 4064} 4065#endif 4066 4067static int cgroup_pressure_show(struct seq_file *seq, void *v) 4068{ 4069 struct cgroup *cgrp = seq_css(seq)->cgroup; 4070 struct psi_group *psi = cgroup_psi(cgrp); 4071 4072 seq_printf(seq, "%d\n", psi->enabled); 4073 4074 return 0; 4075} 4076 4077static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, 4078 char *buf, size_t nbytes, 4079 loff_t off) 4080{ 4081 ssize_t ret; 4082 int enable; 4083 struct cgroup *cgrp; 4084 struct psi_group *psi; 4085 4086 ret = kstrtoint(strstrip(buf), 0, &enable); 4087 if (ret) 4088 return ret; 4089 4090 if (enable < 0 || enable > 1) 4091 return -ERANGE; 4092 4093 cgrp = cgroup_kn_lock_live(of->kn, false); 4094 if (!cgrp) 4095 return -ENOENT; 4096 4097 psi = cgroup_psi(cgrp); 4098 if (psi->enabled != enable) { 4099 int i; 4100 4101 /* show or hide {cpu,memory,io,irq}.pressure files */ 4102 for (i = 0; i < NR_PSI_RESOURCES; i++) 4103 cgroup_file_show(&cgrp->psi_files[i], enable); 4104 4105 psi->enabled = enable; 4106 if (enable) 4107 psi_cgroup_restart(psi); 4108 } 4109 4110 cgroup_kn_unlock(of->kn); 4111 4112 return nbytes; 4113} 4114 4115static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of, 4116 poll_table *pt) 4117{ 4118 struct cgroup_file_ctx *ctx = of->priv; 4119 4120 return psi_trigger_poll(&ctx->psi.trigger, of->file, pt); 4121} 4122 4123static void cgroup_pressure_release(struct kernfs_open_file *of) 4124{ 4125 struct cgroup_file_ctx *ctx = of->priv; 4126 4127 psi_trigger_destroy(ctx->psi.trigger); 4128} 4129 4130bool cgroup_psi_enabled(void) 4131{ 4132 if (static_branch_likely(&psi_disabled)) 4133 return false; 4134 4135 return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0; 4136} 4137 4138#else /* CONFIG_PSI */ 4139bool cgroup_psi_enabled(void) 4140{ 4141 return false; 4142} 4143 4144#endif /* CONFIG_PSI */ 4145 4146static int cgroup_freeze_show(struct seq_file *seq, void *v) 4147{ 4148 struct cgroup *cgrp = seq_css(seq)->cgroup; 4149 4150 seq_printf(seq, "%d\n", cgrp->freezer.freeze); 4151 4152 return 0; 4153} 4154 4155static ssize_t cgroup_freeze_write(struct kernfs_open_file *of, 4156 char *buf, size_t nbytes, loff_t off) 4157{ 4158 struct cgroup *cgrp; 4159 ssize_t ret; 4160 int freeze; 4161 4162 ret = kstrtoint(strstrip(buf), 0, &freeze); 4163 if (ret) 4164 return ret; 4165 4166 if (freeze < 0 || freeze > 1) 4167 return -ERANGE; 4168 4169 cgrp = cgroup_kn_lock_live(of->kn, false); 4170 if (!cgrp) 4171 return -ENOENT; 4172 4173 cgroup_freeze(cgrp, freeze); 4174 4175 cgroup_kn_unlock(of->kn); 4176 4177 return nbytes; 4178} 4179 4180static void __cgroup_kill(struct cgroup *cgrp) 4181{ 4182 struct css_task_iter it; 4183 struct task_struct *task; 4184 4185 lockdep_assert_held(&cgroup_mutex); 4186 4187 spin_lock_irq(&css_set_lock); 4188 cgrp->kill_seq++; 4189 spin_unlock_irq(&css_set_lock); 4190 4191 css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it); 4192 while ((task = css_task_iter_next(&it))) { 4193 /* Ignore kernel threads here. */ 4194 if (task->flags & PF_KTHREAD) 4195 continue; 4196 4197 /* Skip tasks that are already dying. */ 4198 if (__fatal_signal_pending(task)) 4199 continue; 4200 4201 send_sig(SIGKILL, task, 0); 4202 } 4203 css_task_iter_end(&it); 4204} 4205 4206static void cgroup_kill(struct cgroup *cgrp) 4207{ 4208 struct cgroup_subsys_state *css; 4209 struct cgroup *dsct; 4210 4211 lockdep_assert_held(&cgroup_mutex); 4212 4213 cgroup_for_each_live_descendant_pre(dsct, css, cgrp) 4214 __cgroup_kill(dsct); 4215} 4216 4217static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf, 4218 size_t nbytes, loff_t off) 4219{ 4220 ssize_t ret = 0; 4221 int kill; 4222 struct cgroup *cgrp; 4223 4224 ret = kstrtoint(strstrip(buf), 0, &kill); 4225 if (ret) 4226 return ret; 4227 4228 if (kill != 1) 4229 return -ERANGE; 4230 4231 cgrp = cgroup_kn_lock_live(of->kn, false); 4232 if (!cgrp) 4233 return -ENOENT; 4234 4235 /* 4236 * Killing is a process directed operation, i.e. the whole thread-group 4237 * is taken down so act like we do for cgroup.procs and only make this 4238 * writable in non-threaded cgroups. 4239 */ 4240 if (cgroup_is_threaded(cgrp)) 4241 ret = -EOPNOTSUPP; 4242 else 4243 cgroup_kill(cgrp); 4244 4245 cgroup_kn_unlock(of->kn); 4246 4247 return ret ?: nbytes; 4248} 4249 4250static int cgroup_file_open(struct kernfs_open_file *of) 4251{ 4252 struct cftype *cft = of_cft(of); 4253 struct cgroup_file_ctx *ctx; 4254 int ret; 4255 4256 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 4257 if (!ctx) 4258 return -ENOMEM; 4259 4260 ctx->ns = current->nsproxy->cgroup_ns; 4261 get_cgroup_ns(ctx->ns); 4262 of->priv = ctx; 4263 4264 if (!cft->open) 4265 return 0; 4266 4267 ret = cft->open(of); 4268 if (ret) { 4269 put_cgroup_ns(ctx->ns); 4270 kfree(ctx); 4271 } 4272 return ret; 4273} 4274 4275static void cgroup_file_release(struct kernfs_open_file *of) 4276{ 4277 struct cftype *cft = of_cft(of); 4278 struct cgroup_file_ctx *ctx = of->priv; 4279 4280 if (cft->release) 4281 cft->release(of); 4282 put_cgroup_ns(ctx->ns); 4283 kfree(ctx); 4284 of->priv = NULL; 4285} 4286 4287static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf, 4288 size_t nbytes, loff_t off) 4289{ 4290 struct cgroup_file_ctx *ctx = of->priv; 4291 struct cgroup *cgrp = kn_priv(of->kn); 4292 struct cftype *cft = of_cft(of); 4293 struct cgroup_subsys_state *css; 4294 int ret; 4295 4296 if (!nbytes) 4297 return 0; 4298 4299 /* 4300 * If namespaces are delegation boundaries, disallow writes to 4301 * files in an non-init namespace root from inside the namespace 4302 * except for the files explicitly marked delegatable - 4303 * eg. cgroup.procs, cgroup.threads and cgroup.subtree_control. 4304 */ 4305 if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) && 4306 !(cft->flags & CFTYPE_NS_DELEGATABLE) && 4307 ctx->ns != &init_cgroup_ns && ctx->ns->root_cset->dfl_cgrp == cgrp) 4308 return -EPERM; 4309 4310 if (cft->write) 4311 return cft->write(of, buf, nbytes, off); 4312 4313 /* 4314 * kernfs guarantees that a file isn't deleted with operations in 4315 * flight, which means that the matching css is and stays alive and 4316 * doesn't need to be pinned. The RCU locking is not necessary 4317 * either. It's just for the convenience of using cgroup_css(). 4318 */ 4319 rcu_read_lock(); 4320 css = cgroup_css(cgrp, cft->ss); 4321 rcu_read_unlock(); 4322 4323 if (cft->write_u64) { 4324 unsigned long long v; 4325 ret = kstrtoull(buf, 0, &v); 4326 if (!ret) 4327 ret = cft->write_u64(css, cft, v); 4328 } else if (cft->write_s64) { 4329 long long v; 4330 ret = kstrtoll(buf, 0, &v); 4331 if (!ret) 4332 ret = cft->write_s64(css, cft, v); 4333 } else { 4334 ret = -EINVAL; 4335 } 4336 4337 return ret ?: nbytes; 4338} 4339 4340static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt) 4341{ 4342 struct cftype *cft = of_cft(of); 4343 4344 if (cft->poll) 4345 return cft->poll(of, pt); 4346 4347 return kernfs_generic_poll(of, pt); 4348} 4349 4350static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos) 4351{ 4352 return seq_cft(seq)->seq_start(seq, ppos); 4353} 4354 4355static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos) 4356{ 4357 return seq_cft(seq)->seq_next(seq, v, ppos); 4358} 4359 4360static void cgroup_seqfile_stop(struct seq_file *seq, void *v) 4361{ 4362 if (seq_cft(seq)->seq_stop) 4363 seq_cft(seq)->seq_stop(seq, v); 4364} 4365 4366static int cgroup_seqfile_show(struct seq_file *m, void *arg) 4367{ 4368 struct cftype *cft = seq_cft(m); 4369 struct cgroup_subsys_state *css = seq_css(m); 4370 4371 if (cft->seq_show) 4372 return cft->seq_show(m, arg); 4373 4374 if (cft->read_u64) 4375 seq_printf(m, "%llu\n", cft->read_u64(css, cft)); 4376 else if (cft->read_s64) 4377 seq_printf(m, "%lld\n", cft->read_s64(css, cft)); 4378 else 4379 return -EINVAL; 4380 return 0; 4381} 4382 4383static struct kernfs_ops cgroup_kf_single_ops = { 4384 .atomic_write_len = PAGE_SIZE, 4385 .open = cgroup_file_open, 4386 .release = cgroup_file_release, 4387 .write = cgroup_file_write, 4388 .poll = cgroup_file_poll, 4389 .seq_show = cgroup_seqfile_show, 4390}; 4391 4392static struct kernfs_ops cgroup_kf_ops = { 4393 .atomic_write_len = PAGE_SIZE, 4394 .open = cgroup_file_open, 4395 .release = cgroup_file_release, 4396 .write = cgroup_file_write, 4397 .poll = cgroup_file_poll, 4398 .seq_start = cgroup_seqfile_start, 4399 .seq_next = cgroup_seqfile_next, 4400 .seq_stop = cgroup_seqfile_stop, 4401 .seq_show = cgroup_seqfile_show, 4402}; 4403 4404static void cgroup_file_notify_timer(struct timer_list *timer) 4405{ 4406 cgroup_file_notify(container_of(timer, struct cgroup_file, 4407 notify_timer)); 4408} 4409 4410static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp, 4411 struct cftype *cft) 4412{ 4413 char name[CGROUP_FILE_NAME_MAX]; 4414 struct kernfs_node *kn; 4415 struct lock_class_key *key = NULL; 4416 4417#ifdef CONFIG_DEBUG_LOCK_ALLOC 4418 key = &cft->lockdep_key; 4419#endif 4420 kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name), 4421 cgroup_file_mode(cft), 4422 current_fsuid(), current_fsgid(), 4423 0, cft->kf_ops, cft, 4424 NULL, key); 4425 if (IS_ERR(kn)) 4426 return PTR_ERR(kn); 4427 4428 if (cft->file_offset) { 4429 struct cgroup_file *cfile = (void *)css + cft->file_offset; 4430 4431 timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0); 4432 4433 spin_lock_irq(&cgroup_file_kn_lock); 4434 cfile->kn = kn; 4435 spin_unlock_irq(&cgroup_file_kn_lock); 4436 } 4437 4438 return 0; 4439} 4440 4441/** 4442 * cgroup_addrm_files - add or remove files to a cgroup directory 4443 * @css: the target css 4444 * @cgrp: the target cgroup (usually css->cgroup) 4445 * @cfts: array of cftypes to be added 4446 * @is_add: whether to add or remove 4447 * 4448 * Depending on @is_add, add or remove files defined by @cfts on @cgrp. 4449 * For removals, this function never fails. 4450 */ 4451static int cgroup_addrm_files(struct cgroup_subsys_state *css, 4452 struct cgroup *cgrp, struct cftype cfts[], 4453 bool is_add) 4454{ 4455 struct cftype *cft, *cft_end = NULL; 4456 int ret = 0; 4457 4458 lockdep_assert_held(&cgroup_mutex); 4459 4460restart: 4461 for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) { 4462 /* does cft->flags tell us to skip this file on @cgrp? */ 4463 if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp)) 4464 continue; 4465 if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp)) 4466 continue; 4467 if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp)) 4468 continue; 4469 if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp)) 4470 continue; 4471 if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug) 4472 continue; 4473 if (is_add) { 4474 ret = cgroup_add_file(css, cgrp, cft); 4475 if (ret) { 4476 pr_warn("%s: failed to add %s, err=%d\n", 4477 __func__, cft->name, ret); 4478 cft_end = cft; 4479 is_add = false; 4480 goto restart; 4481 } 4482 } else { 4483 cgroup_rm_file(cgrp, cft); 4484 } 4485 } 4486 return ret; 4487} 4488 4489static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add) 4490{ 4491 struct cgroup_subsys *ss = cfts[0].ss; 4492 struct cgroup *root = &ss->root->cgrp; 4493 struct cgroup_subsys_state *css; 4494 int ret = 0; 4495 4496 lockdep_assert_held(&cgroup_mutex); 4497 4498 /* add/rm files for all cgroups created before */ 4499 css_for_each_descendant_pre(css, cgroup_css(root, ss)) { 4500 struct cgroup *cgrp = css->cgroup; 4501 4502 if (!(css->flags & CSS_VISIBLE)) 4503 continue; 4504 4505 ret = cgroup_addrm_files(css, cgrp, cfts, is_add); 4506 if (ret) 4507 break; 4508 } 4509 4510 if (is_add && !ret) 4511 kernfs_activate(root->kn); 4512 return ret; 4513} 4514 4515static void cgroup_exit_cftypes(struct cftype *cfts) 4516{ 4517 struct cftype *cft; 4518 4519 for (cft = cfts; cft->name[0] != '\0'; cft++) { 4520 /* free copy for custom atomic_write_len, see init_cftypes() */ 4521 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) 4522 kfree(cft->kf_ops); 4523 cft->kf_ops = NULL; 4524 cft->ss = NULL; 4525 4526 /* revert flags set by cgroup core while adding @cfts */ 4527 cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL | 4528 __CFTYPE_ADDED); 4529 } 4530} 4531 4532static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4533{ 4534 struct cftype *cft; 4535 int ret = 0; 4536 4537 for (cft = cfts; cft->name[0] != '\0'; cft++) { 4538 struct kernfs_ops *kf_ops; 4539 4540 WARN_ON(cft->ss || cft->kf_ops); 4541 4542 if (cft->flags & __CFTYPE_ADDED) { 4543 ret = -EBUSY; 4544 break; 4545 } 4546 4547 if (cft->seq_start) 4548 kf_ops = &cgroup_kf_ops; 4549 else 4550 kf_ops = &cgroup_kf_single_ops; 4551 4552 /* 4553 * Ugh... if @cft wants a custom max_write_len, we need to 4554 * make a copy of kf_ops to set its atomic_write_len. 4555 */ 4556 if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) { 4557 kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL); 4558 if (!kf_ops) { 4559 ret = -ENOMEM; 4560 break; 4561 } 4562 kf_ops->atomic_write_len = cft->max_write_len; 4563 } 4564 4565 cft->kf_ops = kf_ops; 4566 cft->ss = ss; 4567 cft->flags |= __CFTYPE_ADDED; 4568 } 4569 4570 if (ret) 4571 cgroup_exit_cftypes(cfts); 4572 return ret; 4573} 4574 4575static void cgroup_rm_cftypes_locked(struct cftype *cfts) 4576{ 4577 lockdep_assert_held(&cgroup_mutex); 4578 4579 list_del(&cfts->node); 4580 cgroup_apply_cftypes(cfts, false); 4581 cgroup_exit_cftypes(cfts); 4582} 4583 4584/** 4585 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem 4586 * @cfts: zero-length name terminated array of cftypes 4587 * 4588 * Unregister @cfts. Files described by @cfts are removed from all 4589 * existing cgroups and all future cgroups won't have them either. This 4590 * function can be called anytime whether @cfts' subsys is attached or not. 4591 * 4592 * Returns 0 on successful unregistration, -ENOENT if @cfts is not 4593 * registered. 