include/linux/cgroup-defs.h

   1 /*
   2  * linux/cgroup-defs.h - basic definitions for cgroup
   3  *
   4  * This file provides basic type and interface.  Include this file directly
   5  * only if necessary to avoid cyclic dependencies.
   6  */
   7 #ifndef _LINUX_CGROUP_DEFS_H
   8 #define _LINUX_CGROUP_DEFS_H
   9
  10 #include <linux/limits.h>
  11 #include <linux/list.h>
  12 #include <linux/idr.h>
  13 #include <linux/wait.h>
  14 #include <linux/mutex.h>
  15 #include <linux/rcupdate.h>
  16 #include <linux/refcount.h>
  17 #include <linux/percpu-refcount.h>
  18 #include <linux/percpu-rwsem.h>
  19 #include <linux/workqueue.h>
  20 #include <linux/bpf-cgroup.h>
  21
  22 #ifdef CONFIG_CGROUPS
  23
  24 struct cgroup;
  25 struct cgroup_root;
  26 struct cgroup_subsys;
  27 struct cgroup_taskset;
  28 struct kernfs_node;
  29 struct kernfs_ops;
  30 struct kernfs_open_file;
  31 struct seq_file;
  32
  33 #define MAX_CGROUP_TYPE_NAMELEN 32
  34 #define MAX_CGROUP_ROOT_NAMELEN 64
  35 #define MAX_CFTYPE_NAME         64
  36
  37 /* define the enumeration of all cgroup subsystems */
  38 #define SUBSYS(_x) _x ## _cgrp_id,
  39 enum cgroup_subsys_id {
  40 #include <linux/cgroup_subsys.h>
  41         CGROUP_SUBSYS_COUNT,
  42 };
  43 #undef SUBSYS
  44
  45 /* bits in struct cgroup_subsys_state flags field */
  46 enum {
  47         CSS_NO_REF      = (1 << 0), /* no reference counting for this css */
  48         CSS_ONLINE      = (1 << 1), /* between ->css_online() and ->css_offline() */
  49         CSS_RELEASED    = (1 << 2), /* refcnt reached zero, released */
  50         CSS_VISIBLE     = (1 << 3), /* css is visible to userland */
  51         CSS_DYING       = (1 << 4), /* css is dying */
  52 };
  53
  54 /* bits in struct cgroup flags field */
  55 enum {
  56         /* Control Group requires release notifications to userspace */
  57         CGRP_NOTIFY_ON_RELEASE,
  58         /*
  59          * Clone the parent's configuration when creating a new child
  60          * cpuset cgroup.  For historical reasons, this option can be
  61          * specified at mount time and thus is implemented here.
  62          */
  63         CGRP_CPUSET_CLONE_CHILDREN,
  64 };
  65
  66 /* cgroup_root->flags */
  67 enum {
  68         CGRP_ROOT_NOPREFIX      = (1 << 1), /* mounted subsystems have no named prefix */
  69         CGRP_ROOT_XATTR         = (1 << 2), /* supports extended attributes */
  70
  71         /*
  72          * Consider namespaces as delegation boundaries.  If this flag is
  73          * set, controller specific interface files in a namespace root
  74          * aren't writeable from inside the namespace.
  75          */
  76         CGRP_ROOT_NS_DELEGATE   = (1 << 3),
  77
  78         /*
  79          * Enable cpuset controller in v1 cgroup to use v2 behavior.
  80          */
  81         CGRP_ROOT_CPUSET_V2_MODE = (1 << 4),
  82 };
  83
  84 /* cftype->flags */
  85 enum {
  86         CFTYPE_ONLY_ON_ROOT     = (1 << 0),     /* only create on root cgrp */
  87         CFTYPE_NOT_ON_ROOT      = (1 << 1),     /* don't create on root cgrp */
  88         CFTYPE_NS_DELEGATABLE   = (1 << 2),     /* writeable beyond delegation boundaries */
  89
  90         CFTYPE_NO_PREFIX        = (1 << 3),     /* (DON'T USE FOR NEW FILES) no subsys prefix */
  91         CFTYPE_WORLD_WRITABLE   = (1 << 4),     /* (DON'T USE FOR NEW FILES) S_IWUGO */
  92
  93         /* internal flags, do not use outside cgroup core proper */
  94         __CFTYPE_ONLY_ON_DFL    = (1 << 16),    /* only on default hierarchy */
  95         __CFTYPE_NOT_ON_DFL     = (1 << 17),    /* not on default hierarchy */
  96 };
  97
  98 /*
  99  * cgroup_file is the handle for a file instance created in a cgroup which
 100  * is used, for example, to generate file changed notifications.  This can
 101  * be obtained by setting cftype->file_offset.
