1                                CGROUPS
   2                                -------
   4Written by Paul Menage <> based on
   7Original copyright statements from cpusets.txt:
   8Portions Copyright (C) 2004 BULL SA.
   9Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
  10Modified by Paul Jackson <>
  11Modified by Christoph Lameter <>
  161. Control Groups
  17  1.1 What are cgroups ?
  18  1.2 Why are cgroups needed ?
  19  1.3 How are cgroups implemented ?
  20  1.4 What does notify_on_release do ?
  21  1.5 What does clone_children do ?
  22  1.6 How do I use cgroups ?
  232. Usage Examples and Syntax
  24  2.1 Basic Usage
  25  2.2 Attaching processes
  26  2.3 Mounting hierarchies by name
  273. Kernel API
  28  3.1 Overview
  29  3.2 Synchronization
  30  3.3 Subsystem API
  314. Extended attributes usage
  325. Questions
  341. Control Groups
  371.1 What are cgroups ?
  40Control Groups provide a mechanism for aggregating/partitioning sets of
  41tasks, and all their future children, into hierarchical groups with
  42specialized behaviour.
  46A *cgroup* associates a set of tasks with a set of parameters for one
  47or more subsystems.
  49A *subsystem* is a module that makes use of the task grouping
  50facilities provided by cgroups to treat groups of tasks in
  51particular ways. A subsystem is typically a "resource controller" that
  52schedules a resource or applies per-cgroup limits, but it may be
  53anything that wants to act on a group of processes, e.g. a
  54virtualization subsystem.
  56A *hierarchy* is a set of cgroups arranged in a tree, such that
  57every task in the system is in exactly one of the cgroups in the
  58hierarchy, and a set of subsystems; each subsystem has system-specific
  59state attached to each cgroup in the hierarchy.  Each hierarchy has
  60an instance of the cgroup virtual filesystem associated with it.
  62At any one time there may be multiple active hierarchies of task
  63cgroups. Each hierarchy is a partition of all tasks in the system.
  65User-level code may create and destroy cgroups by name in an
  66instance of the cgroup virtual file system, specify and query to
  67which cgroup a task is assigned, and list the task PIDs assigned to
  68a cgroup. Those creations and assignments only affect the hierarchy
  69associated with that instance of the cgroup file system.
  71On their own, the only use for cgroups is for simple job
  72tracking. The intention is that other subsystems hook into the generic
  73cgroup support to provide new attributes for cgroups, such as
  74accounting/limiting the resources which processes in a cgroup can
  75access. For example, cpusets (see Documentation/cgroups/cpusets.txt) allow
  76you to associate a set of CPUs and a set of memory nodes with the
  77tasks in each cgroup.
  791.2 Why are cgroups needed ?
  82There are multiple efforts to provide process aggregations in the
  83Linux kernel, mainly for resource-tracking purposes. Such efforts
  84include cpusets, CKRM/ResGroups, UserBeanCounters, and virtual server
  85namespaces. These all require the basic notion of a
  86grouping/partitioning of processes, with newly forked processes ending
  87up in the same group (cgroup) as their parent process.
  89The kernel cgroup patch provides the minimum essential kernel
  90mechanisms required to efficiently implement such groups. It has
  91minimal impact on the system fast paths, and provides hooks for
  92specific subsystems such as cpusets to provide additional behaviour as
  95Multiple hierarchy support is provided to allow for situations where
  96the division of tasks into cgroups is distinctly different for
  97different subsystems - having parallel hierarchies allows each
  98hierarchy to be a natural division of tasks, without having to handle
  99complex combinations of tasks that would be present if several
 100unrelated subsystems needed to be forced into the same tree of
 103At one extreme, each resource controller or subsystem could be in a
 104separate hierarchy; at the other extreme, all subsystems
 105would be attached to the same hierarchy.
 107As an example of a scenario (originally proposed by
 108that can benefit from multiple hierarchies, consider a large
 109university server with various users - students, professors, system
 110tasks etc. The resource planning for this server could be along the
 111following lines:
 113       CPU :          "Top cpuset"
 114                       /       \
 115               CPUSet1         CPUSet2
 116                  |               |
 117               (Professors)    (Students)
 119               In addition (system tasks) are attached to topcpuset (so
 120               that they can run anywhere) with a limit of 20%
 122       Memory : Professors (50%), Students (30%), system (20%)
 124       Disk : Professors (50%), Students (30%), system (20%)
 126       Network : WWW browsing (20%), Network File System (60%), others (20%)
 127                               / \
 128               Professors (15%)  students (5%)
 130Browsers like Firefox/Lynx go into the WWW network class, while (k)nfsd goes
 131into the NFS network class.
