linux/kernel/cgroup/cgroup.c
<<
>>
Prefs
   1/*
   2 *  Generic process-grouping system.
   3 *
   4 *  Based originally on the cpuset system, extracted by Paul Menage
   5 *  Copyright (C) 2006 Google, Inc
   6 *
   7 *  Notifications support
   8 *  Copyright (C) 2009 Nokia Corporation
   9 *  Author: Kirill A. Shutemov
  10 *
  11 *  Copyright notices from the original cpuset code:
  12 *  --------------------------------------------------
  13 *  Copyright (C) 2003 BULL SA.
  14 *  Copyright (C) 2004-2006 Silicon Graphics, Inc.
  15 *
  16 *  Portions derived from Patrick Mochel's sysfs code.
  17 *  sysfs is Copyright (c) 2001-3 Patrick Mochel
  18 *
  19 *  2003-10-10 Written by Simon Derr.
  20 *  2003-10-22 Updates by Stephen Hemminger.
  21 *  2004 May-July Rework by Paul Jackson.
  22 *  ---------------------------------------------------
  23 *
  24 *  This file is subject to the terms and conditions of the GNU General Public
  25 *  License.  See the file COPYING in the main directory of the Linux
  26 *  distribution for more details.
  27 */
  28
  29#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  30
  31#include "cgroup-internal.h"
  32
  33#include <linux/cred.h>
  34#include <linux/errno.h>
  35#include <linux/init_task.h>
  36#include <linux/kernel.h>
  37#include <linux/magic.h>
  38#include <linux/mutex.h>
  39#include <linux/mount.h>
  40#include <linux/pagemap.h>
  41#include <linux/proc_fs.h>
  42#include <linux/rcupdate.h>
  43#include <linux/sched.h>
  44#include <linux/sched/task.h>
  45#include <linux/slab.h>
  46#include <linux/spinlock.h>
  47#include <linux/percpu-rwsem.h>
  48#include <linux/string.h>
  49#include <linux/hashtable.h>
  50#include <linux/idr.h>
  51#include <linux/kthread.h>
  52#include <linux/atomic.h>
  53#include <linux/cpuset.h>
  54#include <linux/proc_ns.h>
  55#include <linux/nsproxy.h>
  56#include <linux/file.h>
  57#include <linux/fs_parser.h>
  58#include <linux/sched/cputime.h>
  59#include <linux/psi.h>
  60#include <net/sock.h>
  61
  62#define CREATE_TRACE_POINTS
  63#include <trace/events/cgroup.h>
  64
  65#define CGROUP_FILE_NAME_MAX            (MAX_CGROUP_TYPE_NAMELEN +      \
  66                                         MAX_CFTYPE_NAME + 2)
  67/* let's not notify more than 100 times per second */
  68#define CGROUP_FILE_NOTIFY_MIN_INTV     DIV_ROUND_UP(HZ, 100)
  69
  70/*
  71 * cgroup_mutex is the master lock.  Any modification to cgroup or its
  72 * hierarchy must be performed while holding it.
  73 *
  74 * css_set_lock protects task->cgroups pointer, the list of css_set
  75 * objects, and the chain of tasks off each css_set.
  76 *
  77 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
  78 * cgroup.h can use them for lockdep annotations.
  79 */
  80DEFINE_MUTEX(cgroup_mutex);
  81DEFINE_SPINLOCK(css_set_lock);
  82
  83#ifdef CONFIG_PROVE_RCU
  84EXPORT_SYMBOL_GPL(cgroup_mutex);
  85EXPORT_SYMBOL_GPL(css_set_lock);
  86#endif
  87
  88DEFINE_SPINLOCK(trace_cgroup_path_lock);
  89char trace_cgroup_path[TRACE_CGROUP_PATH_LEN];
  90bool cgroup_debug __read_mostly;
  91
  92/*
  93 * Protects cgroup_idr and css_idr so that IDs can be released without
  94 * grabbing cgroup_mutex.
  95 */
  96static DEFINE_SPINLOCK(cgroup_idr_lock);
  97
  98/*
  99 * Protects cgroup_file->kn for !self csses.  It synchronizes notifications
 100 * against file removal/re-creation across css hiding.
 101 */
 102static DEFINE_SPINLOCK(cgroup_file_kn_lock);
 103
 104DEFINE_PERCPU_RWSEM(cgroup_threadgroup_rwsem);
 105
 106#define cgroup_assert_mutex_or_rcu_locked()                             \
 107        RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&                       \
 108                           !lockdep_is_held(&cgroup_mutex),             \
 109                           "cgroup_mutex or RCU read lock required");
 110
 111/*
 112 * cgroup destruction makes heavy use of work items and there can be a lot
 113 * of concurrent destructions.  Use a separate workqueue so that cgroup
 114 * destruction work items don't end up filling up max_active of system_wq
 115 * which may lead to deadlock.
 116 */
 117static struct workqueue_struct *cgroup_destroy_wq;
 118
 119/* generate an array of cgroup subsystem pointers */
 120#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
 121struct cgroup_subsys *cgroup_subsys[] = {
 122#include <linux/cgroup_subsys.h>
 123};
 124#undef SUBSYS
 125
 126/* array of cgroup subsystem names */
 127#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
 128static const char *cgroup_subsys_name[] = {
 129#include <linux/cgroup_subsys.h>
 130};
 131#undef SUBSYS
 132
 133/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
 134#define SUBSYS(_x)                                                              \
 135        DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key);                 \
 136        DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key);                  \
 137        EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key);                      \
 138        EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
 139#include <linux/cgroup_subsys.h>
 140#undef SUBSYS
 141
 142#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
 143static struct static_key_true *cgroup_subsys_enabled_key[] = {
 144#include <linux/cgroup_subsys.h>
 145};
 146#undef SUBSYS
 147
 148#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
 149static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
 150#include <linux/cgroup_subsys.h>
 151};
 152#undef SUBSYS
 153
 154static DEFINE_PER_CPU(struct cgroup_rstat_cpu, cgrp_dfl_root_rstat_cpu);
 155
 156/* the default hierarchy */
 157struct cgroup_root cgrp_dfl_root = { .cgrp.rstat_cpu = &cgrp_dfl_root_rstat_cpu };
 158EXPORT_SYMBOL_GPL(cgrp_dfl_root);
 159
 160/*
 161 * The default hierarchy always exists but is hidden until mounted for the
 162 * first time.  This is for backward compatibility.
 163 */
 164static bool cgrp_dfl_visible;
 165
 166/* some controllers are not supported in the default hierarchy */
 167static u16 cgrp_dfl_inhibit_ss_mask;
 168
 169/* some controllers are implicitly enabled on the default hierarchy */
 170static u16 cgrp_dfl_implicit_ss_mask;
 171
 172/* some controllers can be threaded on the default hierarchy */
 173static u16 cgrp_dfl_threaded_ss_mask;
 174
 175/* The list of hierarchy roots */
 176LIST_HEAD(cgroup_roots);
 177static int cgroup_root_count;
 178
 179/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
 180static DEFINE_IDR(cgroup_hierarchy_idr);
 181
 182/*
 183 * Assign a monotonically increasing serial number to csses.  It guarantees
 184 * cgroups with bigger numbers are newer than those with smaller numbers.
 185 * Also, as csses are always appended to the parent's ->children list, it
 186 * guarantees that sibling csses are always sorted in the ascending serial
 187 * number order on the list.  Protected by cgroup_mutex.
 188 */
 189static u64 css_serial_nr_next = 1;
 190
 191/*
 192 * These bitmasks identify subsystems with specific features to avoid
 193 * having to do iterative checks repeatedly.
 194 */
 195static u16 have_fork_callback __read_mostly;
 196static u16 have_exit_callback __read_mostly;
 197static u16 have_release_callback __read_mostly;
 198static u16 have_canfork_callback __read_mostly;
 199
 200/* cgroup namespace for init task */
 201struct cgroup_namespace init_cgroup_ns = {
 202        .ns.count       = REFCOUNT_INIT(2),
 203        .user_ns        = &init_user_ns,
 204        .ns.ops         = &cgroupns_operations,
 205        .ns.inum        = PROC_CGROUP_INIT_INO,
 206        .root_cset      = &init_css_set,
 207};
 208
 209static struct file_system_type cgroup2_fs_type;
 210static struct cftype cgroup_base_files[];
 211
 212/* cgroup optional features */
 213enum cgroup_opt_features {
 214#ifdef CONFIG_PSI
 215        OPT_FEATURE_PRESSURE,
 216#endif
 217        OPT_FEATURE_COUNT
 218};
 219
 220static const char *cgroup_opt_feature_names[OPT_FEATURE_COUNT] = {
 221#ifdef CONFIG_PSI
 222        "pressure",
 223#endif
 224};
 225
 226static u16 cgroup_feature_disable_mask __read_mostly;
 227
 228static int cgroup_apply_control(struct cgroup *cgrp);
 229static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
 230static void css_task_iter_skip(struct css_task_iter *it,
 231                               struct task_struct *task);
 232static int cgroup_destroy_locked(struct cgroup *cgrp);
 233static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
 234                                              struct cgroup_subsys *ss);
 235static void css_release(struct percpu_ref *ref);
 236static void kill_css(struct cgroup_subsys_state *css);
 237static int cgroup_addrm_files(struct cgroup_subsys_state *css,
 238                              struct cgroup *cgrp, struct cftype cfts[],
 239                              bool is_add);
 240
 241/**
 242 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
 243 * @ssid: subsys ID of interest
 244 *
 245 * cgroup_subsys_enabled() can only be used with literal subsys names which
 246 * is fine for individual subsystems but unsuitable for cgroup core.  This
 247 * is slower static_key_enabled() based test indexed by @ssid.
 248 */
 249bool cgroup_ssid_enabled(int ssid)
 250{
 251        if (CGROUP_SUBSYS_COUNT == 0)
 252                return false;
 253
 254        return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
 255}
 256
 257/**
 258 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
 259 * @cgrp: the cgroup of interest
 260 *
 261 * The default hierarchy is the v2 interface of cgroup and this function
 262 * can be used to test whether a cgroup is on the default hierarchy for
 263 * cases where a subsystem should behave differently depending on the
 264 * interface version.
 265 *
 266 * List of changed behaviors:
 267 *
 268 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
 269 *   and "name" are disallowed.
 270 *
 271 * - When mounting an existing superblock, mount options should match.
 272 *
 273 * - Remount is disallowed.
 274 *
 275 * - rename(2) is disallowed.
 276 *
 277 * - "tasks" is removed.  Everything should be at process granularity.  Use
 278 *   "cgroup.procs" instead.
 279 *
 280 * - "cgroup.procs" is not sorted.  pids will be unique unless they got
 281 *   recycled in-between reads.
 282 *
 283 * - "release_agent" and "notify_on_release" are removed.  Replacement
 284 *   notification mechanism will be implemented.
 285 *
 286 * - "cgroup.clone_children" is removed.
 287 *
 288 * - "cgroup.subtree_populated" is available.  Its value is 0 if the cgroup
 289 *   and its descendants contain no task; otherwise, 1.  The file also
 290 *   generates kernfs notification which can be monitored through poll and
 291 *   [di]notify when the value of the file changes.
 292 *
 293 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
 294 *   take masks of ancestors with non-empty cpus/mems, instead of being
 295 *   moved to an ancestor.
 296 *
 297 * - cpuset: a task can be moved into an empty cpuset, and again it takes
 298 *   masks of ancestors.
 299 *
 300 * - blkcg: blk-throttle becomes properly hierarchical.
 301 *
 302 * - debug: disallowed on the default hierarchy.
 303 */
 304bool cgroup_on_dfl(const struct cgroup *cgrp)
 305{
 306        return cgrp->root == &cgrp_dfl_root;
 307}
 308
 309/* IDR wrappers which synchronize using cgroup_idr_lock */
 310static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
 311                            gfp_t gfp_mask)
 312{
 313        int ret;
 314
 315        idr_preload(gfp_mask);
 316        spin_lock_bh(&cgroup_idr_lock);
 317        ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
 318        spin_unlock_bh(&cgroup_idr_lock);
 319        idr_preload_end();
 320        return ret;
 321}
 322
 323static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
 324{
 325        void *ret;
 326
 327        spin_lock_bh(&cgroup_idr_lock);
 328        ret = idr_replace(idr, ptr, id);
 329        spin_unlock_bh(&cgroup_idr_lock);
 330        return ret;
 331}
 332
 333static void cgroup_idr_remove(struct idr *idr, int id)
 334{
 335        spin_lock_bh(&cgroup_idr_lock);
 336        idr_remove(idr, id);
 337        spin_unlock_bh(&cgroup_idr_lock);
 338}
 339
 340static bool cgroup_has_tasks(struct cgroup *cgrp)
 341{
 342        return cgrp->nr_populated_csets;
 343}
 344
 345bool cgroup_is_threaded(struct cgroup *cgrp)
 346{
 347        return cgrp->dom_cgrp != cgrp;
 348}
 349
 350/* can @cgrp host both domain and threaded children? */
 351static bool cgroup_is_mixable(struct cgroup *cgrp)
 352{
 353        /*
 354         * Root isn't under domain level resource control exempting it from
 355         * the no-internal-process constraint, so it can serve as a thread
 356         * root and a parent of resource domains at the same time.
 357         */
 358        return !cgroup_parent(cgrp);
 359}
 360
 361/* can @cgrp become a thread root? Should always be true for a thread root */
 362static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
 363{
 364        /* mixables don't care */
 365        if (cgroup_is_mixable(cgrp))
 366                return true;
 367
 368        /* domain roots can't be nested under threaded */
 369        if (cgroup_is_threaded(cgrp))
 370                return false;
 371
 372        /* can only have either domain or threaded children */
 373        if (cgrp->nr_populated_domain_children)
 374                return false;
 375
 376        /* and no domain controllers can be enabled */
 377        if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
 378                return false;
 379
 380        return true;
 381}
 382
 383/* is @cgrp root of a threaded subtree? */
 384bool cgroup_is_thread_root(struct cgroup *cgrp)
 385{
 386        /* thread root should be a domain */
 387        if (cgroup_is_threaded(cgrp))
 388                return false;
 389
 390        /* a domain w/ threaded children is a thread root */
 391        if (cgrp->nr_threaded_children)
 392                return true;
 393
 394        /*
 395         * A domain which has tasks and explicit threaded controllers
 396         * enabled is a thread root.
 397         */
 398        if (cgroup_has_tasks(cgrp) &&
 399            (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
 400                return true;
 401
 402        return false;
 403}
 404
 405/* a domain which isn't connected to the root w/o brekage can't be used */
 406static bool cgroup_is_valid_domain(struct cgroup *cgrp)
 407{
 408        /* the cgroup itself can be a thread root */
 409        if (cgroup_is_threaded(cgrp))
 410                return false;
 411
 412        /* but the ancestors can't be unless mixable */
 413        while ((cgrp = cgroup_parent(cgrp))) {
 414                if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
 415                        return false;
 416                if (cgroup_is_threaded(cgrp))
 417                        return false;
 418        }
 419
 420        return true;
 421}
 422
 423/* subsystems visibly enabled on a cgroup */
 424static u16 cgroup_control(struct cgroup *cgrp)
 425{
 426        struct cgroup *parent = cgroup_parent(cgrp);
 427        u16 root_ss_mask = cgrp->root->subsys_mask;
 428
 429        if (parent) {
 430                u16 ss_mask = parent->subtree_control;
 431
 432                /* threaded cgroups can only have threaded controllers */
 433                if (cgroup_is_threaded(cgrp))
 434                        ss_mask &= cgrp_dfl_threaded_ss_mask;
 435                return ss_mask;
 436        }
 437
 438        if (cgroup_on_dfl(cgrp))
 439                root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
 440                                  cgrp_dfl_implicit_ss_mask);
 441        return root_ss_mask;
 442}
 443
 444/* subsystems enabled on a cgroup */
 445static u16 cgroup_ss_mask(struct cgroup *cgrp)
 446{
 447        struct cgroup *parent = cgroup_parent(cgrp);
 448
 449        if (parent) {
 450                u16 ss_mask = parent->subtree_ss_mask;
 451
 452                /* threaded cgroups can only have threaded controllers */
 453                if (cgroup_is_threaded(cgrp))
 454                        ss_mask &= cgrp_dfl_threaded_ss_mask;
 455                return ss_mask;
 456        }
 457
 458        return cgrp->root->subsys_mask;
 459}
 460
 461/**
 462 * cgroup_css - obtain a cgroup's css for the specified subsystem
 463 * @cgrp: the cgroup of interest
 464 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
 465 *
 466 * Return @cgrp's css (cgroup_subsys_state) associated with @ss.  This
 467 * function must be called either under cgroup_mutex or rcu_read_lock() and
 468 * the caller is responsible for pinning the returned css if it wants to
 469 * keep accessing it outside the said locks.  This function may return
 470 * %NULL if @cgrp doesn't have @subsys_id enabled.
 471 */
 472static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
 473                                              struct cgroup_subsys *ss)
 474{
 475        if (ss)
 476                return rcu_dereference_check(cgrp->subsys[ss->id],
 477                                        lockdep_is_held(&cgroup_mutex));
 478        else
 479                return &cgrp->self;
 480}
 481
 482/**
 483 * cgroup_tryget_css - try to get a cgroup's css for the specified subsystem
 484 * @cgrp: the cgroup of interest
 485 * @ss: the subsystem of interest
 486 *
 487 * Find and get @cgrp's css associated with @ss.  If the css doesn't exist
 488 * or is offline, %NULL is returned.
 489 */
 490static struct cgroup_subsys_state *cgroup_tryget_css(struct cgroup *cgrp,
 491                                                     struct cgroup_subsys *ss)
 492{
 493        struct cgroup_subsys_state *css;
 494
 495        rcu_read_lock();
 496        css = cgroup_css(cgrp, ss);
 497        if (css && !css_tryget_online(css))
 498                css = NULL;
 499        rcu_read_unlock();
 500
 501        return css;
 502}
 503
 504/**
 505 * cgroup_e_css_by_mask - obtain a cgroup's effective css for the specified ss
 506 * @cgrp: the cgroup of interest
 507 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
 508 *
 509 * Similar to cgroup_css() but returns the effective css, which is defined
 510 * as the matching css of the nearest ancestor including self which has @ss
 511 * enabled.  If @ss is associated with the hierarchy @cgrp is on, this
 512 * function is guaranteed to return non-NULL css.
 513 */
 514static struct cgroup_subsys_state *cgroup_e_css_by_mask(struct cgroup *cgrp,
 515                                                        struct cgroup_subsys *ss)
 516{
 517        lockdep_assert_held(&cgroup_mutex);
 518
 519        if (!ss)
 520                return &cgrp->self;
 521
 522        /*
 523         * This function is used while updating css associations and thus
 524         * can't test the csses directly.  Test ss_mask.
 525         */
 526        while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
 527                cgrp = cgroup_parent(cgrp);
 528                if (!cgrp)
 529                        return NULL;
 530        }
 531
 532        return cgroup_css(cgrp, ss);
 533}
 534
 535/**
 536 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
 537 * @cgrp: the cgroup of interest
 538 * @ss: the subsystem of interest
 539 *
 540 * Find and get the effective css of @cgrp for @ss.  The effective css is
 541 * defined as the matching css of the nearest ancestor including self which
 542 * has @ss enabled.  If @ss is not mounted on the hierarchy @cgrp is on,
 543 * the root css is returned, so this function always returns a valid css.
 544 *
 545 * The returned css is not guaranteed to be online, and therefore it is the
 546 * callers responsibility to try get a reference for it.
 547 */
 548struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
 549                                         struct cgroup_subsys *ss)
 550{
 551        struct cgroup_subsys_state *css;
 552
 553        do {
 554                css = cgroup_css(cgrp, ss);
 555
 556                if (css)
 557                        return css;
 558                cgrp = cgroup_parent(cgrp);
 559        } while (cgrp);
 560
 561        return init_css_set.subsys[ss->id];
 562}
 563
 564/**
 565 * cgroup_get_e_css - get 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 * The returned css must be put using css_put().