 102  */
 103 struct cgroup_file {
 104         /* do not access any fields from outside cgroup core */
 105         struct kernfs_node *kn;
 106 };
 107
 108 /*
 109  * Per-subsystem/per-cgroup state maintained by the system.  This is the
 110  * fundamental structural building block that controllers deal with.
 111  *
 112  * Fields marked with "PI:" are public and immutable and may be accessed
 113  * directly without synchronization.
 114  */
 115 struct cgroup_subsys_state {
 116         /* PI: the cgroup that this css is attached to */
 117         struct cgroup *cgroup;
 118
 119         /* PI: the cgroup subsystem that this css is attached to */
 120         struct cgroup_subsys *ss;
 121
 122         /* reference count - access via css_[try]get() and css_put() */
 123         struct percpu_ref refcnt;
 124
 125         /* siblings list anchored at the parent's ->children */
 126         struct list_head sibling;
 127         struct list_head children;
 128
 129         /*
 130          * PI: Subsys-unique ID.  0 is unused and root is always 1.  The
 131          * matching css can be looked up using css_from_id().
 132          */
 133         int id;
 134
 135         unsigned int flags;
 136
 137         /*
 138          * Monotonically increasing unique serial number which defines a
 139          * uniform order among all csses.  It's guaranteed that all
 140          * ->children lists are in the ascending order of ->serial_nr and
 141          * used to allow interrupting and resuming iterations.
 142          */
 143         u64 serial_nr;
 144
 145         /*
 146          * Incremented by online self and children.  Used to guarantee that
 147          * parents are not offlined before their children.
 148          */
 149         atomic_t online_cnt;
 150
 151         /* percpu_ref killing and RCU release */
 152         struct rcu_head rcu_head;
 153         struct work_struct destroy_work;
 154
 155         /*
 156          * PI: the parent css.  Placed here for cache proximity to following
 157          * fields of the containing structure.
 158          */
 159         struct cgroup_subsys_state *parent;
 160 };
 161
 162 /*
 163  * A css_set is a structure holding pointers to a set of
 164  * cgroup_subsys_state objects. This saves space in the task struct
 165  * object and speeds up fork()/exit(), since a single inc/dec and a
 166  * list_add()/del() can bump the reference count on the entire cgroup
 167  * set for a task.
 168  */
 169 struct css_set {
 170         /*
 171          * Set of subsystem states, one for each subsystem. This array is
 172          * immutable after creation apart from the init_css_set during
 173          * subsystem registration (at boot time).
 174          */
 175         struct cgroup_subsys_state *subsys[CGROUP_SUBSYS_COUNT];
 176
 177         /* reference count */
 178         refcount_t refcount;
 179
 180         /*
 181          * For a domain cgroup, the following points to self.  If threaded,
 182          * to the matching cset of the nearest domain ancestor.  The
 183          * dom_cset provides access to the domain cgroup and its csses to
 184          * which domain level resource consumptions should be charged.
 185          */
 186         struct css_set *dom_cset;
 187
 188         /* the default cgroup associated with this css_set */
 189         struct cgroup *dfl_cgrp;
 190
 191         /* internal task count, protected by css_set_lock */
 192         int nr_tasks;
 193
 194         /*
 195          * Lists running through all tasks using this cgroup group.
 196          * mg_tasks lists tasks which belong to this cset but are in the
 197          * process of being migrated out or in.  Protected by
 198          * css_set_rwsem, but, during migration, once tasks are moved to
 199          * mg_tasks, it can be read safely while holding cgroup_mutex.
 200          */
 201         struct list_head tasks;
 202         struct list_head mg_tasks;
 203
 204         /* all css_task_iters currently walking this cset */
 205         struct list_head task_iters;
 206
 207         /*
 208          * On the default hierarhcy, ->subsys[ssid] may point to a css
 209          * attached to an ancestor instead of the cgroup this css_set is
 210          * associated with.  The following node is anchored at
 211          * ->subsys[ssid]->cgroup->e_csets[ssid] and provides a way to
 212          * iterate through all css's attached to a given cgroup.
