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 <net/sock.h>
  58
  59#define CREATE_TRACE_POINTS
  60#include <trace/events/cgroup.h>
  61
  62#define CGROUP_FILE_NAME_MAX            (MAX_CGROUP_TYPE_NAMELEN +      \
  63                                         MAX_CFTYPE_NAME + 2)
  64
  65/*
  66 * cgroup_mutex is the master lock.  Any modification to cgroup or its
  67 * hierarchy must be performed while holding it.
  68 *
  69 * css_set_lock protects task->cgroups pointer, the list of css_set
  70 * objects, and the chain of tasks off each css_set.
  71 *
  72 * These locks are exported if CONFIG_PROVE_RCU so that accessors in
  73 * cgroup.h can use them for lockdep annotations.
  74 */
  75DEFINE_MUTEX(cgroup_mutex);
  76DEFINE_SPINLOCK(css_set_lock);
  77
  78#ifdef CONFIG_PROVE_RCU
  79EXPORT_SYMBOL_GPL(cgroup_mutex);
  80EXPORT_SYMBOL_GPL(css_set_lock);
  81#endif
  82
  83/*
  84 * Protects cgroup_idr and css_idr so that IDs can be released without
  85 * grabbing cgroup_mutex.
  86 */
  87static DEFINE_SPINLOCK(cgroup_idr_lock);
  88
  89/*
  90 * Protects cgroup_file->kn for !self csses.  It synchronizes notifications
  91 * against file removal/re-creation across css hiding.
  92 */
  93static DEFINE_SPINLOCK(cgroup_file_kn_lock);
  94
  95struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
  96
  97#define cgroup_assert_mutex_or_rcu_locked()                             \
  98        RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&                       \
  99                           !lockdep_is_held(&cgroup_mutex),             \
 100                           "cgroup_mutex or RCU read lock required");
 101
 102/*
 103 * cgroup destruction makes heavy use of work items and there can be a lot
 104 * of concurrent destructions.  Use a separate workqueue so that cgroup
 105 * destruction work items don't end up filling up max_active of system_wq
 106 * which may lead to deadlock.
 107 */
 108static struct workqueue_struct *cgroup_destroy_wq;
 109
 110/* generate an array of cgroup subsystem pointers */
 111#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys,
 112struct cgroup_subsys *cgroup_subsys[] = {
 113#include <linux/cgroup_subsys.h>
 114};
 115#undef SUBSYS
 116
 117/* array of cgroup subsystem names */
 118#define SUBSYS(_x) [_x ## _cgrp_id] = #_x,
 119static const char *cgroup_subsys_name[] = {
 120#include <linux/cgroup_subsys.h>
 121};
 122#undef SUBSYS
 123
 124/* array of static_keys for cgroup_subsys_enabled() and cgroup_subsys_on_dfl() */
 125#define SUBSYS(_x)                                                              \
 126        DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_enabled_key);                 \
 127        DEFINE_STATIC_KEY_TRUE(_x ## _cgrp_subsys_on_dfl_key);                  \
 128        EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_enabled_key);                      \
 129        EXPORT_SYMBOL_GPL(_x ## _cgrp_subsys_on_dfl_key);
 130#include <linux/cgroup_subsys.h>
 131#undef SUBSYS
 132
 133#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_enabled_key,
 134static struct static_key_true *cgroup_subsys_enabled_key[] = {
 135#include <linux/cgroup_subsys.h>
 136};
 137#undef SUBSYS
 138
 139#define SUBSYS(_x) [_x ## _cgrp_id] = &_x ## _cgrp_subsys_on_dfl_key,
 140static struct static_key_true *cgroup_subsys_on_dfl_key[] = {
 141#include <linux/cgroup_subsys.h>
 142};
 143#undef SUBSYS
 144
 145/*
 146 * The default hierarchy, reserved for the subsystems that are otherwise
 147 * unattached - it never has more than a single cgroup, and all tasks are
 148 * part of that cgroup.
 149 */
 150struct cgroup_root cgrp_dfl_root;
 151EXPORT_SYMBOL_GPL(cgrp_dfl_root);
 152
 153/*
 154 * The default hierarchy always exists but is hidden until mounted for the
 155 * first time.  This is for backward compatibility.
 156 */
 157static bool cgrp_dfl_visible;
 158
 159/* some controllers are not supported in the default hierarchy */
 160static u16 cgrp_dfl_inhibit_ss_mask;
 161
 162/* some controllers are implicitly enabled on the default hierarchy */
 163static u16 cgrp_dfl_implicit_ss_mask;
 164
 165/* some controllers can be threaded on the default hierarchy */
 166static u16 cgrp_dfl_threaded_ss_mask;
 167
 168/* The list of hierarchy roots */
 169LIST_HEAD(cgroup_roots);
 170static int cgroup_root_count;
 171
 172/* hierarchy ID allocation and mapping, protected by cgroup_mutex */
 173static DEFINE_IDR(cgroup_hierarchy_idr);
 174
 175/*
 176 * Assign a monotonically increasing serial number to csses.  It guarantees
 177 * cgroups with bigger numbers are newer than those with smaller numbers.
 178 * Also, as csses are always appended to the parent's ->children list, it
 179 * guarantees that sibling csses are always sorted in the ascending serial
 180 * number order on the list.  Protected by cgroup_mutex.
 181 */
 182static u64 css_serial_nr_next = 1;
 183
 184/*
 185 * These bitmasks identify subsystems with specific features to avoid
 186 * having to do iterative checks repeatedly.
 187 */
 188static u16 have_fork_callback __read_mostly;
 189static u16 have_exit_callback __read_mostly;
 190static u16 have_free_callback __read_mostly;
 191static u16 have_canfork_callback __read_mostly;
 192
 193/* cgroup namespace for init task */
 194struct cgroup_namespace init_cgroup_ns = {
 195        .count          = REFCOUNT_INIT(2),
 196        .user_ns        = &init_user_ns,
 197        .ns.ops         = &cgroupns_operations,
 198        .ns.inum        = PROC_CGROUP_INIT_INO,
 199        .root_cset      = &init_css_set,
 200};
 201
 202static struct file_system_type cgroup2_fs_type;
 203static struct cftype cgroup_base_files[];
 204
 205static int cgroup_apply_control(struct cgroup *cgrp);
 206static void cgroup_finalize_control(struct cgroup *cgrp, int ret);
 207static void css_task_iter_advance(struct css_task_iter *it);
 208static int cgroup_destroy_locked(struct cgroup *cgrp);
 209static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
 210                                              struct cgroup_subsys *ss);
 211static void css_release(struct percpu_ref *ref);
 212static void kill_css(struct cgroup_subsys_state *css);
 213static int cgroup_addrm_files(struct cgroup_subsys_state *css,
 214                              struct cgroup *cgrp, struct cftype cfts[],
 215                              bool is_add);
 216
 217/**
 218 * cgroup_ssid_enabled - cgroup subsys enabled test by subsys ID
 219 * @ssid: subsys ID of interest
 220 *
 221 * cgroup_subsys_enabled() can only be used with literal subsys names which
 222 * is fine for individual subsystems but unsuitable for cgroup core.  This
 223 * is slower static_key_enabled() based test indexed by @ssid.
 224 */
 225bool cgroup_ssid_enabled(int ssid)
 226{
 227        if (CGROUP_SUBSYS_COUNT == 0)
 228                return false;
 229
 230        return static_key_enabled(cgroup_subsys_enabled_key[ssid]);
 231}
 232
 233/**
 234 * cgroup_on_dfl - test whether a cgroup is on the default hierarchy
 235 * @cgrp: the cgroup of interest
 236 *
 237 * The default hierarchy is the v2 interface of cgroup and this function
 238 * can be used to test whether a cgroup is on the default hierarchy for
 239 * cases where a subsystem should behave differnetly depending on the
 240 * interface version.
 241 *
 242 * The set of behaviors which change on the default hierarchy are still
 243 * being determined and the mount option is prefixed with __DEVEL__.
 244 *
 245 * List of changed behaviors:
 246 *
 247 * - Mount options "noprefix", "xattr", "clone_children", "release_agent"
 248 *   and "name" are disallowed.
 249 *
 250 * - When mounting an existing superblock, mount options should match.
 251 *
 252 * - Remount is disallowed.
 253 *
 254 * - rename(2) is disallowed.
 255 *
 256 * - "tasks" is removed.  Everything should be at process granularity.  Use
 257 *   "cgroup.procs" instead.
 258 *
 259 * - "cgroup.procs" is not sorted.  pids will be unique unless they got
 260 *   recycled inbetween reads.
 261 *
 262 * - "release_agent" and "notify_on_release" are removed.  Replacement
 263 *   notification mechanism will be implemented.
 264 *
 265 * - "cgroup.clone_children" is removed.
 266 *
 267 * - "cgroup.subtree_populated" is available.  Its value is 0 if the cgroup
 268 *   and its descendants contain no task; otherwise, 1.  The file also
 269 *   generates kernfs notification which can be monitored through poll and
 270 *   [di]notify when the value of the file changes.
 271 *
 272 * - cpuset: tasks will be kept in empty cpusets when hotplug happens and
 273 *   take masks of ancestors with non-empty cpus/mems, instead of being
 274 *   moved to an ancestor.
 275 *
 276 * - cpuset: a task can be moved into an empty cpuset, and again it takes
 277 *   masks of ancestors.
 278 *
 279 * - memcg: use_hierarchy is on by default and the cgroup file for the flag
 280 *   is not created.
 281 *
 282 * - blkcg: blk-throttle becomes properly hierarchical.
 283 *
 284 * - debug: disallowed on the default hierarchy.
 285 */
 286bool cgroup_on_dfl(const struct cgroup *cgrp)
 287{
 288        return cgrp->root == &cgrp_dfl_root;
 289}
 290
 291/* IDR wrappers which synchronize using cgroup_idr_lock */
 292static int cgroup_idr_alloc(struct idr *idr, void *ptr, int start, int end,
 293                            gfp_t gfp_mask)
 294{
 295        int ret;
 296
 297        idr_preload(gfp_mask);
 298        spin_lock_bh(&cgroup_idr_lock);
 299        ret = idr_alloc(idr, ptr, start, end, gfp_mask & ~__GFP_DIRECT_RECLAIM);
 300        spin_unlock_bh(&cgroup_idr_lock);
 301        idr_preload_end();
 302        return ret;
 303}
 304
 305static void *cgroup_idr_replace(struct idr *idr, void *ptr, int id)
 306{
 307        void *ret;
 308
 309        spin_lock_bh(&cgroup_idr_lock);
 310        ret = idr_replace(idr, ptr, id);
 311        spin_unlock_bh(&cgroup_idr_lock);
 312        return ret;
 313}
 314
 315static void cgroup_idr_remove(struct idr *idr, int id)
 316{
 317        spin_lock_bh(&cgroup_idr_lock);
 318        idr_remove(idr, id);
 319        spin_unlock_bh(&cgroup_idr_lock);
 320}
 321
 322static bool cgroup_has_tasks(struct cgroup *cgrp)
 323{
 324        return cgrp->nr_populated_csets;
 325}
 326
 327bool cgroup_is_threaded(struct cgroup *cgrp)
 328{
 329        return cgrp->dom_cgrp != cgrp;
 330}
 331
 332/* can @cgrp host both domain and threaded children? */
 333static bool cgroup_is_mixable(struct cgroup *cgrp)
 334{
 335        /*
 336         * Root isn't under domain level resource control exempting it from
 337         * the no-internal-process constraint, so it can serve as a thread
 338         * root and a parent of resource domains at the same time.
 339         */
 340        return !cgroup_parent(cgrp);
 341}
 342
 343/* can @cgrp become a thread root? should always be true for a thread root */
 344static bool cgroup_can_be_thread_root(struct cgroup *cgrp)
 345{
 346        /* mixables don't care */
 347        if (cgroup_is_mixable(cgrp))
 348                return true;
 349
 350        /* domain roots can't be nested under threaded */
 351        if (cgroup_is_threaded(cgrp))
 352                return false;
 353
 354        /* can only have either domain or threaded children */
 355        if (cgrp->nr_populated_domain_children)
 356                return false;
 357
 358        /* and no domain controllers can be enabled */
 359        if (cgrp->subtree_control & ~cgrp_dfl_threaded_ss_mask)
 360                return false;
 361
 362        return true;
 363}
 364
 365/* is @cgrp root of a threaded subtree? */
 366bool cgroup_is_thread_root(struct cgroup *cgrp)
 367{
 368        /* thread root should be a domain */
 369        if (cgroup_is_threaded(cgrp))
 370                return false;
 371
 372        /* a domain w/ threaded children is a thread root */
 373        if (cgrp->nr_threaded_children)
 374                return true;
 375
 376        /*
 377         * A domain which has tasks and explicit threaded controllers
 378         * enabled is a thread root.
 379         */
 380        if (cgroup_has_tasks(cgrp) &&
 381            (cgrp->subtree_control & cgrp_dfl_threaded_ss_mask))
 382                return true;
 383
 384        return false;
 385}
 386
 387/* a domain which isn't connected to the root w/o brekage can't be used */
 388static bool cgroup_is_valid_domain(struct cgroup *cgrp)
 389{
 390        /* the cgroup itself can be a thread root */
 391        if (cgroup_is_threaded(cgrp))
 392                return false;
 393
 394        /* but the ancestors can't be unless mixable */
 395        while ((cgrp = cgroup_parent(cgrp))) {
 396                if (!cgroup_is_mixable(cgrp) && cgroup_is_thread_root(cgrp))
 397                        return false;
 398                if (cgroup_is_threaded(cgrp))
 399                        return false;
 400        }
 401
 402        return true;
 403}
 404
 405/* subsystems visibly enabled on a cgroup */
 406static u16 cgroup_control(struct cgroup *cgrp)
 407{
 408        struct cgroup *parent = cgroup_parent(cgrp);
 409        u16 root_ss_mask = cgrp->root->subsys_mask;
 410
 411        if (parent) {
 412                u16 ss_mask = parent->subtree_control;
 413
 414                /* threaded cgroups can only have threaded controllers */
 415                if (cgroup_is_threaded(cgrp))
 416                        ss_mask &= cgrp_dfl_threaded_ss_mask;
 417                return ss_mask;
 418        }
 419
 420        if (cgroup_on_dfl(cgrp))
 421                root_ss_mask &= ~(cgrp_dfl_inhibit_ss_mask |
 422                                  cgrp_dfl_implicit_ss_mask);
 423        return root_ss_mask;
 424}
 425
 426/* subsystems enabled on a cgroup */
 427static u16 cgroup_ss_mask(struct cgroup *cgrp)
 428{
 429        struct cgroup *parent = cgroup_parent(cgrp);
 430
 431        if (parent) {
 432                u16 ss_mask = parent->subtree_ss_mask;
 433
 434                /* threaded cgroups can only have threaded controllers */
 435                if (cgroup_is_threaded(cgrp))
 436                        ss_mask &= cgrp_dfl_threaded_ss_mask;
 437                return ss_mask;
 438        }
 439
 440        return cgrp->root->subsys_mask;
 441}
 442
 443/**
 444 * cgroup_css - obtain a cgroup's css for the specified subsystem
 445 * @cgrp: the cgroup of interest
 446 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
 447 *
 448 * Return @cgrp's css (cgroup_subsys_state) associated with @ss.  This
 449 * function must be called either under cgroup_mutex or rcu_read_lock() and
 450 * the caller is responsible for pinning the returned css if it wants to
 451 * keep accessing it outside the said locks.  This function may return
 452 * %NULL if @cgrp doesn't have @subsys_id enabled.
 453 */
 454static struct cgroup_subsys_state *cgroup_css(struct cgroup *cgrp,
 455                                              struct cgroup_subsys *ss)
 456{
 457        if (ss)
 458                return rcu_dereference_check(cgrp->subsys[ss->id],
 459                                        lockdep_is_held(&cgroup_mutex));
 460        else
 461                return &cgrp->self;
 462}
 463
 464/**
 465 * cgroup_e_css - obtain a cgroup's effective css for the specified subsystem
 466 * @cgrp: the cgroup of interest
 467 * @ss: the subsystem of interest (%NULL returns @cgrp->self)
 468 *
 469 * Similar to cgroup_css() but returns the effective css, which is defined
 470 * as the matching css of the nearest ancestor including self which has @ss
 471 * enabled.  If @ss is associated with the hierarchy @cgrp is on, this
 472 * function is guaranteed to return non-NULL css.
 473 */
 474static struct cgroup_subsys_state *cgroup_e_css(struct cgroup *cgrp,
 475                                                struct cgroup_subsys *ss)
 476{
 477        lockdep_assert_held(&cgroup_mutex);
 478
 479        if (!ss)
 480                return &cgrp->self;
 481
 482        /*
 483         * This function is used while updating css associations and thus
 484         * can't test the csses directly.  Test ss_mask.
 485         */
 486        while (!(cgroup_ss_mask(cgrp) & (1 << ss->id))) {
 487                cgrp = cgroup_parent(cgrp);
 488                if (!cgrp)
 489                        return NULL;
 490        }
 491
 492        return cgroup_css(cgrp, ss);
 493}
 494
 495/**
 496 * cgroup_get_e_css - get a cgroup's effective css for the specified subsystem
 497 * @cgrp: the cgroup of interest
 498 * @ss: the subsystem of interest
 499 *
 500 * Find and get the effective css of @cgrp for @ss.  The effective css is
 501 * defined as the matching css of the nearest ancestor including self which
 502 * has @ss enabled.  If @ss is not mounted on the hierarchy @cgrp is on,
 503 * the root css is returned, so this function always returns a valid css.
 504 * The returned css must be put using css_put().
 505 */
 506struct cgroup_subsys_state *cgroup_get_e_css(struct cgroup *cgrp,
 507                                             struct cgroup_subsys *ss)
 508{
 509        struct cgroup_subsys_state *css;
 510
 511        rcu_read_lock();
 512
 513        do {
 514                css = cgroup_css(cgrp, ss);
 515
 516                if (css && css_tryget_online(css))
 517                        goto out_unlock;
 518                cgrp = cgroup_parent(cgrp);
 519        } while (cgrp);
 520
 521        css = init_css_set.subsys[ss->id];
 522        css_get(css);
 523out_unlock:
 524        rcu_read_unlock();
 525        return css;
 526}
 527
 528static void cgroup_get_live(struct cgroup *cgrp)
 529{
 530        WARN_ON_ONCE(cgroup_is_dead(cgrp));
 531        css_get(&cgrp->self);
 532}
 533
 534struct cgroup_subsys_state *of_css(struct kernfs_open_file *of)
 535{
 536        struct cgroup *cgrp = of->kn->parent->priv;
 537        struct cftype *cft = of_cft(of);
 538
 539        /*
 540         * This is open and unprotected implementation of cgroup_css().
 541         * seq_css() is only called from a kernfs file operation which has
 542         * an active reference on the file.  Because all the subsystem
 543         * files are drained before a css is disassociated with a cgroup,
 544         * the matching css from the cgroup's subsys table is guaranteed to
 545         * be and stay valid until the enclosing operation is complete.
 546         */
 547        if (cft->ss)
 548                return rcu_dereference_raw(cgrp->subsys[cft->ss->id]);
 549        else
 550                return &cgrp->self;
 551}
 552EXPORT_SYMBOL_GPL(of_css);
 553
 554/**
 555 * for_each_css - iterate all css's of a cgroup
 556 * @css: the iteration cursor
 557 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
 558 * @cgrp: the target cgroup to iterate css's of
 559 *
 560 * Should be called under cgroup_[tree_]mutex.
 561 */
 562#define for_each_css(css, ssid, cgrp)                                   \
 563        for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
 564                if (!((css) = rcu_dereference_check(                    \
 565                                (cgrp)->subsys[(ssid)],                 \
 566                                lockdep_is_held(&cgroup_mutex)))) { }   \
 567                else
 568
 569/**
 570 * for_each_e_css - iterate all effective css's of a cgroup
 571 * @css: the iteration cursor
 572 * @ssid: the index of the subsystem, CGROUP_SUBSYS_COUNT after reaching the end
 573 * @cgrp: the target cgroup to iterate css's of
 574 *
 575 * Should be called under cgroup_[tree_]mutex.
 576 */
 577#define for_each_e_css(css, ssid, cgrp)                                 \
 578        for ((ssid) = 0; (ssid) < CGROUP_SUBSYS_COUNT; (ssid)++)        \
 579                if (!((css) = cgroup_e_css(cgrp, cgroup_subsys[(ssid)]))) \
 580                        ;                                               \
 581                else
 582
 583/**
 584 * do_each_subsys_mask - filter for_each_subsys with a bitmask
 585 * @ss: the iteration cursor
 586 * @ssid: the index of @ss, CGROUP_SUBSYS_COUNT after reaching the end
 587 * @ss_mask: the bitmask
 588 *
 589 * The block will only run for cases where the ssid-th bit (1 << ssid) of
 590 * @ss_mask is set.
 591 */
 592#define do_each_subsys_mask(ss, ssid, ss_mask) do {                     \
 593        unsigned long __ss_mask = (ss_mask);                            \
 594        if (!CGROUP_SUBSYS_COUNT) { /* to avoid spurious gcc warning */ \
 595                (ssid) = 0;                                             \
 596                break;                                                  \
 597        }                                                               \
 598        for_each_set_bit(ssid, &__ss_mask, CGROUP_SUBSYS_COUNT) {       \
 599                (ss) = cgroup_subsys[ssid];                             \
 600                {
 601
 602#define while_each_subsys_mask()                                        \
 603                }                                                       \
 604        }                                                               \
 605} while (false)
 606
 607/* iterate over child cgrps, lock should be held throughout iteration */
 608#define cgroup_for_each_live_child(child, cgrp)                         \
 609        list_for_each_entry((child), &(cgrp)->self.children, self.sibling) \
 610                if (({ lockdep_assert_held(&cgroup_mutex);              \
 611                       cgroup_is_dead(child); }))                       \
 612                        ;                                               \
 613                else
 614
 615/* walk live descendants in preorder */
 616#define cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp)          \
 617        css_for_each_descendant_pre((d_css), cgroup_css((cgrp), NULL))  \
 618                if (({ lockdep_assert_held(&cgroup_mutex);              \
 619                       (dsct) = (d_css)->cgroup;                        \
 620                       cgroup_is_dead(dsct); }))                        \
 621                        ;                                               \
 622                else
 623
 624/* walk live descendants in postorder */
 625#define cgroup_for_each_live_descendant_post(dsct, d_css, cgrp)         \
 626        css_for_each_descendant_post((d_css), cgroup_css((cgrp), NULL)) \
 627                if (({ lockdep_assert_held(&cgroup_mutex);              \
 628                       (dsct) = (d_css)->cgroup;                        \
 629                       cgroup_is_dead(dsct); }))                        \
 630                        ;                                               \
 631                else
 632
 633/*
 634 * The default css_set - used by init and its children prior to any
 635 * hierarchies being mounted. It contains a pointer to the root state
 636 * for each subsystem. Also used to anchor the list of css_sets. Not
 637 * reference-counted, to improve performance when child cgroups
 638 * haven't been created.
 639 */
 640struct css_set init_css_set = {
 641        .refcount               = REFCOUNT_INIT(1),
 642        .dom_cset               = &init_css_set,
 643        .tasks                  = LIST_HEAD_INIT(init_css_set.tasks),
 644        .mg_tasks               = LIST_HEAD_INIT(init_css_set.mg_tasks),
 645        .task_iters             = LIST_HEAD_INIT(init_css_set.task_iters),
 646        .threaded_csets         = LIST_HEAD_INIT(init_css_set.threaded_csets),
 647        .cgrp_links             = LIST_HEAD_INIT(init_css_set.cgrp_links),
 648        .mg_preload_node        = LIST_HEAD_INIT(init_css_set.mg_preload_node),
 649        .mg_node                = LIST_HEAD_INIT(init_css_set.mg_node),
 650};
 651
 652static int css_set_count        = 1;    /* 1 for init_css_set */
 653
 654static bool css_set_threaded(struct css_set *cset)
 655{
 656        return cset->dom_cset != cset;
 657}
 658
 659/**
 660 * css_set_populated - does a css_set contain any tasks?
 661 * @cset: target css_set
 662 *
 663 * css_set_populated() should be the same as !!cset->nr_tasks at steady
 664 * state. However, css_set_populated() can be called while a task is being
 665 * added to or removed from the linked list before the nr_tasks is
 666 * properly updated. Hence, we can't just look at ->nr_tasks here.
 667 */
 668static bool css_set_populated(struct css_set *cset)
 669{
 670        lockdep_assert_held(&css_set_lock);
 671
 672        return !list_empty(&cset->tasks) || !list_empty(&cset->mg_tasks);
 673}
 674
 675/**
 676 * cgroup_update_populated - update the populated count of a cgroup
 677 * @cgrp: the target cgroup
 678 * @populated: inc or dec populated count
 679 *
 680 * One of the css_sets associated with @cgrp is either getting its first
 681 * task or losing the last.  Update @cgrp->nr_populated_* accordingly.  The
 682 * count is propagated towards root so that a given cgroup's
 683 * nr_populated_children is zero iff none of its descendants contain any
 684 * tasks.
 685 *
 686 * @cgrp's interface file "cgroup.populated" is zero if both
 687 * @cgrp->nr_populated_csets and @cgrp->nr_populated_children are zero and
 688 * 1 otherwise.  When the sum changes from or to zero, userland is notified
 689 * that the content of the interface file has changed.  This can be used to
 690 * detect when @cgrp and its descendants become populated or empty.
 691 */
 692static void cgroup_update_populated(struct cgroup *cgrp, bool populated)
 693{
 694        struct cgroup *child = NULL;
 695        int adj = populated ? 1 : -1;
 696
 697        lockdep_assert_held(&css_set_lock);
 698
 699        do {
 700                bool was_populated = cgroup_is_populated(cgrp);
 701
 702                if (!child) {
 703                        cgrp->nr_populated_csets += adj;
 704                } else {
 705                        if (cgroup_is_threaded(child))
 706                                cgrp->nr_populated_threaded_children += adj;
 707                        else
 708                                cgrp->nr_populated_domain_children += adj;
 709                }
 710
 711                if (was_populated == cgroup_is_populated(cgrp))
 712                        break;
 713
 714                cgroup1_check_for_release(cgrp);
 715                cgroup_file_notify(&cgrp->events_file);
 716
 717                child = cgrp;
 718                cgrp = cgroup_parent(cgrp);
 719        } while (cgrp);
 720}
 721
 722/**
 723 * css_set_update_populated - update populated state of a css_set
 724 * @cset: target css_set
 725 * @populated: whether @cset is populated or depopulated
 726 *
 727 * @cset is either getting the first task or losing the last.  Update the
 728 * populated counters of all associated cgroups accordingly.
