Root/Documentation/cgroups/cgroups.txt

1                CGROUPS
2                -------
3
4Written by Paul Menage <menage@google.com> based on
5Documentation/cgroups/cpusets.txt
6
7Original copyright statements from cpusets.txt:
8Portions Copyright (C) 2004 BULL SA.
9Portions Copyright (c) 2004-2006 Silicon Graphics, Inc.
10Modified by Paul Jackson <pj@sgi.com>
11Modified by Christoph Lameter <clameter@sgi.com>
12
13CONTENTS:
14=========
15
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 How do I use cgroups ?
222. Usage Examples and Syntax
23  2.1 Basic Usage
24  2.2 Attaching processes
25  2.3 Mounting hierarchies by name
26  2.4 Notification API
273. Kernel API
28  3.1 Overview
29  3.2 Synchronization
30  3.3 Subsystem API
314. Questions
32
331. Control Groups
34=================
35
361.1 What are cgroups ?
37----------------------
38
39Control Groups provide a mechanism for aggregating/partitioning sets of
40tasks, and all their future children, into hierarchical groups with
41specialized behaviour.
42
43Definitions:
44
45A *cgroup* associates a set of tasks with a set of parameters for one
46or more subsystems.
47
48A *subsystem* is a module that makes use of the task grouping
49facilities provided by cgroups to treat groups of tasks in
50particular ways. A subsystem is typically a "resource controller" that
51schedules a resource or applies per-cgroup limits, but it may be
52anything that wants to act on a group of processes, e.g. a
53virtualization subsystem.
54
55A *hierarchy* is a set of cgroups arranged in a tree, such that
56every task in the system is in exactly one of the cgroups in the
57hierarchy, and a set of subsystems; each subsystem has system-specific
58state attached to each cgroup in the hierarchy. Each hierarchy has
59an instance of the cgroup virtual filesystem associated with it.
60
61At any one time there may be multiple active hierarchies of task
62cgroups. Each hierarchy is a partition of all tasks in the system.
63
64User level code may create and destroy cgroups by name in an
65instance of the cgroup virtual file system, specify and query to
66which cgroup a task is assigned, and list the task pids assigned to
67a cgroup. Those creations and assignments only affect the hierarchy
68associated with that instance of the cgroup file system.
69
70On their own, the only use for cgroups is for simple job
71tracking. The intention is that other subsystems hook into the generic
72cgroup support to provide new attributes for cgroups, such as
73accounting/limiting the resources which processes in a cgroup can
74access. For example, cpusets (see Documentation/cgroups/cpusets.txt) allows
75you to associate a set of CPUs and a set of memory nodes with the
76tasks in each cgroup.
77
781.2 Why are cgroups needed ?
79----------------------------
80
81There are multiple efforts to provide process aggregations in the
82Linux kernel, mainly for resource tracking purposes. Such efforts
83include cpusets, CKRM/ResGroups, UserBeanCounters, and virtual server
84namespaces. These all require the basic notion of a
85grouping/partitioning of processes, with newly forked processes ending
86in the same group (cgroup) as their parent process.
87
88The kernel cgroup patch provides the minimum essential kernel
89mechanisms required to efficiently implement such groups. It has
90minimal impact on the system fast paths, and provides hooks for
91specific subsystems such as cpusets to provide additional behaviour as
92desired.
93
94Multiple hierarchy support is provided to allow for situations where
95the division of tasks into cgroups is distinctly different for
96different subsystems - having parallel hierarchies allows each
97hierarchy to be a natural division of tasks, without having to handle
98complex combinations of tasks that would be present if several
99unrelated subsystems needed to be forced into the same tree of
100cgroups.
101
102At one extreme, each resource controller or subsystem could be in a
103separate hierarchy; at the other extreme, all subsystems
104would be attached to the same hierarchy.