 213          */
 214         struct list_head e_cset_node[CGROUP_SUBSYS_COUNT];
 215
 216         /* all threaded csets whose ->dom_cset points to this cset */
 217         struct list_head threaded_csets;
 218         struct list_head threaded_csets_node;
 219
 220         /*
 221          * List running through all cgroup groups in the same hash
 222          * slot. Protected by css_set_lock
 223          */
 224         struct hlist_node hlist;
 225
 226         /*
 227          * List of cgrp_cset_links pointing at cgroups referenced from this
 228          * css_set.  Protected by css_set_lock.
 229          */
 230         struct list_head cgrp_links;
 231
 232         /*
 233          * List of csets participating in the on-going migration either as
 234          * source or destination.  Protected by cgroup_mutex.
 235          */
 236         struct list_head mg_preload_node;
 237         struct list_head mg_node;
 238
 239         /*
 240          * If this cset is acting as the source of migration the following
 241          * two fields are set.  mg_src_cgrp and mg_dst_cgrp are
 242          * respectively the source and destination cgroups of the on-going
 243          * migration.  mg_dst_cset is the destination cset the target tasks
 244          * on this cset should be migrated to.  Protected by cgroup_mutex.
 245          */
 246         struct cgroup *mg_src_cgrp;
 247         struct cgroup *mg_dst_cgrp;
 248         struct css_set *mg_dst_cset;
 249
 250         /* dead and being drained, ignore for migration */
 251         bool dead;
 252
 253         /* For RCU-protected deletion */
 254         struct rcu_head rcu_head;
 255 };
 256
 257 struct cgroup {
 258         /* self css with NULL ->ss, points back to this cgroup */
 259         struct cgroup_subsys_state self;
 260
 261         unsigned long flags;            /* "unsigned long" so bitops work */
 262
 263         /*
 264          * idr allocated in-hierarchy ID.
 265          *
 266          * ID 0 is not used, the ID of the root cgroup is always 1, and a
 267          * new cgroup will be assigned with a smallest available ID.
 268          *
 269          * Allocating/Removing ID must be protected by cgroup_mutex.
 270          */
 271         int id;
 272
 273         /*
 274          * The depth this cgroup is at.  The root is at depth zero and each
 275          * step down the hierarchy increments the level.  This along with
 276          * ancestor_ids[] can determine whether a given cgroup is a
 277          * descendant of another without traversing the hierarchy.
 278          */
 279         int level;
 280
 281         /* Maximum allowed descent tree depth */
 282         int max_depth;
 283
 284         /*
 285          * Keep track of total numbers of visible and dying descent cgroups.
 286          * Dying cgroups are cgroups which were deleted by a user,
 287          * but are still existing because someone else is holding a reference.
 288          * max_descendants is a maximum allowed number of descent cgroups.
 289          */
 290         int nr_descendants;
 291         int nr_dying_descendants;
 292         int max_descendants;
 293
 294         /*
 295          * Each non-empty css_set associated with this cgroup contributes
 296          * one to nr_populated_csets.  The counter is zero iff this cgroup
 297          * doesn't have any tasks.
 298          *
 299          * All children which have non-zero nr_populated_csets and/or
 300          * nr_populated_children of their own contribute one to either
 301          * nr_populated_domain_children or nr_populated_threaded_children
 302          * depending on their type.  Each counter is zero iff all cgroups
 303          * of the type in the subtree proper don't have any tasks.
 304          */
 305         int nr_populated_csets;
 306         int nr_populated_domain_children;
 307         int nr_populated_threaded_children;
 308
 309         int nr_threaded_children;       /* # of live threaded child cgroups */
 310
 311         struct kernfs_node *kn;         /* cgroup kernfs entry */
 312         struct cgroup_file procs_file;  /* handle for "cgroup.procs" */
 313         struct cgroup_file events_file; /* handle for "cgroup.events" */
 314
 315         /*
 316          * The bitmask of subsystems enabled on the child cgroups.
 317          * ->subtree_control is the one configured through
 318          * "cgroup.subtree_control" while ->child_ss_mask is the effective
 319          * one which may have more subsystems enabled.  Controller knobs
 320          * are made available iff it's enabled in ->subtree_control.