 729 */
 730static void css_set_update_populated(struct css_set *cset, bool populated)
 731{
 732        struct cgrp_cset_link *link;
 733
 734        lockdep_assert_held(&css_set_lock);
 735
 736        list_for_each_entry(link, &cset->cgrp_links, cgrp_link)
 737                cgroup_update_populated(link->cgrp, populated);
 738}
 739
 740/**
 741 * css_set_move_task - move a task from one css_set to another
 742 * @task: task being moved
 743 * @from_cset: css_set @task currently belongs to (may be NULL)
 744 * @to_cset: new css_set @task is being moved to (may be NULL)
 745 * @use_mg_tasks: move to @to_cset->mg_tasks instead of ->tasks
 746 *
 747 * Move @task from @from_cset to @to_cset.  If @task didn't belong to any
 748 * css_set, @from_cset can be NULL.  If @task is being disassociated
 749 * instead of moved, @to_cset can be NULL.
 750 *
 751 * This function automatically handles populated counter updates and
 752 * css_task_iter adjustments but the caller is responsible for managing
 753 * @from_cset and @to_cset's reference counts.
 754 */
 755static void css_set_move_task(struct task_struct *task,
 756                              struct css_set *from_cset, struct css_set *to_cset,
 757                              bool use_mg_tasks)
 758{
 759        lockdep_assert_held(&css_set_lock);
 760
 761        if (to_cset && !css_set_populated(to_cset))
 762                css_set_update_populated(to_cset, true);
 763
 764        if (from_cset) {
 765                struct css_task_iter *it, *pos;
 766
 767                WARN_ON_ONCE(list_empty(&task->cg_list));
 768
 769                /*
 770                 * @task is leaving, advance task iterators which are
 771                 * pointing to it so that they can resume at the next
 772                 * position.  Advancing an iterator might remove it from
 773                 * the list, use safe walk.  See css_task_iter_advance*()
 774                 * for details.
 775                 */
 776                list_for_each_entry_safe(it, pos, &from_cset->task_iters,
 777                                         iters_node)
 778                        if (it->task_pos == &task->cg_list)
 779                                css_task_iter_advance(it);
 780
 781                list_del_init(&task->cg_list);
 782                if (!css_set_populated(from_cset))
 783                        css_set_update_populated(from_cset, false);
 784        } else {
 785                WARN_ON_ONCE(!list_empty(&task->cg_list));
 786        }
 787
 788        if (to_cset) {
 789                /*
 790                 * We are synchronized through cgroup_threadgroup_rwsem
 791                 * against PF_EXITING setting such that we can't race
 792                 * against cgroup_exit() changing the css_set to
 793                 * init_css_set and dropping the old one.
 794                 */
 795                WARN_ON_ONCE(task->flags & PF_EXITING);
 796
 797                rcu_assign_pointer(task->cgroups, to_cset);
 798                list_add_tail(&task->cg_list, use_mg_tasks ? &to_cset->mg_tasks :
 799                                                             &to_cset->tasks);
 800        }
 801}
 802
 803/*
 804 * hash table for cgroup groups. This improves the performance to find
 805 * an existing css_set. This hash doesn't (currently) take into
 806 * account cgroups in empty hierarchies.
 807 */
 808#define CSS_SET_HASH_BITS       7
 809static DEFINE_HASHTABLE(css_set_table, CSS_SET_HASH_BITS);
 810
 811static unsigned long css_set_hash(struct cgroup_subsys_state *css[])
 812{
 813        unsigned long key = 0UL;
 814        struct cgroup_subsys *ss;
 815        int i;
 816
 817        for_each_subsys(ss, i)
 818                key += (unsigned long)css[i];
 819        key = (key >> 16) ^ key;
 820
 821        return key;
 822}
 823
 824void put_css_set_locked(struct css_set *cset)
 825{
 826        struct cgrp_cset_link *link, *tmp_link;
 827        struct cgroup_subsys *ss;
 828        int ssid;
 829
 830        lockdep_assert_held(&css_set_lock);
 831
 832        if (!refcount_dec_and_test(&cset->refcount))
 833                return;
 834
 835        WARN_ON_ONCE(!list_empty(&cset->threaded_csets));
 836
 837        /* This css_set is dead. unlink it and release cgroup and css refs */
 838        for_each_subsys(ss, ssid) {
 839                list_del(&cset->e_cset_node[ssid]);
 840                css_put(cset->subsys[ssid]);
 841        }
 842        hash_del(&cset->hlist);
 843        css_set_count--;
 844
 845        list_for_each_entry_safe(link, tmp_link, &cset->cgrp_links, cgrp_link) {
 846                list_del(&link->cset_link);
 847                list_del(&link->cgrp_link);
 848                if (cgroup_parent(link->cgrp))
 849                        cgroup_put(link->cgrp);
 850                kfree(link);
 851        }
 852
 853        if (css_set_threaded(cset)) {
 854                list_del(&cset->threaded_csets_node);
 855                put_css_set_locked(cset->dom_cset);
 856        }
 857
 858        kfree_rcu(cset, rcu_head);
 859}
 860
 861/**
 862 * compare_css_sets - helper function for find_existing_css_set().
 863 * @cset: candidate css_set being tested
 864 * @old_cset: existing css_set for a task
 865 * @new_cgrp: cgroup that's being entered by the task
 866 * @template: desired set of css pointers in css_set (pre-calculated)
 867 *
 868 * Returns true if "cset" matches "old_cset" except for the hierarchy
 869 * which "new_cgrp" belongs to, for which it should match "new_cgrp".
 870 */
 871static bool compare_css_sets(struct css_set *cset,
 872                             struct css_set *old_cset,
 873                             struct cgroup *new_cgrp,
 874                             struct cgroup_subsys_state *template[])
 875{
 876        struct cgroup *new_dfl_cgrp;
 877        struct list_head *l1, *l2;
 878
 879        /*
 880         * On the default hierarchy, there can be csets which are
 881         * associated with the same set of cgroups but different csses.
 882         * Let's first ensure that csses match.
 883         */
 884        if (memcmp(template, cset->subsys, sizeof(cset->subsys)))
 885                return false;
 886
 887
 888        /* @cset's domain should match the default cgroup's */
 889        if (cgroup_on_dfl(new_cgrp))
 890                new_dfl_cgrp = new_cgrp;
 891        else
 892                new_dfl_cgrp = old_cset->dfl_cgrp;
 893
 894        if (new_dfl_cgrp->dom_cgrp != cset->dom_cset->dfl_cgrp)
 895                return false;
 896
 897        /*
 898         * Compare cgroup pointers in order to distinguish between
 899         * different cgroups in hierarchies.  As different cgroups may
 900         * share the same effective css, this comparison is always
 901         * necessary.
 902         */
 903        l1 = &cset->cgrp_links;
 904        l2 = &old_cset->cgrp_links;
 905        while (1) {
 906                struct cgrp_cset_link *link1, *link2;
 907                struct cgroup *cgrp1, *cgrp2;
 908
 909                l1 = l1->next;
 910                l2 = l2->next;
 911                /* See if we reached the end - both lists are equal length. */
 912                if (l1 == &cset->cgrp_links) {
 913                        BUG_ON(l2 != &old_cset->cgrp_links);
 914                        break;
 915                } else {
 916                        BUG_ON(l2 == &old_cset->cgrp_links);
 917                }
 918                /* Locate the cgroups associated with these links. */
 919                link1 = list_entry(l1, struct cgrp_cset_link, cgrp_link);
 920                link2 = list_entry(l2, struct cgrp_cset_link, cgrp_link);
 921                cgrp1 = link1->cgrp;
 922                cgrp2 = link2->cgrp;
 923                /* Hierarchies should be linked in the same order. */
 924                BUG_ON(cgrp1->root != cgrp2->root);
 925
 926                /*
 927                 * If this hierarchy is the hierarchy of the cgroup
 928                 * that's changing, then we need to check that this
 929                 * css_set points to the new cgroup; if it's any other
 930                 * hierarchy, then this css_set should point to the
 931                 * same cgroup as the old css_set.
 932                 */
 933                if (cgrp1->root == new_cgrp->root) {
 934                        if (cgrp1 != new_cgrp)
 935                                return false;
 936                } else {
 937                        if (cgrp1 != cgrp2)
 938                                return false;
 939                }
 940        }
 941        return true;
 942}
 943
 944/**
 945 * find_existing_css_set - init css array and find the matching css_set
 946 * @old_cset: the css_set that we're using before the cgroup transition
 947 * @cgrp: the cgroup that we're moving into
 948 * @template: out param for the new set of csses, should be clear on entry
 949 */
 950static struct css_set *find_existing_css_set(struct css_set *old_cset,
 951                                        struct cgroup *cgrp,
 952                                        struct cgroup_subsys_state *template[])
 953{
 954        struct cgroup_root *root = cgrp->root;
 955        struct cgroup_subsys *ss;
 956        struct css_set *cset;
 957        unsigned long key;
 958        int i;
 959
 960        /*
 961         * Build the set of subsystem state objects that we want to see in the
 962         * new css_set. while subsystems can change globally, the entries here
 963         * won't change, so no need for locking.
 964         */
 965        for_each_subsys(ss, i) {
 966                if (root->subsys_mask & (1UL << i)) {
 967                        /*
 968                         * @ss is in this hierarchy, so we want the
 969                         * effective css from @cgrp.
 970                         */
 971                        template[i] = cgroup_e_css(cgrp, ss);
 972                } else {
 973                        /*
 974                         * @ss is not in this hierarchy, so we don't want
 975                         * to change the css.
 976                         */
 977                        template[i] = old_cset->subsys[i];
 978                }
 979        }
 980
 981        key = css_set_hash(template);
 982        hash_for_each_possible(css_set_table, cset, hlist, key) {
 983                if (!compare_css_sets(cset, old_cset, cgrp, template))
 984                        continue;
 985
 986                /* This css_set matches what we need */
 987                return cset;
 988        }
 989
 990        /* No existing cgroup group matched */
 991        return NULL;
 992}
 993
 994static void free_cgrp_cset_links(struct list_head *links_to_free)
 995{
 996        struct cgrp_cset_link *link, *tmp_link;
 997
 998        list_for_each_entry_safe(link, tmp_link, links_to_free, cset_link) {
 999                list_del(&link->cset_link);
1000                kfree(link);
1001        }
1002}
1003
1004/**
1005 * allocate_cgrp_cset_links - allocate cgrp_cset_links
1006 * @count: the number of links to allocate
1007 * @tmp_links: list_head the allocated links are put on
1008 *
1009 * Allocate @count cgrp_cset_link structures and chain them on @tmp_links
1010 * through ->cset_link.  Returns 0 on success or -errno.
1011 */
1012static int allocate_cgrp_cset_links(int count, struct list_head *tmp_links)
1013{
1014        struct cgrp_cset_link *link;
1015        int i;
1016
1017        INIT_LIST_HEAD(tmp_links);
1018
1019        for (i = 0; i < count; i++) {
1020                link = kzalloc(sizeof(*link), GFP_KERNEL);
1021                if (!link) {
1022                        free_cgrp_cset_links(tmp_links);
1023                        return -ENOMEM;
1024                }
1025                list_add(&link->cset_link, tmp_links);
1026        }
1027        return 0;
1028}
1029
1030/**
1031 * link_css_set - a helper function to link a css_set to a cgroup
1032 * @tmp_links: cgrp_cset_link objects allocated by allocate_cgrp_cset_links()
1033 * @cset: the css_set to be linked
1034 * @cgrp: the destination cgroup
1035 */
1036static void link_css_set(struct list_head *tmp_links, struct css_set *cset,
1037                         struct cgroup *cgrp)
1038{
1039        struct cgrp_cset_link *link;
1040
1041        BUG_ON(list_empty(tmp_links));
1042
1043        if (cgroup_on_dfl(cgrp))
1044                cset->dfl_cgrp = cgrp;
1045
1046        link = list_first_entry(tmp_links, struct cgrp_cset_link, cset_link);
1047        link->cset = cset;
1048        link->cgrp = cgrp;
1049
1050        /*
1051         * Always add links to the tail of the lists so that the lists are
1052         * in choronological order.
1053         */
1054        list_move_tail(&link->cset_link, &cgrp->cset_links);
1055        list_add_tail(&link->cgrp_link, &cset->cgrp_links);
1056
1057        if (cgroup_parent(cgrp))
1058                cgroup_get_live(cgrp);
1059}
1060
1061/**
1062 * find_css_set - return a new css_set with one cgroup updated
1063 * @old_cset: the baseline css_set
1064 * @cgrp: the cgroup to be updated
1065 *
1066 * Return a new css_set that's equivalent to @old_cset, but with @cgrp
1067 * substituted into the appropriate hierarchy.
1068 */
1069static struct css_set *find_css_set(struct css_set *old_cset,
1070                                    struct cgroup *cgrp)
1071{
1072        struct cgroup_subsys_state *template[CGROUP_SUBSYS_COUNT] = { };
1073        struct css_set *cset;
1074        struct list_head tmp_links;
1075        struct cgrp_cset_link *link;
1076        struct cgroup_subsys *ss;
1077        unsigned long key;
1078        int ssid;
1079
1080        lockdep_assert_held(&cgroup_mutex);
1081
1082        /* First see if we already have a cgroup group that matches
1083         * the desired set */
1084        spin_lock_irq(&css_set_lock);
1085        cset = find_existing_css_set(old_cset, cgrp, template);
1086        if (cset)
1087                get_css_set(cset);
1088        spin_unlock_irq(&css_set_lock);
1089
1090        if (cset)
1091                return cset;
1092
1093        cset = kzalloc(sizeof(*cset), GFP_KERNEL);
1094        if (!cset)
1095                return NULL;
1096
1097        /* Allocate all the cgrp_cset_link objects that we'll need */
1098        if (allocate_cgrp_cset_links(cgroup_root_count, &tmp_links) < 0) {
1099                kfree(cset);
1100                return NULL;
1101        }
1102
1103        refcount_set(&cset->refcount, 1);
1104        cset->dom_cset = cset;
1105        INIT_LIST_HEAD(&cset->tasks);
1106        INIT_LIST_HEAD(&cset->mg_tasks);
1107        INIT_LIST_HEAD(&cset->task_iters);
1108        INIT_LIST_HEAD(&cset->threaded_csets);
1109        INIT_HLIST_NODE(&cset->hlist);
1110        INIT_LIST_HEAD(&cset->cgrp_links);
1111        INIT_LIST_HEAD(&cset->mg_preload_node);
1112        INIT_LIST_HEAD(&cset->mg_node);
1113
1114        /* Copy the set of subsystem state objects generated in
1115         * find_existing_css_set() */
1116        memcpy(cset->subsys, template, sizeof(cset->subsys));
1117
1118        spin_lock_irq(&css_set_lock);
1119        /* Add reference counts and links from the new css_set. */
1120        list_for_each_entry(link, &old_cset->cgrp_links, cgrp_link) {
1121                struct cgroup *c = link->cgrp;
1122
1123                if (c->root == cgrp->root)
1124                        c = cgrp;
1125                link_css_set(&tmp_links, cset, c);
1126        }
1127
1128        BUG_ON(!list_empty(&tmp_links));
1129
1130        css_set_count++;
1131
1132        /* Add @cset to the hash table */
1133        key = css_set_hash(cset->subsys);
1134        hash_add(css_set_table, &cset->hlist, key);
1135
1136        for_each_subsys(ss, ssid) {
1137                struct cgroup_subsys_state *css = cset->subsys[ssid];
1138
1139                list_add_tail(&cset->e_cset_node[ssid],
1140                              &css->cgroup->e_csets[ssid]);
1141                css_get(css);
1142        }
1143
1144        spin_unlock_irq(&css_set_lock);
1145
1146        /*
1147         * If @cset should be threaded, look up the matching dom_cset and
1148         * link them up.  We first fully initialize @cset then look for the
1149         * dom_cset.  It's simpler this way and safe as @cset is guaranteed
1150         * to stay empty until we return.
1151         */
1152        if (cgroup_is_threaded(cset->dfl_cgrp)) {
1153                struct css_set *dcset;
1154
1155                dcset = find_css_set(cset, cset->dfl_cgrp->dom_cgrp);
1156                if (!dcset) {
1157                        put_css_set(cset);
1158                        return NULL;
1159                }
1160
1161                spin_lock_irq(&css_set_lock);
1162                cset->dom_cset = dcset;
1163                list_add_tail(&cset->threaded_csets_node,
1164                              &dcset->threaded_csets);
1165                spin_unlock_irq(&css_set_lock);
1166        }
1167
1168        return cset;
1169}
1170
1171struct cgroup_root *cgroup_root_from_kf(struct kernfs_root *kf_root)
1172{
1173        struct cgroup *root_cgrp = kf_root->kn->priv;
1174
1175        return root_cgrp->root;
1176}
1177
1178static int cgroup_init_root_id(struct cgroup_root *root)
1179{
1180        int id;
1181
1182        lockdep_assert_held(&cgroup_mutex);
1183
1184        id = idr_alloc_cyclic(&cgroup_hierarchy_idr, root, 0, 0, GFP_KERNEL);
1185        if (id < 0)
1186                return id;
1187
1188        root->hierarchy_id = id;
1189        return 0;
1190}
1191
1192static void cgroup_exit_root_id(struct cgroup_root *root)
1193{
1194        lockdep_assert_held(&cgroup_mutex);
1195
1196        idr_remove(&cgroup_hierarchy_idr, root->hierarchy_id);
1197}
1198
1199void cgroup_free_root(struct cgroup_root *root)
1200{
1201        if (root) {
1202                idr_destroy(&root->cgroup_idr);
1203                kfree(root);
1204        }
1205}
1206
1207static void cgroup_destroy_root(struct cgroup_root *root)
1208{
1209        struct cgroup *cgrp = &root->cgrp;
1210        struct cgrp_cset_link *link, *tmp_link;
1211
1212        trace_cgroup_destroy_root(root);
1213
1214        cgroup_lock_and_drain_offline(&cgrp_dfl_root.cgrp);
1215
1216        BUG_ON(atomic_read(&root->nr_cgrps));
1217        BUG_ON(!list_empty(&cgrp->self.children));
1218
1219        /* Rebind all subsystems back to the default hierarchy */
1220        WARN_ON(rebind_subsystems(&cgrp_dfl_root, root->subsys_mask));
1221
1222        /*
1223         * Release all the links from cset_links to this hierarchy's
1224         * root cgroup
1225         */
1226        spin_lock_irq(&css_set_lock);
1227
1228        list_for_each_entry_safe(link, tmp_link, &cgrp->cset_links, cset_link) {
1229                list_del(&link->cset_link);
1230                list_del(&link->cgrp_link);
1231                kfree(link);
1232        }
1233
1234        spin_unlock_irq(&css_set_lock);
1235
1236        if (!list_empty(&root->root_list)) {
1237                list_del(&root->root_list);
1238                cgroup_root_count--;
1239        }
1240
1241        cgroup_exit_root_id(root);
1242
1243        mutex_unlock(&cgroup_mutex);
1244
1245        kernfs_destroy_root(root->kf_root);
1246        cgroup_free_root(root);
1247}
1248
1249/*
1250 * look up cgroup associated with current task's cgroup namespace on the
1251 * specified hierarchy
1252 */
1253static struct cgroup *
1254current_cgns_cgroup_from_root(struct cgroup_root *root)
1255{
1256        struct cgroup *res = NULL;
1257        struct css_set *cset;
1258
1259        lockdep_assert_held(&css_set_lock);
1260
1261        rcu_read_lock();
1262
1263        cset = current->nsproxy->cgroup_ns->root_cset;
1264        if (cset == &init_css_set) {
1265                res = &root->cgrp;
1266        } else {
1267                struct cgrp_cset_link *link;
1268
1269                list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1270                        struct cgroup *c = link->cgrp;
1271
1272                        if (c->root == root) {
1273                                res = c;
1274                                break;
1275                        }
1276                }
1277        }
1278        rcu_read_unlock();
1279
1280        BUG_ON(!res);
1281        return res;
1282}
1283
1284/* look up cgroup associated with given css_set on the specified hierarchy */
1285static struct cgroup *cset_cgroup_from_root(struct css_set *cset,
1286                                            struct cgroup_root *root)
1287{
1288        struct cgroup *res = NULL;
1289
1290        lockdep_assert_held(&cgroup_mutex);
1291        lockdep_assert_held(&css_set_lock);
1292
1293        if (cset == &init_css_set) {
1294                res = &root->cgrp;
1295        } else if (root == &cgrp_dfl_root) {
1296                res = cset->dfl_cgrp;
1297        } else {
1298                struct cgrp_cset_link *link;
1299
1300                list_for_each_entry(link, &cset->cgrp_links, cgrp_link) {
1301                        struct cgroup *c = link->cgrp;
1302
1303                        if (c->root == root) {
1304                                res = c;
1305                                break;
1306                        }
1307                }
1308        }
1309
1310        BUG_ON(!res);
1311        return res;
1312}
1313
1314/*
1315 * Return the cgroup for "task" from the given hierarchy. Must be
1316 * called with cgroup_mutex and css_set_lock held.
1317 */
1318struct cgroup *task_cgroup_from_root(struct task_struct *task,
1319                                     struct cgroup_root *root)
1320{
1321        /*
1322         * No need to lock the task - since we hold cgroup_mutex the
1323         * task can't change groups, so the only thing that can happen
1324         * is that it exits and its css is set back to init_css_set.
1325         */
1326        return cset_cgroup_from_root(task_css_set(task), root);
1327}
1328
1329/*
1330 * A task must hold cgroup_mutex to modify cgroups.
1331 *
1332 * Any task can increment and decrement the count field without lock.
1333 * So in general, code holding cgroup_mutex can't rely on the count
1334 * field not changing.  However, if the count goes to zero, then only
1335 * cgroup_attach_task() can increment it again.  Because a count of zero
1336 * means that no tasks are currently attached, therefore there is no
1337 * way a task attached to that cgroup can fork (the other way to
1338 * increment the count).  So code holding cgroup_mutex can safely
1339 * assume that if the count is zero, it will stay zero. Similarly, if
1340 * a task holds cgroup_mutex on a cgroup with zero count, it
1341 * knows that the cgroup won't be removed, as cgroup_rmdir()
1342 * needs that mutex.
1343 *
1344 * A cgroup can only be deleted if both its 'count' of using tasks
1345 * is zero, and its list of 'children' cgroups is empty.  Since all
1346 * tasks in the system use _some_ cgroup, and since there is always at
1347 * least one task in the system (init, pid == 1), therefore, root cgroup
1348 * always has either children cgroups and/or using tasks.  So we don't
1349 * need a special hack to ensure that root cgroup cannot be deleted.
1350 *
1351 * P.S.  One more locking exception.  RCU is used to guard the
1352 * update of a tasks cgroup pointer by cgroup_attach_task()
1353 */
1354
1355static struct kernfs_syscall_ops cgroup_kf_syscall_ops;
1356
1357static char *cgroup_file_name(struct cgroup *cgrp, const struct cftype *cft,
1358                              char *buf)
1359{
1360        struct cgroup_subsys *ss = cft->ss;
1361
1362        if (cft->ss && !(cft->flags & CFTYPE_NO_PREFIX) &&
1363            !(cgrp->root->flags & CGRP_ROOT_NOPREFIX))
1364                snprintf(buf, CGROUP_FILE_NAME_MAX, "%s.%s",
1365                         cgroup_on_dfl(cgrp) ? ss->name : ss->legacy_name,
1366                         cft->name);
1367        else
1368                strncpy(buf, cft->name, CGROUP_FILE_NAME_MAX);
1369        return buf;
1370}
1371
1372/**
1373 * cgroup_file_mode - deduce file mode of a control file
1374 * @cft: the control file in question
1375 *
1376 * S_IRUGO for read, S_IWUSR for write.
1377 */
1378static umode_t cgroup_file_mode(const struct cftype *cft)
1379{
1380        umode_t mode = 0;
1381
1382        if (cft->read_u64 || cft->read_s64 || cft->seq_show)
1383                mode |= S_IRUGO;
1384
1385        if (cft->write_u64 || cft->write_s64 || cft->write) {
1386                if (cft->flags & CFTYPE_WORLD_WRITABLE)
1387                        mode |= S_IWUGO;
1388                else
1389                        mode |= S_IWUSR;
1390        }
1391
1392        return mode;
1393}
1394
1395/**
1396 * cgroup_calc_subtree_ss_mask - calculate subtree_ss_mask
1397 * @subtree_control: the new subtree_control mask to consider
1398 * @this_ss_mask: available subsystems
1399 *
1400 * On the default hierarchy, a subsystem may request other subsystems to be
1401 * enabled together through its ->depends_on mask.  In such cases, more
1402 * subsystems than specified in "cgroup.subtree_control" may be enabled.
1403 *
1404 * This function calculates which subsystems need to be enabled if
1405 * @subtree_control is to be applied while restricted to @this_ss_mask.
1406 */
1407static u16 cgroup_calc_subtree_ss_mask(u16 subtree_control, u16 this_ss_mask)
1408{
1409        u16 cur_ss_mask = subtree_control;
1410        struct cgroup_subsys *ss;
1411        int ssid;
1412
1413        lockdep_assert_held(&cgroup_mutex);
1414
1415        cur_ss_mask |= cgrp_dfl_implicit_ss_mask;
1416
1417        while (true) {
1418                u16 new_ss_mask = cur_ss_mask;
1419
1420                do_each_subsys_mask(ss, ssid, cur_ss_mask) {
1421                        new_ss_mask |= ss->depends_on;
1422                } while_each_subsys_mask();
1423
1424                /*
1425                 * Mask out subsystems which aren't available.  This can
1426                 * happen only if some depended-upon subsystems were bound
1427                 * to non-default hierarchies.
1428                 */
1429                new_ss_mask &= this_ss_mask;
1430
1431                if (new_ss_mask == cur_ss_mask)
1432                        break;
1433                cur_ss_mask = new_ss_mask;
1434        }
1435
1436        return cur_ss_mask;
1437}
1438
1439/**
1440 * cgroup_kn_unlock - unlocking helper for cgroup kernfs methods
1441 * @kn: the kernfs_node being serviced
1442 *
1443 * This helper undoes cgroup_kn_lock_live() and should be invoked before
1444 * the method finishes if locking succeeded.  Note that once this function
1445 * returns the cgroup returned by cgroup_kn_lock_live() may become
1446 * inaccessible any time.  If the caller intends to continue to access the
1447 * cgroup, it should pin it before invoking this function.