105
106As an example of a scenario (originally proposed by vatsa@in.ibm.com)
107that can benefit from multiple hierarchies, consider a large
108university server with various users - students, professors, system
109tasks etc. The resource planning for this server could be along the
110following lines:
111
112       CPU : Top cpuset
113                       / \
114               CPUSet1 CPUSet2
115                  | |
116               (Profs) (Students)
117
118               In addition (system tasks) are attached to topcpuset (so
119               that they can run anywhere) with a limit of 20%
120
121       Memory : Professors (50%), students (30%), system (20%)
122
123       Disk : Prof (50%), students (30%), system (20%)
124
125       Network : WWW browsing (20%), Network File System (60%), others (20%)
126                               / \
127                       Prof (15%) students (5%)
128
129Browsers like Firefox/Lynx go into the WWW network class, while (k)nfsd go
130into NFS network class.
131
132At the same time Firefox/Lynx will share an appropriate CPU/Memory class
133depending on who launched it (prof/student).
134
135With the ability to classify tasks differently for different resources
136(by putting those resource subsystems in different hierarchies) then
137the admin can easily set up a script which receives exec notifications
138and depending on who is launching the browser he can
139
140       # echo browser_pid > /mnt/<restype>/<userclass>/tasks
141
142With only a single hierarchy, he now would potentially have to create
143a separate cgroup for every browser launched and associate it with
144approp network and other resource class. This may lead to
145proliferation of such cgroups.
146
147Also lets say that the administrator would like to give enhanced network
148access temporarily to a student's browser (since it is night and the user
149wants to do online gaming :)) OR give one of the students simulation
150apps enhanced CPU power,
151
152With ability to write pids directly to resource classes, it's just a
153matter of :
154
155       # echo pid > /mnt/network/<new_class>/tasks
156       (after some time)
157       # echo pid > /mnt/network/<orig_class>/tasks
158
159Without this ability, he would have to split the cgroup into
160multiple separate ones and then associate the new cgroups with the
161new resource classes.
162
163
164
1651.3 How are cgroups implemented ?
166---------------------------------
167
168Control Groups extends the kernel as follows:
169
170 - Each task in the system has a reference-counted pointer to a
171   css_set.
172
173 - A css_set contains a set of reference-counted pointers to
174   cgroup_subsys_state objects, one for each cgroup subsystem
175   registered in the system. There is no direct link from a task to
176   the cgroup of which it's a member in each hierarchy, but this
177   can be determined by following pointers through the
178   cgroup_subsys_state objects. This is because accessing the
179   subsystem state is something that's expected to happen frequently
180   and in performance-critical code, whereas operations that require a
181   task's actual cgroup assignments (in particular, moving between
182   cgroups) are less common. A linked list runs through the cg_list
183   field of each task_struct using the css_set, anchored at
184   css_set->tasks.
185
186 - A cgroup hierarchy filesystem can be mounted for browsing and
187   manipulation from user space.
188
189 - You can list all the tasks (by pid) attached to any cgroup.
190
191The implementation of cgroups requires a few, simple hooks
192into the rest of the kernel, none in performance critical paths:
193
194 - in init/main.c, to initialize the root cgroups and initial
195   css_set at system boot.
196
197 - in fork and exit, to attach and detach a task from its css_set.
198
199In addition a new file system, of type "cgroup" may be mounted, to
200enable browsing and modifying the cgroups presently known to the
201kernel. When mounting a cgroup hierarchy, you may specify a
202comma-separated list of subsystems to mount as the filesystem mount
203options. By default, mounting the cgroup filesystem attempts to
204mount a hierarchy containing all registered subsystems.
205
206If an active hierarchy with exactly the same set of subsystems already
207exists, it will be reused for the new mount. If no existing hierarchy
208matches, and any of the requested subsystems are in use in an existing
209hierarchy, the mount will fail with -EBUSY. Otherwise, a new hierarchy
210is activated, associated with the requested subsystems.
211
212It's not currently possible to bind a new subsystem to an active
213cgroup hierarchy, or to unbind a subsystem from an active cgroup
214hierarchy. This may be possible in future, but is fraught with nasty
215error-recovery issues.