 321          */
 322         u16 subtree_control;
 323         u16 subtree_ss_mask;
 324         u16 old_subtree_control;
 325         u16 old_subtree_ss_mask;
 326
 327         /* Private pointers for each registered subsystem */
 328         struct cgroup_subsys_state __rcu *subsys[CGROUP_SUBSYS_COUNT];
 329
 330         struct cgroup_root *root;
 331
 332         /*
 333          * List of cgrp_cset_links pointing at css_sets with tasks in this
 334          * cgroup.  Protected by css_set_lock.
 335          */
 336         struct list_head cset_links;
 337
 338         /*
 339          * On the default hierarchy, a css_set for a cgroup with some
 340          * susbsys disabled will point to css's which are associated with
 341          * the closest ancestor which has the subsys enabled.  The
 342          * following lists all css_sets which point to this cgroup's css
 343          * for the given subsystem.
 344          */
 345         struct list_head e_csets[CGROUP_SUBSYS_COUNT];
 346
 347         /*
 348          * If !threaded, self.  If threaded, it points to the nearest
 349          * domain ancestor.  Inside a threaded subtree, cgroups are exempt
 350          * from process granularity and no-internal-task constraint.
 351          * Domain level resource consumptions which aren't tied to a
 352          * specific task are charged to the dom_cgrp.
 353          */
 354         struct cgroup *dom_cgrp;
 355
 356         /*
 357          * list of pidlists, up to two for each namespace (one for procs, one
 358          * for tasks); created on demand.
 359          */
 360         struct list_head pidlists;
 361         struct mutex pidlist_mutex;
 362
 363         /* used to wait for offlining of csses */
 364         wait_queue_head_t offline_waitq;
 365
 366         /* used to schedule release agent */
 367         struct work_struct release_agent_work;
 368
 369         /* used to store eBPF programs */
 370         struct cgroup_bpf bpf;
 371
 372         /* ids of the ancestors at each level including self */
 373         int ancestor_ids[];
 374 };
 375
 376 /*
 377  * A cgroup_root represents the root of a cgroup hierarchy, and may be
 378  * associated with a kernfs_root to form an active hierarchy.  This is
 379  * internal to cgroup core.  Don't access directly from controllers.
 380  */
 381 struct cgroup_root {
 382         struct kernfs_root *kf_root;
 383
 384         /* The bitmask of subsystems attached to this hierarchy */
 385         unsigned int subsys_mask;
 386
 387         /* Unique id for this hierarchy. */
 388         int hierarchy_id;
 389
 390         /* The root cgroup.  Root is destroyed on its release. */
 391         struct cgroup cgrp;
 392
 393         /* for cgrp->ancestor_ids[0] */
 394         int cgrp_ancestor_id_storage;
 395
 396         /* Number of cgroups in the hierarchy, used only for /proc/cgroups */
 397         atomic_t nr_cgrps;
 398
 399         /* A list running through the active hierarchies */
 400         struct list_head root_list;
 401
 402         /* Hierarchy-specific flags */
 403         unsigned int flags;
 404
 405         /* IDs for cgroups in this hierarchy */
 406         struct idr cgroup_idr;
 407
 408         /* The path to use for release notifications. */
 409         char release_agent_path[PATH_MAX];
 410
 411         /* The name for this hierarchy - may be empty */
 412         char name[MAX_CGROUP_ROOT_NAMELEN];
 413 };
 414
 415 /*
 416  * struct cftype: handler definitions for cgroup control files
 417  *
 418  * When reading/writing to a file:
 419  *      - the cgroup to use is file->f_path.dentry->d_parent->d_fsdata
 420  *      - the 'cftype' of the file is file->f_path.dentry->d_fsdata
 421  */
 422 struct cftype {
 423         /*
 424          * By convention, the name should begin with the name of the
 425          * subsystem, followed by a period.  Zero length string indicates
 426          * end of cftype array.
 427          */
 428         char name[MAX_CFTYPE_NAME];
 429         unsigned long private;
 430
 431         /*
 432          * The maximum length of string, excluding trailing nul, that can
 433          * be passed to write.  If < PAGE_SIZE-1, PAGE_SIZE-1 is assumed.