4594 */ 4595int cgroup_rm_cftypes(struct cftype *cfts) 4596{ 4597 if (!cfts || cfts[0].name[0] == '\0') 4598 return 0; 4599 4600 if (!(cfts[0].flags & __CFTYPE_ADDED)) 4601 return -ENOENT; 4602 4603 cgroup_lock(); 4604 cgroup_rm_cftypes_locked(cfts); 4605 cgroup_unlock(); 4606 return 0; 4607} 4608 4609/** 4610 * cgroup_add_cftypes - add an array of cftypes to a subsystem 4611 * @ss: target cgroup subsystem 4612 * @cfts: zero-length name terminated array of cftypes 4613 * 4614 * Register @cfts to @ss. Files described by @cfts are created for all 4615 * existing cgroups to which @ss is attached and all future cgroups will 4616 * have them too. This function can be called anytime whether @ss is 4617 * attached or not. 4618 * 4619 * Returns 0 on successful registration, -errno on failure. Note that this 4620 * function currently returns 0 as long as @cfts registration is successful 4621 * even if some file creation attempts on existing cgroups fail. 4622 */ 4623int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4624{ 4625 int ret; 4626 4627 if (!cgroup_ssid_enabled(ss->id)) 4628 return 0; 4629 4630 if (!cfts || cfts[0].name[0] == '\0') 4631 return 0; 4632 4633 ret = cgroup_init_cftypes(ss, cfts); 4634 if (ret) 4635 return ret; 4636 4637 cgroup_lock(); 4638 4639 list_add_tail(&cfts->node, &ss->cfts); 4640 ret = cgroup_apply_cftypes(cfts, true); 4641 if (ret) 4642 cgroup_rm_cftypes_locked(cfts); 4643 4644 cgroup_unlock(); 4645 return ret; 4646} 4647 4648/** 4649 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy 4650 * @ss: target cgroup subsystem 4651 * @cfts: zero-length name terminated array of cftypes 4652 * 4653 * Similar to cgroup_add_cftypes() but the added files are only used for 4654 * the default hierarchy. 4655 */ 4656int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4657{ 4658 struct cftype *cft; 4659 4660 for (cft = cfts; cft && cft->name[0] != '\0'; cft++) 4661 cft->flags |= __CFTYPE_ONLY_ON_DFL; 4662 return cgroup_add_cftypes(ss, cfts); 4663} 4664 4665/** 4666 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies 4667 * @ss: target cgroup subsystem 4668 * @cfts: zero-length name terminated array of cftypes 4669 * 4670 * Similar to cgroup_add_cftypes() but the added files are only used for 4671 * the legacy hierarchies. 4672 */ 4673int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts) 4674{ 4675 struct cftype *cft; 4676 4677 for (cft = cfts; cft && cft->name[0] != '\0'; cft++) 4678 cft->flags |= __CFTYPE_NOT_ON_DFL; 4679 return cgroup_add_cftypes(ss, cfts); 4680} 4681 4682/** 4683 * cgroup_file_notify - generate a file modified event for a cgroup_file 4684 * @cfile: target cgroup_file 4685 * 4686 * @cfile must have been obtained by setting cftype->file_offset. 4687 */ 4688void cgroup_file_notify(struct cgroup_file *cfile) 4689{ 4690 unsigned long flags; 4691 4692 spin_lock_irqsave(&cgroup_file_kn_lock, flags); 4693 if (cfile->kn) { 4694 unsigned long last = cfile->notified_at; 4695 unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV; 4696 4697 if (time_in_range(jiffies, last, next)) { 4698 timer_reduce(&cfile->notify_timer, next); 4699 } else { 4700 kernfs_notify(cfile->kn); 4701 cfile->notified_at = jiffies; 4702 } 4703 } 4704 spin_unlock_irqrestore(&cgroup_file_kn_lock, flags); 4705} 4706EXPORT_SYMBOL_GPL(cgroup_file_notify); 4707 4708/** 4709 * cgroup_file_show - show or hide a hidden cgroup file 4710 * @cfile: target cgroup_file obtained by setting cftype->file_offset 4711 * @show: whether to show or hide 4712 */ 4713void cgroup_file_show(struct cgroup_file *cfile, bool show) 4714{ 4715 struct kernfs_node *kn; 4716 4717 spin_lock_irq(&cgroup_file_kn_lock); 4718 kn = cfile->kn; 4719 kernfs_get(kn); 4720 spin_unlock_irq(&cgroup_file_kn_lock); 4721 4722 if (kn) 4723 kernfs_show(kn, show); 4724 4725 kernfs_put(kn); 4726} 4727 4728/** 4729 * css_next_child - find the next child of a given css 4730 * @pos: the current position (%NULL to initiate traversal) 4731 * @parent: css whose children to walk 4732 * 4733 * This function returns the next child of @parent and should be called 4734 * under either cgroup_mutex or RCU read lock. The only requirement is 4735 * that @parent and @pos are accessible. The next sibling is guaranteed to 4736 * be returned regardless of their states. 4737 * 4738 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4739 * css which finished ->css_online() is guaranteed to be visible in the 4740 * future iterations and will stay visible until the last reference is put. 4741 * A css which hasn't finished ->css_online() or already finished 4742 * ->css_offline() may show up during traversal. It's each subsystem's 4743 * responsibility to synchronize against on/offlining. 4744 */ 4745struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos, 4746 struct cgroup_subsys_state *parent) 4747{ 4748 struct cgroup_subsys_state *next; 4749 4750 cgroup_assert_mutex_or_rcu_locked(); 4751 4752 /* 4753 * @pos could already have been unlinked from the sibling list. 4754 * Once a cgroup is removed, its ->sibling.next is no longer 4755 * updated when its next sibling changes. CSS_RELEASED is set when 4756 * @pos is taken off list, at which time its next pointer is valid, 4757 * and, as releases are serialized, the one pointed to by the next 4758 * pointer is guaranteed to not have started release yet. This 4759 * implies that if we observe !CSS_RELEASED on @pos in this RCU 4760 * critical section, the one pointed to by its next pointer is 4761 * guaranteed to not have finished its RCU grace period even if we 4762 * have dropped rcu_read_lock() in-between iterations. 4763 * 4764 * If @pos has CSS_RELEASED set, its next pointer can't be 4765 * dereferenced; however, as each css is given a monotonically 4766 * increasing unique serial number and always appended to the 4767 * sibling list, the next one can be found by walking the parent's 4768 * children until the first css with higher serial number than 4769 * @pos's. While this path can be slower, it happens iff iteration 4770 * races against release and the race window is very small. 4771 */ 4772 if (!pos) { 4773 next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling); 4774 } else if (likely(!(pos->flags & CSS_RELEASED))) { 4775 next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling); 4776 } else { 4777 list_for_each_entry_rcu(next, &parent->children, sibling, 4778 lockdep_is_held(&cgroup_mutex)) 4779 if (next->serial_nr > pos->serial_nr) 4780 break; 4781 } 4782 4783 /* 4784 * @next, if not pointing to the head, can be dereferenced and is 4785 * the next sibling. 4786 */ 4787 if (&next->sibling != &parent->children) 4788 return next; 4789 return NULL; 4790} 4791 4792/** 4793 * css_next_descendant_pre - find the next descendant for pre-order walk 4794 * @pos: the current position (%NULL to initiate traversal) 4795 * @root: css whose descendants to walk 4796 * 4797 * To be used by css_for_each_descendant_pre(). Find the next descendant 4798 * to visit for pre-order traversal of @root's descendants. @root is 4799 * included in the iteration and the first node to be visited. 4800 * 4801 * While this function requires cgroup_mutex or RCU read locking, it 4802 * doesn't require the whole traversal to be contained in a single critical 4803 * section. Additionally, it isn't necessary to hold onto a reference to @pos. 4804 * This function will return the correct next descendant as long as both @pos 4805 * and @root are accessible and @pos is a descendant of @root. 4806 * 4807 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4808 * css which finished ->css_online() is guaranteed to be visible in the 4809 * future iterations and will stay visible until the last reference is put. 4810 * A css which hasn't finished ->css_online() or already finished 4811 * ->css_offline() may show up during traversal. It's each subsystem's 4812 * responsibility to synchronize against on/offlining. 4813 */ 4814struct cgroup_subsys_state * 4815css_next_descendant_pre(struct cgroup_subsys_state *pos, 4816 struct cgroup_subsys_state *root) 4817{ 4818 struct cgroup_subsys_state *next; 4819 4820 cgroup_assert_mutex_or_rcu_locked(); 4821 4822 /* if first iteration, visit @root */ 4823 if (!pos) 4824 return root; 4825 4826 /* visit the first child if exists */ 4827 next = css_next_child(NULL, pos); 4828 if (next) 4829 return next; 4830 4831 /* no child, visit my or the closest ancestor's next sibling */ 4832 while (pos != root) { 4833 next = css_next_child(pos, pos->parent); 4834 if (next) 4835 return next; 4836 pos = pos->parent; 4837 } 4838 4839 return NULL; 4840} 4841EXPORT_SYMBOL_GPL(css_next_descendant_pre); 4842 4843/** 4844 * css_rightmost_descendant - return the rightmost descendant of a css 4845 * @pos: css of interest 4846 * 4847 * Return the rightmost descendant of @pos. If there's no descendant, @pos 4848 * is returned. This can be used during pre-order traversal to skip 4849 * subtree of @pos. 4850 * 4851 * While this function requires cgroup_mutex or RCU read locking, it 4852 * doesn't require the whole traversal to be contained in a single critical 4853 * section. Additionally, it isn't necessary to hold onto a reference to @pos. 4854 * This function will return the correct rightmost descendant as long as @pos 4855 * is accessible. 4856 */ 4857struct cgroup_subsys_state * 4858css_rightmost_descendant(struct cgroup_subsys_state *pos) 4859{ 4860 struct cgroup_subsys_state *last, *tmp; 4861 4862 cgroup_assert_mutex_or_rcu_locked(); 4863 4864 do { 4865 last = pos; 4866 /* ->prev isn't RCU safe, walk ->next till the end */ 4867 pos = NULL; 4868 css_for_each_child(tmp, last) 4869 pos = tmp; 4870 } while (pos); 4871 4872 return last; 4873} 4874 4875static struct cgroup_subsys_state * 4876css_leftmost_descendant(struct cgroup_subsys_state *pos) 4877{ 4878 struct cgroup_subsys_state *last; 4879 4880 do { 4881 last = pos; 4882 pos = css_next_child(NULL, pos); 4883 } while (pos); 4884 4885 return last; 4886} 4887 4888/** 4889 * css_next_descendant_post - find the next descendant for post-order walk 4890 * @pos: the current position (%NULL to initiate traversal) 4891 * @root: css whose descendants to walk 4892 * 4893 * To be used by css_for_each_descendant_post(). Find the next descendant 4894 * to visit for post-order traversal of @root's descendants. @root is 4895 * included in the iteration and the last node to be visited. 4896 * 4897 * While this function requires cgroup_mutex or RCU read locking, it 4898 * doesn't require the whole traversal to be contained in a single critical 4899 * section. Additionally, it isn't necessary to hold onto a reference to @pos. 4900 * This function will return the correct next descendant as long as both @pos 4901 * and @cgroup are accessible and @pos is a descendant of @cgroup. 4902 * 4903 * If a subsystem synchronizes ->css_online() and the start of iteration, a 4904 * css which finished ->css_online() is guaranteed to be visible in the 4905 * future iterations and will stay visible until the last reference is put. 4906 * A css which hasn't finished ->css_online() or already finished 4907 * ->css_offline() may show up during traversal. It's each subsystem's 4908 * responsibility to synchronize against on/offlining. 4909 */ 4910struct cgroup_subsys_state * 4911css_next_descendant_post(struct cgroup_subsys_state *pos, 4912 struct cgroup_subsys_state *root) 4913{ 4914 struct cgroup_subsys_state *next; 4915 4916 cgroup_assert_mutex_or_rcu_locked(); 4917 4918 /* if first iteration, visit leftmost descendant which may be @root */ 4919 if (!pos) 4920 return css_leftmost_descendant(root); 4921 4922 /* if we visited @root, we're done */ 4923 if (pos == root) 4924 return NULL; 4925 4926 /* if there's an unvisited sibling, visit its leftmost descendant */ 4927 next = css_next_child(pos, pos->parent); 4928 if (next) 4929 return css_leftmost_descendant(next); 4930 4931 /* no sibling left, visit parent */ 4932 return pos->parent; 4933} 4934 4935/** 4936 * css_has_online_children - does a css have online children 4937 * @css: the target css 4938 * 4939 * Returns %true if @css has any online children; otherwise, %false. This 4940 * function can be called from any context but the caller is responsible 4941 * for synchronizing against on/offlining as necessary. 