 133At the same time Firefox/Lynx will share an appropriate CPU/Memory class
 134depending on who launched it (prof/student).
 136With the ability to classify tasks differently for different resources
 137(by putting those resource subsystems in different hierarchies),
 138the admin can easily set up a script which receives exec notifications
 139and depending on who is launching the browser he can
 141    # echo browser_pid > /sys/fs/cgroup/<restype>/<userclass>/tasks
 143With only a single hierarchy, he now would potentially have to create
 144a separate cgroup for every browser launched and associate it with
 145appropriate network and other resource class.  This may lead to
 146proliferation of such cgroups.
 148Also let's say that the administrator would like to give enhanced network
 149access temporarily to a student's browser (since it is night and the user
 150wants to do online gaming :))  OR give one of the student's simulation
 151apps enhanced CPU power.
 153With ability to write PIDs directly to resource classes, it's just a
 154matter of:
 156       # echo pid > /sys/fs/cgroup/network/<new_class>/tasks
 157       (after some time)
 158       # echo pid > /sys/fs/cgroup/network/<orig_class>/tasks
 160Without this ability, the administrator would have to split the cgroup into
 161multiple separate ones and then associate the new cgroups with the
 162new resource classes.
 1661.3 How are cgroups implemented ?
 169Control Groups extends the kernel as follows:
 171 - Each task in the system has a reference-counted pointer to a
 172   css_set.
 174 - A css_set contains a set of reference-counted pointers to
 175   cgroup_subsys_state objects, one for each cgroup subsystem
 176   registered in the system. There is no direct link from a task to
 177   the cgroup of which it's a member in each hierarchy, but this
 178   can be determined by following pointers through the
 179   cgroup_subsys_state objects. This is because accessing the
 180   subsystem state is something that's expected to happen frequently
 181   and in performance-critical code, whereas operations that require a
 182   task's actual cgroup assignments (in particular, moving between
 183   cgroups) are less common. A linked list runs through the cg_list
 184   field of each task_struct using the css_set, anchored at
 185   css_set->tasks.
 187 - A cgroup hierarchy filesystem can be mounted for browsing and
 188   manipulation from user space.
 190 - You can list all the tasks (by PID) attached to any cgroup.
 192The implementation of cgroups requires a few, simple hooks
 193into the rest of the kernel, none in performance-critical paths:
 195 - in init/main.c, to initialize the root cgroups and initial
 196   css_set at system boot.
 198 - in fork and exit, to attach and detach a task from its css_set.
 200In addition, a new file system of type "cgroup" may be mounted, to
 201enable browsing and modifying the cgroups presently known to the
 202kernel.  When mounting a cgroup hierarchy, you may specify a
 203comma-separated list of subsystems to mount as the filesystem mount
 204options.  By default, mounting the cgroup filesystem attempts to
 205mount a hierarchy containing all registered subsystems.
 207If an active hierarchy with exactly the same set of subsystems already
 208exists, it will be reused for the new mount. If no existing hierarchy
 209matches, and any of the requested subsystems are in use in an existing
 210hierarchy, the mount will fail with -EBUSY. Otherwise, a new hierarchy
 211is activated, associated with the requested subsystems.
 213It's not currently possible to bind a new subsystem to an active
 214cgroup hierarchy, or to unbind a subsystem from an active cgroup
 215hierarchy. This may be possible in future, but is fraught with nasty
 216error-recovery issues.
 218When a cgroup filesystem is unmounted, if there are any
 219child cgroups created below the top-level cgroup, that hierarchy
 220will remain active even though unmounted; if there are no
 221child cgroups then the hierarchy will be deactivated.
 223No new system calls are added for cgroups - all support for
 224querying and modifying cgroups is via this cgroup file system.
 226Each task under /proc has an added file named 'cgroup' displaying,
 227for each active hierarchy, the subsystem names and the cgroup name
 228as the path relative to the root of the cgroup file system.
 230Each cgroup is represented by a directory in the cgroup file system
 231containing the following files describing that cgroup:
 233 - tasks: list of tasks (by PID) attached to that cgroup.  This list
 234   is not guaranteed to be sorted.  Writing a thread ID into this file
 235   moves the thread into this cgroup.