 574 */
 575struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
 576                                             struct cgroup_subsys *ss)
 577{
 578        struct cgroup_subsys_state *css;
 579
 580        rcu_read_lock();
 581
 582        do {
 583                css = cgroup_css(cgrp, ss);
 584
 585                if (css && css_tryget_online(css))
 586                        goto out_unlock;
 587                cgrp = cgroup_parent(cgrp);
 588        } while (cgrp);
 589
 590        css = init_css_set.subsys[ss->id];
 591        css_get(css);
 592out_unlock:
 593        rcu_read_unlock();
 594        return css;
 595}
 596EXPORT_SYMBOL_GPL(cgroup_get_e_css);
 597
 598static void cgroup_get_live(struct cgroup *cgrp)
 599{
 600        WARN_ON_ONCE(cgroup_is_dead(cgrp));
 601        css_get(&cgrp->self);
 602}
 603
 604/**
 605 * __cgroup_task_count - count the number of tasks in a cgroup. The caller
 606 * is responsible for taking the css_set_lock.
 607 * @cgrp: the cgroup in question
 608 */
 609int __cgroup_task_count(const struct cgroup *cgrp)
 610{
 611        int count = 0;
 612        struct cgrp_cset_link *link;
 613
 614        lockdep_assert_held(&css_set_lock);
 615
 616        list_for_each_entry(link, &cgrp->cset_links, cset_link)
 617                count += link->cset->nr_tasks;
 618
 619        return count;
 620}
 621
 622/**
 623 * cgroup_task_count - count the number of tasks in a cgroup.
 624 * @cgrp: the cgroup in question
 625 */
 626int cgroup_task_count(const struct cgroup *cgrp)
 627{
 628        int count;
 629
 630        spin_lock_irq(&css_set_lock);
 631        count = __cgroup_task_count(cgrp);
 632        spin_unlock_irq(&css_set_lock);
 633
 634        return count;
 635}
 636
 637struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
 638{
 639        struct cgroup *cgrp = of->kn->parent->priv;
 640        struct cftype *cft = of_cft(of);
 641
 642        /*
 643         * This is open and unprotected implementation of cgroup_css().
 644         * seq_css() is only called from a kernfs file operation which has
 645         * an active reference on the file.  Because all the subsystem
 646         * files are drained before a css is disassociated with a cgroup,
 647         * the matching css from the cgroup's subsys table is guaranteed to
 648         * be and stay valid until the enclosing operation is complete.
 649         */
 650        if (cft->ss)
 651                return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
 652        else
 653                return &cgrp->self;
 654}
 655EXPORT_SYMBOL_GPL(of_css);
 656
 657/**
 658 * for_each_css - iterate all css's of a cgroup
 659 * @css: the iteration cursor
 660 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
 661 * @cgrp: the target cgroup to iterate css's of
 662 *
 663 * Should be called under cgroup_[tree_]mutex.
 664 */
 665#define for_each_css(css, ssid, cgrp)                                   \
 666        for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
 667                if (!((css) = rcu_dereference_check(                    \
 668                                (cgrp)->subsys[(ssid)],                 \
 669                                lockdep_is_held(&cgroup_mutex)))) { }   \
 670                else
 671
 672/**
 673 * for_each_e_css - iterate all effective css's of a cgroup
 674 * @css: the iteration cursor
 675 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
 676 * @cgrp: the target cgroup to iterate css's of
 677 *
 678 * Should be called under cgroup_[tree_]mutex.
 679 */
 680#define for_each_e_css(css, ssid, cgrp)                                     \
 681        for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)            \
 682                if (!((css) = cgroup_e_css_by_mask(cgrp,                    \
 683                                                   cgroup_subsys[(ssid)]))) \
 684                        ;                                                   \
 685                else
 686
 687/**
 688 * do_each_subsys_mask - filter for_each_subsys with a bitmask
 689 * @ss: the iteration cursor
 690 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
 691 * @ss_mask: the bitmask
 692 *
 693 * The block will only run for cases where the ssid-th bit (1 << ssid) of
 694 * @ss_mask is set.
 695 */
 696#define do_each_subsys_mask(ss, ssid, ss_mask) do {                     \
 697        unsigned long __ss_mask = (ss_mask);                            \
 698        if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
 699                (ssid) = 0;                                             \
 700                break;                                                  \
 701        }                                                               \
 702        for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) {       \
 703                (ss) = cgroup_subsys[ssid];                             \
 704                {
 705
 706#define while_each_subsys_mask()                                        \
 707                }                                                       \
 708        }                                                               \
 709} while (false)
 710
 711/* iterate over child cgrps, lock should be held throughout iteration */
 712#define cgroup_for_each_live_child(child, cgrp)                         \
 713        list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
 714                if (({ lockdep_assert_held(&cgroup_mutex);              \
 715                       cgroup_is_dead(child); }))                       \
 716                        ;                                               \
 717                else
 718
 719/* walk live descendants in pre order */
 720#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)          \
 721        css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL))  \
 722                if (({ lockdep_assert_held(&cgroup_mutex);              \
 723                       (dsct) = (d_css)->cgroup;                        \
 724                       cgroup_is_dead(dsct); }))                        \
 725                        ;                                               \
 726                else
 727
 728/* walk live descendants in postorder */
 729#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp)         \
 730        css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
 731                if (({ lockdep_assert_held(&cgroup_mutex);              \
 732                       (dsct) = (d_css)->cgroup;                        \
 733                       cgroup_is_dead(dsct); }))                        \
 734                        ;                                               \
 735                else
 736
 737/*
 738 * The default css_set - used by init and its children prior to any
 739 * hierarchies being mounted. It contains a pointer to the root state
 740 * for each subsystem. Also used to anchor the list of css_sets. Not
 741 * reference-counted, to improve performance when child cgroups
 742 * haven't been created.
 743 */
 744struct css_set init_css_set = {
 745        .refcount               = REFCOUNT_INIT(1),
 746        .dom_cset               = &init_css_set,
 747        .tasks                  = LIST_HEAD_INIT(init_css_set.tasks),
 748        .mg_tasks               = LIST_HEAD_INIT(init_css_set.mg_tasks),
 749        .dying_tasks            = LIST_HEAD_INIT(init_css_set.dying_tasks),
 750        .task_iters             = LIST_HEAD_INIT(init_css_set.task_iters),
 751        .threaded_csets         = LIST_HEAD_INIT(init_css_set.threaded_csets),
 752        .cgrp_links             = LIST_HEAD_INIT(init_css_set.cgrp_links),
 753        .mg_preload_node        = LIST_HEAD_INIT(init_css_set.mg_preload_node),
 754        .mg_node                = LIST_HEAD_INIT(init_css_set.mg_node),
 755
 756        /*
 757         * The following field is re-initialized when this cset gets linked
 758         * in cgroup_init().  However, let's initialize the field
 759         * statically too so that the default cgroup can be accessed safely
 760         * early during boot.
 761         */
 762        .dfl_cgrp               = &cgrp_dfl_root.cgrp,
 763};
 764
 765static int css_set_count        = 1;    /* 1 for init_css_set */
 766
 767static bool css_set_threaded(struct css_set *cset)
 768{
 769        return cset->dom_cset != cset;
 770}
 771
 772/**
 773 * css_set_populated - does a css_set contain any tasks?
 774 * @cset: target css_set
 775 *
 776 * css_set_populated() should be the same as !!cset->nr_tasks at steady
 777 * state. However, css_set_populated() can be called while a task is being
 778 * added to or removed from the linked list before the nr_tasks is
 779 * properly updated. Hence, we can't just look at ->nr_tasks here.
 780 */
 781static bool css_set_populated(struct css_set *cset)
 782{
 783        lockdep_assert_held(&css_set_lock);
 784
 785        return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
 786}
 787
 788/**
 789 * cgroup_update_populated - update the populated count of a cgroup
 790 * @cgrp: the target cgroup
 791 * @populated: inc or dec populated count
 792 *
 793 * One of the css_sets associated with @cgrp is either getting its first
 794 * task or losing the last.  Update @cgrp->nr_populated_* accordingly.  The
 795 * count is propagated towards root so that a given cgroup's
 796 * nr_populated_children is zero iff none of its descendants contain any
 797 * tasks.
 798 *
 799 * @cgrp's interface file "cgroup.populated" is zero if both
 800 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
 801 * 1 otherwise.  When the sum changes from or to zero, userland is notified
 802 * that the content of the interface file has changed.  This can be used to
 803 * detect when @cgrp and its descendants become populated or empty.
 804 */
 805static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
 806{
 807        struct cgroup *child = NULL;
 808        int adj = populated ? 1 : -1;
 809
 810        lockdep_assert_held(&css_set_lock);
 811
 812        do {
 813                bool was_populated = cgroup_is_populated(cgrp);
 814
 815                if (!child) {
 816                        cgrp->nr_populated_csets += adj;
 817                } else {
 818                        if (cgroup_is_threaded(child))
 819                                cgrp->nr_populated_threaded_children += adj;
 820                        else
 821                                cgrp->nr_populated_domain_children += adj;
 822                }
 823
 824                if (was_populated == cgroup_is_populated(cgrp))
 825                        break;
 826
 827                cgroup1_check_for_release(cgrp);
 828                TRACE_CGROUP_PATH(notify_populated, cgrp,
 829                                  cgroup_is_populated(cgrp));
 830                cgroup_file_notify(&cgrp->events_file);
 831
 832                child = cgrp;
 833                cgrp = cgroup_parent(cgrp);
 834        } while (cgrp);
 835}
 836
 837/**
 838 * css_set_update_populated - update populated state of a css_set
 839 * @cset: target css_set
 840 * @populated: whether @cset is populated or depopulated
 841 *
 842 * @cset is either getting the first task or losing the last.  Update the
 843 * populated counters of all associated cgroups accordingly.
 844 */
 845static void css_set_update_populated(struct css_set *cset, bool populated)
 846{
 847        struct cgrp_cset_link *link;
 848
 849        lockdep_assert_held(&css_set_lock);
 850
 851        list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
 852                cgroup_update_populated(link->cgrp, populated);
 853}
 854
 855/*
 856 * @task is leaving, advance task iterators which are pointing to it so
 857 * that they can resume at the next position.  Advancing an iterator might
 858 * remove it from the list, use safe walk.  See css_task_iter_skip() for
 859 * details.
 860 */
 861static void css_set_skip_task_iters(struct css_set *cset,
 862                                    struct task_struct *task)
 863{
 864        struct css_task_iter *it, *pos;
 865
 866        list_for_each_entry_safe(it, pos, &cset->task_iters, iters_node)
 867                css_task_iter_skip(it, task);
 868}
 869
 870/**
 871 * css_set_move_task - move a task from one css_set to another
 872 * @task: task being moved
 873 * @from_cset: css_set @task currently belongs to (may be NULL)
 874 * @to_cset: new css_set @task is being moved to (may be NULL)
 875 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
 876 *
 877 * Move @task from @from_cset to @to_cset.  If @task didn't belong to any
 878 * css_set, @from_cset can be NULL.  If @task is being disassociated
 879 * instead of moved, @to_cset can be NULL.
 880 *
 881 * This function automatically handles populated counter updates and
 882 * css_task_iter adjustments but the caller is responsible for managing
 883 * @from_cset and @to_cset's reference counts.
 884 */
 885static void css_set_move_task(struct task_struct *task,
 886                              struct css_set *from_cset, struct css_set *to_cset,
 887                              bool use_mg_tasks)
 888{
 889        lockdep_assert_held(&css_set_lock);
 890
 891        if (to_cset && !css_set_populated(to_cset))
 892                css_set_update_populated(to_cset, true);
 893
 894        if (from_cset) {
 895                WARN_ON_ONCE(list_empty(&task->cg_list));
 896
 897                css_set_skip_task_iters(from_cset, task);
 898                list_del_init(&task->cg_list);
 899                if (!css_set_populated(from_cset))
 900                        css_set_update_populated(from_cset, false);
 901        } else {
 902                WARN_ON_ONCE(!list_empty(&task->cg_list));
 903        }
 904
 905        if (to_cset) {
 906                /*
 907                 * We are synchronized through cgroup_threadgroup_rwsem
 908                 * against PF_EXITING setting such that we can't race
 909                 * against cgroup_exit()/cgroup_free() dropping the css_set.
 910                 */
 911                WARN_ON_ONCE(task->flags & PF_EXITING);
 912
 913                cgroup_move_task(task, to_cset);
 914                list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
 915                                                             &to_cset->tasks);
 916        }
 917}
 918
 919/*
 920 * hash table for cgroup groups. This improves the performance to find
 921 * an existing css_set. This hash doesn't (currently) take into
 922 * account cgroups in empty hierarchies.
 923 */
 924#define CSS_SET_HASH_BITS       7
 925static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
 926
 927static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
 928{
 929        unsigned long key = 0UL;
 930        struct cgroup_subsys *ss;
 931        int i;
 932
 933        for_each_subsys(ss, i)
 934                key += (unsigned long)css[i];
 935        key = (key >> 16) ^ key;
 936
 937        return key;
 938}
 939
 940void put_css_set_locked(struct css_set *cset)
 941{
 942        struct cgrp_cset_link *link, *tmp_link;
 943        struct cgroup_subsys *ss;
 944        int ssid;
 945
 946        lockdep_assert_held(&css_set_lock);
 947
 948        if (!refcount_dec_and_test(&cset->refcount))
 949                return;
 950
 951        WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
 952
 953        /* This css_set is dead. Unlink it and release cgroup and css refs */
 954        for_each_subsys(ss, ssid) {
 955                list_del(&cset->e_cset_node[ssid]);
 956                css_put(cset->subsys[ssid]);
 957        }
 958        hash_del(&cset->hlist);
 959        css_set_count--;
 960
 961        list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
 962                list_del(&link->cset_link);
 963                list_del(&link->cgrp_link);
 964                if (cgroup_parent(link->cgrp))
 965                        cgroup_put(link->cgrp);
 966                kfree(link);
 967        }
 968
 969        if (css_set_threaded(cset)) {
 970                list_del(&cset->threaded_csets_node);
 971                put_css_set_locked(cset->dom_cset);
 972        }
 973
 974        kfree_rcu(cset, rcu_head);
 975}
 976
 977/**
 978 * compare_css_sets - helper function for find_existing_css_set().
 979 * @cset: candidate css_set being tested
 980 * @old_cset: existing css_set for a task
 981 * @new_cgrp: cgroup that's being entered by the task
 982 * @template: desired set of css pointers in css_set (pre-calculated)
 983 *
 984 * Returns true if "cset" matches "old_cset" except for the hierarchy
 985 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
 986 */
 987static bool compare_css_sets(struct css_set *cset,
 988                             struct css_set *old_cset,
 989                             struct cgroup *new_cgrp,
 990                             struct cgroup_subsys_state *template[])
 991{
 992        struct cgroup *new_dfl_cgrp;
 993        struct list_head *l1, *l2;
 994
 995        /*
 996         * On the default hierarchy, there can be csets which are
 997         * associated with the same set of cgroups but different csses.
 998         * Let's first ensure that csses match.
 999         */
1000        if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
1001                return false;
1002
1003
1004        /* @cset's domain should match the default cgroup's */
1005        if (cgroup_on_dfl(new_cgrp))
1006                new_dfl_cgrp = new_cgrp;
1007        else
1008                new_dfl_cgrp = old_cset->dfl_cgrp;
1009
1010        if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
1011                return false;
1012
1013        /*
1014         * Compare cgroup pointers in order to distinguish between
1015         * different cgroups in hierarchies.  As different cgroups may
1016         * share the same effective css, this comparison is always
1017         * necessary.
1018         */
1019        l1 = &cset->cgrp_links;
1020        l2 = &old_cset->cgrp_links;
1021        while (1) {
1022                struct cgrp_cset_link *link1, *link2;
1023                struct cgroup *cgrp1, *cgrp2;
1024
1025                l1 = l1->next;
1026                l2 = l2->next;
1027                /* See if we reached the end - both lists are equal length. */
1028                if (l1 == &cset->cgrp_links) {
1029                        BUG_ON(l2 != &old_cset->cgrp_links);
1030                        break;
1031                } else {
1032                        BUG_ON(l2 == &old_cset->cgrp_links);
1033                }
1034                /* Locate the cgroups associated with these links. */
1035                link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
1036                link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
1037                cgrp1 = link1->cgrp;
1038                cgrp2 = link2->cgrp;
1039                /* Hierarchies should be linked in the same order. */
1040                BUG_ON(cgrp1->root != cgrp2->root);
1041
1042                /*
1043                 * If this hierarchy is the hierarchy of the cgroup
1044                 * that's changing, then we need to check that this
1045                 * css_set points to the new cgroup; if it's any other
1046                 * hierarchy, then this css_set should point to the
1047                 * same cgroup as the old css_set.
1048                 */
1049                if (cgrp1->root == new_cgrp->root) {
1050                        if (cgrp1 != new_cgrp)
1051                                return false;
1052                } else {
1053                        if (cgrp1 != cgrp2)
1054                                return false;
1055                }
1056        }
1057        return true;
1058}
1059
1060/**
1061 * find_existing_css_set - init css array and find the matching css_set
1062 * @old_cset: the css_set that we're using before the cgroup transition
1063 * @cgrp: the cgroup that we're moving into
1064 * @template: out param for the new set of csses, should be clear on entry
1065 */
1066static struct css_set *find_existing_css_set(struct css_set *old_cset,
1067                                        struct cgroup *cgrp,
1068                                        struct cgroup_subsys_state *template[])
1069{
1070        struct cgroup_root *root = cgrp->root;
1071        struct cgroup_subsys *ss;
1072        struct css_set *cset;
1073        unsigned long key;
1074        int i;
1075
1076        /*
1077         * Build the set of subsystem state objects that we want to see in the
1078         * new css_set. While subsystems can change globally, the entries here
1079         * won't change, so no need for locking.
1080         */
1081        for_each_subsys(ss, i) {
1082                if (root->subsys_mask & (1UL << i)) {
1083                        /*
1084                         * @ss is in this hierarchy, so we want the
1085                         * effective css from @cgrp.
1086                         */
1087                        template[i] = cgroup_e_css_by_mask(cgrp, ss);
1088                } else {
1089                        /*
1090                         * @ss is not in this hierarchy, so we don't want
1091                         * to change the css.
1092                         */
1093                        template[i] = old_cset->subsys[i];
1094                }
1095        }
1096
1097        key = css_set_hash(template);
1098        hash_for_each_possible(css_set_table, cset, hlist, key) {
1099                if (!compare_css_sets(cset, old_cset, cgrp, template))
1100                        continue;
1101
1102                /* This css_set matches what we need */
1103                return cset;
1104        }
1105
1106        /* No existing cgroup group matched */
1107        return NULL;
1108}
1109
1110static void free_cgrp_cset_links(struct list_head *links_to_free)
1111{
1112        struct cgrp_cset_link *link, *tmp_link;
1113
1114        list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
1115                list_del(&link->cset_link);
1116                kfree(link);
1117        }
1118}
1119
1120/**
1121 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1122 * @count: the number of links to allocate
1123 * @tmp_links: list_head the allocated links are put on
1124 *
1125 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1126 * through ->cset_link.  Returns 0 on success or -errno.