4942 */ 4943bool css_has_online_children(struct cgroup_subsys_state *css) 4944{ 4945 struct cgroup_subsys_state *child; 4946 bool ret = false; 4947 4948 rcu_read_lock(); 4949 css_for_each_child(child, css) { 4950 if (child->flags & CSS_ONLINE) { 4951 ret = true; 4952 break; 4953 } 4954 } 4955 rcu_read_unlock(); 4956 return ret; 4957} 4958 4959static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it) 4960{ 4961 struct list_head *l; 4962 struct cgrp_cset_link *link; 4963 struct css_set *cset; 4964 4965 lockdep_assert_held(&css_set_lock); 4966 4967 /* find the next threaded cset */ 4968 if (it->tcset_pos) { 4969 l = it->tcset_pos->next; 4970 4971 if (l != it->tcset_head) { 4972 it->tcset_pos = l; 4973 return container_of(l, struct css_set, 4974 threaded_csets_node); 4975 } 4976 4977 it->tcset_pos = NULL; 4978 } 4979 4980 /* find the next cset */ 4981 l = it->cset_pos; 4982 l = l->next; 4983 if (l == it->cset_head) { 4984 it->cset_pos = NULL; 4985 return NULL; 4986 } 4987 4988 if (it->ss) { 4989 cset = container_of(l, struct css_set, e_cset_node[it->ss->id]); 4990 } else { 4991 link = list_entry(l, struct cgrp_cset_link, cset_link); 4992 cset = link->cset; 4993 } 4994 4995 it->cset_pos = l; 4996 4997 /* initialize threaded css_set walking */ 4998 if (it->flags & CSS_TASK_ITER_THREADED) { 4999 if (it->cur_dcset) 5000 put_css_set_locked(it->cur_dcset); 5001 it->cur_dcset = cset; 5002 get_css_set(cset); 5003 5004 it->tcset_head = &cset->threaded_csets; 5005 it->tcset_pos = &cset->threaded_csets; 5006 } 5007 5008 return cset; 5009} 5010 5011/** 5012 * css_task_iter_advance_css_set - advance a task iterator to the next css_set 5013 * @it: the iterator to advance 5014 * 5015 * Advance @it to the next css_set to walk. 5016 */ 5017static void css_task_iter_advance_css_set(struct css_task_iter *it) 5018{ 5019 struct css_set *cset; 5020 5021 lockdep_assert_held(&css_set_lock); 5022 5023 /* Advance to the next non-empty css_set and find first non-empty tasks list*/ 5024 while ((cset = css_task_iter_next_css_set(it))) { 5025 if (!list_empty(&cset->tasks)) { 5026 it->cur_tasks_head = &cset->tasks; 5027 break; 5028 } else if (!list_empty(&cset->mg_tasks)) { 5029 it->cur_tasks_head = &cset->mg_tasks; 5030 break; 5031 } else if (!list_empty(&cset->dying_tasks)) { 5032 it->cur_tasks_head = &cset->dying_tasks; 5033 break; 5034 } 5035 } 5036 if (!cset) { 5037 it->task_pos = NULL; 5038 return; 5039 } 5040 it->task_pos = it->cur_tasks_head->next; 5041 5042 /* 5043 * We don't keep css_sets locked across iteration steps and thus 5044 * need to take steps to ensure that iteration can be resumed after 5045 * the lock is re-acquired. Iteration is performed at two levels - 5046 * css_sets and tasks in them. 5047 * 5048 * Once created, a css_set never leaves its cgroup lists, so a 5049 * pinned css_set is guaranteed to stay put and we can resume 5050 * iteration afterwards. 5051 * 5052 * Tasks may leave @cset across iteration steps. This is resolved 5053 * by registering each iterator with the css_set currently being 5054 * walked and making css_set_move_task() advance iterators whose 5055 * next task is leaving. 5056 */ 5057 if (it->cur_cset) { 5058 list_del(&it->iters_node); 5059 put_css_set_locked(it->cur_cset); 5060 } 5061 get_css_set(cset); 5062 it->cur_cset = cset; 5063 list_add(&it->iters_node, &cset->task_iters); 5064} 5065 5066static void css_task_iter_skip(struct css_task_iter *it, 5067 struct task_struct *task) 5068{ 5069 lockdep_assert_held(&css_set_lock); 5070 5071 if (it->task_pos == &task->cg_list) { 5072 it->task_pos = it->task_pos->next; 5073 it->flags |= CSS_TASK_ITER_SKIPPED; 5074 } 5075} 5076 5077static void css_task_iter_advance(struct css_task_iter *it) 5078{ 5079 struct task_struct *task; 5080 5081 lockdep_assert_held(&css_set_lock); 5082repeat: 5083 if (it->task_pos) { 5084 /* 5085 * Advance iterator to find next entry. We go through cset 5086 * tasks, mg_tasks and dying_tasks, when consumed we move onto 5087 * the next cset. 5088 */ 5089 if (it->flags & CSS_TASK_ITER_SKIPPED) 5090 it->flags &= ~CSS_TASK_ITER_SKIPPED; 5091 else 5092 it->task_pos = it->task_pos->next; 5093 5094 if (it->task_pos == &it->cur_cset->tasks) { 5095 it->cur_tasks_head = &it->cur_cset->mg_tasks; 5096 it->task_pos = it->cur_tasks_head->next; 5097 } 5098 if (it->task_pos == &it->cur_cset->mg_tasks) { 5099 it->cur_tasks_head = &it->cur_cset->dying_tasks; 5100 it->task_pos = it->cur_tasks_head->next; 5101 } 5102 if (it->task_pos == &it->cur_cset->dying_tasks) 5103 css_task_iter_advance_css_set(it); 5104 } else { 5105 /* called from start, proceed to the first cset */ 5106 css_task_iter_advance_css_set(it); 5107 } 5108 5109 if (!it->task_pos) 5110 return; 5111 5112 task = list_entry(it->task_pos, struct task_struct, cg_list); 5113 /* 5114 * Hide tasks that are exiting but not yet removed. Keep zombie 5115 * leaders with live threads visible. 5116 */ 5117 if ((task->flags & PF_EXITING) && !atomic_read(&task->signal->live)) 5118 goto repeat; 5119 5120 if (it->flags & CSS_TASK_ITER_PROCS) { 5121 /* if PROCS, skip over tasks which aren't group leaders */ 5122 if (!thread_group_leader(task)) 5123 goto repeat; 5124 5125 /* and dying leaders w/o live member threads */ 5126 if (it->cur_tasks_head == &it->cur_cset->dying_tasks && 5127 !atomic_read(&task->signal->live)) 5128 goto repeat; 5129 } else { 5130 /* skip all dying ones */ 5131 if (it->cur_tasks_head == &it->cur_cset->dying_tasks) 5132 goto repeat; 5133 } 5134} 5135 5136/** 5137 * css_task_iter_start - initiate task iteration 5138 * @css: the css to walk tasks of 5139 * @flags: CSS_TASK_ITER_* flags 5140 * @it: the task iterator to use 5141 * 5142 * Initiate iteration through the tasks of @css. The caller can call 5143 * css_task_iter_next() to walk through the tasks until the function 5144 * returns NULL. On completion of iteration, css_task_iter_end() must be 5145 * called. 5146 */ 5147void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags, 5148 struct css_task_iter *it) 5149{ 5150 unsigned long irqflags; 5151 5152 memset(it, 0, sizeof(*it)); 5153 5154 spin_lock_irqsave(&css_set_lock, irqflags); 5155 5156 it->ss = css->ss; 5157 it->flags = flags; 5158 5159 if (CGROUP_HAS_SUBSYS_CONFIG && it->ss) 5160 it->cset_pos = &css->cgroup->e_csets[css->ss->id]; 5161 else 5162 it->cset_pos = &css->cgroup->cset_links; 5163 5164 it->cset_head = it->cset_pos; 5165 5166 css_task_iter_advance(it); 5167 5168 spin_unlock_irqrestore(&css_set_lock, irqflags); 5169} 5170 5171/** 5172 * css_task_iter_next - return the next task for the iterator 5173 * @it: the task iterator being iterated 5174 * 5175 * The "next" function for task iteration. @it should have been 5176 * initialized via css_task_iter_start(). Returns NULL when the iteration 5177 * reaches the end. 5178 */ 5179struct task_struct *css_task_iter_next(struct css_task_iter *it) 5180{ 5181 unsigned long irqflags; 5182 5183 if (it->cur_task) { 5184 put_task_struct(it->cur_task); 5185 it->cur_task = NULL; 5186 } 5187 5188 spin_lock_irqsave(&css_set_lock, irqflags); 5189 5190 /* @it may be half-advanced by skips, finish advancing */ 5191 if (it->flags & CSS_TASK_ITER_SKIPPED) 5192 css_task_iter_advance(it); 5193 5194 if (it->task_pos) { 5195 it->cur_task = list_entry(it->task_pos, struct task_struct, 5196 cg_list); 5197 get_task_struct(it->cur_task); 5198 css_task_iter_advance(it); 5199 } 5200 5201 spin_unlock_irqrestore(&css_set_lock, irqflags); 5202 5203 return it->cur_task; 5204} 5205 5206/** 5207 * css_task_iter_end - finish task iteration 5208 * @it: the task iterator to finish 5209 * 5210 * Finish task iteration started by css_task_iter_start(). 5211 */ 5212void css_task_iter_end(struct css_task_iter *it) 5213{ 5214 unsigned long irqflags; 5215 5216 if (it->cur_cset) { 5217 spin_lock_irqsave(&css_set_lock, irqflags); 5218 list_del(&it->iters_node); 5219 put_css_set_locked(it->cur_cset); 5220 spin_unlock_irqrestore(&css_set_lock, irqflags); 5221 } 5222 5223 if (it->cur_dcset) 5224 put_css_set(it->cur_dcset); 5225 5226 if (it->cur_task) 5227 put_task_struct(it->cur_task); 5228} 5229 5230static void cgroup_procs_release(struct kernfs_open_file *of) 5231{ 5232 struct cgroup_file_ctx *ctx = of->priv; 5233 5234 if (ctx->procs.started) 5235 css_task_iter_end(&ctx->procs.iter); 5236} 5237 5238static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos) 5239{ 5240 struct kernfs_open_file *of = s->private; 5241 struct cgroup_file_ctx *ctx = of->priv; 5242 5243 if (pos) 5244 (*pos)++; 5245 5246 return css_task_iter_next(&ctx->procs.iter); 5247} 5248 5249static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos, 5250 unsigned int iter_flags) 5251{ 5252 struct kernfs_open_file *of = s->private; 5253 struct cgroup *cgrp = seq_css(s)->cgroup; 5254 struct cgroup_file_ctx *ctx = of->priv; 5255 struct css_task_iter *it = &ctx->procs.iter; 5256 5257 /* 5258 * When a seq_file is seeked, it's always traversed sequentially 5259 * from position 0, so we can simply keep iterating on !0 *pos. 5260 */ 5261 if (!ctx->procs.started) { 5262 if (WARN_ON_ONCE((*pos))) 5263 return ERR_PTR(-EINVAL); 5264 css_task_iter_start(&cgrp->self, iter_flags, it); 5265 ctx->procs.started = true; 5266 } else if (!(*pos)) { 5267 css_task_iter_end(it); 5268 css_task_iter_start(&cgrp->self, iter_flags, it); 5269 } else 5270 return it->cur_task; 5271 5272 return cgroup_procs_next(s, NULL, NULL); 5273} 5274 5275static void *cgroup_procs_start(struct seq_file *s, loff_t *pos) 5276{ 5277 struct cgroup *cgrp = seq_css(s)->cgroup; 5278 5279 /* 5280 * All processes of a threaded subtree belong to the domain cgroup 5281 * of the subtree. Only threads can be distributed across the 5282 * subtree. Reject reads on cgroup.procs in the subtree proper. 5283 * They're always empty anyway. 5284 */ 5285 if (cgroup_is_threaded(cgrp)) 5286 return ERR_PTR(-EOPNOTSUPP); 5287 5288 return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS | 5289 CSS_TASK_ITER_THREADED); 5290} 5291 5292static int cgroup_procs_show(struct seq_file *s, void *v) 5293{ 5294 seq_printf(s, "%d\n", task_pid_vnr(v)); 5295 return 0; 5296} 5297 5298static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb) 5299{ 5300 int ret; 5301 struct inode *inode; 5302 5303 lockdep_assert_held(&cgroup_mutex); 5304 5305 inode = kernfs_get_inode(sb, cgrp->procs_file.kn); 5306 if (!inode) 5307 return -ENOMEM; 5308 5309 ret = inode_permission(&nop_mnt_idmap, inode, MAY_WRITE); 5310 iput(inode); 5311 return ret; 5312} 5313 5314static int cgroup_procs_write_permission(struct cgroup *src_cgrp, 5315 struct cgroup *dst_cgrp, 5316 struct super_block *sb, 5317 struct cgroup_namespace *ns) 5318{ 5319 struct cgroup *com_cgrp = src_cgrp; 5320 int ret; 5321 5322 lockdep_assert_held(&cgroup_mutex); 5323 5324 /* find the common ancestor */ 5325 while (!cgroup_is_descendant(dst_cgrp, com_cgrp)) 5326 com_cgrp = cgroup_parent(com_cgrp); 5327 5328 /* %current should be authorized to migrate to the common ancestor */ 5329 ret = cgroup_may_write(com_cgrp, sb); 5330 if (ret) 5331 return ret; 5332 5333 /* 5334 * If namespaces are delegation boundaries, %current must be able 5335 * to see both source and destination cgroups from its namespace. 5336 */ 5337 if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) && 5338 (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) || 5339 !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp))) 5340 return -ENOENT; 5341 5342 return 0; 5343} 5344 5345static int cgroup_attach_permissions(struct cgroup *src_cgrp, 5346 struct cgroup *dst_cgrp, 5347 struct super_block *sb, bool threadgroup, 5348 struct cgroup_namespace *ns) 5349{ 5350 int ret = 0; 5351 5352 ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp, sb, ns); 5353 if (ret) 5354 return ret; 5355 5356 ret = cgroup_migrate_vet_dst(dst_cgrp); 5357 if (ret) 5358 return ret; 5359 5360 if (!threadgroup && (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)) 5361 ret = -EOPNOTSUPP; 5362 5363 return ret; 5364} 5365 5366static ssize_t __cgroup_procs_write(struct kernfs_open_file *of, char *buf, 5367 bool threadgroup) 5368{ 5369 struct cgroup_file_ctx *ctx = of->priv; 5370 struct cgroup *src_cgrp, *dst_cgrp; 5371 struct task_struct *task; 5372 ssize_t ret; 5373 enum cgroup_attach_lock_mode lock_mode; 5374 5375 dst_cgrp = cgroup_kn_lock_live(of->kn, false); 5376 if (!dst_cgrp) 5377 return -ENODEV; 5378 5379 task = cgroup_procs_write_start(buf, threadgroup, &lock_mode); 5380 ret = PTR_ERR_OR_ZERO(task); 5381 if (ret) 5382 goto out_unlock; 5383 5384 /* find the source cgroup */ 5385 spin_lock_irq(&css_set_lock); 5386 src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root); 5387 spin_unlock_irq(&css_set_lock); 5388 5389 /* 5390 * Process and thread migrations follow same delegation rule. Check 5391 * permissions using the credentials from file open to protect against 5392 * inherited fd attacks. 