 434          */
 435         size_t max_write_len;
 436
 437         /* CFTYPE_* flags */
 438         unsigned int flags;
 439
 440         /*
 441          * If non-zero, should contain the offset from the start of css to
 442          * a struct cgroup_file field.  cgroup will record the handle of
 443          * the created file into it.  The recorded handle can be used as
 444          * long as the containing css remains accessible.
 445          */
 446         unsigned int file_offset;
 447
 448         /*
 449          * Fields used for internal bookkeeping.  Initialized automatically
 450          * during registration.
 451          */
 452         struct cgroup_subsys *ss;       /* NULL for cgroup core files */
 453         struct list_head node;          /* anchored at ss->cfts */
 454         struct kernfs_ops *kf_ops;
 455
 456         int (*open)(struct kernfs_open_file *of);
 457         void (*release)(struct kernfs_open_file *of);
 458
 459         /*
 460          * read_u64() is a shortcut for the common case of returning a
 461          * single integer. Use it in place of read()
 462          */
 463         u64 (*read_u64)(struct cgroup_subsys_state *css, struct cftype *cft);
 464         /*
 465          * read_s64() is a signed version of read_u64()
 466          */
 467         s64 (*read_s64)(struct cgroup_subsys_state *css, struct cftype *cft);
 468
 469         /* generic seq_file read interface */
 470         int (*seq_show)(struct seq_file *sf, void *v);
 471
 472         /* optional ops, implement all or none */
 473         void *(*seq_start)(struct seq_file *sf, loff_t *ppos);
 474         void *(*seq_next)(struct seq_file *sf, void *v, loff_t *ppos);
 475         void (*seq_stop)(struct seq_file *sf, void *v);
 476
 477         /*
 478          * write_u64() is a shortcut for the common case of accepting
 479          * a single integer (as parsed by simple_strtoull) from
 480          * userspace. Use in place of write(); return 0 or error.
 481          */
 482         int (*write_u64)(struct cgroup_subsys_state *css, struct cftype *cft,
 483                          u64 val);
 484         /*
 485          * write_s64() is a signed version of write_u64()
 486          */
 487         int (*write_s64)(struct cgroup_subsys_state *css, struct cftype *cft,
 488                          s64 val);
 489
 490         /*
 491          * write() is the generic write callback which maps directly to
 492          * kernfs write operation and overrides all other operations.
 493          * Maximum write size is determined by ->max_write_len.  Use
 494          * of_css/cft() to access the associated css and cft.
 495          */
 496         ssize_t (*write)(struct kernfs_open_file *of,
 497                          char *buf, size_t nbytes, loff_t off);
 498
 499 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 500         struct lock_class_key   lockdep_key;
 501 #endif
 502 };
 503
 504 /*
 505  * Control Group subsystem type.
 506  * See Documentation/cgroups/cgroups.txt for details
 507  */
 508 struct cgroup_subsys {
 509         struct cgroup_subsys_state *(*css_alloc)(struct cgroup_subsys_state *parent_css);
 510         int (*css_online)(struct cgroup_subsys_state *css);
 511         void (*css_offline)(struct cgroup_subsys_state *css);
 512         void (*css_released)(struct cgroup_subsys_state *css);
 513         void (*css_free)(struct cgroup_subsys_state *css);
 514         void (*css_reset)(struct cgroup_subsys_state *css);
 515
 516         int (*can_attach)(struct cgroup_taskset *tset);
 517         void (*cancel_attach)(struct cgroup_taskset *tset);
 518         void (*attach)(struct cgroup_taskset *tset);
 519         void (*post_attach)(void);
 520         int (*can_fork)(struct task_struct *task);
 521         void (*cancel_fork)(struct task_struct *task);
 522         void (*fork)(struct task_struct *task);
 523         void (*exit)(struct task_struct *task);
 524         void (*free)(struct task_struct *task);
 525         void (*bind)(struct cgroup_subsys_state *root_css);
 526
 527         bool early_init:1;
 528
 529         /*
 530          * If %true, the controller, on the default hierarchy, doesn't show
 531          * up in "cgroup.controllers" or "cgroup.subtree_control", is
 532          * implicitly enabled on all cgroups on the default hierarchy, and
 533          * bypasses the "no internal process" constraint.  This is for
 534          * utility type controllers which is transparent to userland.
 535          *
 536          * An implicit controller can be stolen from the default hierarchy
 537          * anytime and thus must be okay with offline csses from previous
 538          * hierarchies coexisting with csses for the current one.