 236 - cgroup.procs: list of thread group IDs in the cgroup.  This list is
 237   not guaranteed to be sorted or free of duplicate TGIDs, and userspace
 238   should sort/uniquify the list if this property is required.
 239   Writing a thread group ID into this file moves all threads in that
 240   group into this cgroup.
 241 - notify_on_release flag: run the release agent on exit?
 242 - release_agent: the path to use for release notifications (this file
 243   exists in the top cgroup only)
 245Other subsystems such as cpusets may add additional files in each
 246cgroup dir.
 248New cgroups are created using the mkdir system call or shell
 249command.  The properties of a cgroup, such as its flags, are
 250modified by writing to the appropriate file in that cgroups
 251directory, as listed above.
 253The named hierarchical structure of nested cgroups allows partitioning
 254a large system into nested, dynamically changeable, "soft-partitions".
 256The attachment of each task, automatically inherited at fork by any
 257children of that task, to a cgroup allows organizing the work load
 258on a system into related sets of tasks.  A task may be re-attached to
 259any other cgroup, if allowed by the permissions on the necessary
 260cgroup file system directories.
 262When a task is moved from one cgroup to another, it gets a new
 263css_set pointer - if there's an already existing css_set with the
 264desired collection of cgroups then that group is reused, otherwise a new
 265css_set is allocated. The appropriate existing css_set is located by
 266looking into a hash table.
 268To allow access from a cgroup to the css_sets (and hence tasks)
 269that comprise it, a set of cg_cgroup_link objects form a lattice;
 270each cg_cgroup_link is linked into a list of cg_cgroup_links for
 271a single cgroup on its cgrp_link_list field, and a list of
 272cg_cgroup_links for a single css_set on its cg_link_list.
 274Thus the set of tasks in a cgroup can be listed by iterating over
 275each css_set that references the cgroup, and sub-iterating over
 276each css_set's task set.
 278The use of a Linux virtual file system (vfs) to represent the
 279cgroup hierarchy provides for a familiar permission and name space
 280for cgroups, with a minimum of additional kernel code.
 2821.4 What does notify_on_release do ?
 285If the notify_on_release flag is enabled (1) in a cgroup, then
 286whenever the last task in the cgroup leaves (exits or attaches to
 287some other cgroup) and the last child cgroup of that cgroup
 288is removed, then the kernel runs the command specified by the contents
 289of the "release_agent" file in that hierarchy's root directory,
 290supplying the pathname (relative to the mount point of the cgroup
 291file system) of the abandoned cgroup.  This enables automatic
 292removal of abandoned cgroups.  The default value of
 293notify_on_release in the root cgroup at system boot is disabled
 294(0).  The default value of other cgroups at creation is the current
 295value of their parents' notify_on_release settings. The default value of
 296a cgroup hierarchy's release_agent path is empty.
 2981.5 What does clone_children do ?
 301This flag only affects the cpuset controller. If the clone_children
 302flag is enabled (1) in a cgroup, a new cpuset cgroup will copy its
 303configuration from the parent during initialization.
 3051.6 How do I use cgroups ?
 308To start a new job that is to be contained within a cgroup, using
 309the "cpuset" cgroup subsystem, the steps are something like:
 311 1) mount -t tmpfs cgroup_root /sys/fs/cgroup
 312 2) mkdir /sys/fs/cgroup/cpuset
 313 3) mount -t cgroup -ocpuset cpuset /sys/fs/cgroup/cpuset
 314 4) Create the new cgroup by doing mkdir's and write's (or echo's) in
 315    the /sys/fs/cgroup virtual file system.
 316 5) Start a task that will be the "founding father" of the new job.
 317 6) Attach that task to the new cgroup by writing its PID to the
 318    /sys/fs/cgroup/cpuset/tasks file for that cgroup.
 319 7) fork, exec or clone the job tasks from this founding father task.