5393 */ 5394 scoped_with_creds(of->file->f_cred) 5395 ret = cgroup_attach_permissions(src_cgrp, dst_cgrp, 5396 of->file->f_path.dentry->d_sb, 5397 threadgroup, ctx->ns); 5398 if (ret) 5399 goto out_finish; 5400 5401 ret = cgroup_attach_task(dst_cgrp, task, threadgroup); 5402 5403out_finish: 5404 cgroup_procs_write_finish(task, lock_mode); 5405out_unlock: 5406 cgroup_kn_unlock(of->kn); 5407 5408 return ret; 5409} 5410 5411static ssize_t cgroup_procs_write(struct kernfs_open_file *of, 5412 char *buf, size_t nbytes, loff_t off) 5413{ 5414 return __cgroup_procs_write(of, buf, true) ?: nbytes; 5415} 5416 5417static void *cgroup_threads_start(struct seq_file *s, loff_t *pos) 5418{ 5419 return __cgroup_procs_start(s, pos, 0); 5420} 5421 5422static ssize_t cgroup_threads_write(struct kernfs_open_file *of, 5423 char *buf, size_t nbytes, loff_t off) 5424{ 5425 return __cgroup_procs_write(of, buf, false) ?: nbytes; 5426} 5427 5428/* cgroup core interface files for the default hierarchy */ 5429static struct cftype cgroup_base_files[] = { 5430 { 5431 .name = "cgroup.type", 5432 .flags = CFTYPE_NOT_ON_ROOT, 5433 .seq_show = cgroup_type_show, 5434 .write = cgroup_type_write, 5435 }, 5436 { 5437 .name = "cgroup.procs", 5438 .flags = CFTYPE_NS_DELEGATABLE, 5439 .file_offset = offsetof(struct cgroup, procs_file), 5440 .release = cgroup_procs_release, 5441 .seq_start = cgroup_procs_start, 5442 .seq_next = cgroup_procs_next, 5443 .seq_show = cgroup_procs_show, 5444 .write = cgroup_procs_write, 5445 }, 5446 { 5447 .name = "cgroup.threads", 5448 .flags = CFTYPE_NS_DELEGATABLE, 5449 .release = cgroup_procs_release, 5450 .seq_start = cgroup_threads_start, 5451 .seq_next = cgroup_procs_next, 5452 .seq_show = cgroup_procs_show, 5453 .write = cgroup_threads_write, 5454 }, 5455 { 5456 .name = "cgroup.controllers", 5457 .seq_show = cgroup_controllers_show, 5458 }, 5459 { 5460 .name = "cgroup.subtree_control", 5461 .flags = CFTYPE_NS_DELEGATABLE, 5462 .seq_show = cgroup_subtree_control_show, 5463 .write = cgroup_subtree_control_write, 5464 }, 5465 { 5466 .name = "cgroup.events", 5467 .flags = CFTYPE_NOT_ON_ROOT, 5468 .file_offset = offsetof(struct cgroup, events_file), 5469 .seq_show = cgroup_events_show, 5470 }, 5471 { 5472 .name = "cgroup.max.descendants", 5473 .seq_show = cgroup_max_descendants_show, 5474 .write = cgroup_max_descendants_write, 5475 }, 5476 { 5477 .name = "cgroup.max.depth", 5478 .seq_show = cgroup_max_depth_show, 5479 .write = cgroup_max_depth_write, 5480 }, 5481 { 5482 .name = "cgroup.stat", 5483 .seq_show = cgroup_stat_show, 5484 }, 5485 { 5486 .name = "cgroup.stat.local", 5487 .flags = CFTYPE_NOT_ON_ROOT, 5488 .seq_show = cgroup_core_local_stat_show, 5489 }, 5490 { 5491 .name = "cgroup.freeze", 5492 .flags = CFTYPE_NOT_ON_ROOT, 5493 .seq_show = cgroup_freeze_show, 5494 .write = cgroup_freeze_write, 5495 }, 5496 { 5497 .name = "cgroup.kill", 5498 .flags = CFTYPE_NOT_ON_ROOT, 5499 .write = cgroup_kill_write, 5500 }, 5501 { 5502 .name = "cpu.stat", 5503 .seq_show = cpu_stat_show, 5504 }, 5505 { 5506 .name = "cpu.stat.local", 5507 .seq_show = cpu_local_stat_show, 5508 }, 5509 { } /* terminate */ 5510}; 5511 5512static struct cftype cgroup_psi_files[] = { 5513#ifdef CONFIG_PSI 5514 { 5515 .name = "io.pressure", 5516 .file_offset = offsetof(struct cgroup, psi_files[PSI_IO]), 5517 .seq_show = cgroup_io_pressure_show, 5518 .write = cgroup_io_pressure_write, 5519 .poll = cgroup_pressure_poll, 5520 .release = cgroup_pressure_release, 5521 }, 5522 { 5523 .name = "memory.pressure", 5524 .file_offset = offsetof(struct cgroup, psi_files[PSI_MEM]), 5525 .seq_show = cgroup_memory_pressure_show, 5526 .write = cgroup_memory_pressure_write, 5527 .poll = cgroup_pressure_poll, 5528 .release = cgroup_pressure_release, 5529 }, 5530 { 5531 .name = "cpu.pressure", 5532 .file_offset = offsetof(struct cgroup, psi_files[PSI_CPU]), 5533 .seq_show = cgroup_cpu_pressure_show, 5534 .write = cgroup_cpu_pressure_write, 5535 .poll = cgroup_pressure_poll, 5536 .release = cgroup_pressure_release, 5537 }, 5538#ifdef CONFIG_IRQ_TIME_ACCOUNTING 5539 { 5540 .name = "irq.pressure", 5541 .file_offset = offsetof(struct cgroup, psi_files[PSI_IRQ]), 5542 .seq_show = cgroup_irq_pressure_show, 5543 .write = cgroup_irq_pressure_write, 5544 .poll = cgroup_pressure_poll, 5545 .release = cgroup_pressure_release, 5546 }, 5547#endif 5548 { 5549 .name = "cgroup.pressure", 5550 .seq_show = cgroup_pressure_show, 5551 .write = cgroup_pressure_write, 5552 }, 5553#endif /* CONFIG_PSI */ 5554 { } /* terminate */ 5555}; 5556 5557/* 5558 * css destruction is four-stage process. 5559 * 5560 * 1. Destruction starts. Killing of the percpu_ref is initiated. 5561 * Implemented in kill_css(). 5562 * 5563 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs 5564 * and thus css_tryget_online() is guaranteed to fail, the css can be 5565 * offlined by invoking offline_css(). After offlining, the base ref is 5566 * put. Implemented in css_killed_work_fn(). 5567 * 5568 * 3. When the percpu_ref reaches zero, the only possible remaining 5569 * accessors are inside RCU read sections. css_release() schedules the 5570 * RCU callback. 5571 * 5572 * 4. After the grace period, the css can be freed. Implemented in 5573 * css_free_rwork_fn(). 5574 * 5575 * It is actually hairier because both step 2 and 4 require process context 5576 * and thus involve punting to css->destroy_work adding two additional 5577 * steps to the already complex sequence. 5578 */ 5579static void css_free_rwork_fn(struct work_struct *work) 5580{ 5581 struct cgroup_subsys_state *css = container_of(to_rcu_work(work), 5582 struct cgroup_subsys_state, destroy_rwork); 5583 struct cgroup_subsys *ss = css->ss; 5584 struct cgroup *cgrp = css->cgroup; 5585 5586 percpu_ref_exit(&css->refcnt); 5587 css_rstat_exit(css); 5588 5589 if (!css_is_self(css)) { 5590 /* css free path */ 5591 struct cgroup_subsys_state *parent = css->parent; 5592 int id = css->id; 5593 5594 ss->css_free(css); 5595 cgroup_idr_remove(&ss->css_idr, id); 5596 cgroup_put(cgrp); 5597 5598 if (parent) 5599 css_put(parent); 5600 } else { 5601 /* cgroup free path */ 5602 atomic_dec(&cgrp->root->nr_cgrps); 5603 if (!cgroup_on_dfl(cgrp)) 5604 cgroup1_pidlist_destroy_all(cgrp); 5605 cancel_work_sync(&cgrp->release_agent_work); 5606 bpf_cgrp_storage_free(cgrp); 5607 5608 if (cgroup_parent(cgrp)) { 5609 /* 5610 * We get a ref to the parent, and put the ref when 5611 * this cgroup is being freed, so it's guaranteed 5612 * that the parent won't be destroyed before its 5613 * children. 5614 */ 5615 cgroup_put(cgroup_parent(cgrp)); 5616 kernfs_put(cgrp->kn); 5617 psi_cgroup_free(cgrp); 5618 kfree(cgrp); 5619 } else { 5620 /* 5621 * This is root cgroup's refcnt reaching zero, 5622 * which indicates that the root should be 5623 * released. 5624 */ 5625 cgroup_destroy_root(cgrp->root); 5626 } 5627 } 5628} 5629 5630static void css_release_work_fn(struct work_struct *work) 5631{ 5632 struct cgroup_subsys_state *css = 5633 container_of(work, struct cgroup_subsys_state, destroy_work); 5634 struct cgroup_subsys *ss = css->ss; 5635 struct cgroup *cgrp = css->cgroup; 5636 5637 cgroup_lock(); 5638 5639 css->flags |= CSS_RELEASED; 5640 list_del_rcu(&css->sibling); 5641 5642 if (!css_is_self(css)) { 5643 struct cgroup *parent_cgrp; 5644 5645 css_rstat_flush(css); 5646 5647 cgroup_idr_replace(&ss->css_idr, NULL, css->id); 5648 if (ss->css_released) 5649 ss->css_released(css); 5650 5651 cgrp->nr_dying_subsys[ss->id]--; 5652 /* 5653 * When a css is released and ready to be freed, its 5654 * nr_descendants must be zero. However, the corresponding 5655 * cgrp->nr_dying_subsys[ss->id] may not be 0 if a subsystem 5656 * is activated and deactivated multiple times with one or 5657 * more of its previous activation leaving behind dying csses. 5658 */ 5659 WARN_ON_ONCE(css->nr_descendants); 5660 parent_cgrp = cgroup_parent(cgrp); 5661 while (parent_cgrp) { 5662 parent_cgrp->nr_dying_subsys[ss->id]--; 5663 parent_cgrp = cgroup_parent(parent_cgrp); 5664 } 5665 } else { 5666 struct cgroup *tcgrp; 5667 5668 /* cgroup release path */ 5669 TRACE_CGROUP_PATH(release, cgrp); 5670 5671 css_rstat_flush(&cgrp->self); 5672 5673 spin_lock_irq(&css_set_lock); 5674 for (tcgrp = cgroup_parent(cgrp); tcgrp; 5675 tcgrp = cgroup_parent(tcgrp)) 5676 tcgrp->nr_dying_descendants--; 5677 spin_unlock_irq(&css_set_lock); 5678 5679 /* 5680 * There are two control paths which try to determine 5681 * cgroup from dentry without going through kernfs - 5682 * cgroupstats_build() and css_tryget_online_from_dir(). 5683 * Those are supported by RCU protecting clearing of 5684 * cgrp->kn->priv backpointer. 5685 */ 5686 if (cgrp->kn) 5687 RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv, 5688 NULL); 5689 } 5690 5691 cgroup_unlock(); 5692 5693 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn); 5694 queue_rcu_work(cgroup_free_wq, &css->destroy_rwork); 5695} 5696 5697static void css_release(struct percpu_ref *ref) 5698{ 5699 struct cgroup_subsys_state *css = 5700 container_of(ref, struct cgroup_subsys_state, refcnt); 5701 5702 INIT_WORK(&css->destroy_work, css_release_work_fn); 5703 queue_work(cgroup_release_wq, &css->destroy_work); 5704} 5705 5706static void init_and_link_css(struct cgroup_subsys_state *css, 5707 struct cgroup_subsys *ss, struct cgroup *cgrp) 5708{ 5709 lockdep_assert_held(&cgroup_mutex); 5710 5711 cgroup_get_live(cgrp); 5712 5713 memset(css, 0, sizeof(*css)); 5714 css->cgroup = cgrp; 5715 css->ss = ss; 5716 css->id = -1; 5717 INIT_LIST_HEAD(&css->sibling); 5718 INIT_LIST_HEAD(&css->children); 5719 css->serial_nr = css_serial_nr_next++; 5720 atomic_set(&css->online_cnt, 0); 5721 5722 if (cgroup_parent(cgrp)) { 5723 css->parent = cgroup_css(cgroup_parent(cgrp), ss); 5724 css_get(css->parent); 5725 } 5726 5727 BUG_ON(cgroup_css(cgrp, ss)); 5728} 5729 5730/* invoke ->css_online() on a new CSS and mark it online if successful */ 5731static int online_css(struct cgroup_subsys_state *css) 5732{ 5733 struct cgroup_subsys *ss = css->ss; 5734 int ret = 0; 5735 5736 lockdep_assert_held(&cgroup_mutex); 5737 5738 if (ss->css_online) 5739 ret = ss->css_online(css); 5740 if (!ret) { 5741 css->flags |= CSS_ONLINE; 5742 rcu_assign_pointer(css->cgroup->subsys[ss->id], css); 5743 5744 atomic_inc(&css->online_cnt); 5745 if (css->parent) { 5746 atomic_inc(&css->parent->online_cnt); 5747 while ((css = css->parent)) 5748 css->nr_descendants++; 5749 } 5750 } 5751 return ret; 5752} 5753 5754/* if the CSS is online, invoke ->css_offline() on it and mark it offline */ 5755static void offline_css(struct cgroup_subsys_state *css) 5756{ 5757 struct cgroup_subsys *ss = css->ss; 5758 5759 lockdep_assert_held(&cgroup_mutex); 5760 5761 if (!(css->flags & CSS_ONLINE)) 5762 return; 5763 5764 if (ss->css_offline) 5765 ss->css_offline(css); 5766 5767 css->flags &= ~CSS_ONLINE; 5768 RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL); 5769 5770 wake_up_all(&css->cgroup->offline_waitq); 5771 5772 css->cgroup->nr_dying_subsys[ss->id]++; 5773 /* 5774 * Parent css and cgroup cannot be freed until after the freeing 5775 * of child css, see css_free_rwork_fn(). 5776 */ 5777 while ((css = css->parent)) { 5778 css->nr_descendants--; 5779 css->cgroup->nr_dying_subsys[ss->id]++; 5780 } 5781} 5782 5783/** 5784 * css_create - create a cgroup_subsys_state 5785 * @cgrp: the cgroup new css will be associated with 5786 * @ss: the subsys of new css 5787 * 5788 * Create a new css associated with @cgrp - @ss pair. On success, the new 5789 * css is online and installed in @cgrp. This function doesn't create the 5790 * interface files. Returns 0 on success, -errno on failure. 5791 */ 5792static struct cgroup_subsys_state *css_create(struct cgroup *cgrp, 5793 struct cgroup_subsys *ss) 5794{ 5795 struct cgroup *parent = cgroup_parent(cgrp); 5796 struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss); 5797 struct cgroup_subsys_state *css; 5798 int err; 5799 5800 lockdep_assert_held(&cgroup_mutex); 5801 5802 css = ss->css_alloc(parent_css); 5803 if (!css) 5804 css = ERR_PTR(-ENOMEM); 5805 if (IS_ERR(css)) 5806 return css; 5807 5808 init_and_link_css(css, ss, cgrp); 5809 5810 err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL); 5811 if (err) 5812 goto err_free_css; 5813 5814 err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL); 5815 if (err < 0) 5816 goto err_free_css; 5817 css->id = err; 5818 5819 err = css_rstat_init(css); 5820 if (err) 5821 goto err_free_css; 5822 5823 /* @css is ready to be brought online now, make it visible */ 5824 list_add_tail_rcu(&css->sibling, &parent_css->children); 5825 cgroup_idr_replace(&ss->css_idr, css, css->id); 5826 5827 err = online_css(css); 5828 if (err) 5829 goto err_list_del; 5830 5831 return css; 5832 5833err_list_del: 5834 list_del_rcu(&css->sibling); 5835err_free_css: 5836 INIT_RCU_WORK(&css->destroy_rwork, css_free_rwork_fn); 5837 queue_rcu_work(cgroup_free_wq, &css->destroy_rwork); 5838 return ERR_PTR(err); 5839} 5840 5841/* 5842 * The returned cgroup is fully initialized including its control mask, but 5843 * it doesn't have the control mask applied. 