1127 */
1128static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1129{
1130        struct cgrp_cset_link *link;
1131        int i;
1132
1133        INIT_LIST_HEAD(tmp_links);
1134
1135        for (i = 0; i < count; i++) {
1136                link = kzalloc(sizeof(*link), GFP_KERNEL);
1137                if (!link) {
1138                        free_cgrp_cset_links(tmp_links);
1139                        return -ENOMEM;
1140                }
1141                list_add(&link->cset_link, tmp_links);
1142        }
1143        return 0;
1144}
1145
1146/**
1147 * link_css_set - a helper function to link a css_set to a cgroup
1148 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1149 * @cset: the css_set to be linked
1150 * @cgrp: the destination cgroup
1151 */
1152static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1153                         struct cgroup *cgrp)
1154{
1155        struct cgrp_cset_link *link;
1156
1157        BUG_ON(list_empty(tmp_links));
1158
1159        if (cgroup_on_dfl(cgrp))
1160                cset->dfl_cgrp = cgrp;
1161
1162        link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1163        link->cset = cset;
1164        link->cgrp = cgrp;
1165
1166        /*
1167         * Always add links to the tail of the lists so that the lists are
1168         * in chronological order.
1169         */
1170        list_move_tail(&link->cset_link, &cgrp->cset_links);
1171        list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1172
1173        if (cgroup_parent(cgrp))
1174                cgroup_get_live(cgrp);
1175}
1176
1177/**
1178 * find_css_set - return a new css_set with one cgroup updated
1179 * @old_cset: the baseline css_set
1180 * @cgrp: the cgroup to be updated
1181 *
1182 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1183 * substituted into the appropriate hierarchy.
1184 */
1185static struct css_set *find_css_set(struct css_set *old_cset,
1186                                    struct cgroup *cgrp)
1187{
1188        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1189        struct css_set *cset;
1190        struct list_head tmp_links;
1191        struct cgrp_cset_link *link;
1192        struct cgroup_subsys *ss;
1193        unsigned long key;
1194        int ssid;
1195
1196        lockdep_assert_held(&cgroup_mutex);
1197
1198        /* First see if we already have a cgroup group that matches
1199         * the desired set */
1200        spin_lock_irq(&css_set_lock);
1201        cset = find_existing_css_set(old_cset, cgrp, template);
1202        if (cset)
1203                get_css_set(cset);
1204        spin_unlock_irq(&css_set_lock);
1205
1206        if (cset)
1207                return cset;
1208
1209        cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1210        if (!cset)
1211                return NULL;
1212
1213        /* Allocate all the cgrp_cset_link objects that we'll need */
1214        if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1215                kfree(cset);
1216                return NULL;
1217        }
1218
1219        refcount_set(&cset->refcount, 1);
1220        cset->dom_cset = cset;
1221        INIT_LIST_HEAD(&cset->tasks);
1222        INIT_LIST_HEAD(&cset->mg_tasks);
1223        INIT_LIST_HEAD(&cset->dying_tasks);
1224        INIT_LIST_HEAD(&cset->task_iters);
1225        INIT_LIST_HEAD(&cset->threaded_csets);
1226        INIT_HLIST_NODE(&cset->hlist);
1227        INIT_LIST_HEAD(&cset->cgrp_links);
1228        INIT_LIST_HEAD(&cset->mg_preload_node);
1229        INIT_LIST_HEAD(&cset->mg_node);
1230
1231        /* Copy the set of subsystem state objects generated in
1232         * find_existing_css_set() */
1233        memcpy(cset->subsys, template, sizeof(cset->subsys));
1234
1235        spin_lock_irq(&css_set_lock);
1236        /* Add reference counts and links from the new css_set. */
1237        list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1238                struct cgroup *c = link->cgrp;
1239
1240                if (c->root == cgrp->root)
1241                        c = cgrp;
1242                link_css_set(&tmp_links, cset, c);
1243        }
1244
1245        BUG_ON(!list_empty(&tmp_links));
1246
1247        css_set_count++;
1248
1249        /* Add @cset to the hash table */
1250        key = css_set_hash(cset->subsys);
1251        hash_add(css_set_table, &cset->hlist, key);
1252
1253        for_each_subsys(ss, ssid) {
1254                struct cgroup_subsys_state *css = cset->subsys[ssid];
1255
1256                list_add_tail(&cset->e_cset_node[ssid],
1257                              &css->cgroup->e_csets[ssid]);
1258                css_get(css);
1259        }
1260
1261        spin_unlock_irq(&css_set_lock);
1262
1263        /*
1264         * If @cset should be threaded, look up the matching dom_cset and
1265         * link them up.  We first fully initialize @cset then look for the
1266         * dom_cset.  It's simpler this way and safe as @cset is guaranteed
1267         * to stay empty until we return.
1268         */
1269        if (cgroup_is_threaded(cset->dfl_cgrp)) {
1270                struct css_set *dcset;
1271
1272                dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1273                if (!dcset) {
1274                        put_css_set(cset);
1275                        return NULL;
1276                }
1277
1278                spin_lock_irq(&css_set_lock);
1279                cset->dom_cset = dcset;
1280                list_add_tail(&cset->threaded_csets_node,
1281                              &dcset->threaded_csets);
1282                spin_unlock_irq(&css_set_lock);
1283        }
1284
1285        return cset;
1286}
1287
1288struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1289{
1290        struct cgroup *root_cgrp = kf_root->kn->priv;
1291
1292        return root_cgrp->root;
1293}
1294
1295static int cgroup_init_root_id(struct cgroup_root *root)
1296{
1297        int id;
1298
1299        lockdep_assert_held(&cgroup_mutex);
1300
1301        id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1302        if (id < 0)
1303                return id;
1304
1305        root->hierarchy_id = id;
1306        return 0;
1307}
1308
1309static void cgroup_exit_root_id(struct cgroup_root *root)
1310{
1311        lockdep_assert_held(&cgroup_mutex);
1312
1313        idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1314}
1315
1316void cgroup_free_root(struct cgroup_root *root)
1317{
1318        kfree(root);
1319}
1320
1321static void cgroup_destroy_root(struct cgroup_root *root)
1322{
1323        struct cgroup *cgrp = &root->cgrp;
1324        struct cgrp_cset_link *link, *tmp_link;
1325
1326        trace_cgroup_destroy_root(root);
1327
1328        cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1329
1330        BUG_ON(atomic_read(&root->nr_cgrps));
1331        BUG_ON(!list_empty(&cgrp->self.children));
1332
1333        /* Rebind all subsystems back to the default hierarchy */
1334        WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1335
1336        /*
1337         * Release all the links from cset_links to this hierarchy's
1338         * root cgroup
1339         */
1340        spin_lock_irq(&css_set_lock);
1341
1342        list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1343                list_del(&link->cset_link);
1344                list_del(&link->cgrp_link);
1345                kfree(link);
1346        }
1347
1348        spin_unlock_irq(&css_set_lock);
1349
1350        if (!list_empty(&root->root_list)) {
1351                list_del(&root->root_list);
1352                cgroup_root_count--;
1353        }
1354
1355        cgroup_exit_root_id(root);
1356
1357        mutex_unlock(&cgroup_mutex);
1358
1359        cgroup_rstat_exit(cgrp);
1360        kernfs_destroy_root(root->kf_root);
1361        cgroup_free_root(root);
1362}
1363
1364/*
1365 * look up cgroup associated with current task's cgroup namespace on the
1366 * specified hierarchy
1367 */
1368static struct cgroup *
1369current_cgns_cgroup_from_root(struct cgroup_root *root)
1370{
1371        struct cgroup *res = NULL;
1372        struct css_set *cset;
1373
1374        lockdep_assert_held(&css_set_lock);
1375
1376        rcu_read_lock();
1377
1378        cset = current->nsproxy->cgroup_ns->root_cset;
1379        if (cset == &init_css_set) {
1380                res = &root->cgrp;
1381        } else if (root == &cgrp_dfl_root) {
1382                res = cset->dfl_cgrp;
1383        } else {
1384                struct cgrp_cset_link *link;
1385
1386                list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1387                        struct cgroup *c = link->cgrp;
1388
1389                        if (c->root == root) {
1390                                res = c;
1391                                break;
1392                        }
1393                }
1394        }
1395        rcu_read_unlock();
1396
1397        BUG_ON(!res);
1398        return res;
1399}
1400
1401/* look up cgroup associated with given css_set on the specified hierarchy */
1402static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1403                                            struct cgroup_root *root)
1404{
1405        struct cgroup *res = NULL;
1406
1407        lockdep_assert_held(&cgroup_mutex);
1408        lockdep_assert_held(&css_set_lock);
1409
1410        if (cset == &init_css_set) {
1411                res = &root->cgrp;
1412        } else if (root == &cgrp_dfl_root) {
1413                res = cset->dfl_cgrp;
1414        } else {
1415                struct cgrp_cset_link *link;
1416
1417                list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1418                        struct cgroup *c = link->cgrp;
1419
1420                        if (c->root == root) {
1421                                res = c;
1422                                break;
1423                        }
1424                }
1425        }
1426
1427        BUG_ON(!res);
1428        return res;
1429}
1430
1431/*
1432 * Return the cgroup for "task" from the given hierarchy. Must be
1433 * called with cgroup_mutex and css_set_lock held.
1434 */
1435struct cgroup *task_cgroup_from_root(struct task_struct *task,
1436                                     struct cgroup_root *root)
1437{
1438        /*
1439         * No need to lock the task - since we hold css_set_lock the
1440         * task can't change groups.
1441         */
1442        return cset_cgroup_from_root(task_css_set(task), root);
1443}
1444
1445/*
1446 * A task must hold cgroup_mutex to modify cgroups.
1447 *
1448 * Any task can increment and decrement the count field without lock.
1449 * So in general, code holding cgroup_mutex can't rely on the count
1450 * field not changing.  However, if the count goes to zero, then only
1451 * cgroup_attach_task() can increment it again.  Because a count of zero
1452 * means that no tasks are currently attached, therefore there is no
1453 * way a task attached to that cgroup can fork (the other way to
1454 * increment the count).  So code holding cgroup_mutex can safely
1455 * assume that if the count is zero, it will stay zero. Similarly, if
1456 * a task holds cgroup_mutex on a cgroup with zero count, it
1457 * knows that the cgroup won't be removed, as cgroup_rmdir()
1458 * needs that mutex.
1459 *
1460 * A cgroup can only be deleted if both its 'count' of using tasks
1461 * is zero, and its list of 'children' cgroups is empty.  Since all
1462 * tasks in the system use _some_ cgroup, and since there is always at
1463 * least one task in the system (init, pid == 1), therefore, root cgroup
1464 * always has either children cgroups and/or using tasks.  So we don't
1465 * need a special hack to ensure that root cgroup cannot be deleted.
1466 *
1467 * P.S.  One more locking exception.  RCU is used to guard the
1468 * update of a tasks cgroup pointer by cgroup_attach_task()
1469 */
1470
1471static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1472
1473static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1474                              char *buf)
1475{
1476        struct cgroup_subsys *ss = cft->ss;
1477
1478        if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1479            !(cgrp->root->flags & CGRP_ROOT_NOPREFIX)) {
1480                const char *dbg = (cft->flags & CFTYPE_DEBUG) ? ".__DEBUG__." : "";
1481
1482                snprintf(buf, CGROUP_FILE_NAME_MAX, "%s%s.%s",
1483                         dbg, cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1484                         cft->name);
1485        } else {
1486                strscpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1487        }
1488        return buf;
1489}
1490
1491/**
1492 * cgroup_file_mode - deduce file mode of a control file
1493 * @cft: the control file in question
1494 *
1495 * S_IRUGO for read, S_IWUSR for write.
1496 */
1497static umode_t cgroup_file_mode(const struct cftype *cft)
1498{
1499        umode_t mode = 0;
1500
1501        if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1502                mode |= S_IRUGO;
1503
1504        if (cft->write_u64 || cft->write_s64 || cft->write) {
1505                if (cft->flags & CFTYPE_WORLD_WRITABLE)
1506                        mode |= S_IWUGO;
1507                else
1508                        mode |= S_IWUSR;
1509        }
1510
1511        return mode;
1512}
1513
1514/**
1515 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1516 * @subtree_control: the new subtree_control mask to consider
1517 * @this_ss_mask: available subsystems
1518 *
1519 * On the default hierarchy, a subsystem may request other subsystems to be
1520 * enabled together through its ->depends_on mask.  In such cases, more
1521 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1522 *
1523 * This function calculates which subsystems need to be enabled if
1524 * @subtree_control is to be applied while restricted to @this_ss_mask.
1525 */
1526static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1527{
1528        u16 cur_ss_mask = subtree_control;
1529        struct cgroup_subsys *ss;
1530        int ssid;
1531
1532        lockdep_assert_held(&cgroup_mutex);
1533
1534        cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1535
1536        while (true) {
1537                u16 new_ss_mask = cur_ss_mask;
1538
1539                do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1540                        new_ss_mask |= ss->depends_on;
1541                } while_each_subsys_mask();
1542
1543                /*
1544                 * Mask out subsystems which aren't available.  This can
1545                 * happen only if some depended-upon subsystems were bound
1546                 * to non-default hierarchies.
1547                 */
1548                new_ss_mask &= this_ss_mask;
1549
1550                if (new_ss_mask == cur_ss_mask)
1551                        break;
1552                cur_ss_mask = new_ss_mask;
1553        }
1554
1555        return cur_ss_mask;
1556}
1557
1558/**
1559 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1560 * @kn: the kernfs_node being serviced
1561 *
1562 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1563 * the method finishes if locking succeeded.  Note that once this function
1564 * returns the cgroup returned by cgroup_kn_lock_live() may become
1565 * inaccessible any time.  If the caller intends to continue to access the
1566 * cgroup, it should pin it before invoking this function.
1567 */
1568void cgroup_kn_unlock(struct kernfs_node *kn)
1569{
1570        struct cgroup *cgrp;
1571
1572        if (kernfs_type(kn) == KERNFS_DIR)
1573                cgrp = kn->priv;
1574        else
1575                cgrp = kn->parent->priv;
1576
1577        mutex_unlock(&cgroup_mutex);
1578
1579        kernfs_unbreak_active_protection(kn);
1580        cgroup_put(cgrp);
1581}
1582
1583/**
1584 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1585 * @kn: the kernfs_node being serviced
1586 * @drain_offline: perform offline draining on the cgroup
1587 *
1588 * This helper is to be used by a cgroup kernfs method currently servicing
1589 * @kn.  It breaks the active protection, performs cgroup locking and
1590 * verifies that the associated cgroup is alive.  Returns the cgroup if
1591 * alive; otherwise, %NULL.  A successful return should be undone by a
1592 * matching cgroup_kn_unlock() invocation.  If @drain_offline is %true, the
1593 * cgroup is drained of offlining csses before return.
1594 *
1595 * Any cgroup kernfs method implementation which requires locking the
1596 * associated cgroup should use this helper.  It avoids nesting cgroup
1597 * locking under kernfs active protection and allows all kernfs operations
1598 * including self-removal.
1599 */
1600struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1601{
1602        struct cgroup *cgrp;
1603
1604        if (kernfs_type(kn) == KERNFS_DIR)
1605                cgrp = kn->priv;
1606        else
1607                cgrp = kn->parent->priv;
1608
1609        /*
1610         * We're gonna grab cgroup_mutex which nests outside kernfs
1611         * active_ref.  cgroup liveliness check alone provides enough
1612         * protection against removal.  Ensure @cgrp stays accessible and
1613         * break the active_ref protection.
1614         */
1615        if (!cgroup_tryget(cgrp))
1616                return NULL;
1617        kernfs_break_active_protection(kn);
1618
1619        if (drain_offline)
1620                cgroup_lock_and_drain_offline(cgrp);
1621        else
1622                mutex_lock(&cgroup_mutex);
1623
1624        if (!cgroup_is_dead(cgrp))
1625                return cgrp;
1626
1627        cgroup_kn_unlock(kn);
1628        return NULL;
1629}
1630
1631static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1632{
1633        char name[CGROUP_FILE_NAME_MAX];
1634
1635        lockdep_assert_held(&cgroup_mutex);
1636
1637        if (cft->file_offset) {
1638                struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1639                struct cgroup_file *cfile = (void *)css + cft->file_offset;
1640
1641                spin_lock_irq(&cgroup_file_kn_lock);
1642                cfile->kn = NULL;
1643                spin_unlock_irq(&cgroup_file_kn_lock);
1644
1645                del_timer_sync(&cfile->notify_timer);
1646        }
1647
1648        kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1649}
1650
1651/**
1652 * css_clear_dir - remove subsys files in a cgroup directory
1653 * @css: target css
1654 */
1655static void css_clear_dir(struct cgroup_subsys_state *css)
1656{
1657        struct cgroup *cgrp = css->cgroup;
1658        struct cftype *cfts;
1659
1660        if (!(css->flags & CSS_VISIBLE))
1661                return;
1662
1663        css->flags &= ~CSS_VISIBLE;
1664
1665        if (!css->ss) {
1666                if (cgroup_on_dfl(cgrp))
1667                        cfts = cgroup_base_files;
1668                else
1669                        cfts = cgroup1_base_files;
1670
1671                cgroup_addrm_files(css, cgrp, cfts, false);
1672        } else {
1673                list_for_each_entry(cfts, &css->ss->cfts, node)
1674                        cgroup_addrm_files(css, cgrp, cfts, false);
1675        }
1676}
1677
1678/**
1679 * css_populate_dir - create subsys files in a cgroup directory
1680 * @css: target css
1681 *
1682 * On failure, no file is added.
1683 */
1684static int css_populate_dir(struct cgroup_subsys_state *css)
1685{
1686        struct cgroup *cgrp = css->cgroup;
1687        struct cftype *cfts, *failed_cfts;
1688        int ret;
1689
1690        if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1691                return 0;
1692
1693        if (!css->ss) {
1694                if (cgroup_on_dfl(cgrp))
1695                        cfts = cgroup_base_files;
1696                else
1697                        cfts = cgroup1_base_files;
1698
1699                ret = cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1700                if (ret < 0)
1701                        return ret;
1702        } else {
1703                list_for_each_entry(cfts, &css->ss->cfts, node) {
1704                        ret = cgroup_addrm_files(css, cgrp, cfts, true);
1705                        if (ret < 0) {
1706                                failed_cfts = cfts;
1707                                goto err;
1708                        }
1709                }
1710        }
1711
1712        css->flags |= CSS_VISIBLE;
1713
1714        return 0;
1715err:
1716        list_for_each_entry(cfts, &css->ss->cfts, node) {
1717                if (cfts == failed_cfts)
1718                        break;
1719                cgroup_addrm_files(css, cgrp, cfts, false);
1720        }
1721        return ret;
1722}
1723
1724int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1725{
1726        struct cgroup *dcgrp = &dst_root->cgrp;
1727        struct cgroup_subsys *ss;
1728        int ssid, i, ret;
1729
1730        lockdep_assert_held(&cgroup_mutex);
1731
1732        do_each_subsys_mask(ss, ssid, ss_mask) {
1733                /*
1734                 * If @ss has non-root csses attached to it, can't move.
1735                 * If @ss is an implicit controller, it is exempt from this
1736                 * rule and can be stolen.