1448 */
1449void cgroup_kn_unlock(struct kernfs_node *kn)
1450{
1451        struct cgroup *cgrp;
1452
1453        if (kernfs_type(kn) == KERNFS_DIR)
1454                cgrp = kn->priv;
1455        else
1456                cgrp = kn->parent->priv;
1457
1458        mutex_unlock(&cgroup_mutex);
1459
1460        kernfs_unbreak_active_protection(kn);
1461        cgroup_put(cgrp);
1462}
1463
1464/**
1465 * cgroup_kn_lock_live - locking helper for cgroup kernfs methods
1466 * @kn: the kernfs_node being serviced
1467 * @drain_offline: perform offline draining on the cgroup
1468 *
1469 * This helper is to be used by a cgroup kernfs method currently servicing
1470 * @kn.  It breaks the active protection, performs cgroup locking and
1471 * verifies that the associated cgroup is alive.  Returns the cgroup if
1472 * alive; otherwise, %NULL.  A successful return should be undone by a
1473 * matching cgroup_kn_unlock() invocation.  If @drain_offline is %true, the
1474 * cgroup is drained of offlining csses before return.
1475 *
1476 * Any cgroup kernfs method implementation which requires locking the
1477 * associated cgroup should use this helper.  It avoids nesting cgroup
1478 * locking under kernfs active protection and allows all kernfs operations
1479 * including self-removal.
1480 */
1481struct cgroup *cgroup_kn_lock_live(struct kernfs_node *kn, bool drain_offline)
1482{
1483        struct cgroup *cgrp;
1484
1485        if (kernfs_type(kn) == KERNFS_DIR)
1486                cgrp = kn->priv;
1487        else
1488                cgrp = kn->parent->priv;
1489
1490        /*
1491         * We're gonna grab cgroup_mutex which nests outside kernfs
1492         * active_ref.  cgroup liveliness check alone provides enough
1493         * protection against removal.  Ensure @cgrp stays accessible and
1494         * break the active_ref protection.
1495         */
1496        if (!cgroup_tryget(cgrp))
1497                return NULL;
1498        kernfs_break_active_protection(kn);
1499
1500        if (drain_offline)
1501                cgroup_lock_and_drain_offline(cgrp);
1502        else
1503                mutex_lock(&cgroup_mutex);
1504
1505        if (!cgroup_is_dead(cgrp))
1506                return cgrp;
1507
1508        cgroup_kn_unlock(kn);
1509        return NULL;
1510}
1511
1512static void cgroup_rm_file(struct cgroup *cgrp, const struct cftype *cft)
1513{
1514        char name[CGROUP_FILE_NAME_MAX];
1515
1516        lockdep_assert_held(&cgroup_mutex);
1517
1518        if (cft->file_offset) {
1519                struct cgroup_subsys_state *css = cgroup_css(cgrp, cft->ss);
1520                struct cgroup_file *cfile = (void *)css + cft->file_offset;
1521
1522                spin_lock_irq(&cgroup_file_kn_lock);
1523                cfile->kn = NULL;
1524                spin_unlock_irq(&cgroup_file_kn_lock);
1525        }
1526
1527        kernfs_remove_by_name(cgrp->kn, cgroup_file_name(cgrp, cft, name));
1528}
1529
1530/**
1531 * css_clear_dir - remove subsys files in a cgroup directory
1532 * @css: taget css
1533 */
1534static void css_clear_dir(struct cgroup_subsys_state *css)
1535{
1536        struct cgroup *cgrp = css->cgroup;
1537        struct cftype *cfts;
1538
1539        if (!(css->flags & CSS_VISIBLE))
1540                return;
1541
1542        css->flags &= ~CSS_VISIBLE;
1543
1544        list_for_each_entry(cfts, &css->ss->cfts, node)
1545                cgroup_addrm_files(css, cgrp, cfts, false);
1546}
1547
1548/**
1549 * css_populate_dir - create subsys files in a cgroup directory
1550 * @css: target css
1551 *
1552 * On failure, no file is added.
1553 */
1554static int css_populate_dir(struct cgroup_subsys_state *css)
1555{
1556        struct cgroup *cgrp = css->cgroup;
1557        struct cftype *cfts, *failed_cfts;
1558        int ret;
1559
1560        if ((css->flags & CSS_VISIBLE) || !cgrp->kn)
1561                return 0;
1562
1563        if (!css->ss) {
1564                if (cgroup_on_dfl(cgrp))
1565                        cfts = cgroup_base_files;
1566                else
1567                        cfts = cgroup1_base_files;
1568
1569                return cgroup_addrm_files(&cgrp->self, cgrp, cfts, true);
1570        }
1571
1572        list_for_each_entry(cfts, &css->ss->cfts, node) {
1573                ret = cgroup_addrm_files(css, cgrp, cfts, true);
1574                if (ret < 0) {
1575                        failed_cfts = cfts;
1576                        goto err;
1577                }
1578        }
1579
1580        css->flags |= CSS_VISIBLE;
1581
1582        return 0;
1583err:
1584        list_for_each_entry(cfts, &css->ss->cfts, node) {
1585                if (cfts == failed_cfts)
1586                        break;
1587                cgroup_addrm_files(css, cgrp, cfts, false);
1588        }
1589        return ret;
1590}
1591
1592int rebind_subsystems(struct cgroup_root *dst_root, u16 ss_mask)
1593{
1594        struct cgroup *dcgrp = &dst_root->cgrp;
1595        struct cgroup_subsys *ss;
1596        int ssid, i, ret;
1597
1598        lockdep_assert_held(&cgroup_mutex);
1599
1600        do_each_subsys_mask(ss, ssid, ss_mask) {
1601                /*
1602                 * If @ss has non-root csses attached to it, can't move.
1603                 * If @ss is an implicit controller, it is exempt from this
1604                 * rule and can be stolen.
1605                 */
1606                if (css_next_child(NULL, cgroup_css(&ss->root->cgrp, ss)) &&
1607                    !ss->implicit_on_dfl)
1608                        return -EBUSY;
1609
1610                /* can't move between two non-dummy roots either */
1611                if (ss->root != &cgrp_dfl_root && dst_root != &cgrp_dfl_root)
1612                        return -EBUSY;
1613        } while_each_subsys_mask();
1614
1615        do_each_subsys_mask(ss, ssid, ss_mask) {
1616                struct cgroup_root *src_root = ss->root;
1617                struct cgroup *scgrp = &src_root->cgrp;
1618                struct cgroup_subsys_state *css = cgroup_css(scgrp, ss);
1619                struct css_set *cset;
1620
1621                WARN_ON(!css || cgroup_css(dcgrp, ss));
1622
1623                /* disable from the source */
1624                src_root->subsys_mask &= ~(1 << ssid);
1625                WARN_ON(cgroup_apply_control(scgrp));
1626                cgroup_finalize_control(scgrp, 0);
1627
1628                /* rebind */
1629                RCU_INIT_POINTER(scgrp->subsys[ssid], NULL);
1630                rcu_assign_pointer(dcgrp->subsys[ssid], css);
1631                ss->root = dst_root;
1632                css->cgroup = dcgrp;
1633
1634                spin_lock_irq(&css_set_lock);
1635                hash_for_each(css_set_table, i, cset, hlist)
1636                        list_move_tail(&cset->e_cset_node[ss->id],
1637                                       &dcgrp->e_csets[ss->id]);
1638                spin_unlock_irq(&css_set_lock);
1639
1640                /* default hierarchy doesn't enable controllers by default */
1641                dst_root->subsys_mask |= 1 << ssid;
1642                if (dst_root == &cgrp_dfl_root) {
1643                        static_branch_enable(cgroup_subsys_on_dfl_key[ssid]);
1644                } else {
1645                        dcgrp->subtree_control |= 1 << ssid;
1646                        static_branch_disable(cgroup_subsys_on_dfl_key[ssid]);
1647                }
1648
1649                ret = cgroup_apply_control(dcgrp);
1650                if (ret)
1651                        pr_warn("partial failure to rebind %s controller (err=%d)\n",
1652                                ss->name, ret);
1653
1654                if (ss->bind)
1655                        ss->bind(css);
1656        } while_each_subsys_mask();
1657
1658        kernfs_activate(dcgrp->kn);
1659        return 0;
1660}
1661
1662int cgroup_show_path(struct seq_file *sf, struct kernfs_node *kf_node,
1663                     struct kernfs_root *kf_root)
1664{
1665        int len = 0;
1666        char *buf = NULL;
1667        struct cgroup_root *kf_cgroot = cgroup_root_from_kf(kf_root);
1668        struct cgroup *ns_cgroup;
1669
1670        buf = kmalloc(PATH_MAX, GFP_KERNEL);
1671        if (!buf)
1672                return -ENOMEM;
1673
1674        spin_lock_irq(&css_set_lock);
1675        ns_cgroup = current_cgns_cgroup_from_root(kf_cgroot);
1676        len = kernfs_path_from_node(kf_node, ns_cgroup->kn, buf, PATH_MAX);
1677        spin_unlock_irq(&css_set_lock);
1678
1679        if (len >= PATH_MAX)
1680                len = -ERANGE;
1681        else if (len > 0) {
1682                seq_escape(sf, buf, " \t\n\\");
1683                len = 0;
1684        }
1685        kfree(buf);
1686        return len;
1687}
1688
1689static int parse_cgroup_root_flags(char *data, unsigned int *root_flags)
1690{
1691        char *token;
1692
1693        *root_flags = 0;
1694
1695        if (!data)
1696                return 0;
1697
1698        while ((token = strsep(&data, ",")) != NULL) {
1699                if (!strcmp(token, "nsdelegate")) {
1700                        *root_flags |= CGRP_ROOT_NS_DELEGATE;
1701                        continue;
1702                }
1703
1704                pr_err("cgroup2: unknown option \"%s\"\n", token);
1705                return -EINVAL;
1706        }
1707
1708        return 0;
1709}
1710
1711static void apply_cgroup_root_flags(unsigned int root_flags)
1712{
1713        if (current->nsproxy->cgroup_ns == &init_cgroup_ns) {
1714                if (root_flags & CGRP_ROOT_NS_DELEGATE)
1715                        cgrp_dfl_root.flags |= CGRP_ROOT_NS_DELEGATE;
1716                else
1717                        cgrp_dfl_root.flags &= ~CGRP_ROOT_NS_DELEGATE;
1718        }
1719}
1720
1721static int cgroup_show_options(struct seq_file *seq, struct kernfs_root *kf_root)
1722{
1723        if (cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE)
1724                seq_puts(seq, ",nsdelegate");
1725        return 0;
1726}
1727
1728static int cgroup_remount(struct kernfs_root *kf_root, int *flags, char *data)
1729{
1730        unsigned int root_flags;
1731        int ret;
1732
1733        ret = parse_cgroup_root_flags(data, &root_flags);
1734        if (ret)
1735                return ret;
1736
1737        apply_cgroup_root_flags(root_flags);
1738        return 0;
1739}
1740
1741/*
1742 * To reduce the fork() overhead for systems that are not actually using
1743 * their cgroups capability, we don't maintain the lists running through
1744 * each css_set to its tasks until we see the list actually used - in other
1745 * words after the first mount.
1746 */
1747static bool use_task_css_set_links __read_mostly;
1748
1749static void cgroup_enable_task_cg_lists(void)
1750{
1751        struct task_struct *p, *g;
1752
1753        spin_lock_irq(&css_set_lock);
1754
1755        if (use_task_css_set_links)
1756                goto out_unlock;
1757
1758        use_task_css_set_links = true;
1759
1760        /*
1761         * We need tasklist_lock because RCU is not safe against
1762         * while_each_thread(). Besides, a forking task that has passed
1763         * cgroup_post_fork() without seeing use_task_css_set_links = 1
1764         * is not guaranteed to have its child immediately visible in the
1765         * tasklist if we walk through it with RCU.
1766         */
1767        read_lock(&tasklist_lock);
1768        do_each_thread(g, p) {
1769                WARN_ON_ONCE(!list_empty(&p->cg_list) ||
1770                             task_css_set(p) != &init_css_set);
1771
1772                /*
1773                 * We should check if the process is exiting, otherwise
1774                 * it will race with cgroup_exit() in that the list
1775                 * entry won't be deleted though the process has exited.
1776                 * Do it while holding siglock so that we don't end up
1777                 * racing against cgroup_exit().
1778                 *
1779                 * Interrupts were already disabled while acquiring
1780                 * the css_set_lock, so we do not need to disable it
1781                 * again when acquiring the sighand->siglock here.
1782                 */
1783                spin_lock(&p->sighand->siglock);
1784                if (!(p->flags & PF_EXITING)) {
1785                        struct css_set *cset = task_css_set(p);
1786
1787                        if (!css_set_populated(cset))
1788                                css_set_update_populated(cset, true);
1789                        list_add_tail(&p->cg_list, &cset->tasks);
1790                        get_css_set(cset);
1791                        cset->nr_tasks++;
1792                }
1793                spin_unlock(&p->sighand->siglock);
1794        } while_each_thread(g, p);
1795        read_unlock(&tasklist_lock);
1796out_unlock:
1797        spin_unlock_irq(&css_set_lock);
1798}
1799
1800static void init_cgroup_housekeeping(struct cgroup *cgrp)
1801{
1802        struct cgroup_subsys *ss;
1803        int ssid;
1804
1805        INIT_LIST_HEAD(&cgrp->self.sibling);
1806        INIT_LIST_HEAD(&cgrp->self.children);
1807        INIT_LIST_HEAD(&cgrp->cset_links);
1808        INIT_LIST_HEAD(&cgrp->pidlists);
1809        mutex_init(&cgrp->pidlist_mutex);
1810        cgrp->self.cgroup = cgrp;
1811        cgrp->self.flags |= CSS_ONLINE;
1812        cgrp->dom_cgrp = cgrp;
1813        cgrp->max_descendants = INT_MAX;
1814        cgrp->max_depth = INT_MAX;
1815
1816        for_each_subsys(ss, ssid)
1817                INIT_LIST_HEAD(&cgrp->e_csets[ssid]);
1818
1819        init_waitqueue_head(&cgrp->offline_waitq);
1820        INIT_WORK(&cgrp->release_agent_work, cgroup1_release_agent);
1821}
1822
1823void init_cgroup_root(struct cgroup_root *root, struct cgroup_sb_opts *opts)
1824{
1825        struct cgroup *cgrp = &root->cgrp;
1826
1827        INIT_LIST_HEAD(&root->root_list);
1828        atomic_set(&root->nr_cgrps, 1);
1829        cgrp->root = root;
1830        init_cgroup_housekeeping(cgrp);
1831        idr_init(&root->cgroup_idr);
1832
1833        root->flags = opts->flags;
1834        if (opts->release_agent)
1835                strcpy(root->release_agent_path, opts->release_agent);
1836        if (opts->name)
1837                strcpy(root->name, opts->name);
1838        if (opts->cpuset_clone_children)
1839                set_bit(CGRP_CPUSET_CLONE_CHILDREN, &root->cgrp.flags);
1840}
1841
1842int cgroup_setup_root(struct cgroup_root *root, u16 ss_mask, int ref_flags)
1843{
1844        LIST_HEAD(tmp_links);
1845        struct cgroup *root_cgrp = &root->cgrp;
1846        struct kernfs_syscall_ops *kf_sops;
1847        struct css_set *cset;
1848        int i, ret;
1849
1850        lockdep_assert_held(&cgroup_mutex);
1851
1852        ret = cgroup_idr_alloc(&root->cgroup_idr, root_cgrp, 1, 2, GFP_KERNEL);
1853        if (ret < 0)
1854                goto out;
1855        root_cgrp->id = ret;
1856        root_cgrp->ancestor_ids[0] = ret;
1857
1858        ret = percpu_ref_init(&root_cgrp->self.refcnt, css_release,
1859                              ref_flags, GFP_KERNEL);
1860        if (ret)
1861                goto out;
1862
1863        /*
1864         * We're accessing css_set_count without locking css_set_lock here,
1865         * but that's OK - it can only be increased by someone holding
1866         * cgroup_lock, and that's us.  Later rebinding may disable
1867         * controllers on the default hierarchy and thus create new csets,
1868         * which can't be more than the existing ones.  Allocate 2x.
1869         */
1870        ret = allocate_cgrp_cset_links(2 * css_set_count, &tmp_links);
1871        if (ret)
1872                goto cancel_ref;
1873
1874        ret = cgroup_init_root_id(root);
1875        if (ret)
1876                goto cancel_ref;
1877
1878        kf_sops = root == &cgrp_dfl_root ?
1879                &cgroup_kf_syscall_ops : &cgroup1_kf_syscall_ops;
1880
1881        root->kf_root = kernfs_create_root(kf_sops,
1882                                           KERNFS_ROOT_CREATE_DEACTIVATED |
1883                                           KERNFS_ROOT_SUPPORT_EXPORTOP,
1884                                           root_cgrp);
1885        if (IS_ERR(root->kf_root)) {
1886                ret = PTR_ERR(root->kf_root);
1887                goto exit_root_id;
1888        }
1889        root_cgrp->kn = root->kf_root->kn;
1890
1891        ret = css_populate_dir(&root_cgrp->self);
1892        if (ret)
1893                goto destroy_root;
1894
1895        ret = rebind_subsystems(root, ss_mask);
1896        if (ret)
1897                goto destroy_root;
1898
1899        trace_cgroup_setup_root(root);
1900
1901        /*
1902         * There must be no failure case after here, since rebinding takes
1903         * care of subsystems' refcounts, which are explicitly dropped in
1904         * the failure exit path.
1905         */
1906        list_add(&root->root_list, &cgroup_roots);
1907        cgroup_root_count++;
1908
1909        /*
1910         * Link the root cgroup in this hierarchy into all the css_set
1911         * objects.
1912         */
1913        spin_lock_irq(&css_set_lock);
1914        hash_for_each(css_set_table, i, cset, hlist) {
1915                link_css_set(&tmp_links, cset, root_cgrp);
1916                if (css_set_populated(cset))
1917                        cgroup_update_populated(root_cgrp, true);
1918        }
1919        spin_unlock_irq(&css_set_lock);
1920
1921        BUG_ON(!list_empty(&root_cgrp->self.children));
1922        BUG_ON(atomic_read(&root->nr_cgrps) != 1);
1923
1924        kernfs_activate(root_cgrp->kn);
1925        ret = 0;
1926        goto out;
1927
1928destroy_root:
1929        kernfs_destroy_root(root->kf_root);
1930        root->kf_root = NULL;
1931exit_root_id:
1932        cgroup_exit_root_id(root);
1933cancel_ref:
1934        percpu_ref_exit(&root_cgrp->self.refcnt);
1935out:
1936        free_cgrp_cset_links(&tmp_links);
1937        return ret;
1938}
1939
1940struct dentry *cgroup_do_mount(struct file_system_type *fs_type, int flags,
1941                               struct cgroup_root *root, unsigned long magic,
1942                               struct cgroup_namespace *ns)
1943{
1944        struct dentry *dentry;
1945        bool new_sb;
1946
1947        dentry = kernfs_mount(fs_type, flags, root->kf_root, magic, &new_sb);
1948
1949        /*
1950         * In non-init cgroup namespace, instead of root cgroup's dentry,
1951         * we return the dentry corresponding to the cgroupns->root_cgrp.
1952         */
1953        if (!IS_ERR(dentry) && ns != &init_cgroup_ns) {
1954                struct dentry *nsdentry;
1955                struct cgroup *cgrp;
1956
1957                mutex_lock(&cgroup_mutex);
1958                spin_lock_irq(&css_set_lock);
1959
1960                cgrp = cset_cgroup_from_root(ns->root_cset, root);
1961
1962                spin_unlock_irq(&css_set_lock);
1963                mutex_unlock(&cgroup_mutex);
1964
1965                nsdentry = kernfs_node_dentry(cgrp->kn, dentry->d_sb);
1966                dput(dentry);
1967                dentry = nsdentry;
1968        }
1969
1970        if (IS_ERR(dentry) || !new_sb)
1971                cgroup_put(&root->cgrp);
1972
1973        return dentry;
1974}
1975
1976static struct dentry *cgroup_mount(struct file_system_type *fs_type,
1977                         int flags, const char *unused_dev_name,
1978                         void *data)
1979{
1980        struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
1981        struct dentry *dentry;
1982        int ret;
1983
1984        get_cgroup_ns(ns);
1985
1986        /* Check if the caller has permission to mount. */
1987        if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN)) {
1988                put_cgroup_ns(ns);
1989                return ERR_PTR(-EPERM);
1990        }
1991
1992        /*
1993         * The first time anyone tries to mount a cgroup, enable the list
1994         * linking each css_set to its tasks and fix up all existing tasks.
1995         */
1996        if (!use_task_css_set_links)
1997                cgroup_enable_task_cg_lists();
1998
1999        if (fs_type == &cgroup2_fs_type) {
2000                unsigned int root_flags;
2001
2002                ret = parse_cgroup_root_flags(data, &root_flags);
2003                if (ret) {
2004                        put_cgroup_ns(ns);
2005                        return ERR_PTR(ret);
2006                }
2007
2008                cgrp_dfl_visible = true;
2009                cgroup_get_live(&cgrp_dfl_root.cgrp);
2010
2011                dentry = cgroup_do_mount(&cgroup2_fs_type, flags, &cgrp_dfl_root,
2012                                         CGROUP2_SUPER_MAGIC, ns);
2013                if (!IS_ERR(dentry))
2014                        apply_cgroup_root_flags(root_flags);
2015        } else {
2016                dentry = cgroup1_mount(&cgroup_fs_type, flags, data,
2017                                       CGROUP_SUPER_MAGIC, ns);
2018        }
2019
2020        put_cgroup_ns(ns);
2021        return dentry;
2022}
2023
2024static void cgroup_kill_sb(struct super_block *sb)
2025{
2026        struct kernfs_root *kf_root = kernfs_root_from_sb(sb);
2027        struct cgroup_root *root = cgroup_root_from_kf(kf_root);
2028
2029        /*
2030         * If @root doesn't have any mounts or children, start killing it.
2031         * This prevents new mounts by disabling percpu_ref_tryget_live().
2032         * cgroup_mount() may wait for @root's release.
2033         *
2034         * And don't kill the default root.
2035         */
2036        if (!list_empty(&root->cgrp.self.children) ||
2037            root == &cgrp_dfl_root)
2038                cgroup_put(&root->cgrp);
2039        else
2040                percpu_ref_kill(&root->cgrp.self.refcnt);
2041
2042        kernfs_kill_sb(sb);
2043}
2044
2045struct file_system_type cgroup_fs_type = {
2046        .name = "cgroup",
2047        .mount = cgroup_mount,
2048        .kill_sb = cgroup_kill_sb,
2049        .fs_flags = FS_USERNS_MOUNT,
2050};
2051
2052static struct file_system_type cgroup2_fs_type = {
2053        .name = "cgroup2",
2054        .mount = cgroup_mount,
2055        .kill_sb = cgroup_kill_sb,
2056        .fs_flags = FS_USERNS_MOUNT,
2057};
2058
2059int cgroup_path_ns_locked(struct cgroup *cgrp, char *buf, size_t buflen,
2060                          struct cgroup_namespace *ns)
2061{
2062        struct cgroup *root = cset_cgroup_from_root(ns->root_cset, cgrp->root);
2063
2064        return kernfs_path_from_node(cgrp->kn, root->kn, buf, buflen);
2065}
2066
2067int cgroup_path_ns(struct cgroup *cgrp, char *buf, size_t buflen,
2068                   struct cgroup_namespace *ns)
2069{
2070        int ret;
2071
2072        mutex_lock(&cgroup_mutex);
2073        spin_lock_irq(&css_set_lock);
2074
2075        ret = cgroup_path_ns_locked(cgrp, buf, buflen, ns);
2076
2077        spin_unlock_irq(&css_set_lock);
2078        mutex_unlock(&cgroup_mutex);
2079
2080        return ret;
2081}
2082EXPORT_SYMBOL_GPL(cgroup_path_ns);
2083
2084/**
2085 * task_cgroup_path - cgroup path of a task in the first cgroup hierarchy
2086 * @task: target task
2087 * @buf: the buffer to write the path into
2088 * @buflen: the length of the buffer
2089 *
2090 * Determine @task's cgroup on the first (the one with the lowest non-zero
2091 * hierarchy_id) cgroup hierarchy and copy its path into @buf.  This
2092 * function grabs cgroup_mutex and shouldn't be used inside locks used by
2093 * cgroup controller callbacks.
2094 *
2095 * Return value is the same as kernfs_path().
2096 */
2097int task_cgroup_path(struct task_struct *task, char *buf, size_t buflen)
2098{
2099        struct cgroup_root *root;
2100        struct cgroup *cgrp;
2101        int hierarchy_id = 1;
2102        int ret;
2103
2104        mutex_lock(&cgroup_mutex);
2105        spin_lock_irq(&css_set_lock);
2106
2107        root = idr_get_next(&cgroup_hierarchy_idr, &hierarchy_id);
2108
2109        if (root) {
2110                cgrp = task_cgroup_from_root(task, root);
2111                ret = cgroup_path_ns_locked(cgrp, buf, buflen, &init_cgroup_ns);
2112        } else {
2113                /* if no hierarchy exists, everyone is in "/" */
2114                ret = strlcpy(buf, "/", buflen);
2115        }
2116
2117        spin_unlock_irq(&css_set_lock);
2118        mutex_unlock(&cgroup_mutex);
2119        return ret;
2120}
2121EXPORT_SYMBOL_GPL(task_cgroup_path);
2122
2123/**
2124 * cgroup_migrate_add_task - add a migration target task to a migration context
2125 * @task: target task
2126 * @mgctx: target migration context
2127 *
2128 * Add @task, which is a migration target, to @mgctx->tset.  This function
2129 * becomes noop if @task doesn't need to be migrated.  @task's css_set
2130 * should have been added as a migration source and @task->cg_list will be
2131 * moved from the css_set's tasks list to mg_tasks one.