5844 */ 5845static struct cgroup *cgroup_create(struct cgroup *parent, const char *name, 5846 umode_t mode) 5847{ 5848 struct cgroup_root *root = parent->root; 5849 struct cgroup *cgrp, *tcgrp; 5850 struct kernfs_node *kn; 5851 int i, level = parent->level + 1; 5852 int ret; 5853 5854 /* allocate the cgroup and its ID, 0 is reserved for the root */ 5855 cgrp = kzalloc(struct_size(cgrp, _low_ancestors, level), GFP_KERNEL); 5856 if (!cgrp) 5857 return ERR_PTR(-ENOMEM); 5858 5859 ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL); 5860 if (ret) 5861 goto out_free_cgrp; 5862 5863 /* create the directory */ 5864 kn = kernfs_create_dir_ns(parent->kn, name, mode, 5865 current_fsuid(), current_fsgid(), 5866 cgrp, NULL); 5867 if (IS_ERR(kn)) { 5868 ret = PTR_ERR(kn); 5869 goto out_cancel_ref; 5870 } 5871 cgrp->kn = kn; 5872 5873 init_cgroup_housekeeping(cgrp); 5874 5875 cgrp->self.parent = &parent->self; 5876 cgrp->root = root; 5877 cgrp->level = level; 5878 5879 /* 5880 * Now that init_cgroup_housekeeping() has been called and cgrp->self 5881 * is setup, it is safe to perform rstat initialization on it. 5882 */ 5883 ret = css_rstat_init(&cgrp->self); 5884 if (ret) 5885 goto out_kernfs_remove; 5886 5887 ret = psi_cgroup_alloc(cgrp); 5888 if (ret) 5889 goto out_stat_exit; 5890 5891 for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) 5892 cgrp->ancestors[tcgrp->level] = tcgrp; 5893 5894 /* 5895 * New cgroup inherits effective freeze counter, and 5896 * if the parent has to be frozen, the child has too. 5897 */ 5898 cgrp->freezer.e_freeze = parent->freezer.e_freeze; 5899 seqcount_spinlock_init(&cgrp->freezer.freeze_seq, &css_set_lock); 5900 if (cgrp->freezer.e_freeze) { 5901 /* 5902 * Set the CGRP_FREEZE flag, so when a process will be 5903 * attached to the child cgroup, it will become frozen. 5904 * At this point the new cgroup is unpopulated, so we can 5905 * consider it frozen immediately. 5906 */ 5907 set_bit(CGRP_FREEZE, &cgrp->flags); 5908 cgrp->freezer.freeze_start_nsec = ktime_get_ns(); 5909 set_bit(CGRP_FROZEN, &cgrp->flags); 5910 } 5911 5912 if (notify_on_release(parent)) 5913 set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags); 5914 5915 if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags)) 5916 set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags); 5917 5918 cgrp->self.serial_nr = css_serial_nr_next++; 5919 5920 ret = blocking_notifier_call_chain_robust(&cgroup_lifetime_notifier, 5921 CGROUP_LIFETIME_ONLINE, 5922 CGROUP_LIFETIME_OFFLINE, cgrp); 5923 ret = notifier_to_errno(ret); 5924 if (ret) 5925 goto out_psi_free; 5926 5927 /* allocation complete, commit to creation */ 5928 spin_lock_irq(&css_set_lock); 5929 for (i = 0; i < level; i++) { 5930 tcgrp = cgrp->ancestors[i]; 5931 tcgrp->nr_descendants++; 5932 5933 /* 5934 * If the new cgroup is frozen, all ancestor cgroups get a new 5935 * frozen descendant, but their state can't change because of 5936 * this. 5937 */ 5938 if (cgrp->freezer.e_freeze) 5939 tcgrp->freezer.nr_frozen_descendants++; 5940 } 5941 spin_unlock_irq(&css_set_lock); 5942 5943 list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children); 5944 atomic_inc(&root->nr_cgrps); 5945 cgroup_get_live(parent); 5946 5947 /* 5948 * On the default hierarchy, a child doesn't automatically inherit 5949 * subtree_control from the parent. Each is configured manually. 5950 */ 5951 if (!cgroup_on_dfl(cgrp)) 5952 cgrp->subtree_control = cgroup_control(cgrp); 5953 5954 cgroup_propagate_control(cgrp); 5955 5956 return cgrp; 5957 5958out_psi_free: 5959 psi_cgroup_free(cgrp); 5960out_stat_exit: 5961 css_rstat_exit(&cgrp->self); 5962out_kernfs_remove: 5963 kernfs_remove(cgrp->kn); 5964out_cancel_ref: 5965 percpu_ref_exit(&cgrp->self.refcnt); 5966out_free_cgrp: 5967 kfree(cgrp); 5968 return ERR_PTR(ret); 5969} 5970 5971static bool cgroup_check_hierarchy_limits(struct cgroup *parent) 5972{ 5973 struct cgroup *cgroup; 5974 int ret = false; 5975 int level = 0; 5976 5977 lockdep_assert_held(&cgroup_mutex); 5978 5979 for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) { 5980 if (cgroup->nr_descendants >= cgroup->max_descendants) 5981 goto fail; 5982 5983 if (level >= cgroup->max_depth) 5984 goto fail; 5985 5986 level++; 5987 } 5988 5989 ret = true; 5990fail: 5991 return ret; 5992} 5993 5994int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode) 5995{ 5996 struct cgroup *parent, *cgrp; 5997 int ret; 5998 5999 /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */ 6000 if (strchr(name, '\n')) 6001 return -EINVAL; 6002 6003 parent = cgroup_kn_lock_live(parent_kn, false); 6004 if (!parent) 6005 return -ENODEV; 6006 6007 if (!cgroup_check_hierarchy_limits(parent)) { 6008 ret = -EAGAIN; 6009 goto out_unlock; 6010 } 6011 6012 cgrp = cgroup_create(parent, name, mode); 6013 if (IS_ERR(cgrp)) { 6014 ret = PTR_ERR(cgrp); 6015 goto out_unlock; 6016 } 6017 6018 /* 6019 * This extra ref will be put in css_free_rwork_fn() and guarantees 6020 * that @cgrp->kn is always accessible. 6021 */ 6022 kernfs_get(cgrp->kn); 6023 6024 ret = css_populate_dir(&cgrp->self); 6025 if (ret) 6026 goto out_destroy; 6027 6028 ret = cgroup_apply_control_enable(cgrp); 6029 if (ret) 6030 goto out_destroy; 6031 6032 TRACE_CGROUP_PATH(mkdir, cgrp); 6033 6034 /* let's create and online css's */ 6035 kernfs_activate(cgrp->kn); 6036 6037 ret = 0; 6038 goto out_unlock; 6039 6040out_destroy: 6041 cgroup_destroy_locked(cgrp); 6042out_unlock: 6043 cgroup_kn_unlock(parent_kn); 6044 return ret; 6045} 6046 6047/* 6048 * This is called when the refcnt of a css is confirmed to be killed. 6049 * css_tryget_online() is now guaranteed to fail. Tell the subsystem to 6050 * initiate destruction and put the css ref from kill_css(). 6051 */ 6052static void css_killed_work_fn(struct work_struct *work) 6053{ 6054 struct cgroup_subsys_state *css = 6055 container_of(work, struct cgroup_subsys_state, destroy_work); 6056 6057 cgroup_lock(); 6058 6059 do { 6060 offline_css(css); 6061 css_put(css); 6062 /* @css can't go away while we're holding cgroup_mutex */ 6063 css = css->parent; 6064 } while (css && atomic_dec_and_test(&css->online_cnt)); 6065 6066 cgroup_unlock(); 6067} 6068 6069/* css kill confirmation processing requires process context, bounce */ 6070static void css_killed_ref_fn(struct percpu_ref *ref) 6071{ 6072 struct cgroup_subsys_state *css = 6073 container_of(ref, struct cgroup_subsys_state, refcnt); 6074 6075 if (atomic_dec_and_test(&css->online_cnt)) { 6076 INIT_WORK(&css->destroy_work, css_killed_work_fn); 6077 queue_work(cgroup_offline_wq, &css->destroy_work); 6078 } 6079} 6080 6081/** 6082 * kill_css - destroy a css 6083 * @css: css to destroy 6084 * 6085 * This function initiates destruction of @css by removing cgroup interface 6086 * files and putting its base reference. ->css_offline() will be invoked 6087 * asynchronously once css_tryget_online() is guaranteed to fail and when 6088 * the reference count reaches zero, @css will be released. 6089 */ 6090static void kill_css(struct cgroup_subsys_state *css) 6091{ 6092 lockdep_assert_held(&cgroup_mutex); 6093 6094 if (css->flags & CSS_DYING) 6095 return; 6096 6097 /* 6098 * Call css_killed(), if defined, before setting the CSS_DYING flag 6099 */ 6100 if (css->ss->css_killed) 6101 css->ss->css_killed(css); 6102 6103 css->flags |= CSS_DYING; 6104 6105 /* 6106 * This must happen before css is disassociated with its cgroup. 6107 * See seq_css() for details. 6108 */ 6109 css_clear_dir(css); 6110 6111 /* 6112 * Killing would put the base ref, but we need to keep it alive 6113 * until after ->css_offline(). 6114 */ 6115 css_get(css); 6116 6117 /* 6118 * cgroup core guarantees that, by the time ->css_offline() is 6119 * invoked, no new css reference will be given out via 6120 * css_tryget_online(). We can't simply call percpu_ref_kill() and 6121 * proceed to offlining css's because percpu_ref_kill() doesn't 6122 * guarantee that the ref is seen as killed on all CPUs on return. 6123 * 6124 * Use percpu_ref_kill_and_confirm() to get notifications as each 6125 * css is confirmed to be seen as killed on all CPUs. 6126 */ 6127 percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn); 6128} 6129 6130/** 6131 * cgroup_destroy_locked - the first stage of cgroup destruction 6132 * @cgrp: cgroup to be destroyed 6133 * 6134 * css's make use of percpu refcnts whose killing latency shouldn't be 6135 * exposed to userland and are RCU protected. Also, cgroup core needs to 6136 * guarantee that css_tryget_online() won't succeed by the time 6137 * ->css_offline() is invoked. To satisfy all the requirements, 6138 * destruction is implemented in the following two steps. 6139 * 6140 * s1. Verify @cgrp can be destroyed and mark it dying. Remove all 6141 * userland visible parts and start killing the percpu refcnts of 6142 * css's. Set up so that the next stage will be kicked off once all 6143 * the percpu refcnts are confirmed to be killed. 6144 * 6145 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the 6146 * rest of destruction. Once all cgroup references are gone, the 6147 * cgroup is RCU-freed. 6148 * 6149 * This function implements s1. After this step, @cgrp is gone as far as 6150 * the userland is concerned and a new cgroup with the same name may be 6151 * created. As cgroup doesn't care about the names internally, this 6152 * doesn't cause any problem. 6153 */ 6154static int cgroup_destroy_locked(struct cgroup *cgrp) 6155 __releases(&cgroup_mutex) __acquires(&cgroup_mutex) 6156{ 6157 struct cgroup *tcgrp, *parent = cgroup_parent(cgrp); 6158 struct cgroup_subsys_state *css; 6159 struct cgrp_cset_link *link; 6160 int ssid, ret; 6161 6162 lockdep_assert_held(&cgroup_mutex); 6163 6164 /* 6165 * Only migration can raise populated from zero and we're already 6166 * holding cgroup_mutex. 6167 */ 6168 if (cgroup_is_populated(cgrp)) 6169 return -EBUSY; 6170 6171 /* 6172 * Make sure there's no live children. We can't test emptiness of 6173 * ->self.children as dead children linger on it while being 6174 * drained; otherwise, "rmdir parent/child parent" may fail. 6175 */ 6176 if (css_has_online_children(&cgrp->self)) 6177 return -EBUSY; 6178 6179 /* 6180 * Mark @cgrp and the associated csets dead. The former prevents 6181 * further task migration and child creation by disabling 6182 * cgroup_kn_lock_live(). The latter makes the csets ignored by 6183 * the migration path. 6184 */ 6185 cgrp->self.flags &= ~CSS_ONLINE; 6186 6187 spin_lock_irq(&css_set_lock); 6188 list_for_each_entry(link, &cgrp->cset_links, cset_link) 6189 link->cset->dead = true; 6190 spin_unlock_irq(&css_set_lock); 6191 6192 /* initiate massacre of all css's */ 6193 for_each_css(css, ssid, cgrp) 6194 kill_css(css); 6195 6196 /* clear and remove @cgrp dir, @cgrp has an extra ref on its kn */ 6197 css_clear_dir(&cgrp->self); 6198 kernfs_remove(cgrp->kn); 6199 6200 if (cgroup_is_threaded(cgrp)) 6201 parent->nr_threaded_children--; 6202 6203 spin_lock_irq(&css_set_lock); 6204 for (tcgrp = parent; tcgrp; tcgrp = cgroup_parent(tcgrp)) { 6205 tcgrp->nr_descendants--; 6206 tcgrp->nr_dying_descendants++; 6207 /* 6208 * If the dying cgroup is frozen, decrease frozen descendants 6209 * counters of ancestor cgroups. 6210 */ 6211 if (test_bit(CGRP_FROZEN, &cgrp->flags)) 6212 tcgrp->freezer.nr_frozen_descendants--; 6213 } 6214 spin_unlock_irq(&css_set_lock); 6215 6216 cgroup1_check_for_release(parent); 6217 6218 ret = blocking_notifier_call_chain(&cgroup_lifetime_notifier, 6219 CGROUP_LIFETIME_OFFLINE, cgrp); 6220 WARN_ON_ONCE(notifier_to_errno(ret)); 6221 6222 /* put the base reference */ 6223 percpu_ref_kill(&cgrp->self.refcnt); 6224 6225 return 0; 6226}; 6227 6228/** 6229 * cgroup_drain_dying - wait for dying tasks to leave before rmdir 6230 * @cgrp: the cgroup being removed 6231 * 6232 * cgroup.procs and cgroup.threads use css_task_iter which filters out 6233 * PF_EXITING tasks so that userspace doesn't see tasks that have already been 6234 * reaped via waitpid(). However, cgroup_has_tasks() - which tests whether the 6235 * cgroup has non-empty css_sets - is only updated when dying tasks pass through 6236 * cgroup_task_dead() in finish_task_switch(). This creates a window where 6237 * cgroup.procs reads empty but cgroup_has_tasks() is still true, making rmdir 6238 * fail with -EBUSY from cgroup_destroy_locked() even though userspace sees no 6239 * tasks. 