216
217When a cgroup filesystem is unmounted, if there are any
218child cgroups created below the top-level cgroup, that hierarchy
219will remain active even though unmounted; if there are no
220child cgroups then the hierarchy will be deactivated.
221
222No new system calls are added for cgroups - all support for
223querying and modifying cgroups is via this cgroup file system.
224
225Each task under /proc has an added file named 'cgroup' displaying,
226for each active hierarchy, the subsystem names and the cgroup name
227as the path relative to the root of the cgroup file system.
228
229Each cgroup is represented by a directory in the cgroup file system
230containing the following files describing that cgroup:
231
232 - tasks: list of tasks (by pid) attached to that cgroup. This list
233   is not guaranteed to be sorted. Writing a thread id into this file
234   moves the thread into this cgroup.
235 - cgroup.procs: list of tgids in the cgroup. This list is not
236   guaranteed to be sorted or free of duplicate tgids, and userspace
237   should sort/uniquify the list if this property is required.
238   This is a read-only file, for now.
239 - notify_on_release flag: run the release agent on exit?
240 - release_agent: the path to use for release notifications (this file
241   exists in the top cgroup only)
242
243Other subsystems such as cpusets may add additional files in each
244cgroup dir.
245
246New cgroups are created using the mkdir system call or shell
247command. The properties of a cgroup, such as its flags, are
248modified by writing to the appropriate file in that cgroups
249directory, as listed above.
250
251The named hierarchical structure of nested cgroups allows partitioning
252a large system into nested, dynamically changeable, "soft-partitions".
253
254The attachment of each task, automatically inherited at fork by any
255children of that task, to a cgroup allows organizing the work load
256on a system into related sets of tasks. A task may be re-attached to
257any other cgroup, if allowed by the permissions on the necessary
258cgroup file system directories.
259
260When a task is moved from one cgroup to another, it gets a new
261css_set pointer - if there's an already existing css_set with the
262desired collection of cgroups then that group is reused, else a new
263css_set is allocated. The appropriate existing css_set is located by
264looking into a hash table.
265
266To allow access from a cgroup to the css_sets (and hence tasks)
267that comprise it, a set of cg_cgroup_link objects form a lattice;
268each cg_cgroup_link is linked into a list of cg_cgroup_links for
269a single cgroup on its cgrp_link_list field, and a list of
270cg_cgroup_links for a single css_set on its cg_link_list.
271
272Thus the set of tasks in a cgroup can be listed by iterating over
273each css_set that references the cgroup, and sub-iterating over
274each css_set's task set.
275
276The use of a Linux virtual file system (vfs) to represent the
277cgroup hierarchy provides for a familiar permission and name space
278for cgroups, with a minimum of additional kernel code.
279
2801.4 What does notify_on_release do ?
281------------------------------------
282
283If the notify_on_release flag is enabled (1) in a cgroup, then
284whenever the last task in the cgroup leaves (exits or attaches to
285some other cgroup) and the last child cgroup of that cgroup
286is removed, then the kernel runs the command specified by the contents
287of the "release_agent" file in that hierarchy's root directory,
288supplying the pathname (relative to the mount point of the cgroup
289file system) of the abandoned cgroup. This enables automatic
290removal of abandoned cgroups. The default value of
291notify_on_release in the root cgroup at system boot is disabled
292(0). The default value of other cgroups at creation is the current
293value of their parents notify_on_release setting. The default value of
294a cgroup hierarchy's release_agent path is empty.
295
2961.5 How do I use cgroups ?
297--------------------------
298
299To start a new job that is to be contained within a cgroup, using
300the "cpuset" cgroup subsystem, the steps are something like:
301
302 1) mkdir /dev/cgroup
303 2) mount -t cgroup -ocpuset cpuset /dev/cgroup
304 3) Create the new cgroup by doing mkdir's and write's (or echo's) in
305    the /dev/cgroup virtual file system.
306 4) Start a task that will be the "founding father" of the new job.
307 5) Attach that task to the new cgroup by writing its pid to the
308    /dev/cgroup tasks file for that cgroup.