2132 */
2133static void cgroup_migrate_add_task(struct task_struct *task,
2134                                    struct cgroup_mgctx *mgctx)
2135{
2136        struct css_set *cset;
2137
2138        lockdep_assert_held(&css_set_lock);
2139
2140        /* @task either already exited or can't exit until the end */
2141        if (task->flags & PF_EXITING)
2142                return;
2143
2144        /* leave @task alone if post_fork() hasn't linked it yet */
2145        if (list_empty(&task->cg_list))
2146                return;
2147
2148        cset = task_css_set(task);
2149        if (!cset->mg_src_cgrp)
2150                return;
2151
2152        mgctx->tset.nr_tasks++;
2153
2154        list_move_tail(&task->cg_list, &cset->mg_tasks);
2155        if (list_empty(&cset->mg_node))
2156                list_add_tail(&cset->mg_node,
2157                              &mgctx->tset.src_csets);
2158        if (list_empty(&cset->mg_dst_cset->mg_node))
2159                list_add_tail(&cset->mg_dst_cset->mg_node,
2160                              &mgctx->tset.dst_csets);
2161}
2162
2163/**
2164 * cgroup_taskset_first - reset taskset and return the first task
2165 * @tset: taskset of interest
2166 * @dst_cssp: output variable for the destination css
2167 *
2168 * @tset iteration is initialized and the first task is returned.
2169 */
2170struct task_struct *cgroup_taskset_first(struct cgroup_taskset *tset,
2171                                         struct cgroup_subsys_state **dst_cssp)
2172{
2173        tset->cur_cset = list_first_entry(tset->csets, struct css_set, mg_node);
2174        tset->cur_task = NULL;
2175
2176        return cgroup_taskset_next(tset, dst_cssp);
2177}
2178
2179/**
2180 * cgroup_taskset_next - iterate to the next task in taskset
2181 * @tset: taskset of interest
2182 * @dst_cssp: output variable for the destination css
2183 *
2184 * Return the next task in @tset.  Iteration must have been initialized
2185 * with cgroup_taskset_first().
2186 */
2187struct task_struct *cgroup_taskset_next(struct cgroup_taskset *tset,
2188                                        struct cgroup_subsys_state **dst_cssp)
2189{
2190        struct css_set *cset = tset->cur_cset;
2191        struct task_struct *task = tset->cur_task;
2192
2193        while (&cset->mg_node != tset->csets) {
2194                if (!task)
2195                        task = list_first_entry(&cset->mg_tasks,
2196                                                struct task_struct, cg_list);
2197                else
2198                        task = list_next_entry(task, cg_list);
2199
2200                if (&task->cg_list != &cset->mg_tasks) {
2201                        tset->cur_cset = cset;
2202                        tset->cur_task = task;
2203
2204                        /*
2205                         * This function may be called both before and
2206                         * after cgroup_taskset_migrate().  The two cases
2207                         * can be distinguished by looking at whether @cset
2208                         * has its ->mg_dst_cset set.
2209                         */
2210                        if (cset->mg_dst_cset)
2211                                *dst_cssp = cset->mg_dst_cset->subsys[tset->ssid];
2212                        else
2213                                *dst_cssp = cset->subsys[tset->ssid];
2214
2215                        return task;
2216                }
2217
2218                cset = list_next_entry(cset, mg_node);
2219                task = NULL;
2220        }
2221
2222        return NULL;
2223}
2224
2225/**
2226 * cgroup_taskset_migrate - migrate a taskset
2227 * @mgctx: migration context
2228 *
2229 * Migrate tasks in @mgctx as setup by migration preparation functions.
2230 * This function fails iff one of the ->can_attach callbacks fails and
2231 * guarantees that either all or none of the tasks in @mgctx are migrated.
2232 * @mgctx is consumed regardless of success.
2233 */
2234static int cgroup_migrate_execute(struct cgroup_mgctx *mgctx)
2235{
2236        struct cgroup_taskset *tset = &mgctx->tset;
2237        struct cgroup_subsys *ss;
2238        struct task_struct *task, *tmp_task;
2239        struct css_set *cset, *tmp_cset;
2240        int ssid, failed_ssid, ret;
2241
2242        /* check that we can legitimately attach to the cgroup */
2243        if (tset->nr_tasks) {
2244                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2245                        if (ss->can_attach) {
2246                                tset->ssid = ssid;
2247                                ret = ss->can_attach(tset);
2248                                if (ret) {
2249                                        failed_ssid = ssid;
2250                                        goto out_cancel_attach;
2251                                }
2252                        }
2253                } while_each_subsys_mask();
2254        }
2255
2256        /*
2257         * Now that we're guaranteed success, proceed to move all tasks to
2258         * the new cgroup.  There are no failure cases after here, so this
2259         * is the commit point.
2260         */
2261        spin_lock_irq(&css_set_lock);
2262        list_for_each_entry(cset, &tset->src_csets, mg_node) {
2263                list_for_each_entry_safe(task, tmp_task, &cset->mg_tasks, cg_list) {
2264                        struct css_set *from_cset = task_css_set(task);
2265                        struct css_set *to_cset = cset->mg_dst_cset;
2266
2267                        get_css_set(to_cset);
2268                        to_cset->nr_tasks++;
2269                        css_set_move_task(task, from_cset, to_cset, true);
2270                        put_css_set_locked(from_cset);
2271                        from_cset->nr_tasks--;
2272                }
2273        }
2274        spin_unlock_irq(&css_set_lock);
2275
2276        /*
2277         * Migration is committed, all target tasks are now on dst_csets.
2278         * Nothing is sensitive to fork() after this point.  Notify
2279         * controllers that migration is complete.
2280         */
2281        tset->csets = &tset->dst_csets;
2282
2283        if (tset->nr_tasks) {
2284                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2285                        if (ss->attach) {
2286                                tset->ssid = ssid;
2287                                ss->attach(tset);
2288                        }
2289                } while_each_subsys_mask();
2290        }
2291
2292        ret = 0;
2293        goto out_release_tset;
2294
2295out_cancel_attach:
2296        if (tset->nr_tasks) {
2297                do_each_subsys_mask(ss, ssid, mgctx->ss_mask) {
2298                        if (ssid == failed_ssid)
2299                                break;
2300                        if (ss->cancel_attach) {
2301                                tset->ssid = ssid;
2302                                ss->cancel_attach(tset);
2303                        }
2304                } while_each_subsys_mask();
2305        }
2306out_release_tset:
2307        spin_lock_irq(&css_set_lock);
2308        list_splice_init(&tset->dst_csets, &tset->src_csets);
2309        list_for_each_entry_safe(cset, tmp_cset, &tset->src_csets, mg_node) {
2310                list_splice_tail_init(&cset->mg_tasks, &cset->tasks);
2311                list_del_init(&cset->mg_node);
2312        }
2313        spin_unlock_irq(&css_set_lock);
2314
2315        /*
2316         * Re-initialize the cgroup_taskset structure in case it is reused
2317         * again in another cgroup_migrate_add_task()/cgroup_migrate_execute()
2318         * iteration.
2319         */
2320        tset->nr_tasks = 0;
2321        tset->csets    = &tset->src_csets;
2322        return ret;
2323}
2324
2325/**
2326 * cgroup_migrate_vet_dst - verify whether a cgroup can be migration destination
2327 * @dst_cgrp: destination cgroup to test
2328 *
2329 * On the default hierarchy, except for the mixable, (possible) thread root
2330 * and threaded cgroups, subtree_control must be zero for migration
2331 * destination cgroups with tasks so that child cgroups don't compete
2332 * against tasks.
2333 */
2334int cgroup_migrate_vet_dst(struct cgroup *dst_cgrp)
2335{
2336        /* v1 doesn't have any restriction */
2337        if (!cgroup_on_dfl(dst_cgrp))
2338                return 0;
2339
2340        /* verify @dst_cgrp can host resources */
2341        if (!cgroup_is_valid_domain(dst_cgrp->dom_cgrp))
2342                return -EOPNOTSUPP;
2343
2344        /* mixables don't care */
2345        if (cgroup_is_mixable(dst_cgrp))
2346                return 0;
2347
2348        /*
2349         * If @dst_cgrp is already or can become a thread root or is
2350         * threaded, it doesn't matter.
2351         */
2352        if (cgroup_can_be_thread_root(dst_cgrp) || cgroup_is_threaded(dst_cgrp))
2353                return 0;
2354
2355        /* apply no-internal-process constraint */
2356        if (dst_cgrp->subtree_control)
2357                return -EBUSY;
2358
2359        return 0;
2360}
2361
2362/**
2363 * cgroup_migrate_finish - cleanup after attach
2364 * @mgctx: migration context
2365 *
2366 * Undo cgroup_migrate_add_src() and cgroup_migrate_prepare_dst().  See
2367 * those functions for details.
2368 */
2369void cgroup_migrate_finish(struct cgroup_mgctx *mgctx)
2370{
2371        LIST_HEAD(preloaded);
2372        struct css_set *cset, *tmp_cset;
2373
2374        lockdep_assert_held(&cgroup_mutex);
2375
2376        spin_lock_irq(&css_set_lock);
2377
2378        list_splice_tail_init(&mgctx->preloaded_src_csets, &preloaded);
2379        list_splice_tail_init(&mgctx->preloaded_dst_csets, &preloaded);
2380
2381        list_for_each_entry_safe(cset, tmp_cset, &preloaded, mg_preload_node) {
2382                cset->mg_src_cgrp = NULL;
2383                cset->mg_dst_cgrp = NULL;
2384                cset->mg_dst_cset = NULL;
2385                list_del_init(&cset->mg_preload_node);
2386                put_css_set_locked(cset);
2387        }
2388
2389        spin_unlock_irq(&css_set_lock);
2390}
2391
2392/**
2393 * cgroup_migrate_add_src - add a migration source css_set
2394 * @src_cset: the source css_set to add
2395 * @dst_cgrp: the destination cgroup
2396 * @mgctx: migration context
2397 *
2398 * Tasks belonging to @src_cset are about to be migrated to @dst_cgrp.  Pin
2399 * @src_cset and add it to @mgctx->src_csets, which should later be cleaned
2400 * up by cgroup_migrate_finish().
2401 *
2402 * This function may be called without holding cgroup_threadgroup_rwsem
2403 * even if the target is a process.  Threads may be created and destroyed
2404 * but as long as cgroup_mutex is not dropped, no new css_set can be put
2405 * into play and the preloaded css_sets are guaranteed to cover all
2406 * migrations.
2407 */
2408void cgroup_migrate_add_src(struct css_set *src_cset,
2409                            struct cgroup *dst_cgrp,
2410                            struct cgroup_mgctx *mgctx)
2411{
2412        struct cgroup *src_cgrp;
2413
2414        lockdep_assert_held(&cgroup_mutex);
2415        lockdep_assert_held(&css_set_lock);
2416
2417        /*
2418         * If ->dead, @src_set is associated with one or more dead cgroups
2419         * and doesn't contain any migratable tasks.  Ignore it early so
2420         * that the rest of migration path doesn't get confused by it.
2421         */
2422        if (src_cset->dead)
2423                return;
2424
2425        src_cgrp = cset_cgroup_from_root(src_cset, dst_cgrp->root);
2426
2427        if (!list_empty(&src_cset->mg_preload_node))
2428                return;
2429
2430        WARN_ON(src_cset->mg_src_cgrp);
2431        WARN_ON(src_cset->mg_dst_cgrp);
2432        WARN_ON(!list_empty(&src_cset->mg_tasks));
2433        WARN_ON(!list_empty(&src_cset->mg_node));
2434
2435        src_cset->mg_src_cgrp = src_cgrp;
2436        src_cset->mg_dst_cgrp = dst_cgrp;
2437        get_css_set(src_cset);
2438        list_add_tail(&src_cset->mg_preload_node, &mgctx->preloaded_src_csets);
2439}
2440
2441/**
2442 * cgroup_migrate_prepare_dst - prepare destination css_sets for migration
2443 * @mgctx: migration context
2444 *
2445 * Tasks are about to be moved and all the source css_sets have been
2446 * preloaded to @mgctx->preloaded_src_csets.  This function looks up and
2447 * pins all destination css_sets, links each to its source, and append them
2448 * to @mgctx->preloaded_dst_csets.
2449 *
2450 * This function must be called after cgroup_migrate_add_src() has been
2451 * called on each migration source css_set.  After migration is performed
2452 * using cgroup_migrate(), cgroup_migrate_finish() must be called on
2453 * @mgctx.
2454 */
2455int cgroup_migrate_prepare_dst(struct cgroup_mgctx *mgctx)
2456{
2457        struct css_set *src_cset, *tmp_cset;
2458
2459        lockdep_assert_held(&cgroup_mutex);
2460
2461        /* look up the dst cset for each src cset and link it to src */
2462        list_for_each_entry_safe(src_cset, tmp_cset, &mgctx->preloaded_src_csets,
2463                                 mg_preload_node) {
2464                struct css_set *dst_cset;
2465                struct cgroup_subsys *ss;
2466                int ssid;
2467
2468                dst_cset = find_css_set(src_cset, src_cset->mg_dst_cgrp);
2469                if (!dst_cset)
2470                        goto err;
2471
2472                WARN_ON_ONCE(src_cset->mg_dst_cset || dst_cset->mg_dst_cset);
2473
2474                /*
2475                 * If src cset equals dst, it's noop.  Drop the src.
2476                 * cgroup_migrate() will skip the cset too.  Note that we
2477                 * can't handle src == dst as some nodes are used by both.
2478                 */
2479                if (src_cset == dst_cset) {
2480                        src_cset->mg_src_cgrp = NULL;
2481                        src_cset->mg_dst_cgrp = NULL;
2482                        list_del_init(&src_cset->mg_preload_node);
2483                        put_css_set(src_cset);
2484                        put_css_set(dst_cset);
2485                        continue;
2486                }
2487
2488                src_cset->mg_dst_cset = dst_cset;
2489
2490                if (list_empty(&dst_cset->mg_preload_node))
2491                        list_add_tail(&dst_cset->mg_preload_node,
2492                                      &mgctx->preloaded_dst_csets);
2493                else
2494                        put_css_set(dst_cset);
2495
2496                for_each_subsys(ss, ssid)
2497                        if (src_cset->subsys[ssid] != dst_cset->subsys[ssid])
2498                                mgctx->ss_mask |= 1 << ssid;
2499        }
2500
2501        return 0;
2502err:
2503        cgroup_migrate_finish(mgctx);
2504        return -ENOMEM;
2505}
2506
2507/**
2508 * cgroup_migrate - migrate a process or task to a cgroup
2509 * @leader: the leader of the process or the task to migrate
2510 * @threadgroup: whether @leader points to the whole process or a single task
2511 * @mgctx: migration context
2512 *
2513 * Migrate a process or task denoted by @leader.  If migrating a process,
2514 * the caller must be holding cgroup_threadgroup_rwsem.  The caller is also
2515 * responsible for invoking cgroup_migrate_add_src() and
2516 * cgroup_migrate_prepare_dst() on the targets before invoking this
2517 * function and following up with cgroup_migrate_finish().
2518 *
2519 * As long as a controller's ->can_attach() doesn't fail, this function is
2520 * guaranteed to succeed.  This means that, excluding ->can_attach()
2521 * failure, when migrating multiple targets, the success or failure can be
2522 * decided for all targets by invoking group_migrate_prepare_dst() before
2523 * actually starting migrating.
2524 */
2525int cgroup_migrate(struct task_struct *leader, bool threadgroup,
2526                   struct cgroup_mgctx *mgctx)
2527{
2528        struct task_struct *task;
2529
2530        /*
2531         * Prevent freeing of tasks while we take a snapshot. Tasks that are
2532         * already PF_EXITING could be freed from underneath us unless we
2533         * take an rcu_read_lock.
2534         */
2535        spin_lock_irq(&css_set_lock);
2536        rcu_read_lock();
2537        task = leader;
2538        do {
2539                cgroup_migrate_add_task(task, mgctx);
2540                if (!threadgroup)
2541                        break;
2542        } while_each_thread(leader, task);
2543        rcu_read_unlock();
2544        spin_unlock_irq(&css_set_lock);
2545
2546        return cgroup_migrate_execute(mgctx);
2547}
2548
2549/**
2550 * cgroup_attach_task - attach a task or a whole threadgroup to a cgroup
2551 * @dst_cgrp: the cgroup to attach to
2552 * @leader: the task or the leader of the threadgroup to be attached
2553 * @threadgroup: attach the whole threadgroup?
2554 *
2555 * Call holding cgroup_mutex and cgroup_threadgroup_rwsem.
2556 */
2557int cgroup_attach_task(struct cgroup *dst_cgrp, struct task_struct *leader,
2558                       bool threadgroup)
2559{
2560        DEFINE_CGROUP_MGCTX(mgctx);
2561        struct task_struct *task;
2562        int ret;
2563
2564        ret = cgroup_migrate_vet_dst(dst_cgrp);
2565        if (ret)
2566                return ret;
2567
2568        /* look up all src csets */
2569        spin_lock_irq(&css_set_lock);
2570        rcu_read_lock();
2571        task = leader;
2572        do {
2573                cgroup_migrate_add_src(task_css_set(task), dst_cgrp, &mgctx);
2574                if (!threadgroup)
2575                        break;
2576        } while_each_thread(leader, task);
2577        rcu_read_unlock();
2578        spin_unlock_irq(&css_set_lock);
2579
2580        /* prepare dst csets and commit */
2581        ret = cgroup_migrate_prepare_dst(&mgctx);
2582        if (!ret)
2583                ret = cgroup_migrate(leader, threadgroup, &mgctx);
2584
2585        cgroup_migrate_finish(&mgctx);
2586
2587        if (!ret)
2588                trace_cgroup_attach_task(dst_cgrp, leader, threadgroup);
2589
2590        return ret;
2591}
2592
2593struct task_struct *cgroup_procs_write_start(char *buf, bool threadgroup)
2594        __acquires(&cgroup_threadgroup_rwsem)
2595{
2596        struct task_struct *tsk;
2597        pid_t pid;
2598
2599        if (kstrtoint(strstrip(buf), 0, &pid) || pid < 0)
2600                return ERR_PTR(-EINVAL);
2601
2602        percpu_down_write(&cgroup_threadgroup_rwsem);
2603
2604        rcu_read_lock();
2605        if (pid) {
2606                tsk = find_task_by_vpid(pid);
2607                if (!tsk) {
2608                        tsk = ERR_PTR(-ESRCH);
2609                        goto out_unlock_threadgroup;
2610                }
2611        } else {
2612                tsk = current;
2613        }
2614
2615        if (threadgroup)
2616                tsk = tsk->group_leader;
2617
2618        /*
2619         * kthreads may acquire PF_NO_SETAFFINITY during initialization.
2620         * If userland migrates such a kthread to a non-root cgroup, it can
2621         * become trapped in a cpuset, or RT kthread may be born in a
2622         * cgroup with no rt_runtime allocated.  Just say no.
2623         */
2624        if (tsk->no_cgroup_migration || (tsk->flags & PF_NO_SETAFFINITY)) {
2625                tsk = ERR_PTR(-EINVAL);
2626                goto out_unlock_threadgroup;
2627        }
2628
2629        get_task_struct(tsk);
2630        goto out_unlock_rcu;
2631
2632out_unlock_threadgroup:
2633        percpu_up_write(&cgroup_threadgroup_rwsem);
2634out_unlock_rcu:
2635        rcu_read_unlock();
2636        return tsk;
2637}
2638
2639void cgroup_procs_write_finish(struct task_struct *task)
2640        __releases(&cgroup_threadgroup_rwsem)
2641{
2642        struct cgroup_subsys *ss;
2643        int ssid;
2644
2645        /* release reference from cgroup_procs_write_start() */
2646        put_task_struct(task);
2647
2648        percpu_up_write(&cgroup_threadgroup_rwsem);
2649        for_each_subsys(ss, ssid)
2650                if (ss->post_attach)
2651                        ss->post_attach();
2652}
2653
2654static void cgroup_print_ss_mask(struct seq_file *seq, u16 ss_mask)
2655{
2656        struct cgroup_subsys *ss;
2657        bool printed = false;
2658        int ssid;
2659
2660        do_each_subsys_mask(ss, ssid, ss_mask) {
2661                if (printed)
2662                        seq_putc(seq, ' ');
2663                seq_printf(seq, "%s", ss->name);
2664                printed = true;
2665        } while_each_subsys_mask();
2666        if (printed)
2667                seq_putc(seq, '\n');
2668}
2669
2670/* show controllers which are enabled from the parent */
2671static int cgroup_controllers_show(struct seq_file *seq, void *v)
2672{
2673        struct cgroup *cgrp = seq_css(seq)->cgroup;
2674
2675        cgroup_print_ss_mask(seq, cgroup_control(cgrp));
2676        return 0;
2677}
2678
2679/* show controllers which are enabled for a given cgroup's children */
2680static int cgroup_subtree_control_show(struct seq_file *seq, void *v)
2681{
2682        struct cgroup *cgrp = seq_css(seq)->cgroup;
2683
2684        cgroup_print_ss_mask(seq, cgrp->subtree_control);
2685        return 0;
2686}
2687
2688/**
2689 * cgroup_update_dfl_csses - update css assoc of a subtree in default hierarchy
2690 * @cgrp: root of the subtree to update csses for
2691 *
2692 * @cgrp's control masks have changed and its subtree's css associations
2693 * need to be updated accordingly.  This function looks up all css_sets
2694 * which are attached to the subtree, creates the matching updated css_sets
2695 * and migrates the tasks to the new ones.
2696 */
2697static int cgroup_update_dfl_csses(struct cgroup *cgrp)
2698{
2699        DEFINE_CGROUP_MGCTX(mgctx);
2700        struct cgroup_subsys_state *d_css;
2701        struct cgroup *dsct;
2702        struct css_set *src_cset;
2703        int ret;
2704
2705        lockdep_assert_held(&cgroup_mutex);
2706
2707        percpu_down_write(&cgroup_threadgroup_rwsem);
2708
2709        /* look up all csses currently attached to @cgrp's subtree */
2710        spin_lock_irq(&css_set_lock);
2711        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2712                struct cgrp_cset_link *link;
2713
2714                list_for_each_entry(link, &dsct->cset_links, cset_link)
2715                        cgroup_migrate_add_src(link->cset, dsct, &mgctx);
2716        }
2717        spin_unlock_irq(&css_set_lock);
2718
2719        /* NULL dst indicates self on default hierarchy */
2720        ret = cgroup_migrate_prepare_dst(&mgctx);
2721        if (ret)
2722                goto out_finish;
2723
2724        spin_lock_irq(&css_set_lock);
2725        list_for_each_entry(src_cset, &mgctx.preloaded_src_csets, mg_preload_node) {
2726                struct task_struct *task, *ntask;
2727
2728                /* all tasks in src_csets need to be migrated */
2729                list_for_each_entry_safe(task, ntask, &src_cset->tasks, cg_list)
2730                        cgroup_migrate_add_task(task, &mgctx);
2731        }
2732        spin_unlock_irq(&css_set_lock);
2733
2734        ret = cgroup_migrate_execute(&mgctx);
2735out_finish:
2736        cgroup_migrate_finish(&mgctx);
2737        percpu_up_write(&cgroup_threadgroup_rwsem);
2738        return ret;
2739}
2740
2741/**
2742 * cgroup_lock_and_drain_offline - lock cgroup_mutex and drain offlined csses
2743 * @cgrp: root of the target subtree
2744 *
2745 * Because css offlining is asynchronous, userland may try to re-enable a
2746 * controller while the previous css is still around.  This function grabs
2747 * cgroup_mutex and drains the previous css instances of @cgrp's subtree.
2748 */
2749void cgroup_lock_and_drain_offline(struct cgroup *cgrp)
2750        __acquires(&cgroup_mutex)
2751{
2752        struct cgroup *dsct;
2753        struct cgroup_subsys_state *d_css;
2754        struct cgroup_subsys *ss;
2755        int ssid;
2756
2757restart:
2758        mutex_lock(&cgroup_mutex);
2759
2760        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2761                for_each_subsys(ss, ssid) {
2762                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2763                        DEFINE_WAIT(wait);
2764
2765                        if (!css || !percpu_ref_is_dying(&css->refcnt))
2766                                continue;
2767
2768                        cgroup_get_live(dsct);
2769                        prepare_to_wait(&dsct->offline_waitq, &wait,
2770                                        TASK_UNINTERRUPTIBLE);
2771
2772                        mutex_unlock(&cgroup_mutex);
2773                        schedule();
2774                        finish_wait(&dsct->offline_waitq, &wait);
2775
2776                        cgroup_put(dsct);
2777                        goto restart;
2778                }
2779        }
2780}
2781
2782/**
2783 * cgroup_save_control - save control masks of a subtree
2784 * @cgrp: root of the target subtree
2785 *
2786 * Save ->subtree_control and ->subtree_ss_mask to the respective old_
2787 * prefixed fields for @cgrp's subtree including @cgrp itself.
2788 */
2789static void cgroup_save_control(struct cgroup *cgrp)
2790{
2791        struct cgroup *dsct;
2792        struct cgroup_subsys_state *d_css;
2793
2794        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2795                dsct->old_subtree_control = dsct->subtree_control;
2796                dsct->old_subtree_ss_mask = dsct->subtree_ss_mask;
2797        }
2798}
2799
2800/**
2801 * cgroup_propagate_control - refresh control masks of a subtree
2802 * @cgrp: root of the target subtree
2803 *
2804 * For @cgrp and its subtree, ensure ->subtree_ss_mask matches
2805 * ->subtree_control and propagate controller availability through the
2806 * subtree so that descendants don't have unavailable controllers enabled.
2807 */
2808static void cgroup_propagate_control(struct cgroup *cgrp)
2809{
2810        struct cgroup *dsct;
2811        struct cgroup_subsys_state *d_css;
2812
2813        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2814                dsct->subtree_control &= cgroup_control(dsct);
2815                dsct->subtree_ss_mask =
2816                        cgroup_calc_subtree_ss_mask(dsct->subtree_control,
2817                                                    cgroup_ss_mask(dsct));
2818        }
2819}
2820
2821/**
2822 * cgroup_restore_control - restore control masks of a subtree
2823 * @cgrp: root of the target subtree
2824 *
2825 * Restore ->subtree_control and ->subtree_ss_mask from the respective old_
2826 * prefixed fields for @cgrp's subtree including @cgrp itself.