6240 * 6241 * This function aligns cgroup_has_tasks() with what userspace can observe. If 6242 * cgroup_has_tasks() but the task iterator sees nothing (all remaining tasks are 6243 * PF_EXITING), we wait for cgroup_task_dead() to finish processing them. As the 6244 * window between PF_EXITING and cgroup_task_dead() is short, the wait is brief. 6245 * 6246 * This function only concerns itself with this cgroup's own dying tasks. 6247 * Whether the cgroup has children is cgroup_destroy_locked()'s problem. 6248 * 6249 * Each cgroup_task_dead() kicks the waitqueue via cset->cgrp_links, and we 6250 * retry the full check from scratch. 6251 * 6252 * Must be called with cgroup_mutex held. 6253 */ 6254static int cgroup_drain_dying(struct cgroup *cgrp) 6255 __releases(&cgroup_mutex) __acquires(&cgroup_mutex) 6256{ 6257 struct css_task_iter it; 6258 struct task_struct *task; 6259 DEFINE_WAIT(wait); 6260 6261 lockdep_assert_held(&cgroup_mutex); 6262retry: 6263 if (!cgroup_has_tasks(cgrp)) 6264 return 0; 6265 6266 /* Same iterator as cgroup.threads - if any task is visible, it's busy */ 6267 css_task_iter_start(&cgrp->self, 0, &it); 6268 task = css_task_iter_next(&it); 6269 css_task_iter_end(&it); 6270 6271 if (task) 6272 return -EBUSY; 6273 6274 /* 6275 * All remaining tasks are PF_EXITING and will pass through 6276 * cgroup_task_dead() shortly. Wait for a kick and retry. 6277 * 6278 * cgroup_has_tasks() can't transition from false to true while we're 6279 * holding cgroup_mutex, but the true to false transition happens 6280 * under css_set_lock (via cgroup_task_dead()). We must retest and 6281 * prepare_to_wait() under css_set_lock. Otherwise, the transition 6282 * can happen between our first test and prepare_to_wait(), and we 6283 * sleep with no one to wake us. 6284 */ 6285 spin_lock_irq(&css_set_lock); 6286 if (!cgroup_has_tasks(cgrp)) { 6287 spin_unlock_irq(&css_set_lock); 6288 return 0; 6289 } 6290 prepare_to_wait(&cgrp->dying_populated_waitq, &wait, 6291 TASK_UNINTERRUPTIBLE); 6292 spin_unlock_irq(&css_set_lock); 6293 mutex_unlock(&cgroup_mutex); 6294 schedule(); 6295 finish_wait(&cgrp->dying_populated_waitq, &wait); 6296 mutex_lock(&cgroup_mutex); 6297 goto retry; 6298} 6299 6300int cgroup_rmdir(struct kernfs_node *kn) 6301{ 6302 struct cgroup *cgrp; 6303 int ret = 0; 6304 6305 cgrp = cgroup_kn_lock_live(kn, false); 6306 if (!cgrp) 6307 return 0; 6308 6309 ret = cgroup_drain_dying(cgrp); 6310 if (!ret) { 6311 ret = cgroup_destroy_locked(cgrp); 6312 if (!ret) 6313 TRACE_CGROUP_PATH(rmdir, cgrp); 6314 } 6315 6316 cgroup_kn_unlock(kn); 6317 return ret; 6318} 6319 6320static struct kernfs_syscall_ops cgroup_kf_syscall_ops = { 6321 .show_options = cgroup_show_options, 6322 .mkdir = cgroup_mkdir, 6323 .rmdir = cgroup_rmdir, 6324 .show_path = cgroup_show_path, 6325}; 6326 6327static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early) 6328{ 6329 struct cgroup_subsys_state *css; 6330 6331 pr_debug("Initializing cgroup subsys %s\n", ss->name); 6332 6333 cgroup_lock(); 6334 6335 idr_init(&ss->css_idr); 6336 INIT_LIST_HEAD(&ss->cfts); 6337 6338 /* Create the root cgroup state for this subsystem */ 6339 ss->root = &cgrp_dfl_root; 6340 css = ss->css_alloc(NULL); 6341 /* We don't handle early failures gracefully */ 6342 BUG_ON(IS_ERR(css)); 6343 init_and_link_css(css, ss, &cgrp_dfl_root.cgrp); 6344 6345 /* 6346 * Root csses are never destroyed and we can't initialize 6347 * percpu_ref during early init. Disable refcnting. 6348 */ 6349 css->flags |= CSS_NO_REF; 6350 6351 if (early) { 6352 /* allocation can't be done safely during early init */ 6353 css->id = 1; 6354 } else { 6355 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL); 6356 BUG_ON(css->id < 0); 6357 6358 BUG_ON(ss_rstat_init(ss)); 6359 BUG_ON(css_rstat_init(css)); 6360 } 6361 6362 /* Update the init_css_set to contain a subsys 6363 * pointer to this state - since the subsystem is 6364 * newly registered, all tasks and hence the 6365 * init_css_set is in the subsystem's root cgroup. */ 6366 init_css_set.subsys[ss->id] = css; 6367 6368 have_fork_callback |= (bool)ss->fork << ss->id; 6369 have_exit_callback |= (bool)ss->exit << ss->id; 6370 have_release_callback |= (bool)ss->release << ss->id; 6371 have_canfork_callback |= (bool)ss->can_fork << ss->id; 6372 6373 /* At system boot, before all subsystems have been 6374 * registered, no tasks have been forked, so we don't 6375 * need to invoke fork callbacks here. */ 6376 BUG_ON(!list_empty(&init_task.tasks)); 6377 6378 BUG_ON(online_css(css)); 6379 6380 cgroup_unlock(); 6381} 6382 6383/** 6384 * cgroup_init_early - cgroup initialization at system boot 6385 * 6386 * Initialize cgroups at system boot, and initialize any 6387 * subsystems that request early init. 6388 */ 6389int __init cgroup_init_early(void) 6390{ 6391 static struct cgroup_fs_context __initdata ctx; 6392 struct cgroup_subsys *ss; 6393 int i; 6394 6395 ctx.root = &cgrp_dfl_root; 6396 init_cgroup_root(&ctx); 6397 cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF; 6398 6399 RCU_INIT_POINTER(init_task.cgroups, &init_css_set); 6400 6401 for_each_subsys(ss, i) { 6402 WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id, 6403 "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n", 6404 i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free, 6405 ss->id, ss->name); 6406 WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN, 6407 "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]); 6408 WARN(ss->early_init && ss->css_rstat_flush, 6409 "cgroup rstat cannot be used with early init subsystem\n"); 6410 6411 ss->id = i; 6412 ss->name = cgroup_subsys_name[i]; 6413 if (!ss->legacy_name) 6414 ss->legacy_name = cgroup_subsys_name[i]; 6415 6416 if (ss->early_init) 6417 cgroup_init_subsys(ss, true); 6418 } 6419 return 0; 6420} 6421 6422/** 6423 * cgroup_init - cgroup initialization 6424 * 6425 * Register cgroup filesystem and /proc file, and initialize 6426 * any subsystems that didn't request early init. 6427 */ 6428int __init cgroup_init(void) 6429{ 6430 struct cgroup_subsys *ss; 6431 int ssid; 6432 6433 BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16); 6434 BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files)); 6435 BUG_ON(cgroup_init_cftypes(NULL, cgroup_psi_files)); 6436 BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files)); 6437 6438 BUG_ON(ss_rstat_init(NULL)); 6439 6440 get_user_ns(init_cgroup_ns.user_ns); 6441 cgroup_rt_init(); 6442 6443 cgroup_lock(); 6444 6445 /* 6446 * Add init_css_set to the hash table so that dfl_root can link to 6447 * it during init. 6448 */ 6449 hash_add(css_set_table, &init_css_set.hlist, 6450 css_set_hash(init_css_set.subsys)); 6451 6452 cgroup_bpf_lifetime_notifier_init(); 6453 6454 BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0)); 6455 6456 cgroup_unlock(); 6457 6458 for_each_subsys(ss, ssid) { 6459 if (ss->early_init) { 6460 struct cgroup_subsys_state *css = 6461 init_css_set.subsys[ss->id]; 6462 6463 css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, 6464 GFP_KERNEL); 6465 BUG_ON(css->id < 0); 6466 } else { 6467 cgroup_init_subsys(ss, false); 6468 } 6469 6470 list_add_tail(&init_css_set.e_cset_node[ssid], 6471 &cgrp_dfl_root.cgrp.e_csets[ssid]); 6472 6473 /* 6474 * Setting dfl_root subsys_mask needs to consider the 6475 * disabled flag and cftype registration needs kmalloc, 6476 * both of which aren't available during early_init. 6477 */ 6478 if (!cgroup_ssid_enabled(ssid)) 6479 continue; 6480 6481 if (cgroup1_ssid_disabled(ssid)) 6482 pr_info("Disabling %s control group subsystem in v1 mounts\n", 6483 ss->legacy_name); 6484 6485 cgrp_dfl_root.subsys_mask |= 1 << ss->id; 6486 6487 /* implicit controllers must be threaded too */ 6488 WARN_ON(ss->implicit_on_dfl && !ss->threaded); 6489 6490 if (ss->implicit_on_dfl) 6491 cgrp_dfl_implicit_ss_mask |= 1 << ss->id; 6492 else if (!ss->dfl_cftypes) 6493 cgrp_dfl_inhibit_ss_mask |= 1 << ss->id; 6494 6495 if (ss->threaded) 6496 cgrp_dfl_threaded_ss_mask |= 1 << ss->id; 6497 6498 if (ss->dfl_cftypes == ss->legacy_cftypes) { 6499 WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes)); 6500 } else { 6501 WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes)); 6502 WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes)); 6503 } 6504 6505 if (ss->bind) 6506 ss->bind(init_css_set.subsys[ssid]); 6507 6508 cgroup_lock(); 6509 css_populate_dir(init_css_set.subsys[ssid]); 6510 cgroup_unlock(); 6511 } 6512 6513 /* init_css_set.subsys[] has been updated, re-hash */ 6514 hash_del(&init_css_set.hlist); 6515 hash_add(css_set_table, &init_css_set.hlist, 6516 css_set_hash(init_css_set.subsys)); 6517 6518 WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup")); 6519 WARN_ON(register_filesystem(&cgroup_fs_type)); 6520 WARN_ON(register_filesystem(&cgroup2_fs_type)); 6521 WARN_ON(!proc_create_single("cgroups", 0, NULL, proc_cgroupstats_show)); 6522#ifdef CONFIG_CPUSETS_V1 6523 WARN_ON(register_filesystem(&cpuset_fs_type)); 6524#endif 6525 6526 ns_tree_add(&init_cgroup_ns); 6527 return 0; 6528} 6529 6530static int __init cgroup_wq_init(void) 6531{ 6532 /* 6533 * There isn't much point in executing destruction path in 6534 * parallel. Good chunk is serialized with cgroup_mutex anyway. 6535 * Use 1 for @max_active. 6536 * 6537 * We would prefer to do this in cgroup_init() above, but that 6538 * is called before init_workqueues(): so leave this until after. 6539 */ 6540 cgroup_offline_wq = alloc_workqueue("cgroup_offline", WQ_PERCPU, 1); 6541 BUG_ON(!cgroup_offline_wq); 6542 6543 cgroup_release_wq = alloc_workqueue("cgroup_release", WQ_PERCPU, 1); 6544 BUG_ON(!cgroup_release_wq); 6545 6546 cgroup_free_wq = alloc_workqueue("cgroup_free", WQ_PERCPU, 1); 6547 BUG_ON(!cgroup_free_wq); 6548 return 0; 6549} 6550core_initcall(cgroup_wq_init); 6551 6552void cgroup_path_from_kernfs_id(u64 id, char *buf, size_t buflen) 6553{ 6554 struct kernfs_node *kn; 6555 6556 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id); 6557 if (!kn) 6558 return; 6559 kernfs_path(kn, buf, buflen); 6560 kernfs_put(kn); 6561} 6562 6563/* 6564 * __cgroup_get_from_id : get the cgroup associated with cgroup id 6565 * @id: cgroup id 6566 * On success return the cgrp or ERR_PTR on failure 6567 * There are no cgroup NS restrictions. 6568 */ 6569struct cgroup *__cgroup_get_from_id(u64 id) 6570{ 6571 struct kernfs_node *kn; 6572 struct cgroup *cgrp; 6573 6574 kn = kernfs_find_and_get_node_by_id(cgrp_dfl_root.kf_root, id); 6575 if (!kn) 6576 return ERR_PTR(-ENOENT); 6577 6578 if (kernfs_type(kn) != KERNFS_DIR) { 6579 kernfs_put(kn); 6580 return ERR_PTR(-ENOENT); 6581 } 6582 6583 rcu_read_lock(); 6584 6585 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); 6586 if (cgrp && !cgroup_tryget(cgrp)) 6587 cgrp = NULL; 6588 6589 rcu_read_unlock(); 6590 kernfs_put(kn); 6591 6592 if (!cgrp) 6593 return ERR_PTR(-ENOENT); 6594 return cgrp; 6595} 6596 6597/* 6598 * cgroup_get_from_id : get the cgroup associated with cgroup id 6599 * @id: cgroup id 6600 * On success return the cgrp or ERR_PTR on failure 6601 * Only cgroups within current task's cgroup NS are valid. 6602 */ 6603struct cgroup *cgroup_get_from_id(u64 id) 6604{ 6605 struct cgroup *cgrp, *root_cgrp; 6606 6607 cgrp = __cgroup_get_from_id(id); 6608 if (IS_ERR(cgrp)) 6609 return cgrp; 6610 6611 root_cgrp = current_cgns_cgroup_dfl(); 6612 if (!cgroup_is_descendant(cgrp, root_cgrp)) { 6613 cgroup_put(cgrp); 6614 return ERR_PTR(-ENOENT); 6615 } 6616 6617 return cgrp; 6618} 6619EXPORT_SYMBOL_GPL(cgroup_get_from_id); 6620 6621/* 6622 * proc_cgroup_show() 6623 * - Print task's cgroup paths into seq_file, one line for each hierarchy 6624 * - Used for /proc/<pid>/cgroup. 