309 6) fork, exec or clone the job tasks from this founding father task.
310
311For example, the following sequence of commands will setup a cgroup
312named "Charlie", containing just CPUs 2 and 3, and Memory Node 1,
313and then start a subshell 'sh' in that cgroup:
314
315  mount -t cgroup cpuset -ocpuset /dev/cgroup
316  cd /dev/cgroup
317  mkdir Charlie
318  cd Charlie
319  /bin/echo 2-3 > cpuset.cpus
320  /bin/echo 1 > cpuset.mems
321  /bin/echo $$ > tasks
322  sh
323  # The subshell 'sh' is now running in cgroup Charlie
324  # The next line should display '/Charlie'
325  cat /proc/self/cgroup
326
3272. Usage Examples and Syntax
328============================
329
3302.1 Basic Usage
331---------------
332
333Creating, modifying, using the cgroups can be done through the cgroup
334virtual filesystem.
335
336To mount a cgroup hierarchy with all available subsystems, type:
337# mount -t cgroup xxx /dev/cgroup
338
339The "xxx" is not interpreted by the cgroup code, but will appear in
340/proc/mounts so may be any useful identifying string that you like.
341
342To mount a cgroup hierarchy with just the cpuset and numtasks
343subsystems, type:
344# mount -t cgroup -o cpuset,memory hier1 /dev/cgroup
345
346To change the set of subsystems bound to a mounted hierarchy, just
347remount with different options:
348# mount -o remount,cpuset,ns hier1 /dev/cgroup
349
350Now memory is removed from the hierarchy and ns is added.
351
352Note this will add ns to the hierarchy but won't remove memory or
353cpuset, because the new options are appended to the old ones:
354# mount -o remount,ns /dev/cgroup
355
356To Specify a hierarchy's release_agent:
357# mount -t cgroup -o cpuset,release_agent="/sbin/cpuset_release_agent" \
358  xxx /dev/cgroup
359
360Note that specifying 'release_agent' more than once will return failure.
361
362Note that changing the set of subsystems is currently only supported
363when the hierarchy consists of a single (root) cgroup. Supporting
364the ability to arbitrarily bind/unbind subsystems from an existing
365cgroup hierarchy is intended to be implemented in the future.
366
367Then under /dev/cgroup you can find a tree that corresponds to the
368tree of the cgroups in the system. For instance, /dev/cgroup
369is the cgroup that holds the whole system.
370
371If you want to change the value of release_agent:
372# echo "/sbin/new_release_agent" > /dev/cgroup/release_agent
373
374It can also be changed via remount.
375
376If you want to create a new cgroup under /dev/cgroup:
377# cd /dev/cgroup
378# mkdir my_cgroup
379
380Now you want to do something with this cgroup.
381# cd my_cgroup
382
383In this directory you can find several files:
384# ls
385cgroup.procs notify_on_release tasks
386(plus whatever files added by the attached subsystems)
387
388Now attach your shell to this cgroup:
389# /bin/echo $$ > tasks
390
391You can also create cgroups inside your cgroup by using mkdir in this
392directory.
393# mkdir my_sub_cs
394
395To remove a cgroup, just use rmdir:
396# rmdir my_sub_cs
397
398This will fail if the cgroup is in use (has cgroups inside, or
399has processes attached, or is held alive by other subsystem-specific
400reference).
401
4022.2 Attaching processes
403-----------------------
404
405# /bin/echo PID > tasks
406
407Note that it is PID, not PIDs. You can only attach ONE task at a time.
408If you have several tasks to attach, you have to do it one after another:
409
410# /bin/echo PID1 > tasks
411# /bin/echo PID2 > tasks
412    ...
413# /bin/echo PIDn > tasks
414
415You can attach the current shell task by echoing 0:
416
417# echo 0 > tasks
418
4192.3 Mounting hierarchies by name
420--------------------------------
421
422Passing the name=<x> option when mounting a cgroups hierarchy
423associates the given name with the hierarchy. This can be used when
424mounting a pre-existing hierarchy, in order to refer to it by name
425rather than by its set of active subsystems. Each hierarchy is either
426nameless, or has a unique name.