1737                 */
1738                if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1739                    !ss->implicit_on_dfl)
1740                        return -EBUSY;
1741
1742                /* can't move between two non-dummy roots either */
1743                if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1744                        return -EBUSY;
1745        } while_each_subsys_mask();
1746
1747        do_each_subsys_mask(ss, ssid, ss_mask) {
1748                struct cgroup_root *src_root = ss->root;
1749                struct cgroup *scgrp = &src_root->cgrp;
1750                struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1751                struct css_set *cset;
1752
1753                WARN_ON(!css || cgroup_css(dcgrp, ss));
1754
1755                /* disable from the source */
1756                src_root->subsys_mask &= ~(1 << ssid);
1757                WARN_ON(cgroup_apply_control(scgrp));
1758                cgroup_finalize_control(scgrp, 0);
1759
1760                /* rebind */
1761                RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1762                rcu_assign_pointer(dcgrp->subsys[ssid], css);
1763                ss->root = dst_root;
1764                css->cgroup = dcgrp;
1765
1766                spin_lock_irq(&css_set_lock);
1767                hash_for_each(css_set_table, i, cset, hlist)
1768                        list_move_tail(&cset->e_cset_node[ss->id],
1769                                       &dcgrp->e_csets[ss->id]);
1770                spin_unlock_irq(&css_set_lock);
1771
1772                if (ss->css_rstat_flush) {
1773                        list_del_rcu(&css->rstat_css_node);
1774                        list_add_rcu(&css->rstat_css_node,
1775                                     &dcgrp->rstat_css_list);
1776                }
1777
1778                /* default hierarchy doesn't enable controllers by default */
1779                dst_root->subsys_mask |= 1 << ssid;
1780                if (dst_root == &cgrp_dfl_root) {
1781                        static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1782                } else {
1783                        dcgrp->subtree_control |= 1 << ssid;
1784                        static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1785                }
1786
1787                ret = cgroup_apply_control(dcgrp);
1788                if (ret)
1789                        pr_warn("partial failure to rebind %s controller (err=%d)\n",
1790                                ss->name, ret);
1791
1792                if (ss->bind)
1793                        ss->bind(css);
1794        } while_each_subsys_mask();
1795
1796        kernfs_activate(dcgrp->kn);
1797        return 0;
1798}
1799
1800int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1801                     struct kernfs_root *kf_root)
1802{
1803        int len = 0;
1804        char *buf = NULL;
1805        struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1806        struct cgroup *ns_cgroup;
1807
1808        buf = kmalloc(PATH_MAX, GFP_KERNEL);
1809        if (!buf)
1810                return -ENOMEM;
1811
1812        spin_lock_irq(&css_set_lock);
1813        ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1814        len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1815        spin_unlock_irq(&css_set_lock);
1816
1817        if (len >= PATH_MAX)
1818                len = -ERANGE;
1819        else if (len > 0) {
1820                seq_escape(sf, buf, " \t\n\\");
1821                len = 0;
1822        }
1823        kfree(buf);
1824        return len;
1825}
1826
1827enum cgroup2_param {
1828        Opt_nsdelegate,
1829        Opt_memory_localevents,
1830        Opt_memory_recursiveprot,
1831        nr__cgroup2_params
1832};
1833
1834static const struct fs_parameter_spec cgroup2_fs_parameters[] = {
1835        fsparam_flag("nsdelegate",              Opt_nsdelegate),
1836        fsparam_flag("memory_localevents",      Opt_memory_localevents),
1837        fsparam_flag("memory_recursiveprot",    Opt_memory_recursiveprot),
1838        {}
1839};
1840
1841static int cgroup2_parse_param(struct fs_context *fc, struct fs_parameter *param)
1842{
1843        struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1844        struct fs_parse_result result;
1845        int opt;
1846
1847        opt = fs_parse(fc, cgroup2_fs_parameters, param, &result);
1848        if (opt < 0)
1849                return opt;
1850
1851        switch (opt) {
1852        case Opt_nsdelegate:
1853                ctx->flags |= CGRP_ROOT_NS_DELEGATE;
1854                return 0;
1855        case Opt_memory_localevents:
1856                ctx->flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1857                return 0;
1858        case Opt_memory_recursiveprot:
1859                ctx->flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1860                return 0;
1861        }
1862        return -EINVAL;
1863}
1864
1865static void apply_cgroup_root_flags(unsigned int root_flags)
1866{
1867        if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1868                if (root_flags & CGRP_ROOT_NS_DELEGATE)
1869                        cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1870                else
1871                        cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1872
1873                if (root_flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1874                        cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1875                else
1876                        cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_LOCAL_EVENTS;
1877
1878                if (root_flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1879                        cgrp_dfl_root.flags |= CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1880                else
1881                        cgrp_dfl_root.flags &= ~CGRP_ROOT_MEMORY_RECURSIVE_PROT;
1882        }
1883}
1884
1885static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1886{
1887        if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1888                seq_puts(seq, ",nsdelegate");
1889        if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_LOCAL_EVENTS)
1890                seq_puts(seq, ",memory_localevents");
1891        if (cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT)
1892                seq_puts(seq, ",memory_recursiveprot");
1893        return 0;
1894}
1895
1896static int cgroup_reconfigure(struct fs_context *fc)
1897{
1898        struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
1899
1900        apply_cgroup_root_flags(ctx->flags);
1901        return 0;
1902}
1903
1904static void init_cgroup_housekeeping(struct cgroup *cgrp)
1905{
1906        struct cgroup_subsys *ss;
1907        int ssid;
1908
1909        INIT_LIST_HEAD(&cgrp->self.sibling);
1910        INIT_LIST_HEAD(&cgrp->self.children);
1911        INIT_LIST_HEAD(&cgrp->cset_links);
1912        INIT_LIST_HEAD(&cgrp->pidlists);
1913        mutex_init(&cgrp->pidlist_mutex);
1914        cgrp->self.cgroup = cgrp;
1915        cgrp->self.flags |= CSS_ONLINE;
1916        cgrp->dom_cgrp = cgrp;
1917        cgrp->max_descendants = INT_MAX;
1918        cgrp->max_depth = INT_MAX;
1919        INIT_LIST_HEAD(&cgrp->rstat_css_list);
1920        prev_cputime_init(&cgrp->prev_cputime);
1921
1922        for_each_subsys(ss, ssid)
1923                INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1924
1925        init_waitqueue_head(&cgrp->offline_waitq);
1926        INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1927}
1928
1929void init_cgroup_root(struct cgroup_fs_context *ctx)
1930{
1931        struct cgroup_root *root = ctx->root;
1932        struct cgroup *cgrp = &root->cgrp;
1933
1934        INIT_LIST_HEAD(&root->root_list);
1935        atomic_set(&root->nr_cgrps, 1);
1936        cgrp->root = root;
1937        init_cgroup_housekeeping(cgrp);
1938
1939        root->flags = ctx->flags;
1940        if (ctx->release_agent)
1941                strscpy(root->release_agent_path, ctx->release_agent, PATH_MAX);
1942        if (ctx->name)
1943                strscpy(root->name, ctx->name, MAX_CGROUP_ROOT_NAMELEN);
1944        if (ctx->cpuset_clone_children)
1945                set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1946}
1947
1948int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask)
1949{
1950        LIST_HEAD(tmp_links);
1951        struct cgroup *root_cgrp = &root->cgrp;
1952        struct kernfs_syscall_ops *kf_sops;
1953        struct css_set *cset;
1954        int i, ret;
1955
1956        lockdep_assert_held(&cgroup_mutex);
1957
1958        ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1959                              0, GFP_KERNEL);
1960        if (ret)
1961                goto out;
1962
1963        /*
1964         * We're accessing css_set_count without locking css_set_lock here,
1965         * but that's OK - it can only be increased by someone holding
1966         * cgroup_lock, and that's us.  Later rebinding may disable
1967         * controllers on the default hierarchy and thus create new csets,
1968         * which can't be more than the existing ones.  Allocate 2x.
1969         */
1970        ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1971        if (ret)
1972                goto cancel_ref;
1973
1974        ret = cgroup_init_root_id(root);
1975        if (ret)
1976                goto cancel_ref;
1977
1978        kf_sops = root == &cgrp_dfl_root ?
1979                &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1980
1981        root->kf_root = kernfs_create_root(kf_sops,
1982                                           KERNFS_ROOT_CREATE_DEACTIVATED |
1983                                           KERNFS_ROOT_SUPPORT_EXPORTOP |
1984                                           KERNFS_ROOT_SUPPORT_USER_XATTR,
1985                                           root_cgrp);
1986        if (IS_ERR(root->kf_root)) {
1987                ret = PTR_ERR(root->kf_root);
1988                goto exit_root_id;
1989        }
1990        root_cgrp->kn = root->kf_root->kn;
1991        WARN_ON_ONCE(cgroup_ino(root_cgrp) != 1);
1992        root_cgrp->ancestor_ids[0] = cgroup_id(root_cgrp);
1993
1994        ret = css_populate_dir(&root_cgrp->self);
1995        if (ret)
1996                goto destroy_root;
1997
1998        ret = cgroup_rstat_init(root_cgrp);
1999        if (ret)
2000                goto destroy_root;
2001
2002        ret = rebind_subsystems(root, ss_mask);
2003        if (ret)
2004                goto exit_stats;
2005
2006        ret = cgroup_bpf_inherit(root_cgrp);
2007        WARN_ON_ONCE(ret);
2008
2009        trace_cgroup_setup_root(root);
2010
2011        /*
2012         * There must be no failure case after here, since rebinding takes
2013         * care of subsystems' refcounts, which are explicitly dropped in
2014         * the failure exit path.
2015         */
2016        list_add(&root->root_list, &cgroup_roots);
2017        cgroup_root_count++;
2018
2019        /*
2020         * Link the root cgroup in this hierarchy into all the css_set
2021         * objects.
2022         */
2023        spin_lock_irq(&css_set_lock);
2024        hash_for_each(css_set_table, i, cset, hlist) {
2025                link_css_set(&tmp_links, cset, root_cgrp);
2026                if (css_set_populated(cset))
2027                        cgroup_update_populated(root_cgrp, true);
2028        }
2029        spin_unlock_irq(&css_set_lock);
2030
2031        BUG_ON(!list_empty(&root_cgrp->self.children));
2032        BUG_ON(atomic_read(&root->nr_cgrps) != 1);
2033
2034        ret = 0;
2035        goto out;
2036
2037exit_stats:
2038        cgroup_rstat_exit(root_cgrp);
2039destroy_root:
2040        kernfs_destroy_root(root->kf_root);
2041        root->kf_root = NULL;
2042exit_root_id:
2043        cgroup_exit_root_id(root);
2044cancel_ref:
2045        percpu_ref_exit(&root_cgrp->self.refcnt);
2046out:
2047        free_cgrp_cset_links(&tmp_links);
2048        return ret;
2049}
2050
2051int cgroup_do_get_tree(struct fs_context *fc)
2052{
2053        struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2054        int ret;
2055
2056        ctx->kfc.root = ctx->root->kf_root;
2057        if (fc->fs_type == &cgroup2_fs_type)
2058                ctx->kfc.magic = CGROUP2_SUPER_MAGIC;
2059        else
2060                ctx->kfc.magic = CGROUP_SUPER_MAGIC;
2061        ret = kernfs_get_tree(fc);
2062
2063        /*
2064         * In non-init cgroup namespace, instead of root cgroup's dentry,
2065         * we return the dentry corresponding to the cgroupns->root_cgrp.
2066         */
2067        if (!ret && ctx->ns != &init_cgroup_ns) {
2068                struct dentry *nsdentry;
2069                struct super_block *sb = fc->root->d_sb;
2070                struct cgroup *cgrp;
2071
2072                mutex_lock(&cgroup_mutex);
2073                spin_lock_irq(&css_set_lock);
2074
2075                cgrp = cset_cgroup_from_root(ctx->ns->root_cset, ctx->root);
2076
2077                spin_unlock_irq(&css_set_lock);
2078                mutex_unlock(&cgroup_mutex);
2079
2080                nsdentry = kernfs_node_dentry(cgrp->kn, sb);
2081                dput(fc->root);
2082                if (IS_ERR(nsdentry)) {
2083                        deactivate_locked_super(sb);
2084                        ret = PTR_ERR(nsdentry);
2085                        nsdentry = NULL;
2086                }
2087                fc->root = nsdentry;
2088        }
2089
2090        if (!ctx->kfc.new_sb_created)
2091                cgroup_put(&ctx->root->cgrp);
2092
2093        return ret;
2094}
2095
2096/*
2097 * Destroy a cgroup filesystem context.
2098 */
2099static void cgroup_fs_context_free(struct fs_context *fc)
2100{
2101        struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2102
2103        kfree(ctx->name);
2104        kfree(ctx->release_agent);
2105        put_cgroup_ns(ctx->ns);
2106        kernfs_free_fs_context(fc);
2107        kfree(ctx);
2108}
2109
2110static int cgroup_get_tree(struct fs_context *fc)
2111{
2112        struct cgroup_fs_context *ctx = cgroup_fc2context(fc);
2113        int ret;
2114
2115        cgrp_dfl_visible = true;
2116        cgroup_get_live(&cgrp_dfl_root.cgrp);
2117        ctx->root = &cgrp_dfl_root;
2118
2119        ret = cgroup_do_get_tree(fc);
2120        if (!ret)
2121                apply_cgroup_root_flags(ctx->flags);
2122        return ret;
2123}
2124
2125static const struct fs_context_operations cgroup_fs_context_ops = {
2126        .free           = cgroup_fs_context_free,
2127        .parse_param    = cgroup2_parse_param,
2128        .get_tree       = cgroup_get_tree,
2129        .reconfigure    = cgroup_reconfigure,
2130};
2131
2132static const struct fs_context_operations cgroup1_fs_context_ops = {
2133        .free           = cgroup_fs_context_free,
2134        .parse_param    = cgroup1_parse_param,
2135        .get_tree       = cgroup1_get_tree,
2136        .reconfigure    = cgroup1_reconfigure,
2137};
2138
2139/*
2140 * Initialise the cgroup filesystem creation/reconfiguration context.  Notably,
2141 * we select the namespace we're going to use.
2142 */
2143static int cgroup_init_fs_context(struct fs_context *fc)
2144{
2145        struct cgroup_fs_context *ctx;
2146
2147        ctx = kzalloc(sizeof(struct cgroup_fs_context), GFP_KERNEL);
2148        if (!ctx)
2149                return -ENOMEM;
2150
2151        ctx->ns = current->nsproxy->cgroup_ns;
2152        get_cgroup_ns(ctx->ns);
2153        fc->fs_private = &ctx->kfc;
2154        if (fc->fs_type == &cgroup2_fs_type)
2155                fc->ops = &cgroup_fs_context_ops;
2156        else
2157                fc->ops = &cgroup1_fs_context_ops;
2158        put_user_ns(fc->user_ns);
2159        fc->user_ns = get_user_ns(ctx->ns->user_ns);
2160        fc->global = true;
2161        return 0;
2162}
2163
2164static void cgroup_kill_sb(struct super_block *sb)
2165{
2166        struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2167        struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2168
2169        /*
2170         * If @root doesn't have any children, start killing it.
2171         * This prevents new mounts by disabling percpu_ref_tryget_live().
2172         * cgroup_mount() may wait for @root's release.
2173         *
2174         * And don't kill the default root.
2175         */
2176        if (list_empty(&root->cgrp.self.children) && root != &cgrp_dfl_root &&
2177            !percpu_ref_is_dying(&root->cgrp.self.refcnt))
2178                percpu_ref_kill(&root->cgrp.self.refcnt);
2179        cgroup_put(&root->cgrp);
2180        kernfs_kill_sb(sb);
2181}
2182
2183struct file_system_type cgroup_fs_type = {
2184        .name                   = "cgroup",
2185        .init_fs_context        = cgroup_init_fs_context,
2186        .parameters             = cgroup1_fs_parameters,
2187        .kill_sb                = cgroup_kill_sb,
2188        .fs_flags               = FS_USERNS_MOUNT,
2189};
2190
2191static struct file_system_type cgroup2_fs_type = {
2192        .name                   = "cgroup2",
2193        .init_fs_context        = cgroup_init_fs_context,
2194        .parameters             = cgroup2_fs_parameters,
2195        .kill_sb                = cgroup_kill_sb,
2196        .fs_flags               = FS_USERNS_MOUNT,
2197};
2198
2199#ifdef CONFIG_CPUSETS
2200static const struct fs_context_operations cpuset_fs_context_ops = {
2201        .get_tree       = cgroup1_get_tree,
2202        .free           = cgroup_fs_context_free,
2203};
2204
2205/*
2206 * This is ugly, but preserves the userspace API for existing cpuset
2207 * users. If someone tries to mount the "cpuset" filesystem, we
2208 * silently switch it to mount "cgroup" instead
2209 */
2210static int cpuset_init_fs_context(struct fs_context *fc)
2211{
2212        char *agent = kstrdup("/sbin/cpuset_release_agent", GFP_USER);
2213        struct cgroup_fs_context *ctx;
2214        int err;
2215
2216        err = cgroup_init_fs_context(fc);
2217        if (err) {
2218                kfree(agent);
2219                return err;
2220        }
2221
2222        fc->ops = &cpuset_fs_context_ops;
2223
2224        ctx = cgroup_fc2context(fc);
2225        ctx->subsys_mask = 1 << cpuset_cgrp_id;
2226        ctx->flags |= CGRP_ROOT_NOPREFIX;
2227        ctx->release_agent = agent;
2228
2229        get_filesystem(&cgroup_fs_type);
2230        put_filesystem(fc->fs_type);
2231        fc->fs_type = &cgroup_fs_type;
2232
2233        return 0;
2234}
2235
2236static struct file_system_type cpuset_fs_type = {
2237        .name                   = "cpuset",
2238        .init_fs_context        = cpuset_init_fs_context,
2239        .fs_flags               = FS_USERNS_MOUNT,
2240};
2241#endif
2242
2243int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2244                          struct cgroup_namespace *ns)
2245{
2246        struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2247
2248        return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2249}
2250
2251int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2252                   struct cgroup_namespace *ns)
2253{
2254        int ret;
2255
2256        mutex_lock(&cgroup_mutex);
2257        spin_lock_irq(&css_set_lock);
2258
2259        ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2260
2261        spin_unlock_irq(&css_set_lock);
2262        mutex_unlock(&cgroup_mutex);
2263
2264        return ret;
2265}
2266EXPORT_SYMBOL_GPL(cgroup_path_ns);
2267
2268/**
2269 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2270 * @task: target task
2271 * @buf: the buffer to write the path into
2272 * @buflen: the length of the buffer
2273 *
2274 * Determine @task's cgroup on the first (the one with the lowest non-zero
2275 * hierarchy_id) cgroup hierarchy and copy its path into @buf.  This
2276 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2277 * cgroup controller callbacks.
2278 *
2279 * Return value is the same as kernfs_path().
2280 */
2281int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2282{
2283        struct cgroup_root *root;
2284        struct cgroup *cgrp;
2285        int hierarchy_id = 1;
2286        int ret;
2287
2288        mutex_lock(&cgroup_mutex);
2289        spin_lock_irq(&css_set_lock);
2290
2291        root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2292
2293        if (root) {
2294                cgrp = task_cgroup_from_root(task, root);
2295                ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2296        } else {
2297                /* if no hierarchy exists, everyone is in "/" */
2298                ret = strlcpy(buf, "/", buflen);
2299        }
2300
2301        spin_unlock_irq(&css_set_lock);
2302        mutex_unlock(&cgroup_mutex);
2303        return ret;
2304}
2305EXPORT_SYMBOL_GPL(task_cgroup_path);
2306
2307/**
2308 * cgroup_migrate_add_task - add a migration target task to a migration context
2309 * @task: target task
2310 * @mgctx: target migration context
2311 *
2312 * Add @task, which is a migration target, to @mgctx->tset.  This function
2313 * becomes noop if @task doesn't need to be migrated.  @task's css_set
2314 * should have been added as a migration source and @task->cg_list will be
2315 * moved from the css_set's tasks list to mg_tasks one.
2316 */
2317static void cgroup_migrate_add_task(struct task_struct *task,
2318                                    struct cgroup_mgctx *mgctx)
2319{
2320        struct css_set *cset;
2321
2322        lockdep_assert_held(&css_set_lock);
2323
2324        /* @task either already exited or can't exit until the end */
2325        if (task->flags & PF_EXITING)
2326                return;
2327
2328        /* cgroup_threadgroup_rwsem protects racing against forks */
2329        WARN_ON_ONCE(list_empty(&task->cg_list));
2330
2331        cset = task_css_set(task);
2332        if (!cset->mg_src_cgrp)
2333                return;
2334
2335        mgctx->tset.nr_tasks++;
2336
2337        list_move_tail(&task->cg_list, &cset->mg_tasks);
2338        if (list_empty(&cset->mg_node))
2339                list_add_tail(&cset->mg_node,
2340                              &mgctx->tset.src_csets);
2341        if (list_empty(&cset->mg_dst_cset->mg_node))
2342                list_add_tail(&cset->mg_dst_cset->mg_node,
2343                              &mgctx->tset.dst_csets);
2344}
2345
2346/**
2347 * cgroup_taskset_first - reset taskset and return the first task
2348 * @tset: taskset of interest
2349 * @dst_cssp: output variable for the destination css
2350 *
2351 * @tset iteration is initialized and the first task is returned.