 539          */
 540         bool implicit_on_dfl:1;
 541
 542         /*
 543          * If %true, the controller, supports threaded mode on the default
 544          * hierarchy.  In a threaded subtree, both process granularity and
 545          * no-internal-process constraint are ignored and a threaded
 546          * controllers should be able to handle that.
 547          *
 548          * Note that as an implicit controller is automatically enabled on
 549          * all cgroups on the default hierarchy, it should also be
 550          * threaded.  implicit && !threaded is not supported.
 551          */
 552         bool threaded:1;
 553
 554         /*
 555          * If %false, this subsystem is properly hierarchical -
 556          * configuration, resource accounting and restriction on a parent
 557          * cgroup cover those of its children.  If %true, hierarchy support
 558          * is broken in some ways - some subsystems ignore hierarchy
 559          * completely while others are only implemented half-way.
 560          *
 561          * It's now disallowed to create nested cgroups if the subsystem is
 562          * broken and cgroup core will emit a warning message on such
 563          * cases.  Eventually, all subsystems will be made properly
 564          * hierarchical and this will go away.
 565          */
 566         bool broken_hierarchy:1;
 567         bool warned_broken_hierarchy:1;
 568
 569         /* the following two fields are initialized automtically during boot */
 570         int id;
 571         const char *name;
 572
 573         /* optional, initialized automatically during boot if not set */
 574         const char *legacy_name;
 575
 576         /* link to parent, protected by cgroup_lock() */
 577         struct cgroup_root *root;
 578
 579         /* idr for css->id */
 580         struct idr css_idr;
 581
 582         /*
 583          * List of cftypes.  Each entry is the first entry of an array
 584          * terminated by zero length name.
 585          */
 586         struct list_head cfts;
 587
 588         /*
 589          * Base cftypes which are automatically registered.  The two can
 590          * point to the same array.
 591          */
 592         struct cftype *dfl_cftypes;     /* for the default hierarchy */
 593         struct cftype *legacy_cftypes;  /* for the legacy hierarchies */
 594
 595         /*
 596          * A subsystem may depend on other subsystems.  When such subsystem
 597          * is enabled on a cgroup, the depended-upon subsystems are enabled
 598          * together if available.  Subsystems enabled due to dependency are
 599          * not visible to userland until explicitly enabled.  The following
 600          * specifies the mask of subsystems that this one depends on.
 601          */
 602         unsigned int depends_on;
 603 };
 604
 605 extern struct percpu_rw_semaphore cgroup_threadgroup_rwsem;
 606
 607 /**
 608  * cgroup_threadgroup_change_begin - threadgroup exclusion for cgroups
 609  * @tsk: target task
 610  *
 611  * Allows cgroup operations to synchronize against threadgroup changes
 612  * using a percpu_rw_semaphore.
 613  */
 614 static inline void cgroup_threadgroup_change_begin(struct task_struct *tsk)
 615 {
 616         percpu_down_read(&cgroup_threadgroup_rwsem);
 617 }
 618
 619 /**
 620  * cgroup_threadgroup_change_end - threadgroup exclusion for cgroups
 621  * @tsk: target task
 622  *
 623  * Counterpart of cgroup_threadcgroup_change_begin().
 624  */
 625 static inline void cgroup_threadgroup_change_end(struct task_struct *tsk)
 626 {
 627         percpu_up_read(&cgroup_threadgroup_rwsem);
 628 }
 629
 630 #else   /* CONFIG_CGROUPS */
 631
 632 #define CGROUP_SUBSYS_COUNT 0
 633
 634 static inline void cgroup_threadgroup_change_begin(struct task_struct *tsk)
 635 {
 636         might_sleep();
 637 }
 638
 639 static inline void cgroup_threadgroup_change_end(struct task_struct *tsk) {}
 640
 641 #endif  /* CONFIG_CGROUPS */
 642
 643 #ifdef CONFIG_SOCK_CGROUP_DATA
 644
 645 /*
 646  * sock_cgroup_data is embedded at sock->sk_cgrp_data and contains
 647  * per-socket cgroup information except for memcg association.
 648  *
 649  * On legacy hierarchies, net_prio and net_cls controllers directly set
 650  * attributes on each sock which can then be tested by the network layer.