2827 */
2828static void cgroup_restore_control(struct cgroup *cgrp)
2829{
2830        struct cgroup *dsct;
2831        struct cgroup_subsys_state *d_css;
2832
2833        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2834                dsct->subtree_control = dsct->old_subtree_control;
2835                dsct->subtree_ss_mask = dsct->old_subtree_ss_mask;
2836        }
2837}
2838
2839static bool css_visible(struct cgroup_subsys_state *css)
2840{
2841        struct cgroup_subsys *ss = css->ss;
2842        struct cgroup *cgrp = css->cgroup;
2843
2844        if (cgroup_control(cgrp) & (1 << ss->id))
2845                return true;
2846        if (!(cgroup_ss_mask(cgrp) & (1 << ss->id)))
2847                return false;
2848        return cgroup_on_dfl(cgrp) && ss->implicit_on_dfl;
2849}
2850
2851/**
2852 * cgroup_apply_control_enable - enable or show csses according to control
2853 * @cgrp: root of the target subtree
2854 *
2855 * Walk @cgrp's subtree and create new csses or make the existing ones
2856 * visible.  A css is created invisible if it's being implicitly enabled
2857 * through dependency.  An invisible css is made visible when the userland
2858 * explicitly enables it.
2859 *
2860 * Returns 0 on success, -errno on failure.  On failure, csses which have
2861 * been processed already aren't cleaned up.  The caller is responsible for
2862 * cleaning up with cgroup_apply_control_disable().
2863 */
2864static int cgroup_apply_control_enable(struct cgroup *cgrp)
2865{
2866        struct cgroup *dsct;
2867        struct cgroup_subsys_state *d_css;
2868        struct cgroup_subsys *ss;
2869        int ssid, ret;
2870
2871        cgroup_for_each_live_descendant_pre(dsct, d_css, cgrp) {
2872                for_each_subsys(ss, ssid) {
2873                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2874
2875                        WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2876
2877                        if (!(cgroup_ss_mask(dsct) & (1 << ss->id)))
2878                                continue;
2879
2880                        if (!css) {
2881                                css = css_create(dsct, ss);
2882                                if (IS_ERR(css))
2883                                        return PTR_ERR(css);
2884                        }
2885
2886                        if (css_visible(css)) {
2887                                ret = css_populate_dir(css);
2888                                if (ret)
2889                                        return ret;
2890                        }
2891                }
2892        }
2893
2894        return 0;
2895}
2896
2897/**
2898 * cgroup_apply_control_disable - kill or hide csses according to control
2899 * @cgrp: root of the target subtree
2900 *
2901 * Walk @cgrp's subtree and kill and hide csses so that they match
2902 * cgroup_ss_mask() and cgroup_visible_mask().
2903 *
2904 * A css is hidden when the userland requests it to be disabled while other
2905 * subsystems are still depending on it.  The css must not actively control
2906 * resources and be in the vanilla state if it's made visible again later.
2907 * Controllers which may be depended upon should provide ->css_reset() for
2908 * this purpose.
2909 */
2910static void cgroup_apply_control_disable(struct cgroup *cgrp)
2911{
2912        struct cgroup *dsct;
2913        struct cgroup_subsys_state *d_css;
2914        struct cgroup_subsys *ss;
2915        int ssid;
2916
2917        cgroup_for_each_live_descendant_post(dsct, d_css, cgrp) {
2918                for_each_subsys(ss, ssid) {
2919                        struct cgroup_subsys_state *css = cgroup_css(dsct, ss);
2920
2921                        WARN_ON_ONCE(css && percpu_ref_is_dying(&css->refcnt));
2922
2923                        if (!css)
2924                                continue;
2925
2926                        if (css->parent &&
2927                            !(cgroup_ss_mask(dsct) & (1 << ss->id))) {
2928                                kill_css(css);
2929                        } else if (!css_visible(css)) {
2930                                css_clear_dir(css);
2931                                if (ss->css_reset)
2932                                        ss->css_reset(css);
2933                        }
2934                }
2935        }
2936}
2937
2938/**
2939 * cgroup_apply_control - apply control mask updates to the subtree
2940 * @cgrp: root of the target subtree
2941 *
2942 * subsystems can be enabled and disabled in a subtree using the following
2943 * steps.
2944 *
2945 * 1. Call cgroup_save_control() to stash the current state.
2946 * 2. Update ->subtree_control masks in the subtree as desired.
2947 * 3. Call cgroup_apply_control() to apply the changes.
2948 * 4. Optionally perform other related operations.
2949 * 5. Call cgroup_finalize_control() to finish up.
2950 *
2951 * This function implements step 3 and propagates the mask changes
2952 * throughout @cgrp's subtree, updates csses accordingly and perform
2953 * process migrations.
2954 */
2955static int cgroup_apply_control(struct cgroup *cgrp)
2956{
2957        int ret;
2958
2959        cgroup_propagate_control(cgrp);
2960
2961        ret = cgroup_apply_control_enable(cgrp);
2962        if (ret)
2963                return ret;
2964
2965        /*
2966         * At this point, cgroup_e_css() results reflect the new csses
2967         * making the following cgroup_update_dfl_csses() properly update
2968         * css associations of all tasks in the subtree.
2969         */
2970        ret = cgroup_update_dfl_csses(cgrp);
2971        if (ret)
2972                return ret;
2973
2974        return 0;
2975}
2976
2977/**
2978 * cgroup_finalize_control - finalize control mask update
2979 * @cgrp: root of the target subtree
2980 * @ret: the result of the update
2981 *
2982 * Finalize control mask update.  See cgroup_apply_control() for more info.
2983 */
2984static void cgroup_finalize_control(struct cgroup *cgrp, int ret)
2985{
2986        if (ret) {
2987                cgroup_restore_control(cgrp);
2988                cgroup_propagate_control(cgrp);
2989        }
2990
2991        cgroup_apply_control_disable(cgrp);
2992}
2993
2994static int cgroup_vet_subtree_control_enable(struct cgroup *cgrp, u16 enable)
2995{
2996        u16 domain_enable = enable & ~cgrp_dfl_threaded_ss_mask;
2997
2998        /* if nothing is getting enabled, nothing to worry about */
2999        if (!enable)
3000                return 0;
3001
3002        /* can @cgrp host any resources? */
3003        if (!cgroup_is_valid_domain(cgrp->dom_cgrp))
3004                return -EOPNOTSUPP;
3005
3006        /* mixables don't care */
3007        if (cgroup_is_mixable(cgrp))
3008                return 0;
3009
3010        if (domain_enable) {
3011                /* can't enable domain controllers inside a thread subtree */
3012                if (cgroup_is_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3013                        return -EOPNOTSUPP;
3014        } else {
3015                /*
3016                 * Threaded controllers can handle internal competitions
3017                 * and are always allowed inside a (prospective) thread
3018                 * subtree.
3019                 */
3020                if (cgroup_can_be_thread_root(cgrp) || cgroup_is_threaded(cgrp))
3021                        return 0;
3022        }
3023
3024        /*
3025         * Controllers can't be enabled for a cgroup with tasks to avoid
3026         * child cgroups competing against tasks.
3027         */
3028        if (cgroup_has_tasks(cgrp))
3029                return -EBUSY;
3030
3031        return 0;
3032}
3033
3034/* change the enabled child controllers for a cgroup in the default hierarchy */
3035static ssize_t cgroup_subtree_control_write(struct kernfs_open_file *of,
3036                                            char *buf, size_t nbytes,
3037                                            loff_t off)
3038{
3039        u16 enable = 0, disable = 0;
3040        struct cgroup *cgrp, *child;
3041        struct cgroup_subsys *ss;
3042        char *tok;
3043        int ssid, ret;
3044
3045        /*
3046         * Parse input - space separated list of subsystem names prefixed
3047         * with either + or -.
3048         */
3049        buf = strstrip(buf);
3050        while ((tok = strsep(&buf, " "))) {
3051                if (tok[0] == '\0')
3052                        continue;
3053                do_each_subsys_mask(ss, ssid, ~cgrp_dfl_inhibit_ss_mask) {
3054                        if (!cgroup_ssid_enabled(ssid) ||
3055                            strcmp(tok + 1, ss->name))
3056                                continue;
3057
3058                        if (*tok == '+') {
3059                                enable |= 1 << ssid;
3060                                disable &= ~(1 << ssid);
3061                        } else if (*tok == '-') {
3062                                disable |= 1 << ssid;
3063                                enable &= ~(1 << ssid);
3064                        } else {
3065                                return -EINVAL;
3066                        }
3067                        break;
3068                } while_each_subsys_mask();
3069                if (ssid == CGROUP_SUBSYS_COUNT)
3070                        return -EINVAL;
3071        }
3072
3073        cgrp = cgroup_kn_lock_live(of->kn, true);
3074        if (!cgrp)
3075                return -ENODEV;
3076
3077        for_each_subsys(ss, ssid) {
3078                if (enable & (1 << ssid)) {
3079                        if (cgrp->subtree_control & (1 << ssid)) {
3080                                enable &= ~(1 << ssid);
3081                                continue;
3082                        }
3083
3084                        if (!(cgroup_control(cgrp) & (1 << ssid))) {
3085                                ret = -ENOENT;
3086                                goto out_unlock;
3087                        }
3088                } else if (disable & (1 << ssid)) {
3089                        if (!(cgrp->subtree_control & (1 << ssid))) {
3090                                disable &= ~(1 << ssid);
3091                                continue;
3092                        }
3093
3094                        /* a child has it enabled? */
3095                        cgroup_for_each_live_child(child, cgrp) {
3096                                if (child->subtree_control & (1 << ssid)) {
3097                                        ret = -EBUSY;
3098                                        goto out_unlock;
3099                                }
3100                        }
3101                }
3102        }
3103
3104        if (!enable && !disable) {
3105                ret = 0;
3106                goto out_unlock;
3107        }
3108
3109        ret = cgroup_vet_subtree_control_enable(cgrp, enable);
3110        if (ret)
3111                goto out_unlock;
3112
3113        /* save and update control masks and prepare csses */
3114        cgroup_save_control(cgrp);
3115
3116        cgrp->subtree_control |= enable;
3117        cgrp->subtree_control &= ~disable;
3118
3119        ret = cgroup_apply_control(cgrp);
3120        cgroup_finalize_control(cgrp, ret);
3121        if (ret)
3122                goto out_unlock;
3123
3124        kernfs_activate(cgrp->kn);
3125out_unlock:
3126        cgroup_kn_unlock(of->kn);
3127        return ret ?: nbytes;
3128}
3129
3130/**
3131 * cgroup_enable_threaded - make @cgrp threaded
3132 * @cgrp: the target cgroup
3133 *
3134 * Called when "threaded" is written to the cgroup.type interface file and
3135 * tries to make @cgrp threaded and join the parent's resource domain.
3136 * This function is never called on the root cgroup as cgroup.type doesn't
3137 * exist on it.
3138 */
3139static int cgroup_enable_threaded(struct cgroup *cgrp)
3140{
3141        struct cgroup *parent = cgroup_parent(cgrp);
3142        struct cgroup *dom_cgrp = parent->dom_cgrp;
3143        int ret;
3144
3145        lockdep_assert_held(&cgroup_mutex);
3146
3147        /* noop if already threaded */
3148        if (cgroup_is_threaded(cgrp))
3149                return 0;
3150
3151        /* we're joining the parent's domain, ensure its validity */
3152        if (!cgroup_is_valid_domain(dom_cgrp) ||
3153            !cgroup_can_be_thread_root(dom_cgrp))
3154                return -EOPNOTSUPP;
3155
3156        /*
3157         * The following shouldn't cause actual migrations and should
3158         * always succeed.
3159         */
3160        cgroup_save_control(cgrp);
3161
3162        cgrp->dom_cgrp = dom_cgrp;
3163        ret = cgroup_apply_control(cgrp);
3164        if (!ret)
3165                parent->nr_threaded_children++;
3166        else
3167                cgrp->dom_cgrp = cgrp;
3168
3169        cgroup_finalize_control(cgrp, ret);
3170        return ret;
3171}
3172
3173static int cgroup_type_show(struct seq_file *seq, void *v)
3174{
3175        struct cgroup *cgrp = seq_css(seq)->cgroup;
3176
3177        if (cgroup_is_threaded(cgrp))
3178                seq_puts(seq, "threaded\n");
3179        else if (!cgroup_is_valid_domain(cgrp))
3180                seq_puts(seq, "domain invalid\n");
3181        else if (cgroup_is_thread_root(cgrp))
3182                seq_puts(seq, "domain threaded\n");
3183        else
3184                seq_puts(seq, "domain\n");
3185
3186        return 0;
3187}
3188
3189static ssize_t cgroup_type_write(struct kernfs_open_file *of, char *buf,
3190                                 size_t nbytes, loff_t off)
3191{
3192        struct cgroup *cgrp;
3193        int ret;
3194
3195        /* only switching to threaded mode is supported */
3196        if (strcmp(strstrip(buf), "threaded"))
3197                return -EINVAL;
3198
3199        cgrp = cgroup_kn_lock_live(of->kn, false);
3200        if (!cgrp)
3201                return -ENOENT;
3202
3203        /* threaded can only be enabled */
3204        ret = cgroup_enable_threaded(cgrp);
3205
3206        cgroup_kn_unlock(of->kn);
3207        return ret ?: nbytes;
3208}
3209
3210static int cgroup_max_descendants_show(struct seq_file *seq, void *v)
3211{
3212        struct cgroup *cgrp = seq_css(seq)->cgroup;
3213        int descendants = READ_ONCE(cgrp->max_descendants);
3214
3215        if (descendants == INT_MAX)
3216                seq_puts(seq, "max\n");
3217        else
3218                seq_printf(seq, "%d\n", descendants);
3219
3220        return 0;
3221}
3222
3223static ssize_t cgroup_max_descendants_write(struct kernfs_open_file *of,
3224                                           char *buf, size_t nbytes, loff_t off)
3225{
3226        struct cgroup *cgrp;
3227        int descendants;
3228        ssize_t ret;
3229
3230        buf = strstrip(buf);
3231        if (!strcmp(buf, "max")) {
3232                descendants = INT_MAX;
3233        } else {
3234                ret = kstrtoint(buf, 0, &descendants);
3235                if (ret)
3236                        return ret;
3237        }
3238
3239        if (descendants < 0)
3240                return -ERANGE;
3241
3242        cgrp = cgroup_kn_lock_live(of->kn, false);
3243        if (!cgrp)
3244                return -ENOENT;
3245
3246        cgrp->max_descendants = descendants;
3247
3248        cgroup_kn_unlock(of->kn);
3249
3250        return nbytes;
3251}
3252
3253static int cgroup_max_depth_show(struct seq_file *seq, void *v)
3254{
3255        struct cgroup *cgrp = seq_css(seq)->cgroup;
3256        int depth = READ_ONCE(cgrp->max_depth);
3257
3258        if (depth == INT_MAX)
3259                seq_puts(seq, "max\n");
3260        else
3261                seq_printf(seq, "%d\n", depth);
3262
3263        return 0;
3264}
3265
3266static ssize_t cgroup_max_depth_write(struct kernfs_open_file *of,
3267                                      char *buf, size_t nbytes, loff_t off)
3268{
3269        struct cgroup *cgrp;
3270        ssize_t ret;
3271        int depth;
3272
3273        buf = strstrip(buf);
3274        if (!strcmp(buf, "max")) {
3275                depth = INT_MAX;
3276        } else {
3277                ret = kstrtoint(buf, 0, &depth);
3278                if (ret)
3279                        return ret;
3280        }
3281
3282        if (depth < 0)
3283                return -ERANGE;
3284
3285        cgrp = cgroup_kn_lock_live(of->kn, false);
3286        if (!cgrp)
3287                return -ENOENT;
3288
3289        cgrp->max_depth = depth;
3290
3291        cgroup_kn_unlock(of->kn);
3292
3293        return nbytes;
3294}
3295
3296static int cgroup_events_show(struct seq_file *seq, void *v)
3297{
3298        seq_printf(seq, "populated %d\n",
3299                   cgroup_is_populated(seq_css(seq)->cgroup));
3300        return 0;
3301}
3302
3303static int cgroup_stat_show(struct seq_file *seq, void *v)
3304{
3305        struct cgroup *cgroup = seq_css(seq)->cgroup;
3306
3307        seq_printf(seq, "nr_descendants %d\n",
3308                   cgroup->nr_descendants);
3309        seq_printf(seq, "nr_dying_descendants %d\n",
3310                   cgroup->nr_dying_descendants);
3311
3312        return 0;
3313}
3314
3315static int cgroup_file_open(struct kernfs_open_file *of)
3316{
3317        struct cftype *cft = of->kn->priv;
3318
3319        if (cft->open)
3320                return cft->open(of);
3321        return 0;
3322}
3323
3324static void cgroup_file_release(struct kernfs_open_file *of)
3325{
3326        struct cftype *cft = of->kn->priv;
3327
3328        if (cft->release)
3329                cft->release(of);
3330}
3331
3332static ssize_t cgroup_file_write(struct kernfs_open_file *of, char *buf,
3333                                 size_t nbytes, loff_t off)
3334{
3335        struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
3336        struct cgroup *cgrp = of->kn->parent->priv;
3337        struct cftype *cft = of->kn->priv;
3338        struct cgroup_subsys_state *css;
3339        int ret;
3340
3341        /*
3342         * If namespaces are delegation boundaries, disallow writes to
3343         * files in an non-init namespace root from inside the namespace
3344         * except for the files explicitly marked delegatable -
3345         * cgroup.procs and cgroup.subtree_control.
3346         */
3347        if ((cgrp->root->flags & CGRP_ROOT_NS_DELEGATE) &&
3348            !(cft->flags & CFTYPE_NS_DELEGATABLE) &&
3349            ns != &init_cgroup_ns && ns->root_cset->dfl_cgrp == cgrp)
3350                return -EPERM;
3351
3352        if (cft->write)
3353                return cft->write(of, buf, nbytes, off);
3354
3355        /*
3356         * kernfs guarantees that a file isn't deleted with operations in
3357         * flight, which means that the matching css is and stays alive and
3358         * doesn't need to be pinned.  The RCU locking is not necessary
3359         * either.  It's just for the convenience of using cgroup_css().
3360         */
3361        rcu_read_lock();
3362        css = cgroup_css(cgrp, cft->ss);
3363        rcu_read_unlock();
3364
3365        if (cft->write_u64) {
3366                unsigned long long v;
3367                ret = kstrtoull(buf, 0, &v);
3368                if (!ret)
3369                        ret = cft->write_u64(css, cft, v);
3370        } else if (cft->write_s64) {
3371                long long v;
3372                ret = kstrtoll(buf, 0, &v);
3373                if (!ret)
3374                        ret = cft->write_s64(css, cft, v);
3375        } else {
3376                ret = -EINVAL;
3377        }
3378
3379        return ret ?: nbytes;
3380}
3381
3382static void *cgroup_seqfile_start(struct seq_file *seq, loff_t *ppos)
3383{
3384        return seq_cft(seq)->seq_start(seq, ppos);
3385}
3386
3387static void *cgroup_seqfile_next(struct seq_file *seq, void *v, loff_t *ppos)
3388{
3389        return seq_cft(seq)->seq_next(seq, v, ppos);
3390}
3391
3392static void cgroup_seqfile_stop(struct seq_file *seq, void *v)
3393{
3394        if (seq_cft(seq)->seq_stop)
3395                seq_cft(seq)->seq_stop(seq, v);
3396}
3397
3398static int cgroup_seqfile_show(struct seq_file *m, void *arg)
3399{
3400        struct cftype *cft = seq_cft(m);
3401        struct cgroup_subsys_state *css = seq_css(m);
3402
3403        if (cft->seq_show)
3404                return cft->seq_show(m, arg);
3405
3406        if (cft->read_u64)
3407                seq_printf(m, "%llu\n", cft->read_u64(css, cft));
3408        else if (cft->read_s64)
3409                seq_printf(m, "%lld\n", cft->read_s64(css, cft));
3410        else
3411                return -EINVAL;
3412        return 0;
3413}
3414
3415static struct kernfs_ops cgroup_kf_single_ops = {
3416        .atomic_write_len       = PAGE_SIZE,
3417        .open                   = cgroup_file_open,
3418        .release                = cgroup_file_release,
3419        .write                  = cgroup_file_write,
3420        .seq_show               = cgroup_seqfile_show,
3421};
3422
3423static struct kernfs_ops cgroup_kf_ops = {
3424        .atomic_write_len       = PAGE_SIZE,
3425        .open                   = cgroup_file_open,
3426        .release                = cgroup_file_release,
3427        .write                  = cgroup_file_write,
3428        .seq_start              = cgroup_seqfile_start,
3429        .seq_next               = cgroup_seqfile_next,
3430        .seq_stop               = cgroup_seqfile_stop,
3431        .seq_show               = cgroup_seqfile_show,
3432};
3433
3434/* set uid and gid of cgroup dirs and files to that of the creator */
3435static int cgroup_kn_set_ugid(struct kernfs_node *kn)
3436{
3437        struct iattr iattr = { .ia_valid = ATTR_UID | ATTR_GID,
3438                               .ia_uid = current_fsuid(),
3439                               .ia_gid = current_fsgid(), };
3440
3441        if (uid_eq(iattr.ia_uid, GLOBAL_ROOT_UID) &&
3442            gid_eq(iattr.ia_gid, GLOBAL_ROOT_GID))
3443                return 0;
3444
3445        return kernfs_setattr(kn, &iattr);
3446}
3447
3448static int cgroup_add_file(struct cgroup_subsys_state *css, struct cgroup *cgrp,
3449                           struct cftype *cft)
3450{
3451        char name[CGROUP_FILE_NAME_MAX];
3452        struct kernfs_node *kn;
3453        struct lock_class_key *key = NULL;
3454        int ret;
3455
3456#ifdef CONFIG_DEBUG_LOCK_ALLOC
3457        key = &cft->lockdep_key;
3458#endif
3459        kn = __kernfs_create_file(cgrp->kn, cgroup_file_name(cgrp, cft, name),
3460                                  cgroup_file_mode(cft), 0, cft->kf_ops, cft,
3461                                  NULL, key);
3462        if (IS_ERR(kn))
3463                return PTR_ERR(kn);
3464
3465        ret = cgroup_kn_set_ugid(kn);
3466        if (ret) {
3467                kernfs_remove(kn);
3468                return ret;
3469        }
3470
3471        if (cft->file_offset) {
3472                struct cgroup_file *cfile = (void *)css + cft->file_offset;
3473
3474                spin_lock_irq(&cgroup_file_kn_lock);
3475                cfile->kn = kn;
3476                spin_unlock_irq(&cgroup_file_kn_lock);
3477        }
3478
3479        return 0;
3480}
3481
3482/**
3483 * cgroup_addrm_files - add or remove files to a cgroup directory
3484 * @css: the target css
3485 * @cgrp: the target cgroup (usually css->cgroup)
3486 * @cfts: array of cftypes to be added
3487 * @is_add: whether to add or remove
3488 *
3489 * Depending on @is_add, add or remove files defined by @cfts on @cgrp.
3490 * For removals, this function never fails.
3491 */
3492static int cgroup_addrm_files(struct cgroup_subsys_state *css,
3493                              struct cgroup *cgrp, struct cftype cfts[],
3494                              bool is_add)
3495{
3496        struct cftype *cft, *cft_end = NULL;
3497        int ret = 0;
3498
3499        lockdep_assert_held(&cgroup_mutex);
3500
3501restart:
3502        for (cft = cfts; cft != cft_end && cft->name[0] != '\0'; cft++) {
3503                /* does cft->flags tell us to skip this file on @cgrp? */
3504                if ((cft->flags & __CFTYPE_ONLY_ON_DFL) && !cgroup_on_dfl(cgrp))
3505                        continue;
3506                if ((cft->flags & __CFTYPE_NOT_ON_DFL) && cgroup_on_dfl(cgrp))
3507                        continue;
3508                if ((cft->flags & CFTYPE_NOT_ON_ROOT) && !cgroup_parent(cgrp))
3509                        continue;
3510                if ((cft->flags & CFTYPE_ONLY_ON_ROOT) && cgroup_parent(cgrp))
3511                        continue;
3512
3513                if (is_add) {
3514                        ret = cgroup_add_file(css, cgrp, cft);
3515                        if (ret) {
3516                                pr_warn("%s: failed to add %s, err=%d\n",
3517                                        __func__, cft->name, ret);
3518                                cft_end = cft;
3519                                is_add = false;
3520                                goto restart;
3521                        }
3522                } else {
3523                        cgroup_rm_file(cgrp, cft);
3524                }
3525        }
3526        return ret;
3527}
3528
3529static int cgroup_apply_cftypes(struct cftype *cfts, bool is_add)
3530{
3531        struct cgroup_subsys *ss = cfts[0].ss;
3532        struct cgroup *root = &ss->root->cgrp;
3533        struct cgroup_subsys_state *css;
3534        int ret = 0;
3535
3536        lockdep_assert_held(&cgroup_mutex);
3537
3538        /* add/rm files for all cgroups created before */
3539        css_for_each_descendant_pre(css, cgroup_css(root, ss)) {
3540                struct cgroup *cgrp = css->cgroup;
3541
3542                if (!(css->flags & CSS_VISIBLE))
3543                        continue;
3544
3545                ret = cgroup_addrm_files(css, cgrp, cfts, is_add);
3546                if (ret)
3547                        break;
3548        }
3549
3550        if (is_add && !ret)
3551                kernfs_activate(root->kn);
3552        return ret;
3553}
3554
3555static void cgroup_exit_cftypes(struct cftype *cfts)
3556{
3557        struct cftype *cft;
3558
3559        for (cft = cfts; cft->name[0] != '\0'; cft++) {
3560                /* free copy for custom atomic_write_len, see init_cftypes() */
3561                if (cft->max_write_len && cft->max_write_len != PAGE_SIZE)
3562                        kfree(cft->kf_ops);
3563                cft->kf_ops = NULL;
3564                cft->ss = NULL;
3565
3566                /* revert flags set by cgroup core while adding @cfts */
3567                cft->flags &= ~(__CFTYPE_ONLY_ON_DFL | __CFTYPE_NOT_ON_DFL);
3568        }
3569}
3570
3571static int cgroup_init_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3572{
3573        struct cftype *cft;
3574
3575        for (cft = cfts; cft->name[0] != '\0'; cft++) {
3576                struct kernfs_ops *kf_ops;
3577
3578                WARN_ON(cft->ss || cft->kf_ops);
3579
3580                if (cft->seq_start)
3581                        kf_ops = &cgroup_kf_ops;
3582                else
3583                        kf_ops = &cgroup_kf_single_ops;
3584
3585                /*
3586                 * Ugh... if @cft wants a custom max_write_len, we need to
3587                 * make a copy of kf_ops to set its atomic_write_len.