427
428The name should match [\w.-]+
429
430When passing a name=<x> option for a new hierarchy, you need to
431specify subsystems manually; the legacy behaviour of mounting all
432subsystems when none are explicitly specified is not supported when
433you give a subsystem a name.
434
435The name of the subsystem appears as part of the hierarchy description
436in /proc/mounts and /proc/<pid>/cgroups.
437
4382.4 Notification API
439--------------------
440
441There is mechanism which allows to get notifications about changing
442status of a cgroup.
443
444To register new notification handler you need:
445 - create a file descriptor for event notification using eventfd(2);
446 - open a control file to be monitored (e.g. memory.usage_in_bytes);
447 - write "<event_fd> <control_fd> <args>" to cgroup.event_control.
448   Interpretation of args is defined by control file implementation;
449
450eventfd will be woken up by control file implementation or when the
451cgroup is removed.
452
453To unregister notification handler just close eventfd.
454
455NOTE: Support of notifications should be implemented for the control
456file. See documentation for the subsystem.
457
4583. Kernel API
459=============
460
4613.1 Overview
462------------
463
464Each kernel subsystem that wants to hook into the generic cgroup
465system needs to create a cgroup_subsys object. This contains
466various methods, which are callbacks from the cgroup system, along
467with a subsystem id which will be assigned by the cgroup system.
468
469Other fields in the cgroup_subsys object include:
470
471- subsys_id: a unique array index for the subsystem, indicating which
472  entry in cgroup->subsys[] this subsystem should be managing.
473
474- name: should be initialized to a unique subsystem name. Should be
475  no longer than MAX_CGROUP_TYPE_NAMELEN.
476
477- early_init: indicate if the subsystem needs early initialization
478  at system boot.
479
480Each cgroup object created by the system has an array of pointers,
481indexed by subsystem id; this pointer is entirely managed by the
482subsystem; the generic cgroup code will never touch this pointer.
483
4843.2 Synchronization
485-------------------
486
487There is a global mutex, cgroup_mutex, used by the cgroup
488system. This should be taken by anything that wants to modify a
489cgroup. It may also be taken to prevent cgroups from being
490modified, but more specific locks may be more appropriate in that
491situation.
492
493See kernel/cgroup.c for more details.
494
495Subsystems can take/release the cgroup_mutex via the functions
496cgroup_lock()/cgroup_unlock().
497
498Accessing a task's cgroup pointer may be done in the following ways:
499- while holding cgroup_mutex
500- while holding the task's alloc_lock (via task_lock())
501- inside an rcu_read_lock() section via rcu_dereference()
502
5033.3 Subsystem API
504-----------------
505
506Each subsystem should:
507
508- add an entry in linux/cgroup_subsys.h
509- define a cgroup_subsys object called <name>_subsys
510
511If a subsystem can be compiled as a module, it should also have in its
512module initcall a call to cgroup_load_subsys(), and in its exitcall a
513call to cgroup_unload_subsys(). It should also set its_subsys.module =
514THIS_MODULE in its .c file.
515
516Each subsystem may export the following methods. The only mandatory
517methods are create/destroy. Any others that are null are presumed to
518be successful no-ops.
519
520struct cgroup_subsys_state *create(struct cgroup_subsys *ss,
521                   struct cgroup *cgrp)
522(cgroup_mutex held by caller)
523
524Called to create a subsystem state object for a cgroup. The
525subsystem should allocate its subsystem state object for the passed
526cgroup, returning a pointer to the new object on success or a
527negative error code. On success, the subsystem pointer should point to
528a structure of type cgroup_subsys_state (typically embedded in a
529larger subsystem-specific object), which will be initialized by the
530cgroup system. Note that this will be called at initialization to
531create the root subsystem state for this subsystem; this case can be
532identified by the passed cgroup object having a NULL parent (since
533it's the root of the hierarchy) and may be an appropriate place for
534initialization code.