 651  * On the default hierarchy, each sock is associated with the cgroup it was
 652  * created in and the networking layer can match the cgroup directly.
 653  *
 654  * To avoid carrying all three cgroup related fields separately in sock,
 655  * sock_cgroup_data overloads (prioidx, classid) and the cgroup pointer.
 656  * On boot, sock_cgroup_data records the cgroup that the sock was created
 657  * in so that cgroup2 matches can be made; however, once either net_prio or
 658  * net_cls starts being used, the area is overriden to carry prioidx and/or
 659  * classid.  The two modes are distinguished by whether the lowest bit is
 660  * set.  Clear bit indicates cgroup pointer while set bit prioidx and
 661  * classid.
 662  *
 663  * While userland may start using net_prio or net_cls at any time, once
 664  * either is used, cgroup2 matching no longer works.  There is no reason to
 665  * mix the two and this is in line with how legacy and v2 compatibility is
 666  * handled.  On mode switch, cgroup references which are already being
 667  * pointed to by socks may be leaked.  While this can be remedied by adding
 668  * synchronization around sock_cgroup_data, given that the number of leaked
 669  * cgroups is bound and highly unlikely to be high, this seems to be the
 670  * better trade-off.
 671  */
 672 struct sock_cgroup_data {
 673         union {
 674 #ifdef __LITTLE_ENDIAN
 675                 struct {
 676                         u8      is_data;
 677                         u8      padding;
 678                         u16     prioidx;
 679                         u32     classid;
 680                 } __packed;
 681 #else
 682                 struct {
 683                         u32     classid;
 684                         u16     prioidx;
 685                         u8      padding;
 686                         u8      is_data;
 687                 } __packed;
 688 #endif
 689                 u64             val;
 690         };
 691 };
 692
 693 /*
 694  * There's a theoretical window where the following accessors race with
 695  * updaters and return part of the previous pointer as the prioidx or
 696  * classid.  Such races are short-lived and the result isn't critical.
 697  */
 698 static inline u16 sock_cgroup_prioidx(struct sock_cgroup_data *skcd)
 699 {
 700         /* fallback to 1 which is always the ID of the root cgroup */
 701         return (skcd->is_data & 1) ? skcd->prioidx : 1;
 702 }
 703
 704 static inline u32 sock_cgroup_classid(struct sock_cgroup_data *skcd)
 705 {
 706         /* fallback to 0 which is the unconfigured default classid */
 707         return (skcd->is_data & 1) ? skcd->classid : 0;
 708 }
 709
 710 /*
 711  * If invoked concurrently, the updaters may clobber each other.  The
 712  * caller is responsible for synchronization.
 713  */
 714 static inline void sock_cgroup_set_prioidx(struct sock_cgroup_data *skcd,
 715                                            u16 prioidx)
 716 {
 717         struct sock_cgroup_data skcd_buf = {{ .val = READ_ONCE(skcd->val) }};
 718
 719         if (sock_cgroup_prioidx(&skcd_buf) == prioidx)
 720                 return;
 721
 722         if (!(skcd_buf.is_data & 1)) {
 723                 skcd_buf.val = 0;
 724                 skcd_buf.is_data = 1;
 725         }
 726
 727         skcd_buf.prioidx = prioidx;
 728         WRITE_ONCE(skcd->val, skcd_buf.val);    /* see sock_cgroup_ptr() */
 729 }
 730
 731 static inline void sock_cgroup_set_classid(struct sock_cgroup_data *skcd,
 732                                            u32 classid)
 733 {
 734         struct sock_cgroup_data skcd_buf = {{ .val = READ_ONCE(skcd->val) }};
 735
 736         if (sock_cgroup_classid(&skcd_buf) == classid)
 737                 return;
 738
 739         if (!(skcd_buf.is_data & 1)) {
 740                 skcd_buf.val = 0;
 741                 skcd_buf.is_data = 1;
 742         }
 743
 744         skcd_buf.classid = classid;
 745         WRITE_ONCE(skcd->val, skcd_buf.val);    /* see sock_cgroup_ptr() */
 746 }
 747
 748 #else   /* CONFIG_SOCK_CGROUP_DATA */
 749
 750 struct sock_cgroup_data {
 751 };
 752
 753 #endif  /* CONFIG_SOCK_CGROUP_DATA */
 754
 755 #endif  /* _LINUX_CGROUP_DEFS_H */