3588                 */
3589                if (cft->max_write_len && cft->max_write_len != PAGE_SIZE) {
3590                        kf_ops = kmemdup(kf_ops, sizeof(*kf_ops), GFP_KERNEL);
3591                        if (!kf_ops) {
3592                                cgroup_exit_cftypes(cfts);
3593                                return -ENOMEM;
3594                        }
3595                        kf_ops->atomic_write_len = cft->max_write_len;
3596                }
3597
3598                cft->kf_ops = kf_ops;
3599                cft->ss = ss;
3600        }
3601
3602        return 0;
3603}
3604
3605static int cgroup_rm_cftypes_locked(struct cftype *cfts)
3606{
3607        lockdep_assert_held(&cgroup_mutex);
3608
3609        if (!cfts || !cfts[0].ss)
3610                return -ENOENT;
3611
3612        list_del(&cfts->node);
3613        cgroup_apply_cftypes(cfts, false);
3614        cgroup_exit_cftypes(cfts);
3615        return 0;
3616}
3617
3618/**
3619 * cgroup_rm_cftypes - remove an array of cftypes from a subsystem
3620 * @cfts: zero-length name terminated array of cftypes
3621 *
3622 * Unregister @cfts.  Files described by @cfts are removed from all
3623 * existing cgroups and all future cgroups won't have them either.  This
3624 * function can be called anytime whether @cfts' subsys is attached or not.
3625 *
3626 * Returns 0 on successful unregistration, -ENOENT if @cfts is not
3627 * registered.
3628 */
3629int cgroup_rm_cftypes(struct cftype *cfts)
3630{
3631        int ret;
3632
3633        mutex_lock(&cgroup_mutex);
3634        ret = cgroup_rm_cftypes_locked(cfts);
3635        mutex_unlock(&cgroup_mutex);
3636        return ret;
3637}
3638
3639/**
3640 * cgroup_add_cftypes - add an array of cftypes to a subsystem
3641 * @ss: target cgroup subsystem
3642 * @cfts: zero-length name terminated array of cftypes
3643 *
3644 * Register @cfts to @ss.  Files described by @cfts are created for all
3645 * existing cgroups to which @ss is attached and all future cgroups will
3646 * have them too.  This function can be called anytime whether @ss is
3647 * attached or not.
3648 *
3649 * Returns 0 on successful registration, -errno on failure.  Note that this
3650 * function currently returns 0 as long as @cfts registration is successful
3651 * even if some file creation attempts on existing cgroups fail.
3652 */
3653static int cgroup_add_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3654{
3655        int ret;
3656
3657        if (!cgroup_ssid_enabled(ss->id))
3658                return 0;
3659
3660        if (!cfts || cfts[0].name[0] == '\0')
3661                return 0;
3662
3663        ret = cgroup_init_cftypes(ss, cfts);
3664        if (ret)
3665                return ret;
3666
3667        mutex_lock(&cgroup_mutex);
3668
3669        list_add_tail(&cfts->node, &ss->cfts);
3670        ret = cgroup_apply_cftypes(cfts, true);
3671        if (ret)
3672                cgroup_rm_cftypes_locked(cfts);
3673
3674        mutex_unlock(&cgroup_mutex);
3675        return ret;
3676}
3677
3678/**
3679 * cgroup_add_dfl_cftypes - add an array of cftypes for default hierarchy
3680 * @ss: target cgroup subsystem
3681 * @cfts: zero-length name terminated array of cftypes
3682 *
3683 * Similar to cgroup_add_cftypes() but the added files are only used for
3684 * the default hierarchy.
3685 */
3686int cgroup_add_dfl_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3687{
3688        struct cftype *cft;
3689
3690        for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3691                cft->flags |= __CFTYPE_ONLY_ON_DFL;
3692        return cgroup_add_cftypes(ss, cfts);
3693}
3694
3695/**
3696 * cgroup_add_legacy_cftypes - add an array of cftypes for legacy hierarchies
3697 * @ss: target cgroup subsystem
3698 * @cfts: zero-length name terminated array of cftypes
3699 *
3700 * Similar to cgroup_add_cftypes() but the added files are only used for
3701 * the legacy hierarchies.
3702 */
3703int cgroup_add_legacy_cftypes(struct cgroup_subsys *ss, struct cftype *cfts)
3704{
3705        struct cftype *cft;
3706
3707        for (cft = cfts; cft && cft->name[0] != '\0'; cft++)
3708                cft->flags |= __CFTYPE_NOT_ON_DFL;
3709        return cgroup_add_cftypes(ss, cfts);
3710}
3711
3712/**
3713 * cgroup_file_notify - generate a file modified event for a cgroup_file
3714 * @cfile: target cgroup_file
3715 *
3716 * @cfile must have been obtained by setting cftype->file_offset.
3717 */
3718void cgroup_file_notify(struct cgroup_file *cfile)
3719{
3720        unsigned long flags;
3721
3722        spin_lock_irqsave(&cgroup_file_kn_lock, flags);
3723        if (cfile->kn)
3724                kernfs_notify(cfile->kn);
3725        spin_unlock_irqrestore(&cgroup_file_kn_lock, flags);
3726}
3727
3728/**
3729 * css_next_child - find the next child of a given css
3730 * @pos: the current position (%NULL to initiate traversal)
3731 * @parent: css whose children to walk
3732 *
3733 * This function returns the next child of @parent and should be called
3734 * under either cgroup_mutex or RCU read lock.  The only requirement is
3735 * that @parent and @pos are accessible.  The next sibling is guaranteed to
3736 * be returned regardless of their states.
3737 *
3738 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3739 * css which finished ->css_online() is guaranteed to be visible in the
3740 * future iterations and will stay visible until the last reference is put.
3741 * A css which hasn't finished ->css_online() or already finished
3742 * ->css_offline() may show up during traversal.  It's each subsystem's
3743 * responsibility to synchronize against on/offlining.
3744 */
3745struct cgroup_subsys_state *css_next_child(struct cgroup_subsys_state *pos,
3746                                           struct cgroup_subsys_state *parent)
3747{
3748        struct cgroup_subsys_state *next;
3749
3750        cgroup_assert_mutex_or_rcu_locked();
3751
3752        /*
3753         * @pos could already have been unlinked from the sibling list.
3754         * Once a cgroup is removed, its ->sibling.next is no longer
3755         * updated when its next sibling changes.  CSS_RELEASED is set when
3756         * @pos is taken off list, at which time its next pointer is valid,
3757         * and, as releases are serialized, the one pointed to by the next
3758         * pointer is guaranteed to not have started release yet.  This
3759         * implies that if we observe !CSS_RELEASED on @pos in this RCU
3760         * critical section, the one pointed to by its next pointer is
3761         * guaranteed to not have finished its RCU grace period even if we
3762         * have dropped rcu_read_lock() inbetween iterations.
3763         *
3764         * If @pos has CSS_RELEASED set, its next pointer can't be
3765         * dereferenced; however, as each css is given a monotonically
3766         * increasing unique serial number and always appended to the
3767         * sibling list, the next one can be found by walking the parent's
3768         * children until the first css with higher serial number than
3769         * @pos's.  While this path can be slower, it happens iff iteration
3770         * races against release and the race window is very small.
3771         */
3772        if (!pos) {
3773                next = list_entry_rcu(parent->children.next, struct cgroup_subsys_state, sibling);
3774        } else if (likely(!(pos->flags & CSS_RELEASED))) {
3775                next = list_entry_rcu(pos->sibling.next, struct cgroup_subsys_state, sibling);
3776        } else {
3777                list_for_each_entry_rcu(next, &parent->children, sibling)
3778                        if (next->serial_nr > pos->serial_nr)
3779                                break;
3780        }
3781
3782        /*
3783         * @next, if not pointing to the head, can be dereferenced and is
3784         * the next sibling.
3785         */
3786        if (&next->sibling != &parent->children)
3787                return next;
3788        return NULL;
3789}
3790
3791/**
3792 * css_next_descendant_pre - find the next descendant for pre-order walk
3793 * @pos: the current position (%NULL to initiate traversal)
3794 * @root: css whose descendants to walk
3795 *
3796 * To be used by css_for_each_descendant_pre().  Find the next descendant
3797 * to visit for pre-order traversal of @root's descendants.  @root is
3798 * included in the iteration and the first node to be visited.
3799 *
3800 * While this function requires cgroup_mutex or RCU read locking, it
3801 * doesn't require the whole traversal to be contained in a single critical
3802 * section.  This function will return the correct next descendant as long
3803 * as both @pos and @root are accessible and @pos is a descendant of @root.
3804 *
3805 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3806 * css which finished ->css_online() is guaranteed to be visible in the
3807 * future iterations and will stay visible until the last reference is put.
3808 * A css which hasn't finished ->css_online() or already finished
3809 * ->css_offline() may show up during traversal.  It's each subsystem's
3810 * responsibility to synchronize against on/offlining.
3811 */
3812struct cgroup_subsys_state *
3813css_next_descendant_pre(struct cgroup_subsys_state *pos,
3814                        struct cgroup_subsys_state *root)
3815{
3816        struct cgroup_subsys_state *next;
3817
3818        cgroup_assert_mutex_or_rcu_locked();
3819
3820        /* if first iteration, visit @root */
3821        if (!pos)
3822                return root;
3823
3824        /* visit the first child if exists */
3825        next = css_next_child(NULL, pos);
3826        if (next)
3827                return next;
3828
3829        /* no child, visit my or the closest ancestor's next sibling */
3830        while (pos != root) {
3831                next = css_next_child(pos, pos->parent);
3832                if (next)
3833                        return next;
3834                pos = pos->parent;
3835        }
3836
3837        return NULL;
3838}
3839
3840/**
3841 * css_rightmost_descendant - return the rightmost descendant of a css
3842 * @pos: css of interest
3843 *
3844 * Return the rightmost descendant of @pos.  If there's no descendant, @pos
3845 * is returned.  This can be used during pre-order traversal to skip
3846 * subtree of @pos.
3847 *
3848 * While this function requires cgroup_mutex or RCU read locking, it
3849 * doesn't require the whole traversal to be contained in a single critical
3850 * section.  This function will return the correct rightmost descendant as
3851 * long as @pos is accessible.
3852 */
3853struct cgroup_subsys_state *
3854css_rightmost_descendant(struct cgroup_subsys_state *pos)
3855{
3856        struct cgroup_subsys_state *last, *tmp;
3857
3858        cgroup_assert_mutex_or_rcu_locked();
3859
3860        do {
3861                last = pos;
3862                /* ->prev isn't RCU safe, walk ->next till the end */
3863                pos = NULL;
3864                css_for_each_child(tmp, last)
3865                        pos = tmp;
3866        } while (pos);
3867
3868        return last;
3869}
3870
3871static struct cgroup_subsys_state *
3872css_leftmost_descendant(struct cgroup_subsys_state *pos)
3873{
3874        struct cgroup_subsys_state *last;
3875
3876        do {
3877                last = pos;
3878                pos = css_next_child(NULL, pos);
3879        } while (pos);
3880
3881        return last;
3882}
3883
3884/**
3885 * css_next_descendant_post - find the next descendant for post-order walk
3886 * @pos: the current position (%NULL to initiate traversal)
3887 * @root: css whose descendants to walk
3888 *
3889 * To be used by css_for_each_descendant_post().  Find the next descendant
3890 * to visit for post-order traversal of @root's descendants.  @root is
3891 * included in the iteration and the last node to be visited.
3892 *
3893 * While this function requires cgroup_mutex or RCU read locking, it
3894 * doesn't require the whole traversal to be contained in a single critical
3895 * section.  This function will return the correct next descendant as long
3896 * as both @pos and @cgroup are accessible and @pos is a descendant of
3897 * @cgroup.
3898 *
3899 * If a subsystem synchronizes ->css_online() and the start of iteration, a
3900 * css which finished ->css_online() is guaranteed to be visible in the
3901 * future iterations and will stay visible until the last reference is put.
3902 * A css which hasn't finished ->css_online() or already finished
3903 * ->css_offline() may show up during traversal.  It's each subsystem's
3904 * responsibility to synchronize against on/offlining.
3905 */
3906struct cgroup_subsys_state *
3907css_next_descendant_post(struct cgroup_subsys_state *pos,
3908                         struct cgroup_subsys_state *root)
3909{
3910        struct cgroup_subsys_state *next;
3911
3912        cgroup_assert_mutex_or_rcu_locked();
3913
3914        /* if first iteration, visit leftmost descendant which may be @root */
3915        if (!pos)
3916                return css_leftmost_descendant(root);
3917
3918        /* if we visited @root, we're done */
3919        if (pos == root)
3920                return NULL;
3921
3922        /* if there's an unvisited sibling, visit its leftmost descendant */
3923        next = css_next_child(pos, pos->parent);
3924        if (next)
3925                return css_leftmost_descendant(next);
3926
3927        /* no sibling left, visit parent */
3928        return pos->parent;
3929}
3930
3931/**
3932 * css_has_online_children - does a css have online children
3933 * @css: the target css
3934 *
3935 * Returns %true if @css has any online children; otherwise, %false.  This
3936 * function can be called from any context but the caller is responsible
3937 * for synchronizing against on/offlining as necessary.
3938 */
3939bool css_has_online_children(struct cgroup_subsys_state *css)
3940{
3941        struct cgroup_subsys_state *child;
3942        bool ret = false;
3943
3944        rcu_read_lock();
3945        css_for_each_child(child, css) {
3946                if (child->flags & CSS_ONLINE) {
3947                        ret = true;
3948                        break;
3949                }
3950        }
3951        rcu_read_unlock();
3952        return ret;
3953}
3954
3955static struct css_set *css_task_iter_next_css_set(struct css_task_iter *it)
3956{
3957        struct list_head *l;
3958        struct cgrp_cset_link *link;
3959        struct css_set *cset;
3960
3961        lockdep_assert_held(&css_set_lock);
3962
3963        /* find the next threaded cset */
3964        if (it->tcset_pos) {
3965                l = it->tcset_pos->next;
3966
3967                if (l != it->tcset_head) {
3968                        it->tcset_pos = l;
3969                        return container_of(l, struct css_set,
3970                                            threaded_csets_node);
3971                }
3972
3973                it->tcset_pos = NULL;
3974        }
3975
3976        /* find the next cset */
3977        l = it->cset_pos;
3978        l = l->next;
3979        if (l == it->cset_head) {
3980                it->cset_pos = NULL;
3981                return NULL;
3982        }
3983
3984        if (it->ss) {
3985                cset = container_of(l, struct css_set, e_cset_node[it->ss->id]);
3986        } else {
3987                link = list_entry(l, struct cgrp_cset_link, cset_link);
3988                cset = link->cset;
3989        }
3990
3991        it->cset_pos = l;
3992
3993        /* initialize threaded css_set walking */
3994        if (it->flags & CSS_TASK_ITER_THREADED) {
3995                if (it->cur_dcset)
3996                        put_css_set_locked(it->cur_dcset);
3997                it->cur_dcset = cset;
3998                get_css_set(cset);
3999
4000                it->tcset_head = &cset->threaded_csets;
4001                it->tcset_pos = &cset->threaded_csets;
4002        }
4003
4004        return cset;
4005}
4006
4007/**
4008 * css_task_iter_advance_css_set - advance a task itererator to the next css_set
4009 * @it: the iterator to advance
4010 *
4011 * Advance @it to the next css_set to walk.
4012 */
4013static void css_task_iter_advance_css_set(struct css_task_iter *it)
4014{
4015        struct css_set *cset;
4016
4017        lockdep_assert_held(&css_set_lock);
4018
4019        /* Advance to the next non-empty css_set */
4020        do {
4021                cset = css_task_iter_next_css_set(it);
4022                if (!cset) {
4023                        it->task_pos = NULL;
4024                        return;
4025                }
4026        } while (!css_set_populated(cset));
4027
4028        if (!list_empty(&cset->tasks))
4029                it->task_pos = cset->tasks.next;
4030        else
4031                it->task_pos = cset->mg_tasks.next;
4032
4033        it->tasks_head = &cset->tasks;
4034        it->mg_tasks_head = &cset->mg_tasks;
4035
4036        /*
4037         * We don't keep css_sets locked across iteration steps and thus
4038         * need to take steps to ensure that iteration can be resumed after
4039         * the lock is re-acquired.  Iteration is performed at two levels -
4040         * css_sets and tasks in them.
4041         *
4042         * Once created, a css_set never leaves its cgroup lists, so a
4043         * pinned css_set is guaranteed to stay put and we can resume
4044         * iteration afterwards.
4045         *
4046         * Tasks may leave @cset across iteration steps.  This is resolved
4047         * by registering each iterator with the css_set currently being
4048         * walked and making css_set_move_task() advance iterators whose
4049         * next task is leaving.
4050         */
4051        if (it->cur_cset) {
4052                list_del(&it->iters_node);
4053                put_css_set_locked(it->cur_cset);
4054        }
4055        get_css_set(cset);
4056        it->cur_cset = cset;
4057        list_add(&it->iters_node, &cset->task_iters);
4058}
4059
4060static void css_task_iter_advance(struct css_task_iter *it)
4061{
4062        struct list_head *l = it->task_pos;
4063
4064        lockdep_assert_held(&css_set_lock);
4065        WARN_ON_ONCE(!l);
4066
4067repeat:
4068        /*
4069         * Advance iterator to find next entry.  cset->tasks is consumed
4070         * first and then ->mg_tasks.  After ->mg_tasks, we move onto the
4071         * next cset.
4072         */
4073        l = l->next;
4074
4075        if (l == it->tasks_head)
4076                l = it->mg_tasks_head->next;
4077
4078        if (l == it->mg_tasks_head)
4079                css_task_iter_advance_css_set(it);
4080        else
4081                it->task_pos = l;
4082
4083        /* if PROCS, skip over tasks which aren't group leaders */
4084        if ((it->flags & CSS_TASK_ITER_PROCS) && it->task_pos &&
4085            !thread_group_leader(list_entry(it->task_pos, struct task_struct,
4086                                            cg_list)))
4087                goto repeat;
4088}
4089
4090/**
4091 * css_task_iter_start - initiate task iteration
4092 * @css: the css to walk tasks of
4093 * @flags: CSS_TASK_ITER_* flags
4094 * @it: the task iterator to use
4095 *
4096 * Initiate iteration through the tasks of @css.  The caller can call
4097 * css_task_iter_next() to walk through the tasks until the function
4098 * returns NULL.  On completion of iteration, css_task_iter_end() must be
4099 * called.
4100 */
4101void css_task_iter_start(struct cgroup_subsys_state *css, unsigned int flags,
4102                         struct css_task_iter *it)
4103{
4104        /* no one should try to iterate before mounting cgroups */
4105        WARN_ON_ONCE(!use_task_css_set_links);
4106
4107        memset(it, 0, sizeof(*it));
4108
4109        spin_lock_irq(&css_set_lock);
4110
4111        it->ss = css->ss;
4112        it->flags = flags;
4113
4114        if (it->ss)
4115                it->cset_pos = &css->cgroup->e_csets[css->ss->id];
4116        else
4117                it->cset_pos = &css->cgroup->cset_links;
4118
4119        it->cset_head = it->cset_pos;
4120
4121        css_task_iter_advance_css_set(it);
4122
4123        spin_unlock_irq(&css_set_lock);
4124}
4125
4126/**
4127 * css_task_iter_next - return the next task for the iterator
4128 * @it: the task iterator being iterated
4129 *
4130 * The "next" function for task iteration.  @it should have been
4131 * initialized via css_task_iter_start().  Returns NULL when the iteration
4132 * reaches the end.
4133 */
4134struct task_struct *css_task_iter_next(struct css_task_iter *it)
4135{
4136        if (it->cur_task) {
4137                put_task_struct(it->cur_task);
4138                it->cur_task = NULL;
4139        }
4140
4141        spin_lock_irq(&css_set_lock);
4142
4143        if (it->task_pos) {
4144                it->cur_task = list_entry(it->task_pos, struct task_struct,
4145                                          cg_list);
4146                get_task_struct(it->cur_task);
4147                css_task_iter_advance(it);
4148        }
4149
4150        spin_unlock_irq(&css_set_lock);
4151
4152        return it->cur_task;
4153}
4154
4155/**
4156 * css_task_iter_end - finish task iteration
4157 * @it: the task iterator to finish
4158 *
4159 * Finish task iteration started by css_task_iter_start().
4160 */
4161void css_task_iter_end(struct css_task_iter *it)
4162{
4163        if (it->cur_cset) {
4164                spin_lock_irq(&css_set_lock);
4165                list_del(&it->iters_node);
4166                put_css_set_locked(it->cur_cset);
4167                spin_unlock_irq(&css_set_lock);
4168        }
4169
4170        if (it->cur_dcset)
4171                put_css_set(it->cur_dcset);
4172
4173        if (it->cur_task)
4174                put_task_struct(it->cur_task);
4175}
4176
4177static void cgroup_procs_release(struct kernfs_open_file *of)
4178{
4179        if (of->priv) {
4180                css_task_iter_end(of->priv);
4181                kfree(of->priv);
4182        }
4183}
4184
4185static void *cgroup_procs_next(struct seq_file *s, void *v, loff_t *pos)
4186{
4187        struct kernfs_open_file *of = s->private;
4188        struct css_task_iter *it = of->priv;
4189
4190        return css_task_iter_next(it);
4191}
4192
4193static void *__cgroup_procs_start(struct seq_file *s, loff_t *pos,
4194                                  unsigned int iter_flags)
4195{
4196        struct kernfs_open_file *of = s->private;
4197        struct cgroup *cgrp = seq_css(s)->cgroup;
4198        struct css_task_iter *it = of->priv;
4199
4200        /*
4201         * When a seq_file is seeked, it's always traversed sequentially
4202         * from position 0, so we can simply keep iterating on !0 *pos.
4203         */
4204        if (!it) {
4205                if (WARN_ON_ONCE((*pos)++))
4206                        return ERR_PTR(-EINVAL);
4207
4208                it = kzalloc(sizeof(*it), GFP_KERNEL);
4209                if (!it)
4210                        return ERR_PTR(-ENOMEM);
4211                of->priv = it;
4212                css_task_iter_start(&cgrp->self, iter_flags, it);
4213        } else if (!(*pos)++) {
4214                css_task_iter_end(it);
4215                css_task_iter_start(&cgrp->self, iter_flags, it);
4216        }
4217
4218        return cgroup_procs_next(s, NULL, NULL);
4219}
4220
4221static void *cgroup_procs_start(struct seq_file *s, loff_t *pos)
4222{
4223        struct cgroup *cgrp = seq_css(s)->cgroup;
4224
4225        /*
4226         * All processes of a threaded subtree belong to the domain cgroup
4227         * of the subtree.  Only threads can be distributed across the
4228         * subtree.  Reject reads on cgroup.procs in the subtree proper.
4229         * They're always empty anyway.
4230         */
4231        if (cgroup_is_threaded(cgrp))
4232                return ERR_PTR(-EOPNOTSUPP);
4233
4234        return __cgroup_procs_start(s, pos, CSS_TASK_ITER_PROCS |
4235                                            CSS_TASK_ITER_THREADED);
4236}
4237
4238static int cgroup_procs_show(struct seq_file *s, void *v)
4239{
4240        seq_printf(s, "%d\n", task_pid_vnr(v));
4241        return 0;
4242}
4243
4244static int cgroup_procs_write_permission(struct cgroup *src_cgrp,
4245                                         struct cgroup *dst_cgrp,
4246                                         struct super_block *sb)
4247{
4248        struct cgroup_namespace *ns = current->nsproxy->cgroup_ns;
4249        struct cgroup *com_cgrp = src_cgrp;
4250        struct inode *inode;
4251        int ret;
4252
4253        lockdep_assert_held(&cgroup_mutex);
4254
4255        /* find the common ancestor */
4256        while (!cgroup_is_descendant(dst_cgrp, com_cgrp))
4257                com_cgrp = cgroup_parent(com_cgrp);
4258
4259        /* %current should be authorized to migrate to the common ancestor */
4260        inode = kernfs_get_inode(sb, com_cgrp->procs_file.kn);
4261        if (!inode)
4262                return -ENOMEM;
4263
4264        ret = inode_permission(inode, MAY_WRITE);
4265        iput(inode);
4266        if (ret)
4267                return ret;
4268
4269        /*
4270         * If namespaces are delegation boundaries, %current must be able
4271         * to see both source and destination cgroups from its namespace.