535
536void destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
537(cgroup_mutex held by caller)
538
539The cgroup system is about to destroy the passed cgroup; the subsystem
540should do any necessary cleanup and free its subsystem state
541object. By the time this method is called, the cgroup has already been
542unlinked from the file system and from the child list of its parent;
543cgroup->parent is still valid. (Note - can also be called for a
544newly-created cgroup if an error occurs after this subsystem's
545create() method has been called for the new cgroup).
546
547int pre_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp);
548
549Called before checking the reference count on each subsystem. This may
550be useful for subsystems which have some extra references even if
551there are not tasks in the cgroup. If pre_destroy() returns error code,
552rmdir() will fail with it. From this behavior, pre_destroy() can be
553called multiple times against a cgroup.
554
555int can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
556           struct task_struct *task, bool threadgroup)
557(cgroup_mutex held by caller)
558
559Called prior to moving a task into a cgroup; if the subsystem
560returns an error, this will abort the attach operation. If a NULL
561task is passed, then a successful result indicates that *any*
562unspecified task can be moved into the cgroup. Note that this isn't
563called on a fork. If this method returns 0 (success) then this should
564remain valid while the caller holds cgroup_mutex and it is ensured that either
565attach() or cancel_attach() will be called in future. If threadgroup is
566true, then a successful result indicates that all threads in the given
567thread's threadgroup can be moved together.
568
569void cancel_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
570           struct task_struct *task, bool threadgroup)
571(cgroup_mutex held by caller)
572
573Called when a task attach operation has failed after can_attach() has succeeded.
574A subsystem whose can_attach() has some side-effects should provide this
575function, so that the subsytem can implement a rollback. If not, not necessary.
576This will be called only about subsystems whose can_attach() operation have
577succeeded.
578
579void attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
580        struct cgroup *old_cgrp, struct task_struct *task,
581        bool threadgroup)
582(cgroup_mutex held by caller)
583
584Called after the task has been attached to the cgroup, to allow any
585post-attachment activity that requires memory allocations or blocking.
586If threadgroup is true, the subsystem should take care of all threads
587in the specified thread's threadgroup. Currently does not support any
588subsystem that might need the old_cgrp for every thread in the group.
589
590void fork(struct cgroup_subsy *ss, struct task_struct *task)
591
592Called when a task is forked into a cgroup.
593
594void exit(struct cgroup_subsys *ss, struct task_struct *task)
595
596Called during task exit.
597
598int populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
599(cgroup_mutex held by caller)
600
601Called after creation of a cgroup to allow a subsystem to populate
602the cgroup directory with file entries. The subsystem should make
603calls to cgroup_add_file() with objects of type cftype (see
604include/linux/cgroup.h for details). Note that although this
605method can return an error code, the error code is currently not
606always handled well.
607
608void post_clone(struct cgroup_subsys *ss, struct cgroup *cgrp)
609(cgroup_mutex held by caller)
610
611Called at the end of cgroup_clone() to do any parameter
612initialization which might be required before a task could attach. For
613example in cpusets, no task may attach before 'cpus' and 'mems' are set
614up.
615
616void bind(struct cgroup_subsys *ss, struct cgroup *root)
617(cgroup_mutex and ss->hierarchy_mutex held by caller)
618
619Called when a cgroup subsystem is rebound to a different hierarchy
620and root cgroup. Currently this will only involve movement between
621the default hierarchy (which never has sub-cgroups) and a hierarchy
622that is being created/destroyed (and hence has no sub-cgroups).
623
6244. Questions
625============
626
627Q: what's up with this '/bin/echo' ?
628A: bash's builtin 'echo' command does not check calls to write() against
629   errors. If you use it in the cgroup file system, you won't be
630   able to tell whether a command succeeded or failed.
631
632Q: When I attach processes, only the first of the line gets really attached !
633A: We can only return one error code per call to write(). So you should also
634   put only ONE pid.
635
636

Archive Download this file



interactive