6625 */ 6626int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns, 6627 struct pid *pid, struct task_struct *tsk) 6628{ 6629 char *buf; 6630 int retval; 6631 struct cgroup_root *root; 6632 6633 retval = -ENOMEM; 6634 buf = kmalloc(PATH_MAX, GFP_KERNEL); 6635 if (!buf) 6636 goto out; 6637 6638 rcu_read_lock(); 6639 spin_lock_irq(&css_set_lock); 6640 6641 for_each_root(root) { 6642 struct cgroup_subsys *ss; 6643 struct cgroup *cgrp; 6644 int ssid, count = 0; 6645 6646 if (root == &cgrp_dfl_root && !READ_ONCE(cgrp_dfl_visible)) 6647 continue; 6648 6649 cgrp = task_cgroup_from_root(tsk, root); 6650 /* The root has already been unmounted. */ 6651 if (!cgrp) 6652 continue; 6653 6654 seq_printf(m, "%d:", root->hierarchy_id); 6655 if (root != &cgrp_dfl_root) 6656 for_each_subsys(ss, ssid) 6657 if (root->subsys_mask & (1 << ssid)) 6658 seq_printf(m, "%s%s", count++ ? "," : "", 6659 ss->legacy_name); 6660 if (strlen(root->name)) 6661 seq_printf(m, "%sname=%s", count ? "," : "", 6662 root->name); 6663 seq_putc(m, ':'); 6664 /* 6665 * On traditional hierarchies, all zombie tasks show up as 6666 * belonging to the root cgroup. On the default hierarchy, 6667 * while a zombie doesn't show up in "cgroup.procs" and 6668 * thus can't be migrated, its /proc/PID/cgroup keeps 6669 * reporting the cgroup it belonged to before exiting. If 6670 * the cgroup is removed before the zombie is reaped, 6671 * " (deleted)" is appended to the cgroup path. 6672 */ 6673 if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) { 6674 retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX, 6675 current->nsproxy->cgroup_ns); 6676 if (retval == -E2BIG) 6677 retval = -ENAMETOOLONG; 6678 if (retval < 0) 6679 goto out_unlock; 6680 6681 seq_puts(m, buf); 6682 } else { 6683 seq_puts(m, "/"); 6684 } 6685 6686 if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp)) 6687 seq_puts(m, " (deleted)\n"); 6688 else 6689 seq_putc(m, '\n'); 6690 } 6691 6692 retval = 0; 6693out_unlock: 6694 spin_unlock_irq(&css_set_lock); 6695 rcu_read_unlock(); 6696 kfree(buf); 6697out: 6698 return retval; 6699} 6700 6701/** 6702 * cgroup_fork - initialize cgroup related fields during copy_process() 6703 * @child: pointer to task_struct of forking parent process. 6704 * 6705 * A task is associated with the init_css_set until cgroup_post_fork() 6706 * attaches it to the target css_set. 6707 */ 6708void cgroup_fork(struct task_struct *child) 6709{ 6710 RCU_INIT_POINTER(child->cgroups, &init_css_set); 6711 INIT_LIST_HEAD(&child->cg_list); 6712} 6713 6714/** 6715 * cgroup_v1v2_get_from_file - get a cgroup pointer from a file pointer 6716 * @f: file corresponding to cgroup_dir 6717 * 6718 * Find the cgroup from a file pointer associated with a cgroup directory. 6719 * Returns a pointer to the cgroup on success. ERR_PTR is returned if the 6720 * cgroup cannot be found. 6721 */ 6722static struct cgroup *cgroup_v1v2_get_from_file(struct file *f) 6723{ 6724 struct cgroup_subsys_state *css; 6725 6726 css = css_tryget_online_from_dir(f->f_path.dentry, NULL); 6727 if (IS_ERR(css)) 6728 return ERR_CAST(css); 6729 6730 return css->cgroup; 6731} 6732 6733/** 6734 * cgroup_get_from_file - same as cgroup_v1v2_get_from_file, but only supports 6735 * cgroup2. 6736 * @f: file corresponding to cgroup2_dir 6737 */ 6738static struct cgroup *cgroup_get_from_file(struct file *f) 6739{ 6740 struct cgroup *cgrp = cgroup_v1v2_get_from_file(f); 6741 6742 if (IS_ERR(cgrp)) 6743 return ERR_CAST(cgrp); 6744 6745 if (!cgroup_on_dfl(cgrp)) { 6746 cgroup_put(cgrp); 6747 return ERR_PTR(-EBADF); 6748 } 6749 6750 return cgrp; 6751} 6752 6753/** 6754 * cgroup_css_set_fork - find or create a css_set for a child process 6755 * @kargs: the arguments passed to create the child process 6756 * 6757 * This functions finds or creates a new css_set which the child 6758 * process will be attached to in cgroup_post_fork(). By default, 6759 * the child process will be given the same css_set as its parent. 6760 * 6761 * If CLONE_INTO_CGROUP is specified this function will try to find an 6762 * existing css_set which includes the requested cgroup and if not create 6763 * a new css_set that the child will be attached to later. If this function 6764 * succeeds it will hold cgroup_threadgroup_rwsem on return. If 6765 * CLONE_INTO_CGROUP is requested this function will grab cgroup mutex 6766 * before grabbing cgroup_threadgroup_rwsem and will hold a reference 6767 * to the target cgroup. 6768 */ 6769static int cgroup_css_set_fork(struct kernel_clone_args *kargs) 6770 __acquires(&cgroup_mutex) __acquires(&cgroup_threadgroup_rwsem) 6771{ 6772 int ret; 6773 struct cgroup *dst_cgrp = NULL; 6774 struct css_set *cset; 6775 struct super_block *sb; 6776 6777 if (kargs->flags & CLONE_INTO_CGROUP) 6778 cgroup_lock(); 6779 6780 cgroup_threadgroup_change_begin(current); 6781 6782 spin_lock_irq(&css_set_lock); 6783 cset = task_css_set(current); 6784 get_css_set(cset); 6785 if (kargs->cgrp) 6786 kargs->kill_seq = kargs->cgrp->kill_seq; 6787 else 6788 kargs->kill_seq = cset->dfl_cgrp->kill_seq; 6789 spin_unlock_irq(&css_set_lock); 6790 6791 if (!(kargs->flags & CLONE_INTO_CGROUP)) { 6792 kargs->cset = cset; 6793 return 0; 6794 } 6795 6796 CLASS(fd_raw, f)(kargs->cgroup); 6797 if (fd_empty(f)) { 6798 ret = -EBADF; 6799 goto err; 6800 } 6801 sb = fd_file(f)->f_path.dentry->d_sb; 6802 6803 dst_cgrp = cgroup_get_from_file(fd_file(f)); 6804 if (IS_ERR(dst_cgrp)) { 6805 ret = PTR_ERR(dst_cgrp); 6806 dst_cgrp = NULL; 6807 goto err; 6808 } 6809 6810 if (cgroup_is_dead(dst_cgrp)) { 6811 ret = -ENODEV; 6812 goto err; 6813 } 6814 6815 /* 6816 * Verify that we the target cgroup is writable for us. This is 6817 * usually done by the vfs layer but since we're not going through 6818 * the vfs layer here we need to do it "manually". 6819 */ 6820 ret = cgroup_may_write(dst_cgrp, sb); 6821 if (ret) 6822 goto err; 6823 6824 /* 6825 * Spawning a task directly into a cgroup works by passing a file 6826 * descriptor to the target cgroup directory. This can even be an O_PATH 6827 * file descriptor. But it can never be a cgroup.procs file descriptor. 6828 * This was done on purpose so spawning into a cgroup could be 6829 * conceptualized as an atomic 6830 * 6831 * fd = openat(dfd_cgroup, "cgroup.procs", ...); 6832 * write(fd, <child-pid>, ...); 6833 * 6834 * sequence, i.e. it's a shorthand for the caller opening and writing 6835 * cgroup.procs of the cgroup indicated by @dfd_cgroup. This allows us 6836 * to always use the caller's credentials. 6837 */ 6838 ret = cgroup_attach_permissions(cset->dfl_cgrp, dst_cgrp, sb, 6839 !(kargs->flags & CLONE_THREAD), 6840 current->nsproxy->cgroup_ns); 6841 if (ret) 6842 goto err; 6843 6844 kargs->cset = find_css_set(cset, dst_cgrp); 6845 if (!kargs->cset) { 6846 ret = -ENOMEM; 6847 goto err; 6848 } 6849 6850 put_css_set(cset); 6851 kargs->cgrp = dst_cgrp; 6852 return ret; 6853 6854err: 6855 cgroup_threadgroup_change_end(current); 6856 cgroup_unlock(); 6857 if (dst_cgrp) 6858 cgroup_put(dst_cgrp); 6859 put_css_set(cset); 6860 if (kargs->cset) 6861 put_css_set(kargs->cset); 6862 return ret; 6863} 6864 6865/** 6866 * cgroup_css_set_put_fork - drop references we took during fork 6867 * @kargs: the arguments passed to create the child process 6868 * 6869 * Drop references to the prepared css_set and target cgroup if 6870 * CLONE_INTO_CGROUP was requested. 6871 */ 6872static void cgroup_css_set_put_fork(struct kernel_clone_args *kargs) 6873 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex) 6874{ 6875 struct cgroup *cgrp = kargs->cgrp; 6876 struct css_set *cset = kargs->cset; 6877 6878 cgroup_threadgroup_change_end(current); 6879 6880 if (cset) { 6881 put_css_set(cset); 6882 kargs->cset = NULL; 6883 } 6884 6885 if (kargs->flags & CLONE_INTO_CGROUP) { 6886 cgroup_unlock(); 6887 if (cgrp) { 6888 cgroup_put(cgrp); 6889 kargs->cgrp = NULL; 6890 } 6891 } 6892} 6893 6894/** 6895 * cgroup_can_fork - called on a new task before the process is exposed 6896 * @child: the child process 6897 * @kargs: the arguments passed to create the child process 6898 * 6899 * This prepares a new css_set for the child process which the child will 6900 * be attached to in cgroup_post_fork(). 6901 * This calls the subsystem can_fork() callbacks. If the cgroup_can_fork() 6902 * callback returns an error, the fork aborts with that error code. This 6903 * allows for a cgroup subsystem to conditionally allow or deny new forks. 6904 */ 6905int cgroup_can_fork(struct task_struct *child, struct kernel_clone_args *kargs) 6906{ 6907 struct cgroup_subsys *ss; 6908 int i, j, ret; 6909 6910 ret = cgroup_css_set_fork(kargs); 6911 if (ret) 6912 return ret; 6913 6914 do_each_subsys_mask(ss, i, have_canfork_callback) { 6915 ret = ss->can_fork(child, kargs->cset); 6916 if (ret) 6917 goto out_revert; 6918 } while_each_subsys_mask(); 6919 6920 return 0; 6921 6922out_revert: 6923 for_each_subsys(ss, j) { 6924 if (j >= i) 6925 break; 6926 if (ss->cancel_fork) 6927 ss->cancel_fork(child, kargs->cset); 6928 } 6929 6930 cgroup_css_set_put_fork(kargs); 6931 6932 return ret; 6933} 6934 6935/** 6936 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork() 6937 * @child: the child process 6938 * @kargs: the arguments passed to create the child process 6939 * 6940 * This calls the cancel_fork() callbacks if a fork failed *after* 6941 * cgroup_can_fork() succeeded and cleans up references we took to 6942 * prepare a new css_set for the child process in cgroup_can_fork(). 6943 */ 6944void cgroup_cancel_fork(struct task_struct *child, 6945 struct kernel_clone_args *kargs) 6946{ 6947 struct cgroup_subsys *ss; 6948 int i; 6949 6950 for_each_subsys(ss, i) 6951 if (ss->cancel_fork) 6952 ss->cancel_fork(child, kargs->cset); 6953 6954 cgroup_css_set_put_fork(kargs); 6955} 6956 6957/** 6958 * cgroup_post_fork - finalize cgroup setup for the child process 6959 * @child: the child process 6960 * @kargs: the arguments passed to create the child process 6961 * 6962 * Attach the child process to its css_set calling the subsystem fork() 6963 * callbacks. 6964 */ 6965void cgroup_post_fork(struct task_struct *child, 6966 struct kernel_clone_args *kargs) 6967 __releases(&cgroup_threadgroup_rwsem) __releases(&cgroup_mutex) 6968{ 6969 unsigned int cgrp_kill_seq = 0; 6970 unsigned long cgrp_flags = 0; 6971 bool kill = false; 6972 struct cgroup_subsys *ss; 6973 struct css_set *cset; 6974 int i; 6975 6976 cset = kargs->cset; 6977 kargs->cset = NULL; 6978 6979 spin_lock_irq(&css_set_lock); 6980 6981 /* init tasks are special, only link regular threads */ 6982 if (likely(child->pid)) { 6983 if (kargs->cgrp) { 6984 cgrp_flags = kargs->cgrp->flags; 6985 cgrp_kill_seq = kargs->cgrp->kill_seq; 6986 } else { 6987 cgrp_flags = cset->dfl_cgrp->flags; 6988 cgrp_kill_seq = cset->dfl_cgrp->kill_seq; 6989 } 6990 6991 WARN_ON_ONCE(!list_empty(&child->cg_list)); 6992 cset->nr_tasks++; 6993 css_set_move_task(child, NULL, cset, false); 6994 } else { 6995 put_css_set(cset); 6996 cset = NULL; 6997 } 6998 6999 if (!(child->flags & PF_KTHREAD)) { 7000 if (unlikely(test_bit(CGRP_FREEZE, &cgrp_flags))) { 7001 /* 7002 * If the cgroup has to be frozen, the new task has 7003 * too. Let's set the JOBCTL_TRAP_FREEZE jobctl bit to 7004 * get the task into the frozen state. 7005 */ 7006 spin_lock(&child->sighand->siglock); 7007 WARN_ON_ONCE(child->frozen); 7008 child->jobctl |= JOBCTL_TRAP_FREEZE; 7009 spin_unlock(&child->sighand->siglock); 7010 7011 /* 7012 * Calling cgroup_update_frozen() isn't required here, 7013 * because it will be called anyway a bit later from 7014 * do_freezer_trap(). So we avoid cgroup's transient 7015 * switch from the frozen state and back. 7016 */ 7017 } 7018 7019 /* 7020 * If the cgroup is to be killed notice it now and take the 7021 * child down right after we finished preparing it for 7022 * userspace. 7023 */ 7024 kill = kargs->kill_seq != cgrp_kill_seq; 7025 } 7026 7027 spin_unlock_irq(&css_set_lock); 7028 7029 /* 7030 * Call ss->fork(). This must happen after @child is linked on 7031 * css_set; otherwise, @child might change state between ->fork() 7032 * and addition to css_set. 7033 */ 7034 do_each_subsys_mask(ss, i, have_fork_callback) { 7035 ss->fork(child); 7036 } while_each_subsys_mask(); 7037 7038 /* Make the new cset the root_cset of the new cgroup namespace. */ 7039 if (kargs->flags & CLONE_NEWCGROUP) { 7040 struct css_set *rcset = child->nsproxy->cgroup_ns->root_cset; 7041 7042 get_css_set(cset); 7043 child->nsproxy->cgroup_ns->root_cset = cset; 7044 put_css_set(rcset); 7045 } 7046 7047 /* Cgroup has to be killed so take down child immediately. */ 7048 if (unlikely(kill)) 7049 do_send_sig_info(SIGKILL, SEND_SIG_NOINFO, child, PIDTYPE_TGID); 7050 7051 cgroup_css_set_put_fork(kargs); 7052} 7053 7054/** 7055 * cgroup_task_exit - detach cgroup from exiting task 7056 * @tsk: pointer to task_struct of exiting process 7057 * 7058 * Description: Detach cgroup from @tsk. 7059 * 7060 */ 7061void cgroup_task_exit(struct task_struct *tsk) 7062{ 7063 struct cgroup_subsys *ss; 7064 int i; 7065 7066 /* see cgroup_post_fork() for details */ 7067 do_each_subsys_mask(ss, i, have_exit_callback) { 7068 ss->exit(tsk); 7069 } while_each_subsys_mask(); 7070} 7071 7072static void do_cgroup_task_dead(struct task_struct *tsk) 7073{ 7074 struct cgrp_cset_link *link; 7075 struct css_set *cset; 7076 unsigned long flags; 7077 7078 spin_lock_irqsave(&css_set_lock, flags); 7079 7080 WARN_ON_ONCE(list_empty(&tsk->cg_list)); 7081 cset = task_css_set(tsk); 7082 css_set_move_task(tsk, cset, NULL, false); 7083 cset->nr_tasks--; 7084 /* matches the signal->live check in css_task_iter_advance() */ 7085 if (thread_group_leader(tsk) && atomic_read(&tsk->signal->live)) 7086 list_add_tail(&tsk->cg_list, &cset->dying_tasks); 7087 7088 /* kick cgroup_drain_dying() waiters, see cgroup_rmdir() */ 7089 list_for_each_entry(link, &cset->cgrp_links, cgrp_link) 7090 if (waitqueue_active(&link->cgrp->dying_populated_waitq)) 7091 wake_up(&link->cgrp->dying_populated_waitq); 7092 7093 if (dl_task(tsk)) 7094 dec_dl_tasks_cs(tsk); 7095 7096 WARN_ON_ONCE(cgroup_task_frozen(tsk)); 7097 if (unlikely(!