2352 */
2353struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2354                                         struct cgroup_subsys_state **dst_cssp)
2355{
2356        tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2357        tset->cur_task = NULL;
2358
2359        return cgroup_taskset_next(tset, dst_cssp);
2360}
2361
2362/**
2363 * cgroup_taskset_next - iterate to the next task in taskset
2364 * @tset: taskset of interest
2365 * @dst_cssp: output variable for the destination css
2366 *
2367 * Return the next task in @tset.  Iteration must have been initialized
2368 * with cgroup_taskset_first().
2369 */
2370struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2371                                        struct cgroup_subsys_state **dst_cssp)
2372{
2373        struct css_set *cset = tset->cur_cset;
2374        struct task_struct *task = tset->cur_task;
2375
2376        while (&cset->mg_node != tset->csets) {
2377                if (!task)
2378                        task = list_first_entry(&cset->mg_tasks,
2379                                                struct task_struct, cg_list);
2380                else
2381                        task = list_next_entry(task, cg_list);
2382
2383                if (&task->cg_list != &cset->mg_tasks) {
2384                        tset->cur_cset = cset;
2385                        tset->cur_task = task;
2386
2387                        /*
2388                         * This function may be called both before and
2389                         * after cgroup_taskset_migrate().  The two cases
2390                         * can be distinguished by looking at whether @cset
2391                         * has its ->mg_dst_cset set.
2392                         */
2393                        if (cset->mg_dst_cset)
2394                                *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2395                        else
2396                                *dst_cssp = cset->subsys[tset->ssid];
2397
2398                        return task;
2399                }
2400
2401                cset = list_next_entry(cset, mg_node);
2402                task = NULL;
2403        }
2404
2405        return NULL;
2406}
2407
2408/**
2409 * cgroup_migrate_execute - migrate a taskset
2410 * @mgctx: migration context
2411 *
2412 * Migrate tasks in @mgctx as setup by migration preparation functions.
2413 * This function fails iff one of the ->can_attach callbacks fails and
2414 * guarantees that either all or none of the tasks in @mgctx are migrated.
2415 * @mgctx is consumed regardless of success.
2416 */
2417static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2418{
2419        struct cgroup_taskset *tset = &mgctx->tset;
2420        struct cgroup_subsys *ss;
2421        struct task_struct *task, *tmp_task;
2422        struct css_set *cset, *tmp_cset;
2423        int ssid, failed_ssid, ret;
2424
2425        /* check that we can legitimately attach to the cgroup */
2426        if (tset->nr_tasks) {
2427                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2428                        if (ss->can_attach) {
2429                                tset->ssid = ssid;
2430                                ret = ss->can_attach(tset);
2431                                if (ret) {
2432                                        failed_ssid = ssid;
2433                                        goto out_cancel_attach;
2434                                }
2435                        }
2436                } while_each_subsys_mask();
2437        }
2438
2439        /*
2440         * Now that we're guaranteed success, proceed to move all tasks to
2441         * the new cgroup.  There are no failure cases after here, so this
2442         * is the commit point.
2443         */
2444        spin_lock_irq(&css_set_lock);
2445        list_for_each_entry(cset, &tset->src_csets, mg_node) {
2446                list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2447                        struct css_set *from_cset = task_css_set(task);
2448                        struct css_set *to_cset = cset->mg_dst_cset;
2449
2450                        get_css_set(to_cset);
2451                        to_cset->nr_tasks++;
2452                        css_set_move_task(task, from_cset, to_cset, true);
2453                        from_cset->nr_tasks--;
2454                        /*
2455                         * If the source or destination cgroup is frozen,
2456                         * the task might require to change its state.
2457                         */
2458                        cgroup_freezer_migrate_task(task, from_cset->dfl_cgrp,
2459                                                    to_cset->dfl_cgrp);
2460                        put_css_set_locked(from_cset);
2461
2462                }
2463        }
2464        spin_unlock_irq(&css_set_lock);
2465
2466        /*
2467         * Migration is committed, all target tasks are now on dst_csets.
2468         * Nothing is sensitive to fork() after this point.  Notify
2469         * controllers that migration is complete.
2470         */
2471        tset->csets = &tset->dst_csets;
2472
2473        if (tset->nr_tasks) {
2474                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2475                        if (ss->attach) {
2476                                tset->ssid = ssid;
2477                                ss->attach(tset);
2478                        }
2479                } while_each_subsys_mask();
2480        }
2481
2482        ret = 0;
2483        goto out_release_tset;
2484
2485out_cancel_attach:
2486        if (tset->nr_tasks) {
2487                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2488                        if (ssid == failed_ssid)
2489                                break;
2490                        if (ss->cancel_attach) {
2491                                tset->ssid = ssid;
2492                                ss->cancel_attach(tset);
2493                        }
2494                } while_each_subsys_mask();
2495        }
2496out_release_tset:
2497        spin_lock_irq(&css_set_lock);
2498        list_splice_init(&tset->dst_csets, &tset->src_csets);
2499        list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2500                list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2501                list_del_init(&cset->mg_node);
2502        }
2503        spin_unlock_irq(&css_set_lock);
2504
2505        /*
2506         * Re-initialize the cgroup_taskset structure in case it is reused
2507         * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2508         * iteration.
2509         */
2510        tset->nr_tasks = 0;
2511        tset->csets    = &tset->src_csets;
2512        return ret;
2513}
2514
2515/**
2516 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2517 * @dst_cgrp: destination cgroup to test
2518 *
2519 * On the default hierarchy, except for the mixable, (possible) thread root
2520 * and threaded cgroups, subtree_control must be zero for migration
2521 * destination cgroups with tasks so that child cgroups don't compete
2522 * against tasks.
2523 */
2524int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2525{
2526        /* v1 doesn't have any restriction */
2527        if (!cgroup_on_dfl(dst_cgrp))
2528                return 0;
2529
2530        /* verify @dst_cgrp can host resources */
2531        if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2532                return -EOPNOTSUPP;
2533
2534        /* mixables don't care */
2535        if (cgroup_is_mixable(dst_cgrp))
2536                return 0;
2537
2538        /*
2539         * If @dst_cgrp is already or can become a thread root or is
2540         * threaded, it doesn't matter.
2541         */
2542        if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2543                return 0;
2544
2545        /* apply no-internal-process constraint */
2546        if (dst_cgrp->subtree_control)
2547                return -EBUSY;
2548
2549        return 0;
2550}
2551
2552/**
2553 * cgroup_migrate_finish - cleanup after attach
2554 * @mgctx: migration context
2555 *
2556 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst().  See
2557 * those functions for details.
2558 */
2559void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2560{
2561        LIST_HEAD(preloaded);
2562        struct css_set *cset, *tmp_cset;
2563
2564        lockdep_assert_held(&cgroup_mutex);
2565
2566        spin_lock_irq(&css_set_lock);
2567
2568        list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2569        list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2570
2571        list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2572                cset->mg_src_cgrp = NULL;
2573                cset->mg_dst_cgrp = NULL;
2574                cset->mg_dst_cset = NULL;
2575                list_del_init(&cset->mg_preload_node);
2576                put_css_set_locked(cset);
2577        }
2578
2579        spin_unlock_irq(&css_set_lock);
2580}
2581
2582/**
2583 * cgroup_migrate_add_src - add a migration source css_set
2584 * @src_cset: the source css_set to add
2585 * @dst_cgrp: the destination cgroup
2586 * @mgctx: migration context
2587 *
2588 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp.  Pin
2589 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2590 * up by cgroup_migrate_finish().
2591 *
2592 * This function may be called without holding cgroup_threadgroup_rwsem
2593 * even if the target is a process.  Threads may be created and destroyed
2594 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2595 * into play and the preloaded css_sets are guaranteed to cover all
2596 * migrations.
2597 */
2598void cgroup_migrate_add_src(struct css_set *src_cset,
2599                            struct cgroup *dst_cgrp,
2600                            struct cgroup_mgctx *mgctx)
2601{
2602        struct cgroup *src_cgrp;
2603
2604        lockdep_assert_held(&cgroup_mutex);
2605        lockdep_assert_held(&css_set_lock);
2606
2607        /*
2608         * If ->dead, @src_set is associated with one or more dead cgroups
2609         * and doesn't contain any migratable tasks.  Ignore it early so
2610         * that the rest of migration path doesn't get confused by it.
2611         */
2612        if (src_cset->dead)
2613                return;
2614
2615        src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2616
2617        if (!list_empty(&src_cset->mg_preload_node))
2618                return;
2619
2620        WARN_ON(src_cset->mg_src_cgrp);
2621        WARN_ON(src_cset->mg_dst_cgrp);
2622        WARN_ON(!list_empty(&src_cset->mg_tasks));
2623        WARN_ON(!list_empty(&src_cset->mg_node));
2624
2625        src_cset->mg_src_cgrp = src_cgrp;
2626        src_cset->mg_dst_cgrp = dst_cgrp;
2627        get_css_set(src_cset);
2628        list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2629}
2630
2631/**
2632 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2633 * @mgctx: migration context
2634 *
2635 * Tasks are about to be moved and all the source css_sets have been
2636 * preloaded to @mgctx->preloaded_src_csets.  This function looks up and
2637 * pins all destination css_sets, links each to its source, and append them
2638 * to @mgctx->preloaded_dst_csets.
2639 *
2640 * This function must be called after cgroup_migrate_add_src() has been
2641 * called on each migration source css_set.  After migration is performed
2642 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2643 * @mgctx.
2644 */
2645int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2646{
2647        struct css_set *src_cset, *tmp_cset;
2648
2649        lockdep_assert_held(&cgroup_mutex);
2650
2651        /* look up the dst cset for each src cset and link it to src */
2652        list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2653                                 mg_preload_node) {
2654                struct css_set *dst_cset;
2655                struct cgroup_subsys *ss;
2656                int ssid;
2657
2658                dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2659                if (!dst_cset)
2660                        return -ENOMEM;
2661
2662                WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2663
2664                /*
2665                 * If src cset equals dst, it's noop.  Drop the src.
2666                 * cgroup_migrate() will skip the cset too.  Note that we
2667                 * can't handle src == dst as some nodes are used by both.
2668                 */
2669                if (src_cset == dst_cset) {
2670                        src_cset->mg_src_cgrp = NULL;
2671                        src_cset->mg_dst_cgrp = NULL;
2672                        list_del_init(&src_cset->mg_preload_node);
2673                        put_css_set(src_cset);
2674                        put_css_set(dst_cset);
2675                        continue;
2676                }
2677
2678                src_cset->mg_dst_cset = dst_cset;
2679
2680                if (list_empty(&dst_cset->mg_preload_node))
2681                        list_add_tail(&dst_cset->mg_preload_node,
2682                                      &mgctx->preloaded_dst_csets);
2683                else
2684                        put_css_set(dst_cset);
2685
2686                for_each_subsys(ss, ssid)
2687                        if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2688                                mgctx->ss_mask |= 1 << ssid;
2689        }
2690
2691        return 0;
2692}
2693
2694/**
2695 * cgroup_migrate - migrate a process or task to a cgroup
2696 * @leader: the leader of the process or the task to migrate
2697 * @threadgroup: whether @leader points to the whole process or a single task
2698 * @mgctx: migration context
2699 *
2700 * Migrate a process or task denoted by @leader.  If migrating a process,
2701 * the caller must be holding cgroup_threadgroup_rwsem.  The caller is also
2702 * responsible for invoking cgroup_migrate_add_src() and
2703 * cgroup_migrate_prepare_dst() on the targets before invoking this
2704 * function and following up with cgroup_migrate_finish().
2705 *
2706 * As long as a controller's ->can_attach() doesn't fail, this function is
2707 * guaranteed to succeed.  This means that, excluding ->can_attach()
2708 * failure, when migrating multiple targets, the success or failure can be
2709 * decided for all targets by invoking group_migrate_prepare_dst() before
2710 * actually starting migrating.
2711 */
2712int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2713                   struct cgroup_mgctx *mgctx)
2714{
2715        struct task_struct *task;
2716
2717        /*
2718         * Prevent freeing of tasks while we take a snapshot. Tasks that are
2719         * already PF_EXITING could be freed from underneath us unless we
2720         * take an rcu_read_lock.
2721         */
2722        spin_lock_irq(&css_set_lock);
2723        rcu_read_lock();
2724        task = leader;
2725        do {
2726                cgroup_migrate_add_task(task, mgctx);
2727                if (!threadgroup)
2728                        break;
2729        } while_each_thread(leader, task);
2730        rcu_read_unlock();
2731        spin_unlock_irq(&css_set_lock);
2732
2733        return cgroup_migrate_execute(mgctx);
2734}
2735
2736/**
2737 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2738 * @dst_cgrp: the cgroup to attach to
2739 * @leader: the task or the leader of the threadgroup to be attached
2740 * @threadgroup: attach the whole threadgroup?
2741 *
2742 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2743 */
2744int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2745                       bool threadgroup)
2746{
2747        DEFINE_CGROUP_MGCTX(mgctx);
2748        struct task_struct *task;
2749        int ret = 0;
2750
2751        /* look up all src csets */
2752        spin_lock_irq(&css_set_lock);
2753        rcu_read_lock();
2754        task = leader;
2755        do {
2756                cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2757                if (!threadgroup)
2758                        break;
2759        } while_each_thread(leader, task);
2760        rcu_read_unlock();
2761        spin_unlock_irq(&css_set_lock);
2762
2763        /* prepare dst csets and commit */
2764        ret = cgroup_migrate_prepare_dst(&mgctx);
2765        if (!ret)
2766                ret = cgroup_migrate(leader, threadgroup, &mgctx);
2767
2768        cgroup_migrate_finish(&mgctx);
2769
2770        if (!ret)
2771                TRACE_CGROUP_PATH(attach_task, dst_cgrp, leader, threadgroup);
2772
2773        return ret;
2774}
2775
2776struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup,
2777                                             bool *locked)
2778        __acquires(&cgroup_threadgroup_rwsem)
2779{
2780        struct task_struct *tsk;
2781        pid_t pid;
2782
2783        if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2784                return ERR_PTR(-EINVAL);
2785
2786        /*
2787         * If we migrate a single thread, we don't care about threadgroup
2788         * stability. If the thread is `current`, it won't exit(2) under our
2789         * hands or change PID through exec(2). We exclude
2790         * cgroup_update_dfl_csses and other cgroup_{proc,thread}s_write
2791         * callers by cgroup_mutex.
2792         * Therefore, we can skip the global lock.
2793         */
2794        lockdep_assert_held(&cgroup_mutex);
2795        if (pid || threadgroup) {
2796                percpu_down_write(&cgroup_threadgroup_rwsem);
2797                *locked = true;
2798        } else {
2799                *locked = false;
2800        }
2801
2802        rcu_read_lock();
2803        if (pid) {
2804                tsk = find_task_by_vpid(pid);
2805                if (!tsk) {
2806                        tsk = ERR_PTR(-ESRCH);
2807                        goto out_unlock_threadgroup;
2808                }
2809        } else {
2810                tsk = current;
2811        }
2812
2813        if (threadgroup)
2814                tsk = tsk->group_leader;
2815
2816        /*
2817         * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2818         * If userland migrates such a kthread to a non-root cgroup, it can
2819         * become trapped in a cpuset, or RT kthread may be born in a
2820         * cgroup with no rt_runtime allocated.  Just say no.
2821         */
2822        if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2823                tsk = ERR_PTR(-EINVAL);
2824                goto out_unlock_threadgroup;
2825        }
2826
2827        get_task_struct(tsk);
2828        goto out_unlock_rcu;
2829
2830out_unlock_threadgroup:
2831        if (*locked) {
2832                percpu_up_write(&cgroup_threadgroup_rwsem);
2833                *locked = false;
2834        }
2835out_unlock_rcu:
2836        rcu_read_unlock();
2837        return tsk;
2838}
2839
2840void cgroup_procs_write_finish(struct task_struct *task, bool locked)
2841        __releases(&cgroup_threadgroup_rwsem)
2842{
2843        struct cgroup_subsys *ss;
2844        int ssid;
2845
2846        /* release reference from cgroup_procs_write_start() */
2847        put_task_struct(task);
2848
2849        if (locked)
2850                percpu_up_write(&cgroup_threadgroup_rwsem);
2851        for_each_subsys(ss, ssid)
2852                if (ss->post_attach)
2853                        ss->post_attach();
2854}
2855
2856static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2857{
2858        struct cgroup_subsys *ss;
2859        bool printed = false;
2860        int ssid;
2861
2862        do_each_subsys_mask(ss, ssid, ss_mask) {
2863                if (printed)
2864                        seq_putc(seq, ' ');
2865                seq_puts(seq, ss->name);
2866                printed = true;
2867        } while_each_subsys_mask();
2868        if (printed)
2869                seq_putc(seq, '\n');
2870}
2871
2872/* show controllers which are enabled from the parent */
2873static int cgroup_controllers_show(struct seq_file *seq, void *v)
2874{
2875        struct cgroup *cgrp = seq_css(seq)->cgroup;
2876
2877        cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2878        return 0;
2879}
2880
2881/* show controllers which are enabled for a given cgroup's children */
2882static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2883{
2884        struct cgroup *cgrp = seq_css(seq)->cgroup;
2885
2886        cgroup_print_ss_mask(seq, cgrp->subtree_control);
2887        return 0;
2888}
2889
2890/**
2891 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2892 * @cgrp: root of the subtree to update csses for
2893 *
2894 * @cgrp's control masks have changed and its subtree's css associations
2895 * need to be updated accordingly.  This function looks up all css_sets
2896 * which are attached to the subtree, creates the matching updated css_sets
2897 * and migrates the tasks to the new ones.
2898 */
2899static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2900{
2901        DEFINE_CGROUP_MGCTX(mgctx);
2902        struct cgroup_subsys_state *d_css;
2903        struct cgroup *dsct;
2904        struct css_set *src_cset;
2905        int ret;
2906
2907        lockdep_assert_held(&cgroup_mutex);
2908
2909        percpu_down_write(&cgroup_threadgroup_rwsem);
2910
2911        /* look up all csses currently attached to @cgrp's subtree */
2912        spin_lock_irq(&css_set_lock);
2913        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2914                struct cgrp_cset_link *link;
2915
2916                list_for_each_entry(link, &dsct->cset_links, cset_link)
2917                        cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2918        }
2919        spin_unlock_irq(&css_set_lock);
2920
2921        /* NULL dst indicates self on default hierarchy */
2922        ret = cgroup_migrate_prepare_dst(&mgctx);
2923        if (ret)
2924                goto out_finish;
2925
2926        spin_lock_irq(&css_set_lock);
2927        list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2928                struct task_struct *task, *ntask;
2929
2930                /* all tasks in src_csets need to be migrated */
2931                list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2932                        cgroup_migrate_add_task(task, &mgctx);
2933        }
2934        spin_unlock_irq(&css_set_lock);
2935
2936        ret = cgroup_migrate_execute(&mgctx);
2937out_finish:
2938        cgroup_migrate_finish(&mgctx);
2939        percpu_up_write(&cgroup_threadgroup_rwsem);
2940        return ret;
2941}
2942
2943/**
2944 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2945 * @cgrp: root of the target subtree
2946 *
2947 * Because css offlining is asynchronous, userland may try to re-enable a
2948 * controller while the previous css is still around.  This function grabs
2949 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2950 */
2951void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2952        __acquires(&cgroup_mutex)
2953{
2954        struct cgroup *dsct;
2955        struct cgroup_subsys_state *d_css;
2956        struct cgroup_subsys *ss;
2957        int ssid;
2958
2959restart:
2960        mutex_lock(&cgroup_mutex);
2961
2962        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2963                for_each_subsys(ss, ssid) {
2964                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2965                        DEFINE_WAIT(wait);
2966
2967                        if (!css || !percpu_ref_is_dying(&css->refcnt))
2968                                continue;
2969
2970                        cgroup_get_live(dsct);
2971                        prepare_to_wait(&dsct->offline_waitq, &wait,
2972                                        TASK_UNINTERRUPTIBLE);
2973
2974                        mutex_unlock(&cgroup_mutex);
2975                        schedule();
2976                        finish_wait(&dsct->offline_waitq, &wait);
2977
2978                        cgroup_put(dsct);
2979                        goto restart;
2980                }
2981        }
2982}
2983
2984/**
2985 * cgroup_save_control - save control masks and dom_cgrp of a subtree
2986 * @cgrp: root of the target subtree
2987 *
2988 * Save ->subtree_control, ->subtree_ss_mask and ->dom_cgrp to the
2989 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
2990 * itself.