 321For example, the following sequence of commands will setup a cgroup
 322named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
 323and then start a subshell 'sh' in that cgroup:
 325  mount -t tmpfs cgroup_root /sys/fs/cgroup
 326  mkdir /sys/fs/cgroup/cpuset
 327  mount -t cgroup cpuset -ocpuset /sys/fs/cgroup/cpuset
 328  cd /sys/fs/cgroup/cpuset
 329  mkdir Charlie
 330  cd Charlie
 331  /bin/echo 2-3 > cpuset.cpus
 332  /bin/echo 1 > cpuset.mems
 333  /bin/echo $$ > tasks
 334  sh
 335  # The subshell 'sh' is now running in cgroup Charlie
 336  # The next line should display '/Charlie'
 337  cat /proc/self/cgroup
 3392. Usage Examples and Syntax
 3422.1 Basic Usage
 345Creating, modifying, using cgroups can be done through the cgroup
 346virtual filesystem.
 348To mount a cgroup hierarchy with all available subsystems, type:
 349# mount -t cgroup xxx /sys/fs/cgroup
 351The "xxx" is not interpreted by the cgroup code, but will appear in
 352/proc/mounts so may be any useful identifying string that you like.
 354Note: Some subsystems do not work without some user input first.  For instance,
 355if cpusets are enabled the user will have to populate the cpus and mems files
 356for each new cgroup created before that group can be used.
 358As explained in section `1.2 Why are cgroups needed?' you should create
 359different hierarchies of cgroups for each single resource or group of
 360resources you want to control. Therefore, you should mount a tmpfs on
 361/sys/fs/cgroup and create directories for each cgroup resource or resource
 364# mount -t tmpfs cgroup_root /sys/fs/cgroup
 365# mkdir /sys/fs/cgroup/rg1
 367To mount a cgroup hierarchy with just the cpuset and memory
 368subsystems, type:
 369# mount -t cgroup -o cpuset,memory hier1 /sys/fs/cgroup/rg1
 371While remounting cgroups is currently supported, it is not recommend
 372to use it. Remounting allows changing bound subsystems and
 373release_agent. Rebinding is hardly useful as it only works when the
 374hierarchy is empty and release_agent itself should be replaced with
 375conventional fsnotify. The support for remounting will be removed in
 376the future.
 378To Specify a hierarchy's release_agent:
 379# mount -t cgroup -o cpuset,release_agent="/sbin/cpuset_release_agent" \
 380  xxx /sys/fs/cgroup/rg1
 382Note that specifying 'release_agent' more than once will return failure.
 384Note that changing the set of subsystems is currently only supported
 385when the hierarchy consists of a single (root) cgroup. Supporting
 386the ability to arbitrarily bind/unbind subsystems from an existing
 387cgroup hierarchy is intended to be implemented in the future.
 389Then under /sys/fs/cgroup/rg1 you can find a tree that corresponds to the
 390tree of the cgroups in the system. For instance, /sys/fs/cgroup/rg1
 391is the cgroup that holds the whole system.
 393If you want to change the value of release_agent:
 394# echo "/sbin/new_release_agent" > /sys/fs/cgroup/rg1/release_agent
 396It can also be changed via remount.
 398If you want to create a new cgroup under /sys/fs/cgroup/rg1:
 399# cd /sys/fs/cgroup/rg1
 400# mkdir my_cgroup
 402Now you want to do something with this cgroup.
 403# cd my_cgroup
 405In this directory you can find several files:
 406# ls
 407cgroup.procs notify_on_release tasks
 408(plus whatever files added by the attached subsystems)
 410Now attach your shell to this cgroup:
 411# /bin/echo $$ > tasks
 413You can also create cgroups inside your cgroup by using mkdir in this
 415# mkdir my_sub_cs
 417To remove a cgroup, just use rmdir:
 418# rmdir my_sub_cs
 420This will fail if the cgroup is in use (has cgroups inside, or
 421has processes attached, or is held alive by other subsystem-specific
 4242.2 Attaching processes
 427# /bin/echo PID > tasks
 429Note that it is PID, not PIDs. You can only attach ONE task at a time.
 430If you have several tasks to attach, you have to do it one after another:
 432# /bin/echo PID1 > tasks
 433# /bin/echo PID2 > tasks
 434        ...
 435# /bin/echo PIDn > tasks
 437You can attach the current shell task by echoing 0:
 439# echo 0 > tasks
 441You can use the cgroup.procs file instead of the tasks file to move all
 442threads in a threadgroup at once. Echoing the PID of any task in a
 443threadgroup to cgroup.procs causes all tasks in that threadgroup to be
 444attached to the cgroup. Writing 0 to cgroup.procs moves all tasks
 445in the writing task's threadgroup.