(tsk->flags & PF_KTHREAD) && 7098 test_bit(CGRP_FREEZE, &task_dfl_cgroup(tsk)->flags))) 7099 cgroup_update_frozen(task_dfl_cgroup(tsk)); 7100 7101 spin_unlock_irqrestore(&css_set_lock, flags); 7102} 7103 7104#ifdef CONFIG_PREEMPT_RT 7105/* 7106 * cgroup_task_dead() is called from finish_task_switch() which doesn't allow 7107 * scheduling even in RT. As the task_dead path requires grabbing css_set_lock, 7108 * this lead to sleeping in the invalid context warning bug. css_set_lock is too 7109 * big to become a raw_spinlock. The task_dead path doesn't need to run 7110 * synchronously but can't be delayed indefinitely either as the dead task pins 7111 * the cgroup and task_struct can be pinned indefinitely. Bounce through lazy 7112 * irq_work to allow batching while ensuring timely completion. 7113 */ 7114static DEFINE_PER_CPU(struct llist_head, cgrp_dead_tasks); 7115static DEFINE_PER_CPU(struct irq_work, cgrp_dead_tasks_iwork); 7116 7117static void cgrp_dead_tasks_iwork_fn(struct irq_work *iwork) 7118{ 7119 struct llist_node *lnode; 7120 struct task_struct *task, *next; 7121 7122 lnode = llist_del_all(this_cpu_ptr(&cgrp_dead_tasks)); 7123 llist_for_each_entry_safe(task, next, lnode, cg_dead_lnode) { 7124 do_cgroup_task_dead(task); 7125 put_task_struct(task); 7126 } 7127} 7128 7129static void __init cgroup_rt_init(void) 7130{ 7131 int cpu; 7132 7133 for_each_possible_cpu(cpu) { 7134 init_llist_head(per_cpu_ptr(&cgrp_dead_tasks, cpu)); 7135 per_cpu(cgrp_dead_tasks_iwork, cpu) = 7136 IRQ_WORK_INIT_LAZY(cgrp_dead_tasks_iwork_fn); 7137 } 7138} 7139 7140void cgroup_task_dead(struct task_struct *task) 7141{ 7142 get_task_struct(task); 7143 llist_add(&task->cg_dead_lnode, this_cpu_ptr(&cgrp_dead_tasks)); 7144 irq_work_queue(this_cpu_ptr(&cgrp_dead_tasks_iwork)); 7145} 7146#else /* CONFIG_PREEMPT_RT */ 7147static void __init cgroup_rt_init(void) {} 7148 7149void cgroup_task_dead(struct task_struct *task) 7150{ 7151 do_cgroup_task_dead(task); 7152} 7153#endif /* CONFIG_PREEMPT_RT */ 7154 7155void cgroup_task_release(struct task_struct *task) 7156{ 7157 struct cgroup_subsys *ss; 7158 int ssid; 7159 7160 do_each_subsys_mask(ss, ssid, have_release_callback) { 7161 ss->release(task); 7162 } while_each_subsys_mask(); 7163} 7164 7165void cgroup_task_free(struct task_struct *task) 7166{ 7167 struct css_set *cset = task_css_set(task); 7168 7169 if (!list_empty(&task->cg_list)) { 7170 spin_lock_irq(&css_set_lock); 7171 css_set_skip_task_iters(task_css_set(task), task); 7172 list_del_init(&task->cg_list); 7173 spin_unlock_irq(&css_set_lock); 7174 } 7175 7176 put_css_set(cset); 7177} 7178 7179static int __init cgroup_disable(char *str) 7180{ 7181 struct cgroup_subsys *ss; 7182 char *token; 7183 int i; 7184 7185 while ((token = strsep(&str, ",")) != NULL) { 7186 if (!*token) 7187 continue; 7188 7189 for_each_subsys(ss, i) { 7190 if (strcmp(token, ss->name) && 7191 strcmp(token, ss->legacy_name)) 7192 continue; 7193 7194 static_branch_disable(cgroup_subsys_enabled_key[i]); 7195 pr_info("Disabling %s control group subsystem\n", 7196 ss->name); 7197 } 7198 7199 for (i = 0; i < OPT_FEATURE_COUNT; i++) { 7200 if (strcmp(token, cgroup_opt_feature_names[i])) 7201 continue; 7202 cgroup_feature_disable_mask |= 1 << i; 7203 pr_info("Disabling %s control group feature\n", 7204 cgroup_opt_feature_names[i]); 7205 break; 7206 } 7207 } 7208 return 1; 7209} 7210__setup("cgroup_disable=", cgroup_disable); 7211 7212void __init __weak enable_debug_cgroup(void) { } 7213 7214static int __init enable_cgroup_debug(char *str) 7215{ 7216 cgroup_debug = true; 7217 enable_debug_cgroup(); 7218 return 1; 7219} 7220__setup("cgroup_debug", enable_cgroup_debug); 7221 7222static int __init cgroup_favordynmods_setup(char *str) 7223{ 7224 return (kstrtobool(str, &have_favordynmods) == 0); 7225} 7226__setup("cgroup_favordynmods=", cgroup_favordynmods_setup); 7227 7228/** 7229 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry 7230 * @dentry: directory dentry of interest 7231 * @ss: subsystem of interest 7232 * 7233 * If @dentry is a directory for a cgroup which has @ss enabled on it, try 7234 * to get the corresponding css and return it. If such css doesn't exist 7235 * or can't be pinned, an ERR_PTR value is returned. 7236 */ 7237struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry, 7238 struct cgroup_subsys *ss) 7239{ 7240 struct kernfs_node *kn = kernfs_node_from_dentry(dentry); 7241 struct file_system_type *s_type = dentry->d_sb->s_type; 7242 struct cgroup_subsys_state *css = NULL; 7243 struct cgroup *cgrp; 7244 7245 /* is @dentry a cgroup dir? */ 7246 if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) || 7247 !kn || kernfs_type(kn) != KERNFS_DIR) 7248 return ERR_PTR(-EBADF); 7249 7250 rcu_read_lock(); 7251 7252 /* 7253 * This path doesn't originate from kernfs and @kn could already 7254 * have been or be removed at any point. @kn->priv is RCU 7255 * protected for this access. See css_release_work_fn() for details. 7256 */ 7257 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); 7258 if (cgrp) 7259 css = cgroup_css(cgrp, ss); 7260 7261 if (!css || !css_tryget_online(css)) 7262 css = ERR_PTR(-ENOENT); 7263 7264 rcu_read_unlock(); 7265 return css; 7266} 7267 7268/** 7269 * css_from_id - lookup css by id 7270 * @id: the cgroup id 7271 * @ss: cgroup subsys to be looked into 7272 * 7273 * Returns the css if there's valid one with @id, otherwise returns NULL. 7274 * Should be called under rcu_read_lock(). 7275 */ 7276struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss) 7277{ 7278 WARN_ON_ONCE(!rcu_read_lock_held()); 7279 return idr_find(&ss->css_idr, id); 7280} 7281 7282/** 7283 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path 7284 * @path: path on the default hierarchy 7285 * 7286 * Find the cgroup at @path on the default hierarchy, increment its 7287 * reference count and return it. Returns pointer to the found cgroup on 7288 * success, ERR_PTR(-ENOENT) if @path doesn't exist or if the cgroup has already 7289 * been released and ERR_PTR(-ENOTDIR) if @path points to a non-directory. 7290 */ 7291struct cgroup *cgroup_get_from_path(const char *path) 7292{ 7293 struct kernfs_node *kn; 7294 struct cgroup *cgrp = ERR_PTR(-ENOENT); 7295 struct cgroup *root_cgrp; 7296 7297 root_cgrp = current_cgns_cgroup_dfl(); 7298 kn = kernfs_walk_and_get(root_cgrp->kn, path); 7299 if (!kn) 7300 goto out; 7301 7302 if (kernfs_type(kn) != KERNFS_DIR) { 7303 cgrp = ERR_PTR(-ENOTDIR); 7304 goto out_kernfs; 7305 } 7306 7307 rcu_read_lock(); 7308 7309 cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv); 7310 if (!cgrp || !cgroup_tryget(cgrp)) 7311 cgrp = ERR_PTR(-ENOENT); 7312 7313 rcu_read_unlock(); 7314 7315out_kernfs: 7316 kernfs_put(kn); 7317out: 7318 return cgrp; 7319} 7320EXPORT_SYMBOL_GPL(cgroup_get_from_path); 7321 7322/** 7323 * cgroup_v1v2_get_from_fd - get a cgroup pointer from a fd 7324 * @fd: fd obtained by open(cgroup_dir) 7325 * 7326 * Find the cgroup from a fd which should be obtained 7327 * by opening a cgroup directory. Returns a pointer to the 7328 * cgroup on success. ERR_PTR is returned if the cgroup 7329 * cannot be found. 7330 */ 7331struct cgroup *cgroup_v1v2_get_from_fd(int fd) 7332{ 7333 CLASS(fd_raw, f)(fd); 7334 if (fd_empty(f)) 7335 return ERR_PTR(-EBADF); 7336 7337 return cgroup_v1v2_get_from_file(fd_file(f)); 7338} 7339 7340/** 7341 * cgroup_get_from_fd - same as cgroup_v1v2_get_from_fd, but only supports 7342 * cgroup2. 7343 * @fd: fd obtained by open(cgroup2_dir) 7344 */ 7345struct cgroup *cgroup_get_from_fd(int fd) 7346{ 7347 struct cgroup *cgrp = cgroup_v1v2_get_from_fd(fd); 7348 7349 if (IS_ERR(cgrp)) 7350 return ERR_CAST(cgrp); 7351 7352 if (!cgroup_on_dfl(cgrp)) { 7353 cgroup_put(cgrp); 7354 return ERR_PTR(-EBADF); 7355 } 7356 return cgrp; 7357} 7358EXPORT_SYMBOL_GPL(cgroup_get_from_fd); 7359 7360static u64 power_of_ten(int power) 7361{ 7362 u64 v = 1; 7363 while (power--) 7364 v *= 10; 7365 return v; 7366} 7367 7368/** 7369 * cgroup_parse_float - parse a floating number 7370 * @input: input string 7371 * @dec_shift: number of decimal digits to shift 7372 * @v: output 7373 * 7374 * Parse a decimal floating point number in @input and store the result in 7375 * @v with decimal point right shifted @dec_shift times. For example, if 7376 * @input is "12.3456" and @dec_shift is 3, *@v will be set to 12345. 7377 * Returns 0 on success, -errno otherwise. 7378 * 7379 * There's nothing cgroup specific about this function except that it's 7380 * currently the only user. 7381 */ 7382int cgroup_parse_float(const char *input, unsigned dec_shift, s64 *v) 7383{ 7384 s64 whole, frac = 0; 7385 int fstart = 0, fend = 0, flen; 7386 7387 if (!sscanf(input, "%lld.%n%lld%n", &whole, &fstart, &frac, &fend)) 7388 return -EINVAL; 7389 if (frac < 0) 7390 return -EINVAL; 7391 7392 flen = fend > fstart ? fend - fstart : 0; 7393 if (flen < dec_shift) 7394 frac *= power_of_ten(dec_shift - flen); 7395 else 7396 frac = DIV_ROUND_CLOSEST_ULL(frac, power_of_ten(flen - dec_shift)); 7397 7398 *v = whole * power_of_ten(dec_shift) + frac; 7399 return 0; 7400} 7401 7402/* 7403 * sock->sk_cgrp_data handling. For more info, see sock_cgroup_data 7404 * definition in cgroup-defs.h. 7405 */ 7406#ifdef CONFIG_SOCK_CGROUP_DATA 7407 7408void cgroup_sk_alloc(struct sock_cgroup_data *skcd) 7409{ 7410 struct cgroup *cgroup; 7411 7412 rcu_read_lock(); 7413 /* Don't associate the sock with unrelated interrupted task's cgroup. */ 7414 if (in_interrupt()) { 7415 cgroup = &cgrp_dfl_root.cgrp; 7416 cgroup_get(cgroup); 7417 goto out; 7418 } 7419 7420 while (true) { 7421 struct css_set *cset; 7422 7423 cset = task_css_set(current); 7424 if (likely(cgroup_tryget(cset->dfl_cgrp))) { 7425 cgroup = cset->dfl_cgrp; 7426 break; 7427 } 7428 cpu_relax(); 7429 } 7430out: 7431 skcd->cgroup = cgroup; 7432 cgroup_bpf_get(cgroup); 7433 rcu_read_unlock(); 7434} 7435 7436void cgroup_sk_clone(struct sock_cgroup_data *skcd) 7437{ 7438 struct cgroup *cgrp = sock_cgroup_ptr(skcd); 7439 7440 /* 7441 * We might be cloning a socket which is left in an empty 7442 * cgroup and the cgroup might have already been rmdir'd. 7443 * Don't use cgroup_get_live(). 7444 */ 7445 cgroup_get(cgrp); 7446 cgroup_bpf_get(cgrp); 7447} 7448 7449void cgroup_sk_free(struct sock_cgroup_data *skcd) 7450{ 7451 struct cgroup *cgrp = sock_cgroup_ptr(skcd); 7452 7453 cgroup_bpf_put(cgrp); 7454 cgroup_put(cgrp); 7455} 7456 7457#endif /* CONFIG_SOCK_CGROUP_DATA */ 7458 7459#ifdef CONFIG_SYSFS 7460static ssize_t show_delegatable_files(struct cftype *files, char *buf, 7461 ssize_t size, const char *prefix) 7462{ 7463 struct cftype *cft; 7464 ssize_t ret = 0; 7465 7466 for (cft = files; cft && cft->name[0] != '\0'; cft++) { 7467 if (!(cft->flags & CFTYPE_NS_DELEGATABLE)) 7468 continue; 7469 7470 if (prefix) 7471 ret += snprintf(buf + ret, size - ret, "%s.", prefix); 7472 7473 ret += snprintf(buf + ret, size - ret, "%s\n", cft->name); 7474 7475 if (WARN_ON(ret >= size)) 7476 break; 7477 } 7478 7479 return ret; 7480} 7481 7482static ssize_t delegate_show(struct kobject *kobj, struct kobj_attribute *attr, 7483 char *buf) 7484{ 7485 struct cgroup_subsys *ss; 7486 int ssid; 7487 ssize_t ret = 0; 7488 7489 ret = show_delegatable_files(cgroup_base_files, buf + ret, 7490 PAGE_SIZE - ret, NULL); 7491 if (cgroup_psi_enabled()) 7492 ret += show_delegatable_files(cgroup_psi_files, buf + ret, 7493 PAGE_SIZE - ret, NULL); 7494 7495 for_each_subsys(ss, ssid) 7496 ret += show_delegatable_files(ss->dfl_cftypes, buf + ret, 7497 PAGE_SIZE - ret, 7498 cgroup_subsys_name[ssid]); 7499 7500 return ret; 7501} 7502static struct kobj_attribute cgroup_delegate_attr = __ATTR_RO(delegate); 7503 7504static ssize_t features_show(struct kobject *kobj, struct kobj_attribute *attr, 7505 char *buf) 7506{ 7507 return snprintf(buf, PAGE_SIZE, 7508 "nsdelegate\n" 7509 "favordynmods\n" 7510 "memory_localevents\n" 7511 "memory_recursiveprot\n" 7512 "memory_hugetlb_accounting\n" 7513 "pids_localevents\n"); 7514} 7515static struct kobj_attribute cgroup_features_attr = __ATTR_RO(features); 7516 7517static struct attribute *cgroup_sysfs_attrs[] = { 7518 &cgroup_delegate_attr.attr, 7519 &cgroup_features_attr.attr, 7520 NULL, 7521}; 7522 7523static const struct attribute_group cgroup_sysfs_attr_group = { 7524 .attrs = cgroup_sysfs_attrs, 7525 .name = "cgroup", 7526}; 7527 7528static int __init cgroup_sysfs_init(void) 7529{ 7530 return sysfs_create_group(kernel_kobj, &cgroup_sysfs_attr_group); 7531} 7532subsys_initcall(cgroup_sysfs_init); 7533 7534#endif /* CONFIG_SYSFS */
~/
Read-only file preview.