2991 */
2992static void cgroup_save_control(struct cgroup *cgrp)
2993{
2994        struct cgroup *dsct;
2995        struct cgroup_subsys_state *d_css;
2996
2997        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2998                dsct->old_subtree_control = dsct->subtree_control;
2999                dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
3000                dsct->old_dom_cgrp = dsct->dom_cgrp;
3001        }
3002}
3003
3004/**
3005 * cgroup_propagate_control - refresh control masks of a subtree
3006 * @cgrp: root of the target subtree
3007 *
3008 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
3009 * ->subtree_control and propagate controller availability through the
3010 * subtree so that descendants don't have unavailable controllers enabled.
3011 */
3012static void cgroup_propagate_control(struct cgroup *cgrp)
3013{
3014        struct cgroup *dsct;
3015        struct cgroup_subsys_state *d_css;
3016
3017        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3018                dsct->subtree_control &= cgroup_control(dsct);
3019                dsct->subtree_ss_mask =
3020                        cgroup_calc_subtree_ss_mask(dsct->subtree_control,
3021                                                    cgroup_ss_mask(dsct));
3022        }
3023}
3024
3025/**
3026 * cgroup_restore_control - restore control masks and dom_cgrp of a subtree
3027 * @cgrp: root of the target subtree
3028 *
3029 * Restore ->subtree_control, ->subtree_ss_mask and ->dom_cgrp from the
3030 * respective old_ prefixed fields for @cgrp's subtree including @cgrp
3031 * itself.
3032 */
3033static void cgroup_restore_control(struct cgroup *cgrp)
3034{
3035        struct cgroup *dsct;
3036        struct cgroup_subsys_state *d_css;
3037
3038        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3039                dsct->subtree_control = dsct->old_subtree_control;
3040                dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
3041                dsct->dom_cgrp = dsct->old_dom_cgrp;
3042        }
3043}
3044
3045static bool css_visible(struct cgroup_subsys_state *css)
3046{
3047        struct cgroup_subsys *ss = css->ss;
3048        struct cgroup *cgrp = css->cgroup;
3049
3050        if (cgroup_control(cgrp) & (1 << ss->id))
3051                return true;
3052        if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
3053                return false;
3054        return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
3055}
3056
3057/**
3058 * cgroup_apply_control_enable - enable or show csses according to control
3059 * @cgrp: root of the target subtree
3060 *
3061 * Walk @cgrp's subtree and create new csses or make the existing ones
3062 * visible.  A css is created invisible if it's being implicitly enabled
3063 * through dependency.  An invisible css is made visible when the userland
3064 * explicitly enables it.
3065 *
3066 * Returns 0 on success, -errno on failure.  On failure, csses which have
3067 * been processed already aren't cleaned up.  The caller is responsible for
3068 * cleaning up with cgroup_apply_control_disable().
3069 */
3070static int cgroup_apply_control_enable(struct cgroup *cgrp)
3071{
3072        struct cgroup *dsct;
3073        struct cgroup_subsys_state *d_css;
3074        struct cgroup_subsys *ss;
3075        int ssid, ret;
3076
3077        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
3078                for_each_subsys(ss, ssid) {
3079                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3080
3081                        if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
3082                                continue;
3083
3084                        if (!css) {
3085                                css = css_create(dsct, ss);
3086                                if (IS_ERR(css))
3087                                        return PTR_ERR(css);
3088                        }
3089
3090                        WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3091
3092                        if (css_visible(css)) {
3093                                ret = css_populate_dir(css);
3094                                if (ret)
3095                                        return ret;
3096                        }
3097                }
3098        }
3099
3100        return 0;
3101}
3102
3103/**
3104 * cgroup_apply_control_disable - kill or hide csses according to control
3105 * @cgrp: root of the target subtree
3106 *
3107 * Walk @cgrp's subtree and kill and hide csses so that they match
3108 * cgroup_ss_mask() and cgroup_visible_mask().
3109 *
3110 * A css is hidden when the userland requests it to be disabled while other
3111 * subsystems are still depending on it.  The css must not actively control
3112 * resources and be in the vanilla state if it's made visible again later.
3113 * Controllers which may be depended upon should provide ->css_reset() for
3114 * this purpose.
3115 */
3116static void cgroup_apply_control_disable(struct cgroup *cgrp)
3117{
3118        struct cgroup *dsct;
3119        struct cgroup_subsys_state *d_css;
3120        struct cgroup_subsys *ss;
3121        int ssid;
3122
3123        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
3124                for_each_subsys(ss, ssid) {
3125                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
3126
3127                        if (!css)
3128                                continue;
3129
3130                        WARN_ON_ONCE(percpu_ref_is_dying(&css->refcnt));
3131
3132                        if (css->parent &&
3133                            !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
3134                                kill_css(css);
3135                        } else if (!css_visible(css)) {
3136                                css_clear_dir(css);
3137                                if (ss->css_reset)
3138                                        ss->css_reset(css);
3139                        }
3140                }
3141        }
3142}
3143
3144/**
3145 * cgroup_apply_control - apply control mask updates to the subtree
3146 * @cgrp: root of the target subtree
3147 *
3148 * subsystems can be enabled and disabled in a subtree using the following
3149 * steps.
3150 *
3151 * 1. Call cgroup_save_control() to stash the current state.
3152 * 2. Update ->subtree_control masks in the subtree as desired.
3153 * 3. Call cgroup_apply_control() to apply the changes.
3154 * 4. Optionally perform other related operations.
3155 * 5. Call cgroup_finalize_control() to finish up.
3156 *
3157 * This function implements step 3 and propagates the mask changes
3158 * throughout @cgrp's subtree, updates csses accordingly and perform
3159 * process migrations.
3160 */
3161static int cgroup_apply_control(struct cgroup *cgrp)
3162{
3163        int ret;
3164
3165        cgroup_propagate_control(cgrp);
3166
3167        ret = cgroup_apply_control_enable(cgrp);
3168        if (ret)
3169                return ret;
3170
3171        /*
3172         * At this point, cgroup_e_css_by_mask() results reflect the new csses
3173         * making the following cgroup_update_dfl_csses() properly update
3174         * css associations of all tasks in the subtree.
3175         */
3176        ret = cgroup_update_dfl_csses(cgrp);
3177        if (ret)
3178                return ret;
3179
3180        return 0;
3181}
3182
3183/**
3184 * cgroup_finalize_control - finalize control mask update
3185 * @cgrp: root of the target subtree
3186 * @ret: the result of the update
3187 *
3188 * Finalize control mask update.  See cgroup_apply_control() for more info.
3189 */
3190static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
3191{
3192        if (ret) {
3193                cgroup_restore_control(cgrp);
3194                cgroup_propagate_control(cgrp);
3195        }
3196
3197        cgroup_apply_control_disable(cgrp);
3198}
3199
3200static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
3201{
3202        u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
3203
3204        /* if nothing is getting enabled, nothing to worry about */
3205        if (!enable)
3206                return 0;
3207
3208        /* can @cgrp host any resources? */
3209        if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3210                return -EOPNOTSUPP;
3211
3212        /* mixables don't care */
3213        if (cgroup_is_mixable(cgrp))
3214                return 0;
3215
3216        if (domain_enable) {
3217                /* can't enable domain controllers inside a thread subtree */
3218                if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3219                        return -EOPNOTSUPP;
3220        } else {
3221                /*
3222                 * Threaded controllers can handle internal competitions
3223                 * and are always allowed inside a (prospective) thread
3224                 * subtree.
3225                 */
3226                if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3227                        return 0;
3228        }
3229
3230        /*
3231         * Controllers can't be enabled for a cgroup with tasks to avoid
3232         * child cgroups competing against tasks.
3233         */
3234        if (cgroup_has_tasks(cgrp))
3235                return -EBUSY;
3236
3237        return 0;
3238}
3239
3240/* change the enabled child controllers for a cgroup in the default hierarchy */
3241static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3242                                            char *buf, size_t nbytes,
3243                                            loff_t off)
3244{
3245        u16 enable = 0, disable = 0;
3246        struct cgroup *cgrp, *child;
3247        struct cgroup_subsys *ss;
3248        char *tok;
3249        int ssid, ret;
3250
3251        /*
3252         * Parse input - space separated list of subsystem names prefixed
3253         * with either + or -.
3254         */
3255        buf = strstrip(buf);
3256        while ((tok = strsep(&buf, " "))) {
3257                if (tok[0] == '\0')
3258                        continue;
3259                do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3260                        if (!cgroup_ssid_enabled(ssid) ||
3261                            strcmp(tok + 1, ss->name))
3262                                continue;
3263
3264                        if (*tok == '+') {
3265                                enable |= 1 << ssid;
3266                                disable &= ~(1 << ssid);
3267                        } else if (*tok == '-') {
3268                                disable |= 1 << ssid;
3269                                enable &= ~(1 << ssid);
3270                        } else {
3271                                return -EINVAL;
3272                        }
3273                        break;
3274                } while_each_subsys_mask();
3275                if (ssid == CGROUP_SUBSYS_COUNT)
3276                        return -EINVAL;
3277        }
3278
3279        cgrp = cgroup_kn_lock_live(of->kn, true);
3280        if (!cgrp)
3281                return -ENODEV;
3282
3283        for_each_subsys(ss, ssid) {
3284                if (enable & (1 << ssid)) {
3285                        if (cgrp->subtree_control & (1 << ssid)) {
3286                                enable &= ~(1 << ssid);
3287                                continue;
3288                        }
3289
3290                        if (!(cgroup_control(cgrp) & (1 << ssid))) {
3291                                ret = -ENOENT;
3292                                goto out_unlock;
3293                        }
3294                } else if (disable & (1 << ssid)) {
3295                        if (!(cgrp->subtree_control & (1 << ssid))) {
3296                                disable &= ~(1 << ssid);
3297                                continue;
3298                        }
3299
3300                        /* a child has it enabled? */
3301                        cgroup_for_each_live_child(child, cgrp) {
3302                                if (child->subtree_control & (1 << ssid)) {
3303                                        ret = -EBUSY;
3304                                        goto out_unlock;
3305                                }
3306                        }
3307                }
3308        }
3309
3310        if (!enable && !disable) {
3311                ret = 0;
3312                goto out_unlock;
3313        }
3314
3315        ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3316        if (ret)
3317                goto out_unlock;
3318
3319        /* save and update control masks and prepare csses */
3320        cgroup_save_control(cgrp);
3321
3322        cgrp->subtree_control |= enable;
3323        cgrp->subtree_control &= ~disable;
3324
3325        ret = cgroup_apply_control(cgrp);
3326        cgroup_finalize_control(cgrp, ret);
3327        if (ret)
3328                goto out_unlock;
3329
3330        kernfs_activate(cgrp->kn);
3331out_unlock:
3332        cgroup_kn_unlock(of->kn);
3333        return ret ?: nbytes;
3334}
3335
3336/**
3337 * cgroup_enable_threaded - make @cgrp threaded
3338 * @cgrp: the target cgroup
3339 *
3340 * Called when "threaded" is written to the cgroup.type interface file and
3341 * tries to make @cgrp threaded and join the parent's resource domain.
3342 * This function is never called on the root cgroup as cgroup.type doesn't
3343 * exist on it.
3344 */
3345static int cgroup_enable_threaded(struct cgroup *cgrp)
3346{
3347        struct cgroup *parent = cgroup_parent(cgrp);
3348        struct cgroup *dom_cgrp = parent->dom_cgrp;
3349        struct cgroup *dsct;
3350        struct cgroup_subsys_state *d_css;
3351        int ret;
3352
3353        lockdep_assert_held(&cgroup_mutex);
3354
3355        /* noop if already threaded */
3356        if (cgroup_is_threaded(cgrp))
3357                return 0;
3358
3359        /*
3360         * If @cgroup is populated or has domain controllers enabled, it
3361         * can't be switched.  While the below cgroup_can_be_thread_root()
3362         * test can catch the same conditions, that's only when @parent is
3363         * not mixable, so let's check it explicitly.
3364         */
3365        if (cgroup_is_populated(cgrp) ||
3366            cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
3367                return -EOPNOTSUPP;
3368
3369        /* we're joining the parent's domain, ensure its validity */
3370        if (!cgroup_is_valid_domain(dom_cgrp) ||
3371            !cgroup_can_be_thread_root(dom_cgrp))
3372                return -EOPNOTSUPP;
3373
3374        /*
3375         * The following shouldn't cause actual migrations and should
3376         * always succeed.
3377         */
3378        cgroup_save_control(cgrp);
3379
3380        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)
3381                if (dsct == cgrp || cgroup_is_threaded(dsct))
3382                        dsct->dom_cgrp = dom_cgrp;
3383
3384        ret = cgroup_apply_control(cgrp);
3385        if (!ret)
3386                parent->nr_threaded_children++;
3387
3388        cgroup_finalize_control(cgrp, ret);
3389        return ret;
3390}
3391
3392static int cgroup_type_show(struct seq_file *seq, void *v)
3393{
3394        struct cgroup *cgrp = seq_css(seq)->cgroup;
3395
3396        if (cgroup_is_threaded(cgrp))
3397                seq_puts(seq, "threaded\n");
3398        else if (!cgroup_is_valid_domain(cgrp))
3399                seq_puts(seq, "domain invalid\n");
3400        else if (cgroup_is_thread_root(cgrp))
3401                seq_puts(seq, "domain threaded\n");
3402        else
3403                seq_puts(seq, "domain\n");
3404
3405        return 0;
3406}
3407
3408static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3409                                 size_t nbytes, loff_t off)
3410{
3411        struct cgroup *cgrp;
3412        int ret;
3413
3414        /* only switching to threaded mode is supported */
3415        if (strcmp(strstrip(buf), "threaded"))
3416                return -EINVAL;
3417
3418        /* drain dying csses before we re-apply (threaded) subtree control */
3419        cgrp = cgroup_kn_lock_live(of->kn, true);
3420        if (!cgrp)
3421                return -ENOENT;
3422
3423        /* threaded can only be enabled */
3424        ret = cgroup_enable_threaded(cgrp);
3425
3426        cgroup_kn_unlock(of->kn);
3427        return ret ?: nbytes;
3428}
3429
3430static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3431{
3432        struct cgroup *cgrp = seq_css(seq)->cgroup;
3433        int descendants = READ_ONCE(cgrp->max_descendants);
3434
3435        if (descendants == INT_MAX)
3436                seq_puts(seq, "max\n");
3437        else
3438                seq_printf(seq, "%d\n", descendants);
3439
3440        return 0;
3441}
3442
3443static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3444                                           char *buf, size_t nbytes, loff_t off)
3445{
3446        struct cgroup *cgrp;
3447        int descendants;
3448        ssize_t ret;
3449
3450        buf = strstrip(buf);
3451        if (!strcmp(buf, "max")) {
3452                descendants = INT_MAX;
3453        } else {
3454                ret = kstrtoint(buf, 0, &descendants);
3455                if (ret)
3456                        return ret;
3457        }
3458
3459        if (descendants < 0)
3460                return -ERANGE;
3461
3462        cgrp = cgroup_kn_lock_live(of->kn, false);
3463        if (!cgrp)
3464                return -ENOENT;
3465
3466        cgrp->max_descendants = descendants;
3467
3468        cgroup_kn_unlock(of->kn);
3469
3470        return nbytes;
3471}
3472
3473static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3474{
3475        struct cgroup *cgrp = seq_css(seq)->cgroup;
3476        int depth = READ_ONCE(cgrp->max_depth);
3477
3478        if (depth == INT_MAX)
3479                seq_puts(seq, "max\n");
3480        else
3481                seq_printf(seq, "%d\n", depth);
3482
3483        return 0;
3484}
3485
3486static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3487                                      char *buf, size_t nbytes, loff_t off)
3488{
3489        struct cgroup *cgrp;
3490        ssize_t ret;
3491        int depth;
3492
3493        buf = strstrip(buf);
3494        if (!strcmp(buf, "max")) {
3495                depth = INT_MAX;
3496        } else {
3497                ret = kstrtoint(buf, 0, &depth);
3498                if (ret)
3499                        return ret;
3500        }
3501
3502        if (depth < 0)
3503                return -ERANGE;
3504
3505        cgrp = cgroup_kn_lock_live(of->kn, false);
3506        if (!cgrp)
3507                return -ENOENT;
3508
3509        cgrp->max_depth = depth;
3510
3511        cgroup_kn_unlock(of->kn);
3512
3513        return nbytes;
3514}
3515
3516static int cgroup_events_show(struct seq_file *seq, void *v)
3517{
3518        struct cgroup *cgrp = seq_css(seq)->cgroup;
3519
3520        seq_printf(seq, "populated %d\n", cgroup_is_populated(cgrp));
3521        seq_printf(seq, "frozen %d\n", test_bit(CGRP_FROZEN, &cgrp->flags));
3522
3523        return 0;
3524}
3525
3526static int cgroup_stat_show(struct seq_file *seq, void *v)
3527{
3528        struct cgroup *cgroup = seq_css(seq)->cgroup;
3529
3530        seq_printf(seq, "nr_descendants %d\n",
3531                   cgroup->nr_descendants);
3532        seq_printf(seq, "nr_dying_descendants %d\n",
3533                   cgroup->nr_dying_descendants);
3534
3535        return 0;
3536}
3537
3538static int __maybe_unused cgroup_extra_stat_show(struct seq_file *seq,
3539                                                 struct cgroup *cgrp, int ssid)
3540{
3541        struct cgroup_subsys *ss = cgroup_subsys[ssid];
3542        struct cgroup_subsys_state *css;
3543        int ret;
3544
3545        if (!ss->css_extra_stat_show)
3546                return 0;
3547
3548        css = cgroup_tryget_css(cgrp, ss);
3549        if (!css)
3550                return 0;
3551
3552        ret = ss->css_extra_stat_show(seq, css);
3553        css_put(css);
3554        return ret;
3555}
3556
3557static int cpu_stat_show(struct seq_file *seq, void *v)
3558{
3559        struct cgroup __maybe_unused *cgrp = seq_css(seq)->cgroup;
3560        int ret = 0;
3561
3562        cgroup_base_stat_cputime_show(seq);
3563#ifdef CONFIG_CGROUP_SCHED
3564        ret = cgroup_extra_stat_show(seq, cgrp, cpu_cgrp_id);
3565#endif
3566        return ret;
3567}
3568
3569#ifdef CONFIG_PSI
3570static int cgroup_io_pressure_show(struct seq_file *seq, void *v)
3571{
3572        struct cgroup *cgrp = seq_css(seq)->cgroup;
3573        struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3574
3575        return psi_show(seq, psi, PSI_IO);
3576}
3577static int cgroup_memory_pressure_show(struct seq_file *seq, void *v)
3578{
3579        struct cgroup *cgrp = seq_css(seq)->cgroup;
3580        struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3581
3582        return psi_show(seq, psi, PSI_MEM);
3583}
3584static int cgroup_cpu_pressure_show(struct seq_file *seq, void *v)
3585{
3586        struct cgroup *cgrp = seq_css(seq)->cgroup;
3587        struct psi_group *psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3588