 447Note: Since every task is always a member of exactly one cgroup in each
 448mounted hierarchy, to remove a task from its current cgroup you must
 449move it into a new cgroup (possibly the root cgroup) by writing to the
 450new cgroup's tasks file.
 452Note: Due to some restrictions enforced by some cgroup subsystems, moving
 453a process to another cgroup can fail.
 4552.3 Mounting hierarchies by name
 458Passing the name=<x> option when mounting a cgroups hierarchy
 459associates the given name with the hierarchy.  This can be used when
 460mounting a pre-existing hierarchy, in order to refer to it by name
 461rather than by its set of active subsystems.  Each hierarchy is either
 462nameless, or has a unique name.
 464The name should match [\w.-]+
 466When passing a name=<x> option for a new hierarchy, you need to
 467specify subsystems manually; the legacy behaviour of mounting all
 468subsystems when none are explicitly specified is not supported when
 469you give a subsystem a name.
 471The name of the subsystem appears as part of the hierarchy description
 472in /proc/mounts and /proc/<pid>/cgroups.
 4753. Kernel API
 4783.1 Overview
 481Each kernel subsystem that wants to hook into the generic cgroup
 482system needs to create a cgroup_subsys object. This contains
 483various methods, which are callbacks from the cgroup system, along
 484with a subsystem ID which will be assigned by the cgroup system.
 486Other fields in the cgroup_subsys object include:
 488- subsys_id: a unique array index for the subsystem, indicating which
 489  entry in cgroup->subsys[] this subsystem should be managing.
 491- name: should be initialized to a unique subsystem name. Should be
 492  no longer than MAX_CGROUP_TYPE_NAMELEN.
 494- early_init: indicate if the subsystem needs early initialization
 495  at system boot.
 497Each cgroup object created by the system has an array of pointers,
 498indexed by subsystem ID; this pointer is entirely managed by the
 499subsystem; the generic cgroup code will never touch this pointer.
 5013.2 Synchronization
 504There is a global mutex, cgroup_mutex, used by the cgroup
 505system. This should be taken by anything that wants to modify a
 506cgroup. It may also be taken to prevent cgroups from being
 507modified, but more specific locks may be more appropriate in that
 510See kernel/cgroup.c for more details.
 512Subsystems can take/release the cgroup_mutex via the functions
 515Accessing a task's cgroup pointer may be done in the following ways:
 516- while holding cgroup_mutex
 517- while holding the task's alloc_lock (via task_lock())
 518- inside an rcu_read_lock() section via rcu_dereference()
 5203.3 Subsystem API
 523Each subsystem should:
 525- add an entry in linux/cgroup_subsys.h
 526- define a cgroup_subsys object called <name>_subsys
 528If a subsystem can be compiled as a module, it should also have in its
 529module initcall a call to cgroup_load_subsys(), and in its exitcall a
 530call to cgroup_unload_subsys(). It should also set its_subsys.module =
 531THIS_MODULE in its .c file.
 533Each subsystem may export the following methods. The only mandatory
 534methods are css_alloc/free. Any others that are null are presumed to
 535be successful no-ops.
 537struct cgroup_subsys_state *css_alloc(struct cgroup *cgrp)
 538(cgroup_mutex held by caller)
 540Called to allocate a subsystem state object for a cgroup. The
 541subsystem should allocate its subsystem state object for the passed
 542cgroup, returning a pointer to the new object on success or a
 543ERR_PTR() value. On success, the subsystem pointer should point to
 544a structure of type cgroup_subsys_state (typically embedded in a
 545larger subsystem-specific object), which will be initialized by the
 546cgroup system. Note that this will be called at initialization to
 547create the root subsystem state for this subsystem; this case can be
 548identified by the passed cgroup object having a NULL parent (since
 549it's the root of the hierarchy) and may be an appropriate place for
 550initialization code.
 552int css_online(struct cgroup *cgrp)
 553(cgroup_mutex held by caller)
 555Called after @cgrp successfully completed all allocations and made
 556visible to cgroup_for_each_child/descendant_*() iterators. The
 557subsystem may choose to fail creation by returning -errno. This
 558callback can be used to implement reliable state sharing and
 559propagation along the hierarchy. See the comment on
 560cgroup_for_each_descendant_pre() for details.