4272         */
4273        if ((cgrp_dfl_root.flags & CGRP_ROOT_NS_DELEGATE) &&
4274            (!cgroup_is_descendant(src_cgrp, ns->root_cset->dfl_cgrp) ||
4275             !cgroup_is_descendant(dst_cgrp, ns->root_cset->dfl_cgrp)))
4276                return -ENOENT;
4277
4278        return 0;
4279}
4280
4281static ssize_t cgroup_procs_write(struct kernfs_open_file *of,
4282                                  char *buf, size_t nbytes, loff_t off)
4283{
4284        struct cgroup *src_cgrp, *dst_cgrp;
4285        struct task_struct *task;
4286        ssize_t ret;
4287
4288        dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4289        if (!dst_cgrp)
4290                return -ENODEV;
4291
4292        task = cgroup_procs_write_start(buf, true);
4293        ret = PTR_ERR_OR_ZERO(task);
4294        if (ret)
4295                goto out_unlock;
4296
4297        /* find the source cgroup */
4298        spin_lock_irq(&css_set_lock);
4299        src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4300        spin_unlock_irq(&css_set_lock);
4301
4302        ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4303                                            of->file->f_path.dentry->d_sb);
4304        if (ret)
4305                goto out_finish;
4306
4307        ret = cgroup_attach_task(dst_cgrp, task, true);
4308
4309out_finish:
4310        cgroup_procs_write_finish(task);
4311out_unlock:
4312        cgroup_kn_unlock(of->kn);
4313
4314        return ret ?: nbytes;
4315}
4316
4317static void *cgroup_threads_start(struct seq_file *s, loff_t *pos)
4318{
4319        return __cgroup_procs_start(s, pos, 0);
4320}
4321
4322static ssize_t cgroup_threads_write(struct kernfs_open_file *of,
4323                                    char *buf, size_t nbytes, loff_t off)
4324{
4325        struct cgroup *src_cgrp, *dst_cgrp;
4326        struct task_struct *task;
4327        ssize_t ret;
4328
4329        buf = strstrip(buf);
4330
4331        dst_cgrp = cgroup_kn_lock_live(of->kn, false);
4332        if (!dst_cgrp)
4333                return -ENODEV;
4334
4335        task = cgroup_procs_write_start(buf, false);
4336        ret = PTR_ERR_OR_ZERO(task);
4337        if (ret)
4338                goto out_unlock;
4339
4340        /* find the source cgroup */
4341        spin_lock_irq(&css_set_lock);
4342        src_cgrp = task_cgroup_from_root(task, &cgrp_dfl_root);
4343        spin_unlock_irq(&css_set_lock);
4344
4345        /* thread migrations follow the cgroup.procs delegation rule */
4346        ret = cgroup_procs_write_permission(src_cgrp, dst_cgrp,
4347                                            of->file->f_path.dentry->d_sb);
4348        if (ret)
4349                goto out_finish;
4350
4351        /* and must be contained in the same domain */
4352        ret = -EOPNOTSUPP;
4353        if (src_cgrp->dom_cgrp != dst_cgrp->dom_cgrp)
4354                goto out_finish;
4355
4356        ret = cgroup_attach_task(dst_cgrp, task, false);
4357
4358out_finish:
4359        cgroup_procs_write_finish(task);
4360out_unlock:
4361        cgroup_kn_unlock(of->kn);
4362
4363        return ret ?: nbytes;
4364}
4365
4366/* cgroup core interface files for the default hierarchy */
4367static struct cftype cgroup_base_files[] = {
4368        {
4369                .name = "cgroup.type",
4370                .flags = CFTYPE_NOT_ON_ROOT,
4371                .seq_show = cgroup_type_show,
4372                .write = cgroup_type_write,
4373        },
4374        {
4375                .name = "cgroup.procs",
4376                .flags = CFTYPE_NS_DELEGATABLE,
4377                .file_offset = offsetof(struct cgroup, procs_file),
4378                .release = cgroup_procs_release,
4379                .seq_start = cgroup_procs_start,
4380                .seq_next = cgroup_procs_next,
4381                .seq_show = cgroup_procs_show,
4382                .write = cgroup_procs_write,
4383        },
4384        {
4385                .name = "cgroup.threads",
4386                .release = cgroup_procs_release,
4387                .seq_start = cgroup_threads_start,
4388                .seq_next = cgroup_procs_next,
4389                .seq_show = cgroup_procs_show,
4390                .write = cgroup_threads_write,
4391        },
4392        {
4393                .name = "cgroup.controllers",
4394                .seq_show = cgroup_controllers_show,
4395        },
4396        {
4397                .name = "cgroup.subtree_control",
4398                .flags = CFTYPE_NS_DELEGATABLE,
4399                .seq_show = cgroup_subtree_control_show,
4400                .write = cgroup_subtree_control_write,
4401        },
4402        {
4403                .name = "cgroup.events",
4404                .flags = CFTYPE_NOT_ON_ROOT,
4405                .file_offset = offsetof(struct cgroup, events_file),
4406                .seq_show = cgroup_events_show,
4407        },
4408        {
4409                .name = "cgroup.max.descendants",
4410                .seq_show = cgroup_max_descendants_show,
4411                .write = cgroup_max_descendants_write,
4412        },
4413        {
4414                .name = "cgroup.max.depth",
4415                .seq_show = cgroup_max_depth_show,
4416                .write = cgroup_max_depth_write,
4417        },
4418        {
4419                .name = "cgroup.stat",
4420                .seq_show = cgroup_stat_show,
4421        },
4422        { }     /* terminate */
4423};
4424
4425/*
4426 * css destruction is four-stage process.
4427 *
4428 * 1. Destruction starts.  Killing of the percpu_ref is initiated.
4429 *    Implemented in kill_css().
4430 *
4431 * 2. When the percpu_ref is confirmed to be visible as killed on all CPUs
4432 *    and thus css_tryget_online() is guaranteed to fail, the css can be
4433 *    offlined by invoking offline_css().  After offlining, the base ref is
4434 *    put.  Implemented in css_killed_work_fn().
4435 *
4436 * 3. When the percpu_ref reaches zero, the only possible remaining
4437 *    accessors are inside RCU read sections.  css_release() schedules the
4438 *    RCU callback.
4439 *
4440 * 4. After the grace period, the css can be freed.  Implemented in
4441 *    css_free_work_fn().
4442 *
4443 * It is actually hairier because both step 2 and 4 require process context
4444 * and thus involve punting to css->destroy_work adding two additional
4445 * steps to the already complex sequence.
4446 */
4447static void css_free_work_fn(struct work_struct *work)
4448{
4449        struct cgroup_subsys_state *css =
4450                container_of(work, struct cgroup_subsys_state, destroy_work);
4451        struct cgroup_subsys *ss = css->ss;
4452        struct cgroup *cgrp = css->cgroup;
4453
4454        percpu_ref_exit(&css->refcnt);
4455
4456        if (ss) {
4457                /* css free path */
4458                struct cgroup_subsys_state *parent = css->parent;
4459                int id = css->id;
4460
4461                ss->css_free(css);
4462                cgroup_idr_remove(&ss->css_idr, id);
4463                cgroup_put(cgrp);
4464
4465                if (parent)
4466                        css_put(parent);
4467        } else {
4468                /* cgroup free path */
4469                atomic_dec(&cgrp->root->nr_cgrps);
4470                cgroup1_pidlist_destroy_all(cgrp);
4471                cancel_work_sync(&cgrp->release_agent_work);
4472
4473                if (cgroup_parent(cgrp)) {
4474                        /*
4475                         * We get a ref to the parent, and put the ref when
4476                         * this cgroup is being freed, so it's guaranteed
4477                         * that the parent won't be destroyed before its
4478                         * children.
4479                         */
4480                        cgroup_put(cgroup_parent(cgrp));
4481                        kernfs_put(cgrp->kn);
4482                        kfree(cgrp);
4483                } else {
4484                        /*
4485                         * This is root cgroup's refcnt reaching zero,
4486                         * which indicates that the root should be
4487                         * released.
4488                         */
4489                        cgroup_destroy_root(cgrp->root);
4490                }
4491        }
4492}
4493
4494static void css_free_rcu_fn(struct rcu_head *rcu_head)
4495{
4496        struct cgroup_subsys_state *css =
4497                container_of(rcu_head, struct cgroup_subsys_state, rcu_head);
4498
4499        INIT_WORK(&css->destroy_work, css_free_work_fn);
4500        queue_work(cgroup_destroy_wq, &css->destroy_work);
4501}
4502
4503static void css_release_work_fn(struct work_struct *work)
4504{
4505        struct cgroup_subsys_state *css =
4506                container_of(work, struct cgroup_subsys_state, destroy_work);
4507        struct cgroup_subsys *ss = css->ss;
4508        struct cgroup *cgrp = css->cgroup;
4509
4510        mutex_lock(&cgroup_mutex);
4511
4512        css->flags |= CSS_RELEASED;
4513        list_del_rcu(&css->sibling);
4514
4515        if (ss) {
4516                /* css release path */
4517                cgroup_idr_replace(&ss->css_idr, NULL, css->id);
4518                if (ss->css_released)
4519                        ss->css_released(css);
4520        } else {
4521                struct cgroup *tcgrp;
4522
4523                /* cgroup release path */
4524                trace_cgroup_release(cgrp);
4525
4526                for (tcgrp = cgroup_parent(cgrp); tcgrp;
4527                     tcgrp = cgroup_parent(tcgrp))
4528                        tcgrp->nr_dying_descendants--;
4529
4530                cgroup_idr_remove(&cgrp->root->cgroup_idr, cgrp->id);
4531                cgrp->id = -1;
4532
4533                /*
4534                 * There are two control paths which try to determine
4535                 * cgroup from dentry without going through kernfs -
4536                 * cgroupstats_build() and css_tryget_online_from_dir().
4537                 * Those are supported by RCU protecting clearing of
4538                 * cgrp->kn->priv backpointer.
4539                 */
4540                if (cgrp->kn)
4541                        RCU_INIT_POINTER(*(void __rcu __force **)&cgrp->kn->priv,
4542                                         NULL);
4543
4544                cgroup_bpf_put(cgrp);
4545        }
4546
4547        mutex_unlock(&cgroup_mutex);
4548
4549        call_rcu(&css->rcu_head, css_free_rcu_fn);
4550}
4551
4552static void css_release(struct percpu_ref *ref)
4553{
4554        struct cgroup_subsys_state *css =
4555                container_of(ref, struct cgroup_subsys_state, refcnt);
4556
4557        INIT_WORK(&css->destroy_work, css_release_work_fn);
4558        queue_work(cgroup_destroy_wq, &css->destroy_work);
4559}
4560
4561static void init_and_link_css(struct cgroup_subsys_state *css,
4562                              struct cgroup_subsys *ss, struct cgroup *cgrp)
4563{
4564        lockdep_assert_held(&cgroup_mutex);
4565
4566        cgroup_get_live(cgrp);
4567
4568        memset(css, 0, sizeof(*css));
4569        css->cgroup = cgrp;
4570        css->ss = ss;
4571        css->id = -1;
4572        INIT_LIST_HEAD(&css->sibling);
4573        INIT_LIST_HEAD(&css->children);
4574        css->serial_nr = css_serial_nr_next++;
4575        atomic_set(&css->online_cnt, 0);
4576
4577        if (cgroup_parent(cgrp)) {
4578                css->parent = cgroup_css(cgroup_parent(cgrp), ss);
4579                css_get(css->parent);
4580        }
4581
4582        BUG_ON(cgroup_css(cgrp, ss));
4583}
4584
4585/* invoke ->css_online() on a new CSS and mark it online if successful */
4586static int online_css(struct cgroup_subsys_state *css)
4587{
4588        struct cgroup_subsys *ss = css->ss;
4589        int ret = 0;
4590
4591        lockdep_assert_held(&cgroup_mutex);
4592
4593        if (ss->css_online)
4594                ret = ss->css_online(css);
4595        if (!ret) {
4596                css->flags |= CSS_ONLINE;
4597                rcu_assign_pointer(css->cgroup->subsys[ss->id], css);
4598
4599                atomic_inc(&css->online_cnt);
4600                if (css->parent)
4601                        atomic_inc(&css->parent->online_cnt);
4602        }
4603        return ret;
4604}
4605
4606/* if the CSS is online, invoke ->css_offline() on it and mark it offline */
4607static void offline_css(struct cgroup_subsys_state *css)
4608{
4609        struct cgroup_subsys *ss = css->ss;
4610
4611        lockdep_assert_held(&cgroup_mutex);
4612
4613        if (!(css->flags & CSS_ONLINE))
4614                return;
4615
4616        if (ss->css_offline)
4617                ss->css_offline(css);
4618
4619        css->flags &= ~CSS_ONLINE;
4620        RCU_INIT_POINTER(css->cgroup->subsys[ss->id], NULL);
4621
4622        wake_up_all(&css->cgroup->offline_waitq);
4623}
4624
4625/**
4626 * css_create - create a cgroup_subsys_state
4627 * @cgrp: the cgroup new css will be associated with
4628 * @ss: the subsys of new css
4629 *
4630 * Create a new css associated with @cgrp - @ss pair.  On success, the new
4631 * css is online and installed in @cgrp.  This function doesn't create the
4632 * interface files.  Returns 0 on success, -errno on failure.
4633 */
4634static struct cgroup_subsys_state *css_create(struct cgroup *cgrp,
4635                                              struct cgroup_subsys *ss)
4636{
4637        struct cgroup *parent = cgroup_parent(cgrp);
4638        struct cgroup_subsys_state *parent_css = cgroup_css(parent, ss);
4639        struct cgroup_subsys_state *css;
4640        int err;
4641
4642        lockdep_assert_held(&cgroup_mutex);
4643
4644        css = ss->css_alloc(parent_css);
4645        if (!css)
4646                css = ERR_PTR(-ENOMEM);
4647        if (IS_ERR(css))
4648                return css;
4649
4650        init_and_link_css(css, ss, cgrp);
4651
4652        err = percpu_ref_init(&css->refcnt, css_release, 0, GFP_KERNEL);
4653        if (err)
4654                goto err_free_css;
4655
4656        err = cgroup_idr_alloc(&ss->css_idr, NULL, 2, 0, GFP_KERNEL);
4657        if (err < 0)
4658                goto err_free_css;
4659        css->id = err;
4660
4661        /* @css is ready to be brought online now, make it visible */
4662        list_add_tail_rcu(&css->sibling, &parent_css->children);
4663        cgroup_idr_replace(&ss->css_idr, css, css->id);
4664
4665        err = online_css(css);
4666        if (err)
4667                goto err_list_del;
4668
4669        if (ss->broken_hierarchy && !ss->warned_broken_hierarchy &&
4670            cgroup_parent(parent)) {
4671                pr_warn("%s (%d) created nested cgroup for controller \"%s\" which has incomplete hierarchy support. Nested cgroups may change behavior in the future.\n",
4672                        current->comm, current->pid, ss->name);
4673                if (!strcmp(ss->name, "memory"))
4674                        pr_warn("\"memory\" requires setting use_hierarchy to 1 on the root\n");
4675                ss->warned_broken_hierarchy = true;
4676        }
4677
4678        return css;
4679
4680err_list_del:
4681        list_del_rcu(&css->sibling);
4682err_free_css:
4683        call_rcu(&css->rcu_head, css_free_rcu_fn);
4684        return ERR_PTR(err);
4685}
4686
4687/*
4688 * The returned cgroup is fully initialized including its control mask, but
4689 * it isn't associated with its kernfs_node and doesn't have the control
4690 * mask applied.
4691 */
4692static struct cgroup *cgroup_create(struct cgroup *parent)
4693{
4694        struct cgroup_root *root = parent->root;
4695        struct cgroup *cgrp, *tcgrp;
4696        int level = parent->level + 1;
4697        int ret;
4698
4699        /* allocate the cgroup and its ID, 0 is reserved for the root */
4700        cgrp = kzalloc(sizeof(*cgrp) +
4701                       sizeof(cgrp->ancestor_ids[0]) * (level + 1), GFP_KERNEL);
4702        if (!cgrp)
4703                return ERR_PTR(-ENOMEM);
4704
4705        ret = percpu_ref_init(&cgrp->self.refcnt, css_release, 0, GFP_KERNEL);
4706        if (ret)
4707                goto out_free_cgrp;
4708
4709        /*
4710         * Temporarily set the pointer to NULL, so idr_find() won't return
4711         * a half-baked cgroup.
4712         */
4713        cgrp->id = cgroup_idr_alloc(&root->cgroup_idr, NULL, 2, 0, GFP_KERNEL);
4714        if (cgrp->id < 0) {
4715                ret = -ENOMEM;
4716                goto out_cancel_ref;
4717        }
4718
4719        init_cgroup_housekeeping(cgrp);
4720
4721        cgrp->self.parent = &parent->self;
4722        cgrp->root = root;
4723        cgrp->level = level;
4724
4725        for (tcgrp = cgrp; tcgrp; tcgrp = cgroup_parent(tcgrp)) {
4726                cgrp->ancestor_ids[tcgrp->level] = tcgrp->id;
4727
4728                if (tcgrp != cgrp)
4729                        tcgrp->nr_descendants++;
4730        }
4731
4732        if (notify_on_release(parent))
4733                set_bit(CGRP_NOTIFY_ON_RELEASE, &cgrp->flags);
4734
4735        if (test_bit(CGRP_CPUSET_CLONE_CHILDREN, &parent->flags))
4736                set_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags);
4737
4738        cgrp->self.serial_nr = css_serial_nr_next++;
4739
4740        /* allocation complete, commit to creation */
4741        list_add_tail_rcu(&cgrp->self.sibling, &cgroup_parent(cgrp)->self.children);
4742        atomic_inc(&root->nr_cgrps);
4743        cgroup_get_live(parent);
4744
4745        /*
4746         * @cgrp is now fully operational.  If something fails after this
4747         * point, it'll be released via the normal destruction path.
4748         */
4749        cgroup_idr_replace(&root->cgroup_idr, cgrp, cgrp->id);
4750
4751        /*
4752         * On the default hierarchy, a child doesn't automatically inherit
4753         * subtree_control from the parent.  Each is configured manually.
4754         */
4755        if (!cgroup_on_dfl(cgrp))
4756                cgrp->subtree_control = cgroup_control(cgrp);
4757
4758        if (parent)
4759                cgroup_bpf_inherit(cgrp, parent);
4760
4761        cgroup_propagate_control(cgrp);
4762
4763        return cgrp;
4764
4765out_cancel_ref:
4766        percpu_ref_exit(&cgrp->self.refcnt);
4767out_free_cgrp:
4768        kfree(cgrp);
4769        return ERR_PTR(ret);
4770}
4771
4772static bool cgroup_check_hierarchy_limits(struct cgroup *parent)
4773{
4774        struct cgroup *cgroup;
4775        int ret = false;
4776        int level = 1;
4777
4778        lockdep_assert_held(&cgroup_mutex);
4779
4780        for (cgroup = parent; cgroup; cgroup = cgroup_parent(cgroup)) {
4781                if (cgroup->nr_descendants >= cgroup->max_descendants)
4782                        goto fail;
4783
4784                if (level > cgroup->max_depth)
4785                        goto fail;
4786
4787                level++;
4788        }
4789
4790        ret = true;
4791fail:
4792        return ret;
4793}
4794
4795int cgroup_mkdir(struct kernfs_node *parent_kn, const char *name, umode_t mode)
4796{
4797        struct cgroup *parent, *cgrp;
4798        struct kernfs_node *kn;
4799        int ret;
4800
4801        /* do not accept '\n' to prevent making /proc/<pid>/cgroup unparsable */
4802        if (strchr(name, '\n'))
4803                return -EINVAL;
4804
4805        parent = cgroup_kn_lock_live(parent_kn, false);
4806        if (!parent)
4807                return -ENODEV;
4808
4809        if (!cgroup_check_hierarchy_limits(parent)) {
4810                ret = -EAGAIN;
4811                goto out_unlock;
4812        }
4813
4814        cgrp = cgroup_create(parent);
4815        if (IS_ERR(cgrp)) {
4816                ret = PTR_ERR(cgrp);
4817                goto out_unlock;
4818        }
4819
4820        /* create the directory */
4821        kn = kernfs_create_dir(parent->kn, name, mode, cgrp);
4822        if (IS_ERR(kn)) {
4823                ret = PTR_ERR(kn);
4824                goto out_destroy;
4825        }
4826        cgrp->kn = kn;
4827
4828        /*
4829         * This extra ref will be put in cgroup_free_fn() and guarantees
4830         * that @cgrp->kn is always accessible.
4831         */
4832        kernfs_get(kn);
4833
4834        ret = cgroup_kn_set_ugid(kn);
4835        if (ret)
4836                goto out_destroy;
4837
4838        ret = css_populate_dir(&cgrp->self);
4839        if (ret)
4840                goto out_destroy;
4841
4842        ret = cgroup_apply_control_enable(cgrp);
4843        if (ret)
4844                goto out_destroy;
4845
4846        trace_cgroup_mkdir(cgrp);
4847
4848        /* let's create and online css's */
4849        kernfs_activate(kn);
4850
4851        ret = 0;
4852        goto out_unlock;
4853
4854out_destroy:
4855        cgroup_destroy_locked(cgrp);
4856out_unlock:
4857        cgroup_kn_unlock(parent_kn);
4858        return ret;
4859}
4860
4861/*
4862 * This is called when the refcnt of a css is confirmed to be killed.
4863 * css_tryget_online() is now guaranteed to fail.  Tell the subsystem to
4864 * initate destruction and put the css ref from kill_css().
4865 */
4866static void css_killed_work_fn(struct work_struct *work)
4867{
4868        struct cgroup_subsys_state *css =
4869                container_of(work, struct cgroup_subsys_state, destroy_work);
4870
4871        mutex_lock(&cgroup_mutex);
4872
4873        do {
4874                offline_css(css);
4875                css_put(css);
4876                /* @css can't go away while we're holding cgroup_mutex */
4877                css = css->parent;
4878        } while (css && atomic_dec_and_test(&css->online_cnt));
4879
4880        mutex_unlock(&cgroup_mutex);
4881}
4882
4883/* css kill confirmation processing requires process context, bounce */
4884static void css_killed_ref_fn(struct percpu_ref *ref)
4885{
4886        struct cgroup_subsys_state *css =
4887                container_of(ref, struct cgroup_subsys_state, refcnt);
4888
4889        if (atomic_dec_and_test(&css->online_cnt)) {
4890                INIT_WORK(&css->destroy_work, css_killed_work_fn);
4891                queue_work(cgroup_destroy_wq, &css->destroy_work);
4892        }
4893}
4894
4895/**
4896 * kill_css - destroy a css
4897 * @css: css to destroy
4898 *
4899 * This function initiates destruction of @css by removing cgroup interface
4900 * files and putting its base reference.  ->css_offline() will be invoked
4901 * asynchronously once css_tryget_online() is guaranteed to fail and when
4902 * the reference count reaches zero, @css will be released.
4903 */
4904static void kill_css(struct cgroup_subsys_state *css)
4905{
4906        lockdep_assert_held(&cgroup_mutex);
4907
4908        if (css->flags & CSS_DYING)
4909                return;
4910
4911        css->flags |= CSS_DYING;
4912
4913        /*
4914         * This must happen before css is disassociated with its cgroup.
4915         * See seq_css() for details.
4916         */
4917        css_clear_dir(css);
4918
4919        /*
4920         * Killing would put the base ref, but we need to keep it alive
4921         * until after ->css_offline().
4922         */
4923        css_get(css);
4924
4925        /*
4926         * cgroup core guarantees that, by the time ->css_offline() is
4927         * invoked, no new css reference will be given out via
4928         * css_tryget_online().  We can't simply call percpu_ref_kill() and
4929         * proceed to offlining css's because percpu_ref_kill() doesn't
4930         * guarantee that the ref is seen as killed on all CPUs on return.
4931         *
4932         * Use percpu_ref_kill_and_confirm() to get notifications as each
4933         * css is confirmed to be seen as killed on all CPUs.
4934         */
4935        percpu_ref_kill_and_confirm(&css->refcnt, css_killed_ref_fn);
4936}
4937
4938/**
4939 * cgroup_destroy_locked - the first stage of cgroup destruction
4940 * @cgrp: cgroup to be destroyed
4941 *
4942 * css's make use of percpu refcnts whose killing latency shouldn't be
4943 * exposed to userland and are RCU protected.  Also, cgroup core needs to
4944 * guarantee that css_tryget_online() won't succeed by the time
4945 * ->css_offline() is invoked.  To satisfy all the requirements,
4946 * destruction is implemented in the following two steps.
4947 *
4948 * s1. Verify @cgrp can be destroyed and mark it dying.  Remove all
4949 *     userland visible parts and start killing the percpu refcnts of
4950 *     css's.  Set up so that the next stage will be kicked off once all
4951 *     the percpu refcnts are confirmed to be killed.
4952 *
4953 * s2. Invoke ->css_offline(), mark the cgroup dead and proceed with the
4954 *     rest of destruction.  Once all cgroup references are gone, the
4955 *     cgroup is RCU-freed.
4956 *
4957 * This function implements s1.  After this step, @cgrp is gone as far as
4958 * the userland is concerned and a new cgroup with the same name may be
4959 * created.  As cgroup doesn't care about the names internally, this
4960 * doesn't cause any problem.
4961 */
4962static int cgroup_destroy_locked(struct cgroup *cgrp)
4963        __releases(&cgroup_mutex) __acquires(&cgroup_mutex)
4964{
4965        struct cgroup *tcgrp, *parent = cgroup_parent(cgrp);
4966        struct cgroup_subsys_state *css;
4967        struct cgrp_cset_link *link;
4968        int ssid;
4969
4970        lockdep_assert_held(&cgroup_mutex);
4971
4972        /*
4973         * Only migration can raise populated from zero and we're already
4974         * holding cgroup_mutex.
4975         */
4976        if (cgroup_is_populated(cgrp))
4977                return -EBUSY;
4978
4979        /*
4980         * Make sure there's no live children.  We can't test emptiness of
4981         * ->self.children as dead children linger on it while being
4982         * drained; otherwise, "rmdir parent/child parent" may fail.
4983         */
4984        if (css_has_online_children(&cgrp->self))
4985                return -EBUSY;
4986
4987        /*
4988         * Mark @cgrp and the associated csets dead.  The former prevents
4989         * further task migration and child creation by disabling
4990         * cgroup_lock_live_group().  The latter makes the csets ignored by
4991         * the migration path.
4992         */
4993        cgrp->self.flags &= ~CSS_ONLINE;
4994
4995        spin_lock_irq(&css_set_lock);
4996        list_for_each_entry(link, &cgrp->cset_links, cset_link)
4997                link->cset->dead = true;
4998        spin_unlock_irq(&css_set_lock);
4999
5000        /* initiate massacre of all css's */
5001        for_each_css(css, ssid, cgrp)
5002                kill_css(css);
5003
5004        /*
5005         * Remove @cgrp directory along with the base files.  @cgrp has an
5006         * extra ref on its kn.
5007         */
5008        kernfs_remove(cgrp->kn);
5009
5010        if (parent && cgroup_is_threaded(cgrp))
5011                parent->nr_threaded_children--;
5012
5013        for (tcgrp = cgroup_parent(cgrp); tcgrp; tcgrp = cgroup_parent(tcgrp)) {
5014                tcgrp->nr_descendants--;
5015                tcgrp->nr_dying_descendants++;
5016        }
5017
5018        cgroup1_check_for_release(parent);
5019
5020        /* put the base reference */
5021        percpu_ref_kill(&cgrp->self.refcnt);
5022
5023        return 0;
5024};
5025
5026int cgroup_rmdir(struct kernfs_node *kn)
5027{
5028        struct cgroup *cgrp;
5029        int ret = 0;
5030
5031        cgrp = cgroup_kn_lock_live(kn, false);
5032        if (!cgrp)
5033                return 0;
5034
5035        ret = cgroup_destroy_locked(cgrp);
5036
5037        if (!ret)
5038                trace_cgroup_rmdir(cgrp);
5039
5040        cgroup_kn_unlock(kn);
5041        return ret;
5042}
5043
5044static struct kernfs_syscall_ops cgroup_kf_syscall_ops = {
5045        .show_options           = cgroup_show_options,
5046        .remount_fs             = cgroup_remount,
5047        .mkdir                  = cgroup_mkdir,
5048        .rmdir                  = cgroup_rmdir,
5049        .show_path              = cgroup_show_path,
5050};
5051
5052static void __init cgroup_init_subsys(struct cgroup_subsys *ss, bool early)
5053{
5054        struct cgroup_subsys_state *css;
5055
5056        pr_debug("Initializing cgroup subsys %s\n", ss->name);
5057
5058        mutex_lock(&cgroup_mutex);
5059
5060        idr_init(&ss->css_idr);
5061        INIT_LIST_HEAD(&ss->cfts);
5062
5063        /* Create the root cgroup state for this subsystem */
5064        ss->root = &cgrp_dfl_root;
5065        css = ss->css_alloc(cgroup_css(&cgrp_dfl_root.cgrp, ss));
5066        /* We don't handle early failures gracefully */
5067        BUG_ON(IS_ERR(css));
5068        init_and_link_css(css, ss, &cgrp_dfl_root.cgrp);
5069
5070        /*
5071         * Root csses are never destroyed and we can't initialize
5072         * percpu_ref during early init.  Disable refcnting.