3589        return psi_show(seq, psi, PSI_CPU);
3590}
3591
3592static ssize_t cgroup_pressure_write(struct kernfs_open_file *of, char *buf,
3593                                          size_t nbytes, enum psi_res res)
3594{
3595        struct psi_trigger *new;
3596        struct cgroup *cgrp;
3597        struct psi_group *psi;
3598
3599        cgrp = cgroup_kn_lock_live(of->kn, false);
3600        if (!cgrp)
3601                return -ENODEV;
3602
3603        cgroup_get(cgrp);
3604        cgroup_kn_unlock(of->kn);
3605
3606        psi = cgroup_ino(cgrp) == 1 ? &psi_system : &cgrp->psi;
3607        new = psi_trigger_create(psi, buf, nbytes, res);
3608        if (IS_ERR(new)) {
3609                cgroup_put(cgrp);
3610                return PTR_ERR(new);
3611        }
3612
3613        psi_trigger_replace(&of->priv, new);
3614
3615        cgroup_put(cgrp);
3616
3617        return nbytes;
3618}
3619
3620static ssize_t cgroup_io_pressure_write(struct kernfs_open_file *of,
3621                                          char *buf, size_t nbytes,
3622                                          loff_t off)
3623{
3624        return cgroup_pressure_write(of, buf, nbytes, PSI_IO);
3625}
3626
3627static ssize_t cgroup_memory_pressure_write(struct kernfs_open_file *of,
3628                                          char *buf, size_t nbytes,
3629                                          loff_t off)
3630{
3631        return cgroup_pressure_write(of, buf, nbytes, PSI_MEM);
3632}
3633
3634static ssize_t cgroup_cpu_pressure_write(struct kernfs_open_file *of,
3635                                          char *buf, size_t nbytes,
3636                                          loff_t off)
3637{
3638        return cgroup_pressure_write(of, buf, nbytes, PSI_CPU);
3639}
3640
3641static __poll_t cgroup_pressure_poll(struct kernfs_open_file *of,
3642                                          poll_table *pt)
3643{
3644        return psi_trigger_poll(&of->priv, of->file, pt);
3645}
3646
3647static void cgroup_pressure_release(struct kernfs_open_file *of)
3648{
3649        psi_trigger_replace(&of->priv, NULL);
3650}
3651
3652bool cgroup_psi_enabled(void)
3653{
3654        return (cgroup_feature_disable_mask & (1 << OPT_FEATURE_PRESSURE)) == 0;
3655}
3656
3657#else /* CONFIG_PSI */
3658bool cgroup_psi_enabled(void)
3659{
3660        return false;
3661}
3662
3663#endif /* CONFIG_PSI */
3664
3665static int cgroup_freeze_show(struct seq_file *seq, void *v)
3666{
3667        struct cgroup *cgrp = seq_css(seq)->cgroup;
3668
3669        seq_printf(seq, "%d\n", cgrp->freezer.freeze);
3670
3671        return 0;
3672}
3673
3674static ssize_t cgroup_freeze_write(struct kernfs_open_file *of,
3675                                   char *buf, size_t nbytes, loff_t off)
3676{
3677        struct cgroup *cgrp;
3678        ssize_t ret;
3679        int freeze;
3680
3681        ret = kstrtoint(strstrip(buf), 0, &freeze);
3682        if (ret)
3683                return ret;
3684
3685        if (freeze < 0 || freeze > 1)
3686                return -ERANGE;
3687
3688        cgrp = cgroup_kn_lock_live(of->kn, false);
3689        if (!cgrp)
3690                return -ENOENT;
3691
3692        cgroup_freeze(cgrp, freeze);
3693
3694        cgroup_kn_unlock(of->kn);
3695
3696        return nbytes;
3697}
3698
3699static void __cgroup_kill(struct cgroup *cgrp)
3700{
3701        struct css_task_iter it;
3702        struct task_struct *task;
3703
3704        lockdep_assert_held(&cgroup_mutex);
3705
3706        spin_lock_irq(&css_set_lock);
3707        set_bit(CGRP_KILL, &cgrp->flags);
3708        spin_unlock_irq(&css_set_lock);
3709
3710        css_task_iter_start(&cgrp->self, CSS_TASK_ITER_PROCS | CSS_TASK_ITER_THREADED, &it);
3711        while ((task = css_task_iter_next(&it))) {
3712                /* Ignore kernel threads here. */
3713                if (task->flags & PF_KTHREAD)
3714                        continue;
3715
3716                /* Skip tasks that are already dying. */
3717                if (__fatal_signal_pending(task))
3718                        continue;
3719
3720                send_sig(SIGKILL, task, 0);
3721        }
3722        css_task_iter_end(&it);
3723
3724        spin_lock_irq(&css_set_lock);
3725        clear_bit(CGRP_KILL, &cgrp->flags);
3726        spin_unlock_irq(&css_set_lock);
3727}
3728
3729static void cgroup_kill(struct cgroup *cgrp)
3730{
3731        struct cgroup_subsys_state *css;
3732        struct cgroup *dsct;
3733
3734        lockdep_assert_held(&cgroup_mutex);
3735
3736        cgroup_for_each_live_descendant_pre(dsct, css, cgrp)
3737                __cgroup_kill(dsct);
3738}
3739
3740static ssize_t cgroup_kill_write(struct kernfs_open_file *of, char *buf,
3741                                 size_t nbytes, loff_t off)
3742{
3743        ssize_t ret = 0;
3744        int kill;
3745        struct cgroup *cgrp;
3746
3747        ret = kstrtoint(strstrip(buf), 0, &kill);
3748        if (ret)
3749                return ret;
3750
3751        if (kill != 1)
3752                return -ERANGE;
3753
3754        cgrp = cgroup_kn_lock_live(of->kn, false);
3755        if (!cgrp)
3756                return -ENOENT;
3757
3758        /*
3759         * Killing is a process directed operation, i.e. the whole thread-group
3760         * is taken down so act like we do for cgroup.procs and only make this
3761         * writable in non-threaded cgroups.
3762         */
3763        if (cgroup_is_threaded(cgrp))
3764                ret = -EOPNOTSUPP;
3765        else
3766                cgroup_kill(cgrp);
3767
3768        cgroup_kn_unlock(of->kn);
3769
3770        return ret ?: nbytes;
3771}
3772
3773static int cgroup_file_open(struct kernfs_open_file *of)
3774{
3775        struct cftype *cft = of_cft(of);
3776
3777        if (cft->open)
3778                return cft->open(of);
3779        return 0;
3780}
3781
3782static void cgroup_file_release(struct kernfs_open_file *of)
3783{
3784        struct cftype *cft = of_cft(of);
3785
3786        if (cft->release)
3787                cft->release(of);
3788}
3789
3790static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3791                                 size_t nbytes, loff_t off)
3792{
3793        struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3794        struct cgroup *cgrp = of->kn->parent->priv;
3795        struct cftype *cft = of_cft(of);
3796        struct cgroup_subsys_state *css;
3797        int ret;
3798
3799        if (!nbytes)
3800                return 0;
3801
3802        /*
3803         * If namespaces are delegation boundaries, disallow writes to
3804         * files in an non-init namespace root from inside the namespace
3805         * except for the files explicitly marked delegatable -
3806         * cgroup.procs and cgroup.subtree_control.
3807         */
3808        if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3809            !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3810            ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3811                return -EPERM;
3812
3813        if (cft->write)
3814                return cft->write(of, buf, nbytes, off);
3815
3816        /*
3817         * kernfs guarantees that a file isn't deleted with operations in
3818         * flight, which means that the matching css is and stays alive and
3819         * doesn't need to be pinned.  The RCU locking is not necessary
3820         * either.  It's just for the convenience of using cgroup_css().
3821         */
3822        rcu_read_lock();
3823        css = cgroup_css(cgrp, cft->ss);
3824        rcu_read_unlock();
3825
3826        if (cft->write_u64) {
3827                unsigned long long v;
3828                ret = kstrtoull(buf, 0, &v);
3829                if (!ret)
3830                        ret = cft->write_u64(css, cft, v);
3831        } else if (cft->write_s64) {
3832                long long v;
3833                ret = kstrtoll(buf, 0, &v);
3834                if (!ret)
3835                        ret = cft->write_s64(css, cft, v);
3836        } else {
3837                ret = -EINVAL;
3838        }
3839
3840        return ret ?: nbytes;
3841}
3842
3843static __poll_t cgroup_file_poll(struct kernfs_open_file *of, poll_table *pt)
3844{
3845        struct cftype *cft = of_cft(of);
3846
3847        if (cft->poll)
3848                return cft->poll(of, pt);
3849
3850        return kernfs_generic_poll(of, pt);
3851}
3852
3853static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3854{
3855        return seq_cft(seq)->seq_start(seq, ppos);
3856}
3857
3858static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3859{
3860        return seq_cft(seq)->seq_next(seq, v, ppos);
3861}
3862
3863static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3864{
3865        if (seq_cft(seq)->seq_stop)
3866                seq_cft(seq)->seq_stop(seq, v);
3867}
3868
3869static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3870{
3871        struct cftype *cft = seq_cft(m);
3872        struct cgroup_subsys_state *css = seq_css(m);
3873
3874        if (cft->seq_show)
3875                return cft->seq_show(m, arg);
3876
3877        if (cft->read_u64)
3878                seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3879        else if (cft->read_s64)
3880                seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3881        else
3882                return -EINVAL;
3883        return 0;
3884}
3885
3886static struct kernfs_ops cgroup_kf_single_ops = {
3887        .atomic_write_len       = PAGE_SIZE,
3888        .open                   = cgroup_file_open,
3889        .release                = cgroup_file_release,
3890        .write                  = cgroup_file_write,
3891        .poll                   = cgroup_file_poll,
3892        .seq_show               = cgroup_seqfile_show,
3893};
3894
3895static struct kernfs_ops cgroup_kf_ops = {
3896        .atomic_write_len       = PAGE_SIZE,
3897        .open                   = cgroup_file_open,
3898        .release                = cgroup_file_release,
3899        .write                  = cgroup_file_write,
3900        .poll                   = cgroup_file_poll,
3901        .seq_start              = cgroup_seqfile_start,
3902        .seq_next               = cgroup_seqfile_next,
3903        .seq_stop               = cgroup_seqfile_stop,
3904        .seq_show               = cgroup_seqfile_show,
3905};
3906
3907/* set uid and gid of cgroup dirs and files to that of the creator */
3908static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3909{
3910        struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3911                               .ia_uid = current_fsuid(),
3912                               .ia_gid = current_fsgid(), };
3913
3914        if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3915            gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3916                return 0;
3917
3918        return kernfs_setattr(kn, &iattr);
3919}
3920
3921static void cgroup_file_notify_timer(struct timer_list *timer)
3922{
3923        cgroup_file_notify(container_of(timer, struct cgroup_file,
3924                                        notify_timer));
3925}
3926
3927static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3928                           struct cftype *cft)
3929{
3930        char name[CGROUP_FILE_NAME_MAX];
3931        struct kernfs_node *kn;
3932        struct lock_class_key *key = NULL;
3933        int ret;
3934
3935#ifdef CONFIG_DEBUG_LOCK_ALLOC
3936        key = &cft->lockdep_key;
3937#endif
3938        kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3939                                  cgroup_file_mode(cft),
3940                                  GLOBAL_ROOT_UID, GLOBAL_ROOT_GID,
3941                                  0, cft->kf_ops, cft,
3942                                  NULL, key);
3943        if (IS_ERR(kn))
3944                return PTR_ERR(kn);
3945
3946        ret = cgroup_kn_set_ugid(kn);
3947        if (ret) {
3948                kernfs_remove(kn);
3949                return ret;
3950        }
3951
3952        if (cft->file_offset) {
3953                struct cgroup_file *cfile = (void *)css + cft->file_offset;
3954
3955                timer_setup(&cfile->notify_timer, cgroup_file_notify_timer, 0);
3956
3957                spin_lock_irq(&cgroup_file_kn_lock);
3958                cfile->kn = kn;
3959                spin_unlock_irq(&cgroup_file_kn_lock);
3960        }
3961
3962        return 0;
3963}
3964
3965/**
3966 * cgroup_addrm_files - add or remove files to a cgroup directory
3967 * @css: the target css
3968 * @cgrp: the target cgroup (usually css->cgroup)
3969 * @cfts: array of cftypes to be added
3970 * @is_add: whether to add or remove
3971 *
3972 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3973 * For removals, this function never fails.
3974 */
3975static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3976                              struct cgroup *cgrp, struct cftype cfts[],
3977                              bool is_add)
3978{
3979        struct cftype *cft, *cft_end = NULL;
3980        int ret = 0;
3981
3982        lockdep_assert_held(&cgroup_mutex);
3983
3984restart:
3985        for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3986                /* does cft->flags tell us to skip this file on @cgrp? */
3987                if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
3988                        continue;
3989                if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3990                        continue;
3991                if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3992                        continue;
3993                if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3994                        continue;
3995                if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3996                        continue;
3997                if ((cft->flags & CFTYPE_DEBUG) && !cgroup_debug)
3998                        continue;
3999                if (is_add) {
4000                        ret = cgroup_add_file(css, cgrp, cft);
4001                        if (ret) {
4002                                pr_warn("%s: failed to add %s, err=%d\n",
4003                                        __func__, cft->name, ret);
4004                                cft_end = cft;
4005                                is_add = false;
4006                                goto restart;
4007                        }
4008                } else {
4009                        cgroup_rm_file(cgrp, cft);
4010                }
4011        }
4012        return ret;
4013}
4014
4015static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
4016{
4017        struct cgroup_subsys *ss = cfts[0].ss;
4018        struct cgroup *root = &ss->root->cgrp;
4019        struct cgroup_subsys_state *css;
4020        int ret = 0;
4021
4022        lockdep_assert_held(&cgroup_mutex);
4023
4024        /* add/rm files for all cgroups created before */
4025        css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
4026                struct cgroup *cgrp = css->cgroup;
4027
4028                if (!(css->flags & CSS_VISIBLE))
4029                        continue;
4030
4031                ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
4032                if (ret)
4033                        break;
4034        }
4035
4036        if (is_add && !ret)
4037                kernfs_activate(root->kn);
4038        return ret;
4039}
4040
4041static void cgroup_exit_cftypes(struct cftype *cfts)
4042{
4043        struct cftype *cft;
4044
4045        for (cft = cfts; cft->name[0] != '\0'; cft++) {
4046                /* free copy for custom atomic_write_len, see init_cftypes() */
4047                if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
4048                        kfree(cft->kf_ops);
4049                cft->kf_ops = NULL;
4050                cft->ss = NULL;
4051
4052                /* revert flags set by cgroup core while adding @cfts */
4053                cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
4054        }
4055}
4056
4057static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4058{
4059        struct cftype *cft;
4060
4061        for (cft = cfts; cft->name[0] != '\0'; cft++) {
4062                struct kernfs_ops *kf_ops;
4063
4064                WARN_ON(cft->ss || cft->kf_ops);
4065
4066                if ((cft->flags & CFTYPE_PRESSURE) && !cgroup_psi_enabled())
4067                        continue;
4068
4069                if (cft->seq_start)
4070                        kf_ops = &cgroup_kf_ops;
4071                else
4072                        kf_ops = &cgroup_kf_single_ops;
4073
4074                /*
4075                 * Ugh... if @cft wants a custom max_write_len, we need to
4076                 * make a copy of kf_ops to set its atomic_write_len.
4077                 */
4078                if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
4079                        kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
4080                        if (!kf_ops) {
4081                                cgroup_exit_cftypes(cfts);
4082                                return -ENOMEM;
4083                        }
4084                        kf_ops->atomic_write_len = cft->max_write_len;
4085                }
4086
4087                cft->kf_ops = kf_ops;
4088                cft->ss = ss;
4089        }
4090
4091        return 0;
4092}
4093
4094static int cgroup_rm_cftypes_locked(struct cftype *cfts)
4095{
4096        lockdep_assert_held(&cgroup_mutex);
4097
4098        if (!cfts || !cfts[0].ss)
4099                return -ENOENT;
4100
4101        list_del(&cfts->node);
4102        cgroup_apply_cftypes(cfts, false);
4103        cgroup_exit_cftypes(cfts);
4104        return 0;
4105}
4106
4107/**
4108 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
4109 * @cfts: zero-length name terminated array of cftypes
4110 *
4111 * Unregister @cfts.  Files described by @cfts are removed from all
4112 * existing cgroups and all future cgroups won't have them either.  This
4113 * function can be called anytime whether @cfts' subsys is attached or not.
4114 *
4115 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
4116 * registered.
4117 */
4118int cgroup_rm_cftypes(struct cftype *cfts)
4119{
4120        int ret;
4121
4122        mutex_lock(&cgroup_mutex);
4123        ret = cgroup_rm_cftypes_locked(cfts);
4124        mutex_unlock(&cgroup_mutex);
4125        return ret;
4126}
4127
4128/**
4129 * cgroup_add_cftypes - add an array of cftypes to a subsystem
4130 * @ss: target cgroup subsystem
4131 * @cfts: zero-length name terminated array of cftypes
4132 *
4133 * Register @cfts to @ss.  Files described by @cfts are created for all
4134 * existing cgroups to which @ss is attached and all future cgroups will
4135 * have them too.  This function can be called anytime whether @ss is
4136 * attached or not.
4137 *
4138 * Returns 0 on successful registration, -errno on failure.  Note that this
4139 * function currently returns 0 as long as @cfts registration is successful
4140 * even if some file creation attempts on existing cgroups fail.
4141 */
4142static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4143{
4144        int ret;
4145
4146        if (!cgroup_ssid_enabled(ss->id))
4147                return 0;
4148
4149        if (!cfts || cfts[0].name[0] == '\0')
4150                return 0;
4151
4152        ret = cgroup_init_cftypes(ss, cfts);
4153        if (ret)
4154                return ret;
4155
4156        mutex_lock(&cgroup_mutex);
4157
4158        list_add_tail(&cfts->node, &ss->cfts);
4159        ret = cgroup_apply_cftypes(cfts, true);
4160        if (ret)
4161                cgroup_rm_cftypes_locked(cfts);
4162
4163        mutex_unlock(&cgroup_mutex);
4164        return ret;
4165}
4166
4167/**
4168 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
4169 * @ss: target cgroup subsystem
4170 * @cfts: zero-length name terminated array of cftypes
4171 *
4172 * Similar to cgroup_add_cftypes() but the added files are only used for
4173 * the default hierarchy.
4174 */
4175int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4176{
4177        struct cftype *cft;
4178
4179        for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4180                cft->flags |= __CFTYPE_ONLY_ON_DFL;
4181        return cgroup_add_cftypes(ss, cfts);
4182}
4183
4184/**
4185 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
4186 * @ss: target cgroup subsystem
4187 * @cfts: zero-length name terminated array of cftypes
4188 *
4189 * Similar to cgroup_add_cftypes() but the added files are only used for
4190 * the legacy hierarchies.
4191 */
4192int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
4193{
4194        struct cftype *cft;
4195
4196        for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
4197                cft->flags |= __CFTYPE_NOT_ON_DFL;
4198        return cgroup_add_cftypes(ss, cfts);
4199}
4200
4201/**
4202 * cgroup_file_notify - generate a file modified event for a cgroup_file
4203 * @cfile: target cgroup_file
4204 *
4205 * @cfile must have been obtained by setting cftype->file_offset.