 562void css_offline(struct cgroup *cgrp);
 563(cgroup_mutex held by caller)
 565This is the counterpart of css_online() and called iff css_online()
 566has succeeded on @cgrp. This signifies the beginning of the end of
 567@cgrp. @cgrp is being removed and the subsystem should start dropping
 568all references it's holding on @cgrp. When all references are dropped,
 569cgroup removal will proceed to the next step - css_free(). After this
 570callback, @cgrp should be considered dead to the subsystem.
 572void css_free(struct cgroup *cgrp)
 573(cgroup_mutex held by caller)
 575The cgroup system is about to free @cgrp; the subsystem should free
 576its subsystem state object. By the time this method is called, @cgrp
 577is completely unused; @cgrp->parent is still valid. (Note - can also
 578be called for a newly-created cgroup if an error occurs after this
 579subsystem's create() method has been called for the new cgroup).
 581int can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 582(cgroup_mutex held by caller)
 584Called prior to moving one or more tasks into a cgroup; if the
 585subsystem returns an error, this will abort the attach operation.
 586@tset contains the tasks to be attached and is guaranteed to have at
 587least one task in it.
 589If there are multiple tasks in the taskset, then:
 590  - it's guaranteed that all are from the same thread group
 591  - @tset contains all tasks from the thread group whether or not
 592    they're switching cgroups
 593  - the first task is the leader
 595Each @tset entry also contains the task's old cgroup and tasks which
 596aren't switching cgroup can be skipped easily using the
 597cgroup_taskset_for_each() iterator. Note that this isn't called on a
 598fork. If this method returns 0 (success) then this should remain valid
 599while the caller holds cgroup_mutex and it is ensured that either
 600attach() or cancel_attach() will be called in future.
 602void css_reset(struct cgroup_subsys_state *css)
 603(cgroup_mutex held by caller)
 605An optional operation which should restore @css's configuration to the
 606initial state.  This is currently only used on the unified hierarchy
 607when a subsystem is disabled on a cgroup through
 608"cgroup.subtree_control" but should remain enabled because other
 609subsystems depend on it.  cgroup core makes such a css invisible by
 610removing the associated interface files and invokes this callback so
 611that the hidden subsystem can return to the initial neutral state.
 612This prevents unexpected resource control from a hidden css and
 613ensures that the configuration is in the initial state when it is made
 614visible again later.
 616void cancel_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 617(cgroup_mutex held by caller)
 619Called when a task attach operation has failed after can_attach() has succeeded.
 620A subsystem whose can_attach() has some side-effects should provide this
 621function, so that the subsystem can implement a rollback. If not, not necessary.
 622This will be called only about subsystems whose can_attach() operation have
 623succeeded. The parameters are identical to can_attach().
 625void attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
 626(cgroup_mutex held by caller)
 628Called after the task has been attached to the cgroup, to allow any
 629post-attachment activity that requires memory allocations or blocking.
 630The parameters are identical to can_attach().
 632void fork(struct task_struct *task)
 634Called when a task is forked into a cgroup.
 636void exit(struct task_struct *task)
 638Called during task exit.
 640void bind(struct cgroup *root)
 641(cgroup_mutex held by caller)
 643Called when a cgroup subsystem is rebound to a different hierarchy
 644and root cgroup. Currently this will only involve movement between
 645the default hierarchy (which never has sub-cgroups) and a hierarchy
 646that is being created/destroyed (and hence has no sub-cgroups).
 6484. Extended attribute usage
 651cgroup filesystem supports certain types of extended attributes in its
 652directories and files.  The current supported types are:
 653        - Trusted (XATTR_TRUSTED)
 654        - Security (XATTR_SECURITY)
 656Both require CAP_SYS_ADMIN capability to set.
 658Like in tmpfs, the extended attributes in cgroup filesystem are stored
 659using kernel memory and it's advised to keep the usage at minimum.  This
 660is the reason why user defined extended attributes are not supported, since
 661any user can do it and there's no limit in the value size.
 663The current known users for this feature are SELinux to limit cgroup usage
 664in containers and systemd for assorted meta data like main PID in a cgroup
 665(systemd creates a cgroup per service).
 6675. Questions
 670Q: what's up with this '/bin/echo' ?
 671A: bash's builtin 'echo' command does not check calls to write() against
 672   errors. If you use it in the cgroup file system, you won't be
 673   able to tell whether a command succeeded or failed.
 675Q: When I attach processes, only the first of the line gets really attached !
 676A: We can only return one error code per call to write(). So you should also
 677   put only ONE PID.