5073         */
5074        css->flags |= CSS_NO_REF;
5075
5076        if (early) {
5077                /* allocation can't be done safely during early init */
5078                css->id = 1;
5079        } else {
5080                css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2, GFP_KERNEL);
5081                BUG_ON(css->id < 0);
5082        }
5083
5084        /* Update the init_css_set to contain a subsys
5085         * pointer to this state - since the subsystem is
5086         * newly registered, all tasks and hence the
5087         * init_css_set is in the subsystem's root cgroup. */
5088        init_css_set.subsys[ss->id] = css;
5089
5090        have_fork_callback |= (bool)ss->fork << ss->id;
5091        have_exit_callback |= (bool)ss->exit << ss->id;
5092        have_free_callback |= (bool)ss->free << ss->id;
5093        have_canfork_callback |= (bool)ss->can_fork << ss->id;
5094
5095        /* At system boot, before all subsystems have been
5096         * registered, no tasks have been forked, so we don't
5097         * need to invoke fork callbacks here. */
5098        BUG_ON(!list_empty(&init_task.tasks));
5099
5100        BUG_ON(online_css(css));
5101
5102        mutex_unlock(&cgroup_mutex);
5103}
5104
5105/**
5106 * cgroup_init_early - cgroup initialization at system boot
5107 *
5108 * Initialize cgroups at system boot, and initialize any
5109 * subsystems that request early init.
5110 */
5111int __init cgroup_init_early(void)
5112{
5113        static struct cgroup_sb_opts __initdata opts;
5114        struct cgroup_subsys *ss;
5115        int i;
5116
5117        init_cgroup_root(&cgrp_dfl_root, &opts);
5118        cgrp_dfl_root.cgrp.self.flags |= CSS_NO_REF;
5119
5120        RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
5121
5122        for_each_subsys(ss, i) {
5123                WARN(!ss->css_alloc || !ss->css_free || ss->name || ss->id,
5124                     "invalid cgroup_subsys %d:%s css_alloc=%p css_free=%p id:name=%d:%s\n",
5125                     i, cgroup_subsys_name[i], ss->css_alloc, ss->css_free,
5126                     ss->id, ss->name);
5127                WARN(strlen(cgroup_subsys_name[i]) > MAX_CGROUP_TYPE_NAMELEN,
5128                     "cgroup_subsys_name %s too long\n", cgroup_subsys_name[i]);
5129
5130                ss->id = i;
5131                ss->name = cgroup_subsys_name[i];
5132                if (!ss->legacy_name)
5133                        ss->legacy_name = cgroup_subsys_name[i];
5134
5135                if (ss->early_init)
5136                        cgroup_init_subsys(ss, true);
5137        }
5138        return 0;
5139}
5140
5141static u16 cgroup_disable_mask __initdata;
5142
5143/**
5144 * cgroup_init - cgroup initialization
5145 *
5146 * Register cgroup filesystem and /proc file, and initialize
5147 * any subsystems that didn't request early init.
5148 */
5149int __init cgroup_init(void)
5150{
5151        struct cgroup_subsys *ss;
5152        int ssid;
5153
5154        BUILD_BUG_ON(CGROUP_SUBSYS_COUNT > 16);
5155        BUG_ON(percpu_init_rwsem(&cgroup_threadgroup_rwsem));
5156        BUG_ON(cgroup_init_cftypes(NULL, cgroup_base_files));
5157        BUG_ON(cgroup_init_cftypes(NULL, cgroup1_base_files));
5158
5159        /*
5160         * The latency of the synchronize_sched() is too high for cgroups,
5161         * avoid it at the cost of forcing all readers into the slow path.
5162         */
5163        rcu_sync_enter_start(&cgroup_threadgroup_rwsem.rss);
5164
5165        get_user_ns(init_cgroup_ns.user_ns);
5166
5167        mutex_lock(&cgroup_mutex);
5168
5169        /*
5170         * Add init_css_set to the hash table so that dfl_root can link to
5171         * it during init.
5172         */
5173        hash_add(css_set_table, &init_css_set.hlist,
5174                 css_set_hash(init_css_set.subsys));
5175
5176        BUG_ON(cgroup_setup_root(&cgrp_dfl_root, 0, 0));
5177
5178        mutex_unlock(&cgroup_mutex);
5179
5180        for_each_subsys(ss, ssid) {
5181                if (ss->early_init) {
5182                        struct cgroup_subsys_state *css =
5183                                init_css_set.subsys[ss->id];
5184
5185                        css->id = cgroup_idr_alloc(&ss->css_idr, css, 1, 2,
5186                                                   GFP_KERNEL);
5187                        BUG_ON(css->id < 0);
5188                } else {
5189                        cgroup_init_subsys(ss, false);
5190                }
5191
5192                list_add_tail(&init_css_set.e_cset_node[ssid],
5193                              &cgrp_dfl_root.cgrp.e_csets[ssid]);
5194
5195                /*
5196                 * Setting dfl_root subsys_mask needs to consider the
5197                 * disabled flag and cftype registration needs kmalloc,
5198                 * both of which aren't available during early_init.
5199                 */
5200                if (cgroup_disable_mask & (1 << ssid)) {
5201                        static_branch_disable(cgroup_subsys_enabled_key[ssid]);
5202                        printk(KERN_INFO "Disabling %s control group subsystem\n",
5203                               ss->name);
5204                        continue;
5205                }
5206
5207                if (cgroup1_ssid_disabled(ssid))
5208                        printk(KERN_INFO "Disabling %s control group subsystem in v1 mounts\n",
5209                               ss->name);
5210
5211                cgrp_dfl_root.subsys_mask |= 1 << ss->id;
5212
5213                /* implicit controllers must be threaded too */
5214                WARN_ON(ss->implicit_on_dfl && !ss->threaded);
5215
5216                if (ss->implicit_on_dfl)
5217                        cgrp_dfl_implicit_ss_mask |= 1 << ss->id;
5218                else if (!ss->dfl_cftypes)
5219                        cgrp_dfl_inhibit_ss_mask |= 1 << ss->id;
5220
5221                if (ss->threaded)
5222                        cgrp_dfl_threaded_ss_mask |= 1 << ss->id;
5223
5224                if (ss->dfl_cftypes == ss->legacy_cftypes) {
5225                        WARN_ON(cgroup_add_cftypes(ss, ss->dfl_cftypes));
5226                } else {
5227                        WARN_ON(cgroup_add_dfl_cftypes(ss, ss->dfl_cftypes));
5228                        WARN_ON(cgroup_add_legacy_cftypes(ss, ss->legacy_cftypes));
5229                }
5230
5231                if (ss->bind)
5232                        ss->bind(init_css_set.subsys[ssid]);
5233
5234                mutex_lock(&cgroup_mutex);
5235                css_populate_dir(init_css_set.subsys[ssid]);
5236                mutex_unlock(&cgroup_mutex);
5237        }
5238
5239        /* init_css_set.subsys[] has been updated, re-hash */
5240        hash_del(&init_css_set.hlist);
5241        hash_add(css_set_table, &init_css_set.hlist,
5242                 css_set_hash(init_css_set.subsys));
5243
5244        WARN_ON(sysfs_create_mount_point(fs_kobj, "cgroup"));
5245        WARN_ON(register_filesystem(&cgroup_fs_type));
5246        WARN_ON(register_filesystem(&cgroup2_fs_type));
5247        WARN_ON(!proc_create("cgroups", 0, NULL, &proc_cgroupstats_operations));
5248
5249        return 0;
5250}
5251
5252static int __init cgroup_wq_init(void)
5253{
5254        /*
5255         * There isn't much point in executing destruction path in
5256         * parallel.  Good chunk is serialized with cgroup_mutex anyway.
5257         * Use 1 for @max_active.
5258         *
5259         * We would prefer to do this in cgroup_init() above, but that
5260         * is called before init_workqueues(): so leave this until after.
5261         */
5262        cgroup_destroy_wq = alloc_workqueue("cgroup_destroy", 0, 1);
5263        BUG_ON(!cgroup_destroy_wq);
5264        return 0;
5265}
5266core_initcall(cgroup_wq_init);
5267
5268void cgroup_path_from_kernfs_id(const union kernfs_node_id *id,
5269                                        char *buf, size_t buflen)
5270{
5271        struct kernfs_node *kn;
5272
5273        kn = kernfs_get_node_by_id(cgrp_dfl_root.kf_root, id);
5274        if (!kn)
5275                return;
5276        kernfs_path(kn, buf, buflen);
5277        kernfs_put(kn);
5278}
5279
5280/*
5281 * proc_cgroup_show()
5282 *  - Print task's cgroup paths into seq_file, one line for each hierarchy
5283 *  - Used for /proc/<pid>/cgroup.
5284 */
5285int proc_cgroup_show(struct seq_file *m, struct pid_namespace *ns,
5286                     struct pid *pid, struct task_struct *tsk)
5287{
5288        char *buf;
5289        int retval;
5290        struct cgroup_root *root;
5291
5292        retval = -ENOMEM;
5293        buf = kmalloc(PATH_MAX, GFP_KERNEL);
5294        if (!buf)
5295                goto out;
5296
5297        mutex_lock(&cgroup_mutex);
5298        spin_lock_irq(&css_set_lock);
5299
5300        for_each_root(root) {
5301                struct cgroup_subsys *ss;
5302                struct cgroup *cgrp;
5303                int ssid, count = 0;
5304
5305                if (root == &cgrp_dfl_root && !cgrp_dfl_visible)
5306                        continue;
5307
5308                seq_printf(m, "%d:", root->hierarchy_id);
5309                if (root != &cgrp_dfl_root)
5310                        for_each_subsys(ss, ssid)
5311                                if (root->subsys_mask & (1 << ssid))
5312                                        seq_printf(m, "%s%s", count++ ? "," : "",
5313                                                   ss->legacy_name);
5314                if (strlen(root->name))
5315                        seq_printf(m, "%sname=%s", count ? "," : "",
5316                                   root->name);
5317                seq_putc(m, ':');
5318
5319                cgrp = task_cgroup_from_root(tsk, root);
5320
5321                /*
5322                 * On traditional hierarchies, all zombie tasks show up as
5323                 * belonging to the root cgroup.  On the default hierarchy,
5324                 * while a zombie doesn't show up in "cgroup.procs" and
5325                 * thus can't be migrated, its /proc/PID/cgroup keeps
5326                 * reporting the cgroup it belonged to before exiting.  If
5327                 * the cgroup is removed before the zombie is reaped,
5328                 * " (deleted)" is appended to the cgroup path.
5329                 */
5330                if (cgroup_on_dfl(cgrp) || !(tsk->flags & PF_EXITING)) {
5331                        retval = cgroup_path_ns_locked(cgrp, buf, PATH_MAX,
5332                                                current->nsproxy->cgroup_ns);
5333                        if (retval >= PATH_MAX)
5334                                retval = -ENAMETOOLONG;
5335                        if (retval < 0)
5336                                goto out_unlock;
5337
5338                        seq_puts(m, buf);
5339                } else {
5340                        seq_puts(m, "/");
5341                }
5342
5343                if (cgroup_on_dfl(cgrp) && cgroup_is_dead(cgrp))
5344                        seq_puts(m, " (deleted)\n");
5345                else
5346                        seq_putc(m, '\n');
5347        }
5348
5349        retval = 0;
5350out_unlock:
5351        spin_unlock_irq(&css_set_lock);
5352        mutex_unlock(&cgroup_mutex);
5353        kfree(buf);
5354out:
5355        return retval;
5356}
5357
5358/**
5359 * cgroup_fork - initialize cgroup related fields during copy_process()
5360 * @child: pointer to task_struct of forking parent process.
5361 *
5362 * A task is associated with the init_css_set until cgroup_post_fork()
5363 * attaches it to the parent's css_set.  Empty cg_list indicates that
5364 * @child isn't holding reference to its css_set.
5365 */
5366void cgroup_fork(struct task_struct *child)
5367{
5368        RCU_INIT_POINTER(child->cgroups, &init_css_set);
5369        INIT_LIST_HEAD(&child->cg_list);
5370}
5371
5372/**
5373 * cgroup_can_fork - called on a new task before the process is exposed
5374 * @child: the task in question.
5375 *
5376 * This calls the subsystem can_fork() callbacks. If the can_fork() callback
5377 * returns an error, the fork aborts with that error code. This allows for
5378 * a cgroup subsystem to conditionally allow or deny new forks.
5379 */
5380int cgroup_can_fork(struct task_struct *child)
5381{
5382        struct cgroup_subsys *ss;
5383        int i, j, ret;
5384
5385        do_each_subsys_mask(ss, i, have_canfork_callback) {
5386                ret = ss->can_fork(child);
5387                if (ret)
5388                        goto out_revert;
5389        } while_each_subsys_mask();
5390
5391        return 0;
5392
5393out_revert:
5394        for_each_subsys(ss, j) {
5395                if (j >= i)
5396                        break;
5397                if (ss->cancel_fork)
5398                        ss->cancel_fork(child);
5399        }
5400
5401        return ret;
5402}
5403
5404/**
5405 * cgroup_cancel_fork - called if a fork failed after cgroup_can_fork()
5406 * @child: the task in question
5407 *
5408 * This calls the cancel_fork() callbacks if a fork failed *after*
5409 * cgroup_can_fork() succeded.
5410 */
5411void cgroup_cancel_fork(struct task_struct *child)
5412{
5413        struct cgroup_subsys *ss;
5414        int i;
5415
5416        for_each_subsys(ss, i)
5417                if (ss->cancel_fork)
5418                        ss->cancel_fork(child);
5419}
5420
5421/**
5422 * cgroup_post_fork - called on a new task after adding it to the task list
5423 * @child: the task in question
5424 *
5425 * Adds the task to the list running through its css_set if necessary and
5426 * call the subsystem fork() callbacks.  Has to be after the task is
5427 * visible on the task list in case we race with the first call to
5428 * cgroup_task_iter_start() - to guarantee that the new task ends up on its
5429 * list.
5430 */
5431void cgroup_post_fork(struct task_struct *child)
5432{
5433        struct cgroup_subsys *ss;
5434        int i;
5435
5436        /*
5437         * This may race against cgroup_enable_task_cg_lists().  As that
5438         * function sets use_task_css_set_links before grabbing
5439         * tasklist_lock and we just went through tasklist_lock to add
5440         * @child, it's guaranteed that either we see the set
5441         * use_task_css_set_links or cgroup_enable_task_cg_lists() sees
5442         * @child during its iteration.
5443         *
5444         * If we won the race, @child is associated with %current's
5445         * css_set.  Grabbing css_set_lock guarantees both that the
5446         * association is stable, and, on completion of the parent's
5447         * migration, @child is visible in the source of migration or
5448         * already in the destination cgroup.  This guarantee is necessary
5449         * when implementing operations which need to migrate all tasks of
5450         * a cgroup to another.
5451         *
5452         * Note that if we lose to cgroup_enable_task_cg_lists(), @child
5453         * will remain in init_css_set.  This is safe because all tasks are
5454         * in the init_css_set before cg_links is enabled and there's no
5455         * operation which transfers all tasks out of init_css_set.
5456         */
5457        if (use_task_css_set_links) {
5458                struct css_set *cset;
5459
5460                spin_lock_irq(&css_set_lock);
5461                cset = task_css_set(current);
5462                if (list_empty(&child->cg_list)) {
5463                        get_css_set(cset);
5464                        cset->nr_tasks++;
5465                        css_set_move_task(child, NULL, cset, false);
5466                }
5467                spin_unlock_irq(&css_set_lock);
5468        }
5469
5470        /*
5471         * Call ss->fork().  This must happen after @child is linked on
5472         * css_set; otherwise, @child might change state between ->fork()
5473         * and addition to css_set.
5474         */
5475        do_each_subsys_mask(ss, i, have_fork_callback) {
5476                ss->fork(child);
5477        } while_each_subsys_mask();
5478}
5479
5480/**
5481 * cgroup_exit - detach cgroup from exiting task
5482 * @tsk: pointer to task_struct of exiting process
5483 *
5484 * Description: Detach cgroup from @tsk and release it.
5485 *
5486 * Note that cgroups marked notify_on_release force every task in
5487 * them to take the global cgroup_mutex mutex when exiting.
5488 * This could impact scaling on very large systems.  Be reluctant to
5489 * use notify_on_release cgroups where very high task exit scaling
5490 * is required on large systems.
5491 *
5492 * We set the exiting tasks cgroup to the root cgroup (top_cgroup).  We
5493 * call cgroup_exit() while the task is still competent to handle
5494 * notify_on_release(), then leave the task attached to the root cgroup in
5495 * each hierarchy for the remainder of its exit.  No need to bother with
5496 * init_css_set refcnting.  init_css_set never goes away and we can't race
5497 * with migration path - PF_EXITING is visible to migration path.
5498 */
5499void cgroup_exit(struct task_struct *tsk)
5500{
5501        struct cgroup_subsys *ss;
5502        struct css_set *cset;
5503        int i;
5504
5505        /*
5506         * Unlink from @tsk from its css_set.  As migration path can't race
5507         * with us, we can check css_set and cg_list without synchronization.
5508         */
5509        cset = task_css_set(tsk);
5510
5511        if (!list_empty(&tsk->cg_list)) {
5512                spin_lock_irq(&css_set_lock);
5513                css_set_move_task(tsk, cset, NULL, false);
5514                cset->nr_tasks--;
5515                spin_unlock_irq(&css_set_lock);
5516        } else {
5517                get_css_set(cset);
5518        }
5519
5520        /* see cgroup_post_fork() for details */
5521        do_each_subsys_mask(ss, i, have_exit_callback) {
5522                ss->exit(tsk);
5523        } while_each_subsys_mask();
5524}
5525
5526void cgroup_free(struct task_struct *task)
5527{
5528        struct css_set *cset = task_css_set(task);
5529        struct cgroup_subsys *ss;
5530        int ssid;
5531
5532        do_each_subsys_mask(ss, ssid, have_free_callback) {
5533                ss->free(task);
5534        } while_each_subsys_mask();
5535
5536        put_css_set(cset);
5537}
5538
5539static int __init cgroup_disable(char *str)
5540{
5541        struct cgroup_subsys *ss;
5542        char *token;
5543        int i;
5544
5545        while ((token = strsep(&str, ",")) != NULL) {
5546                if (!*token)
5547                        continue;
5548
5549                for_each_subsys(ss, i) {
5550                        if (strcmp(token, ss->name) &&
5551                            strcmp(token, ss->legacy_name))
5552                                continue;
5553                        cgroup_disable_mask |= 1 << i;
5554                }
5555        }
5556        return 1;
5557}
5558__setup("cgroup_disable=", cgroup_disable);
5559
5560/**
5561 * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
5562 * @dentry: directory dentry of interest
5563 * @ss: subsystem of interest
5564 *
5565 * If @dentry is a directory for a cgroup which has @ss enabled on it, try
5566 * to get the corresponding css and return it.  If such css doesn't exist
5567 * or can't be pinned, an ERR_PTR value is returned.
5568 */
5569struct cgroup_subsys_state *css_tryget_online_from_dir(struct dentry *dentry,
5570                                                       struct cgroup_subsys *ss)
5571{
5572        struct kernfs_node *kn = kernfs_node_from_dentry(dentry);
5573        struct file_system_type *s_type = dentry->d_sb->s_type;
5574        struct cgroup_subsys_state *css = NULL;
5575        struct cgroup *cgrp;
5576
5577        /* is @dentry a cgroup dir? */
5578        if ((s_type != &cgroup_fs_type && s_type != &cgroup2_fs_type) ||
5579            !kn || kernfs_type(kn) != KERNFS_DIR)
5580                return ERR_PTR(-EBADF);
5581
5582        rcu_read_lock();
5583
5584        /*
5585         * This path doesn't originate from kernfs and @kn could already
5586         * have been or be removed at any point.  @kn->priv is RCU
5587         * protected for this access.  See css_release_work_fn() for details.
5588         */
5589        cgrp = rcu_dereference(*(void __rcu __force **)&kn->priv);
5590        if (cgrp)
5591                css = cgroup_css(cgrp, ss);
5592
5593        if (!css || !css_tryget_online(css))
5594                css = ERR_PTR(-ENOENT);
5595
5596        rcu_read_unlock();
5597        return css;
5598}
5599
5600/**
5601 * css_from_id - lookup css by id
5602 * @id: the cgroup id
5603 * @ss: cgroup subsys to be looked into
5604 *
5605 * Returns the css if there's valid one with @id, otherwise returns NULL.
5606 * Should be called under rcu_read_lock().
5607 */
5608struct cgroup_subsys_state *css_from_id(int id, struct cgroup_subsys *ss)
5609{
5610        WARN_ON_ONCE(!rcu_read_lock_held());
5611        return idr_find(&ss->css_idr, id);
5612}
5613
5614/**
5615 * cgroup_get_from_path - lookup and get a cgroup from its default hierarchy path
5616 * @path: path on the default hierarchy
5617 *
5618 * Find the cgroup at @path on the default hierarchy, increment its
5619 * reference count and return it.  Returns pointer to the found cgroup on
5620 * success, ERR_PTR(-ENOENT) if @path doens't exist and ERR_PTR(-ENOTDIR)
5621 * if @path points to a non-directory.
5622 */
5623struct cgroup *cgroup_get_from_path(const char *path)
5624{
5625        struct kernfs_node *kn;
5626        struct cgroup *cgrp;
5627
5628        mutex_lock(&cgroup_mutex);
5629
5630        kn = kernfs_walk_and_get(cgrp_dfl_root.cgrp.kn, path);
5631        if (kn) {
5632                if (kernfs_type(kn) == KERNFS_DIR) {
5633                        cgrp = kn->priv;
5634                        cgroup_get_live(cgrp);
5635                } else {
5636                        cgrp = ERR_PTR(-ENOTDIR);
5637                }
5638                kernfs_put(kn);
5639        } else {
5640                cgrp = ERR_PTR(-ENOENT);
5641        }
5642
5643        mutex_unlock(&cgroup_mutex);
5644        return cgrp;
5645}
5646EXPORT_SYMBOL_GPL(cgroup_get_from_path);
5647
5648/**
5649 * cgroup_get_from_fd - get a cgroup pointer from a fd
5650 * @fd: fd obtained by open(cgroup2_dir)
5651 *
5652 * Find the cgroup from a fd which should be obtained
5653 * by opening a cgroup directory.  Returns a pointer to the
5654 * cgroup on success. ERR_PTR is returned if the cgroup
5655 * cannot be found.
5656 */
5657struct cgroup *cgroup_get_from_fd(int fd)
5658{
5659        struct cgroup_subsys_state *css;
5660        struct cgroup *cgrp;
5661        struct file *f;
5662
5663        f = fget_raw(fd);
5664        if (!f)
5665                return ERR_PTR(-EBADF);
5666
5667        css = css_tryget_online_from_dir(f->f_path.dentry, NULL);
5668        fput(f);
5669        if (IS_ERR(css))
5670                return ERR_CAST(css);
5671
5672        cgrp = css->cgroup;
5673        if (!cgroup_on_dfl(cgrp)) {
5674                cgroup_put(cgrp);
5675                return ERR_PTR(-EBADF);
5676        }
5677
5678        return cgrp;
5679}
5680EXPORT_SYMBOL_GPL(cgroup_get_from_fd);
5681
5682/*
5683 * sock->sk_cgrp_data handling.  For more info, see sock_cgroup_data
5684 * definition in cgroup-defs.h.
5685 */
5686#ifdef CONFIG_SOCK_CGROUP_DATA
5687
5688#if defined(CONFIG_CGROUP_NET_PRIO) || defined(CONFIG_CGROUP_NET_CLASSID)
5689
5690DEFINE_SPINLOCK(cgroup_sk_update_lock);
5691static bool cgroup_sk_alloc_disabled __read_mostly;
5692
5693void cgroup_sk_alloc_disable(void)
5694{
5695        if (cgroup_sk_alloc_disabled)
5696                return;
5697        pr_info("cgroup: disabling cgroup2 socket matching due to net_prio or net_cls activation\n");
5698        cgroup_sk_alloc_disabled = true;
5699}
5700
5701#else
5702
5703#define cgroup_sk_alloc_disabled        false
5704
5705#endif
5706
5707void cgroup_sk_alloc(struct sock_cgroup_data *skcd)
5708{
5709        if (cgroup_sk_alloc_disabled)
5710                return;
5711
5712        /* Socket clone path */
5713        if (skcd->val) {
5714                /*
5715                 * We might be cloning a socket which is left in an empty
5716                 * cgroup and the cgroup might have already been rmdir'd.
5717                 * Don't use cgroup_get_live().
5718                 */
5719                cgroup_get(sock_cgroup_ptr(skcd));
5720                return;
5721        }
5722
5723        rcu_read_lock();
5724
5725        while (true) {
5726                struct css_set *cset;
5727
5728                cset = task_css_set(current);
5729                if (likely(cgroup_tryget(cset->dfl_cgrp))) {
5730                        skcd->val = (unsigned long)cset->dfl_cgrp;
5731                        break;
5732                }
5733                cpu_relax();
5734        }
5735
5736        rcu_read_unlock();
5737}
5738
5739void cgroup_sk_free(struct sock_cgroup_data *skcd)
5740{
5741        cgroup_put(sock_cgroup_ptr(skcd));
5742}
5743
5744#endif  /* CONFIG_SOCK_CGROUP_DATA */
5745
5746#ifdef CONFIG_CGROUP_BPF
5747int cgroup_bpf_update(struct cgroup *cgrp, struct bpf_prog *prog,
5748                      enum bpf_attach_type type, bool overridable)
5749{
5750        struct cgroup *parent = cgroup_parent(cgrp);
5751        int ret;
5752
5753        mutex_lock(&cgroup_mutex);
5754        ret = __cgroup_bpf_update(cgrp, parent, prog, type, overridable);
5755        mutex_unlock(&cgroup_mutex);
5756        return ret;
5757}
5758#endif /* CONFIG_CGROUP_BPF */
5759