4206 */
4207void cgroup_file_notify(struct cgroup_file *cfile)
4208{
4209        unsigned long flags;
4210
4211        spin_lock_irqsave(&cgroup_file_kn_lock, flags);
4212        if (cfile->kn) {
4213                unsigned long last = cfile->notified_at;
4214                unsigned long next = last + CGROUP_FILE_NOTIFY_MIN_INTV;
4215
4216                if (time_in_range(jiffies, last, next)) {
4217                        timer_reduce(&cfile->notify_timer, next);
4218                } else {
4219                        kernfs_notify(cfile->kn);
4220                        cfile->notified_at = jiffies;
4221                }
4222        }
4223        spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
4224}
4225
4226/**
4227 * css_next_child - find the next child of a given css
4228 * @pos: the current position (%NULL to initiate traversal)
4229 * @parent: css whose children to walk
4230 *
4231 * This function returns the next child of @parent and should be called
4232 * under either cgroup_mutex or RCU read lock.  The only requirement is
4233 * that @parent and @pos are accessible.  The next sibling is guaranteed to
4234 * be returned regardless of their states.
4235 *
4236 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4237 * css which finished ->css_online() is guaranteed to be visible in the
4238 * future iterations and will stay visible until the last reference is put.
4239 * A css which hasn't finished ->css_online() or already finished
4240 * ->css_offline() may show up during traversal.  It's each subsystem's
4241 * responsibility to synchronize against on/offlining.
4242 */
4243struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
4244                                           struct cgroup_subsys_state *parent)
4245{
4246        struct cgroup_subsys_state *next;
4247
4248        cgroup_assert_mutex_or_rcu_locked();
4249
4250        /*
4251         * @pos could already have been unlinked from the sibling list.
4252         * Once a cgroup is removed, its ->sibling.next is no longer
4253         * updated when its next sibling changes.  CSS_RELEASED is set when
4254         * @pos is taken off list, at which time its next pointer is valid,
4255         * and, as releases are serialized, the one pointed to by the next
4256         * pointer is guaranteed to not have started release yet.  This
4257         * implies that if we observe !CSS_RELEASED on @pos in this RCU
4258         * critical section, the one pointed to by its next pointer is
4259         * guaranteed to not have finished its RCU grace period even if we
4260         * have dropped rcu_read_lock() in-between iterations.
4261         *
4262         * If @pos has CSS_RELEASED set, its next pointer can't be
4263         * dereferenced; however, as each css is given a monotonically
4264         * increasing unique serial number and always appended to the
4265         * sibling list, the next one can be found by walking the parent's
4266         * children until the first css with higher serial number than
4267         * @pos's.  While this path can be slower, it happens iff iteration
4268         * races against release and the race window is very small.
4269         */
4270        if (!pos) {
4271                next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
4272        } else if (likely(!(pos->flags & CSS_RELEASED))) {
4273                next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
4274        } else {
4275                list_for_each_entry_rcu(next, &parent->children, sibling,
4276                                        lockdep_is_held(&cgroup_mutex))
4277                        if (next->serial_nr > pos->serial_nr)
4278                                break;
4279        }
4280
4281        /*
4282         * @next, if not pointing to the head, can be dereferenced and is
4283         * the next sibling.
4284         */
4285        if (&next->sibling != &parent->children)
4286                return next;
4287        return NULL;
4288}
4289
4290/**
4291 * css_next_descendant_pre - find the next descendant for pre-order walk
4292 * @pos: the current position (%NULL to initiate traversal)
4293 * @root: css whose descendants to walk
4294 *
4295 * To be used by css_for_each_descendant_pre().  Find the next descendant
4296 * to visit for pre-order traversal of @root's descendants.  @root is
4297 * included in the iteration and the first node to be visited.
4298 *
4299 * While this function requires cgroup_mutex or RCU read locking, it
4300 * doesn't require the whole traversal to be contained in a single critical
4301 * section.  This function will return the correct next descendant as long
4302 * as both @pos and @root are accessible and @pos is a descendant of @root.
4303 *
4304 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4305 * css which finished ->css_online() is guaranteed to be visible in the
4306 * future iterations and will stay visible until the last reference is put.
4307 * A css which hasn't finished ->css_online() or already finished
4308 * ->css_offline() may show up during traversal.  It's each subsystem's
4309 * responsibility to synchronize against on/offlining.
4310 */
4311struct cgroup_subsys_state *
4312css_next_descendant_pre(struct cgroup_subsys_state *pos,
4313                        struct cgroup_subsys_state *root)
4314{
4315        struct cgroup_subsys_state *next;
4316
4317        cgroup_assert_mutex_or_rcu_locked();
4318
4319        /* if first iteration, visit @root */
4320        if (!pos)
4321                return root;
4322
4323        /* visit the first child if exists */
4324        next = css_next_child(NULL, pos);
4325        if (next)
4326                return next;
4327
4328        /* no child, visit my or the closest ancestor's next sibling */
4329        while (pos != root) {
4330                next = css_next_child(pos, pos->parent);
4331                if (next)
4332                        return next;
4333                pos = pos->parent;
4334        }
4335
4336        return NULL;
4337}
4338EXPORT_SYMBOL_GPL(css_next_descendant_pre);
4339
4340/**
4341 * css_rightmost_descendant - return the rightmost descendant of a css
4342 * @pos: css of interest
4343 *
4344 * Return the rightmost descendant of @pos.  If there's no descendant, @pos
4345 * is returned.  This can be used during pre-order traversal to skip
4346 * subtree of @pos.
4347 *
4348 * While this function requires cgroup_mutex or RCU read locking, it
4349 * doesn't require the whole traversal to be contained in a single critical
4350 * section.  This function will return the correct rightmost descendant as
4351 * long as @pos is accessible.
4352 */
4353struct cgroup_subsys_state *
4354css_rightmost_descendant(struct cgroup_subsys_state *pos)
4355{
4356        struct cgroup_subsys_state *last, *tmp;
4357
4358        cgroup_assert_mutex_or_rcu_locked();
4359
4360        do {
4361                last = pos;
4362                /* ->prev isn't RCU safe, walk ->next till the end */
4363                pos = NULL;
4364                css_for_each_child(tmp, last)
4365                        pos = tmp;
4366        } while (pos);
4367
4368        return last;
4369}
4370
4371static struct cgroup_subsys_state *
4372css_leftmost_descendant(struct cgroup_subsys_state *pos)
4373{
4374        struct cgroup_subsys_state *last;
4375
4376        do {
4377                last = pos;
4378                pos = css_next_child(NULL, pos);
4379        } while (pos);
4380
4381        return last;
4382}
4383
4384/**
4385 * css_next_descendant_post - find the next descendant for post-order walk
4386 * @pos: the current position (%NULL to initiate traversal)
4387 * @root: css whose descendants to walk
4388 *
4389 * To be used by css_for_each_descendant_post().  Find the next descendant
4390 * to visit for post-order traversal of @root's descendants.  @root is
4391 * included in the iteration and the last node to be visited.
4392 *
4393 * While this function requires cgroup_mutex or RCU read locking, it
4394 * doesn't require the whole traversal to be contained in a single critical
4395 * section.  This function will return the correct next descendant as long
4396 * as both @pos and @cgroup are accessible and @pos is a descendant of
4397 * @cgroup.
4398 *
4399 * If a subsystem synchronizes ->css_online() and the start of iteration, a
4400 * css which finished ->css_online() is guaranteed to be visible in the
4401 * future iterations and will stay visible until the last reference is put.
4402 * A css which hasn't finished ->css_online() or already finished
4403 * ->css_offline() may show up during traversal.  It's each subsystem's
4404 * responsibility to synchronize against on/offlining.
4405 */
4406struct cgroup_subsys_state *
4407css_next_descendant_post(struct cgroup_subsys_state *pos,
4408                         struct cgroup_subsys_state *root)
4409{
4410        struct cgroup_subsys_state *next;
4411
4412        cgroup_assert_mutex_or_rcu_locked();
4413
4414        /* if first iteration, visit leftmost descendant which may be @root */
4415        if (!pos)
4416                return css_leftmost_descendant(root);
4417
4418        /* if we visited @root, we're done */
4419        if (pos == root)
4420                return NULL;
4421
4422        /* if there's an unvisited sibling, visit its leftmost descendant */
4423        next = css_next_child(pos, pos->parent);
4424        if (next)
4425                return css_leftmost_descendant(next);
4426
4427        /* no sibling left, visit parent */
4428        return pos->parent;
4429}
4430
4431/**
4432 * css_has_online_children - does a css have online children
4433 * @css: the target css
4434 *
4435 * Returns %true if @css has any online children; otherwise, %false.  This
4436 * function can be called from any context but the caller is responsible
4437 * for synchronizing against on/offlining as necessary.
4438 */
4439bool css_has_online_children(struct cgroup_subsys_state *css)
4440{
4441        struct cgroup_subsys_state *child;
4442        bool ret = false;
4443
4444        rcu_read_lock();
4445        css_for_each_child(child, css) {
4446                if (child->flags & CSS_ONLINE) {
4447                        ret = true;
4448                        break;
4449                }
4450        }
4451        rcu_read_unlock();
4452        return ret;
4453}
4454
4455static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
4456{
4457        struct list_head *l;
4458        struct cgrp_cset_link *link;
4459        struct css_set *cset;
4460
4461        lockdep_assert_held(&css_set_lock);
4462
4463        /* find the next threaded cset */
4464        if (it->tcset_pos) {
4465                l = it->tcset_pos->next;
4466
4467                if (l != it->tcset_head) {
4468                        it->tcset_pos = l;
4469                        return container_of(l, struct css_set,
4470                                            threaded_csets_node);
4471                }
4472
4473                it->tcset_pos = NULL;
4474        }
4475
4476        /* find the next cset */
4477        l = it->cset_pos;
4478        l = l->next;
4479        if (l == it->cset_head) {
4480                it->cset_pos = NULL;
4481                return NULL;
4482        }
4483
4484        if (it->ss) {
4485                cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
4486        } else {
4487                link = list_entry(l, struct cgrp_cset_link, cset_link);
4488                cset = link->cset;
4489        }
4490
4491        it->cset_pos = l;
4492
4493        /* initialize threaded css_set walking */
4494        if (it->flags & CSS_TASK_ITER_THREADED) {
4495                if (it->cur_dcset)
4496                        put_css_set_locked(it->cur_dcset);
4497                it->cur_dcset = cset;
4498                get_css_set(cset);
4499
4500                it->tcset_head = &cset->threaded_csets;
4501                it->tcset_pos = &cset->threaded_csets;
4502        }
4503
4504        return cset;
4505}
4506
4507/**
4508 * css_task_iter_advance_css_set - advance a task iterator to the next css_set
4509 * @it: the iterator to advance
4510 *
4511 * Advance @it to the next css_set to walk.
4512 */
4513static void css_task_iter_advance_css_set(struct css_task_iter *it)
4514{
4515        struct css_set *cset;
4516
4517        lockdep_assert_held(&css_set_lock);
4518
4519        /* Advance to the next non-empty css_set and find first non-empty tasks list*/
4520        while ((cset = css_task_iter_next_css_set(it))) {
4521                if (!list_empty(&cset->tasks)) {
4522                        it->cur_tasks_head = &cset->tasks;
4523                        break;
4524                } else if (!list_empty(&cset->mg_tasks)) {
4525                        it->cur_tasks_head = &cset->mg_tasks;
4526                        break;
4527                } else if (!list_empty(&cset->dying_tasks)) {
4528                        it->cur_tasks_head = &cset->dying_tasks;
4529                        break;
4530                }
4531        }
4532        if (!cset) {
4533                it->task_pos = NULL;
4534                return;
4535        }
4536        it->task_pos = it->cur_tasks_head->next;
4537
4538        /*
4539         * We don't keep css_sets locked across iteration steps and thus
4540         * need to take steps to ensure that iteration can be resumed after
4541         * the lock is re-acquired.  Iteration is performed at two levels -
4542         * css_sets and tasks in them.
4543         *
4544         * Once created, a css_set never leaves its cgroup lists, so a
4545         * pinned css_set is guaranteed to stay put and we can resume
4546         * iteration afterwards.
4547         *
4548         * Tasks may leave @cset across iteration steps.  This is resolved
4549         * by registering each iterator with the css_set currently being
4550         * walked and making css_set_move_task() advance iterators whose
4551         * next task is leaving.
4552         */
4553        if (it->cur_cset) {
4554                list_del(&it->iters_node);
4555                put_css_set_locked(it->cur_cset);
4556        }
4557        get_css_set(cset);
4558        it->cur_cset = cset;
4559        list_add(&it->iters_node, &cset->task_iters);
4560}
4561
4562static void css_task_iter_skip(struct css_task_iter *it,
4563                               struct task_struct *task)
4564{
4565        lockdep_assert_held(&css_set_lock);
4566
4567        if (it->task_pos == &task->cg_list) {
4568                it->task_pos = it->task_pos->next;
4569                it->flags |= CSS_TASK_ITER_SKIPPED;
4570        }
4571}
4572
4573static void css_task_iter_advance(struct css_task_iter *it)
4574{
4575        struct task_struct *task;
4576
4577        lockdep_assert_held(&css_set_lock);
4578repeat:
4579        if (it->task_pos) {
4580                /*
4581                 * Advance iterator to find next entry. We go through cset
4582                 * tasks, mg_tasks and dying_tasks, when consumed we move onto
4583                 * the next cset.
4584                 */
4585                if (it->flags & CSS_TASK_ITER_SKIPPED)
4586                        it->flags &= ~CSS_TASK_ITER_SKIPPED;
4587                else
4588                        it->task_pos = it->task_pos->next;
4589
4590                if (it->task_pos == &it->cur_cset->tasks) {
4591                        it->cur_tasks_head = &it->cur_cset->mg_tasks;
4592                        it->task_pos = it->cur_tasks_head->next;
4593                }
4594                if (it->task_pos == &it->cur_cset->mg_tasks) {
4595                        it->cur_tasks_head = &it->cur_cset->dying_tasks;
4596                        it->task_pos = it->cur_tasks_head->next;
4597                }
4598                if (it->task_pos == &it->cur_cset->dying_tasks)
4599                        css_task_iter_advance_css_set(it);
4600        } else {
4601                /* called from start, proceed to the first cset */
4602                css_task_iter_advance_css_set(it);
4603        }
4604
4605        if (!it->task_pos)
4606                return;
4607
4608        task = list_entry(it->task_pos, struct task_struct, cg_list);
4609
4610        if (it->flags & CSS_TASK_ITER_PROCS) {
4611                /* if PROCS, skip over tasks which aren't group leaders */
4612                if (!thread_group_leader(task))
4613                        goto repeat;
4614
4615                /* and dying leaders w/o live member threads */
4616                if (it->cur_tasks_head == &it->cur_cset->dying_tasks &&
4617                    !atomic_read(&task->signal->live))
4618                        goto repeat;
4619        } else {
4620                /* skip all dying ones */
4621                if (it->cur_tasks_head == &it->cur_cset->dying_tasks)
4622                        goto repeat;
4623        }
4624}
4625
4626/**
4627 * css_task_iter_start - initiate task iteration
4628 * @css: the css to walk tasks of
4629 * @flags: CSS_TASK_ITER_* flags
4630 * @it: the task iterator to use
4631 *
4632 * Initiate iteration through the tasks of @css.  The caller can call
4633 * css_task_iter_next() to walk through the tasks until the function
4634 * returns NULL.  On completion of iteration, css_task_iter_end() must be
4635 * called.
4636 */
4637void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4638                         struct css_task_iter *it)
4639{
4640        memset(it, 0, sizeof(*it));
4641
4642        spin_lock_irq(&css_set_lock);
4643
4644        it->ss = css->ss;
4645        it->flags = flags;
4646
4647        if (it->ss)
4648                it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4649        else
4650                it->cset_pos = &css->cgroup->cset_links;
4651
4652        it->cset_head = it->cset_pos;
4653
4654        css_task_iter_advance(it);
4655
4656        spin_unlock_irq(&css_set_lock);
4657}
4658
4659/**
4660 * css_task_iter_next - return the next task for the iterator
4661 * @it: the task iterator being iterated
4662 *
4663 * The "next" function for task iteration.  @it should have been
4664 * initialized via css_task_iter_start().  Returns NULL when the iteration
4665 * reaches the end.
4666 */
4667struct task_struct *css_task_iter_next(struct css_task_iter *it)
4668{
4669        if (it->cur_task) {
4670                put_task_struct(it->cur_task);
4671                it->cur_task = NULL;
4672        }
4673
4674        spin_lock_irq(&css_set_lock);
4675
4676        /* @it may be half-advanced by skips, finish advancing */
4677        if (it->flags & CSS_TASK_ITER_SKIPPED)
4678                css_task_iter_advance(it);
4679
4680        if (it->task_pos) {
4681                it->cur_task = list_entry(it->task_pos, struct task_struct,
4682                                          cg_list);
4683                get_task_struct(it->cur_task);
4684                css_task_iter_advance(it);
4685        }
4686
4687        spin_unlock_irq(&css_set_lock);
4688
4689        return it->cur_task;
4690}
4691
4692/**
4693 * css_task_iter_end - finish task iteration
4694 * @it: the task iterator to finish
4695 *
4696 * Finish task iteration started by css_task_iter_start().
4697 */
4698void css_task_iter_end(struct css_task_iter *it)
4699{
4700        if (it->cur_cset) {
4701                spin_lock_irq(&css_set_lock);
4702                list_del(&it->iters_node);
4703                put_css_set_locked(it->cur_cset);
4704                spin_unlock_irq(&css_set_lock);
4705        }
4706
4707        if (it->cur_dcset)
4708                put_css_set(it->cur_dcset);
4709
4710        if (it->cur_task)
4711                put_task_struct(it->cur_task);
4712}
4713
4714static void cgroup_procs_release(struct kernfs_open_file *of)
4715{
4716        if (of->priv) {
4717                css_task_iter_end(of->priv);
4718                kfree(of->priv);
4719        }
4720}
4721
4722static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4723{
4724        struct kernfs_open_file *of = s->private;
4725        struct css_task_iter *it = of->priv;
4726
4727        if (pos)
4728                (*pos)++;
4729
4730        return css_task_iter_next(it);
4731}
4732
4733static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4734                                  unsigned int iter_flags)
4735{
4736        struct kernfs_open_file *of = s->private;
4737        struct cgroup *cgrp = seq_css(s)->cgroup;
4738        struct css_task_iter *it = of->priv;
4739
4740        /*
4741         * When a seq_file is seeked, it's always traversed sequentially
4742         * from position 0, so we can simply keep iterating on !0 *pos.
4743         */
4744        if (!it) {
4745                if (WARN_ON_ONCE((*pos)))
4746                        return ERR_PTR(-EINVAL);
4747
4748                it = kzalloc(sizeof(*it), GFP_KERNEL);
4749                if (!it)
4750                        return ERR_PTR(-ENOMEM);
4751                of->priv = it;
4752                css_task_iter_start(&cgrp->self, iter_flags, it);
4753        } else if (!(*pos)) {
4754                css_task_iter_end(it);
4755                css_task_iter_start(&cgrp->self, iter_flags, it);
4756        } else
4757                return it->cur_task;
4758
4759        return cgroup_procs_next(s, NULL, NULL);
4760}
4761
4762static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4763{
4764        struct cgroup *cgrp = seq_css(s)->cgroup;
4765
4766        /*
4767         * All processes of a threaded subtree belong to the domain cgroup
4768         * of the subtree.  Only threads can be distributed across the
4769         * subtree.  Reject reads on cgroup.procs in the subtree proper.
4770         * They're always empty anyway.
4771         */
4772        if (cgroup_is_threaded(cgrp))
4773                return ERR_PTR(-EOPNOTSUPP);
4774
4775        return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4776                                            CSS_TASK_ITER_THREADED);
4777}
4778